AI Science Tracker

AI discoveries in biology

326 discoveries from 2020 to 2026

Discoveries

Antibody design across observed and forecasted antigenic evolution

The Germinal Multi preprint abstract reports antibodies designed against several antigen variants at once. Experimental screens of tens of designs per campaign found binders to resistance-associated EGFR variants and neutralizers of historical SARS-CoV-2 variants, including variants withheld from design. Redesign of BD55-1205 restored neutralization of a forecast-guided engineered escape variant while retaining parental-variant activity. The escape result concerns an engineered variant, not a demonstrated future circulating strain. Review is limited to the abstract and original-server metadata.

Technology: Germinal Multi (computational antibody design framework); evolutionary forecasting model

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: Tens of designs per campaign were screened experimentally; designs bound EGFR variants and neutralized SARS-CoV-2 variants, including variants withheld from the design process.
  • Still unknown: Review limited to the indexed abstract and publication metadata; full-paper methods and additional limitations were not assessed.
  • AI role: Germinal Multi computationally optimized one antibody paratope against multiple antigen variants, and an evolutionary forecasting model guided construction of an escape variant and antibody redesign.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A%2210.64898%2F2026.11.10.765734%22&resultType=core&format=json · https://api.biorxiv.org/details/biorxiv/10.64898/2026.11.10.765734

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Amsterdam UMC

Machine learning reveals incomplete concordance between DNASE1L3 genotype and cfDNA phenotype

Classifiers and UMAP analyzed 129,676 prenatal-test cfDNA samples, with genotype checks in 169 selected samples. The authors report fragmentomic profiles resembling DNASE1L3 R206C homozygotes in some non-homozygotes, while some confirmed homozygotes initially lacked the profile and acquired it in later pregnancies. This extends an already-known genotype association with phenotype discordance and within-person variation; it does not establish the mechanism or clinical utility.

Technology: Supervised classifiers; UMAP

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Existing-cohort analysis with selected ddPCR genotype validation and a small longitudinal subset.
  • Still unknown: The validation cohort was enriched for discordance; general-population accuracy and causal interpretation remain unresolved.
  • AI role: Classifiers and low-dimensional embeddings identified fragmentomic subgroups and discordant cases.

Evidence sources: https://www.frontiersin.org/journals/immunology/articles/10.3389/fimmu.2026.1887185/full

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Cross-state representation learning enables prospective discovery of dynamic protein allosteric sites

The preprint abstract reports that AllositeUltra recovered allosteric sites among its top three predictions in 65.2% of apo structures and 70.6% of substrate-bound structures. More importantly, application to MAGI2 identified a previously unrecognized interdomain pocket, tested by small-molecule screening and X-ray crystallography. This within-study experimental finding qualifies separately from the benchmark. Review is limited to the abstract and original-server metadata.

Technology: AllositeUltra (structure-aware protein language model combined with physical dynamics; residue-guided candidate supplementation)

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: The predicted MAGI2 interdomain pocket was validated experimentally through small-molecule screening and X-ray crystallography; benchmark recovery was measured on apo and substrate-bound structures.
  • Still unknown: Review limited to the indexed abstract and publication metadata; full-paper methods and additional limitations were not assessed.
  • AI role: A deep-learning model predicted allosteric sites on apo and substrate-bound structures and was applied prospectively to MAGI2 and proteome-wide.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A%2210.64898%2F2026.11.09.765238%22&resultType=core&format=json · https://api.biorxiv.org/details/biorxiv/10.64898/2026.11.09.765238

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Generative design of de novo AAV vectors with unique serological profiles

The preprint abstract reports a variational autoencoder trained on natural AAV sequences to generate capsids. More than 255 designs were synthesized and tested in vitro. Seven transduced human cells at over 50% efficiency; four of those avoided neutralization by every AAV-reactive human serum sample tested, at the reported NT50 threshold below 1:50. This establishes in-vitro capsid properties, not in-vivo delivery or clinical efficacy. Review is limited to the abstract and original-server metadata.

Technology: Variational autoencoder (generative model of AAV capsid sequences)

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: More than 255 AI-generated capsid designs were synthesized and tested in vitro; seven transduced human cells above 50% efficiency and four avoided neutralizing antibodies in all reactive human sera.
  • Still unknown: Review limited to the indexed abstract and publication metadata; full-paper methods and additional limitations were not assessed.
  • AI role: A variational autoencoder learned AAV sequence constraints from natural sequences and generated functional capsid proteins de novo, with a second design campaign exploring uncertain regions.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A%2210.64898%2F2026.11.06.764881%22&resultType=core&format=json · https://api.biorxiv.org/details/biorxiv/10.64898/2026.11.06.764881

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AlphaFold-Multimer predicts 2,895 interactions among human mitochondrial proteins

Texas A&M AgriLife Research screened 630,003 protein pairs from about 1,100 human mitochondrial proteins with AlphaFold-Multimer and identified 2,895 likely interactions, more than half not previously reported, giving 85 uncharacterized proteins their first probable partners. The paper appeared in Nature Communications on 25 August 2026 and the institutional release on 5 November. Lab experiments in yeast and human cells confirmed several of the partnerships; most predictions remain untested, and the disease links are potential targets, not established ones. Predictions are shared through the MitoMatch web tool.

Technology: AlphaFold-Multimer

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: AlphaFold-Multimer screened about 630,000 protein pairs; lab experiments in yeast and human cells confirmed several of the predicted partnerships, while the release does not say how many of the 2,895 were tested.
  • Still unknown: Only several predictions were confirmed, most remain unverified, the interactions are described as likely rather than established, and disease links are potential targets.
  • AI role: AlphaFold-Multimer predicted pairwise interactions among mitochondrial proteins; the predictions were released as the MitoMatch web resource.

Evidence sources: https://www.eurekalert.org/news-releases/1150930

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Anthropic

First 3D structure of a prokaryotic vault particle, with Claude

Claude searched protein databases, designed DNA constructs, reduced purification from about three days to about 90 minutes, and processed cryo-EM data into an initial model in 34 hours. The result is a particle made of 96 copies of a bacteriophage protein, with about 25% more internal volume than the mammalian version. Its biological function is unknown; this is a preprint, and Claude made errors that humans corrected.

Technology: Claude

Family Claude: profile · models page

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: The 3D structure was determined by cryo-EM; cryo-EM confirmed the vault shape for four of seven particle-forming proteins, and fifteen altered protein versions were tested, with a reviewer checking the structure.
  • Still unknown: The biological function is unknown, vaults were made only in E. coli, and an unidentified pocket molecule resembling coenzyme A is unconfirmed.
  • AI role: Claude searched protein databases, designed DNA constructs, optimized purification, and ran RELION cryo-EM processing to an initial model in 34 hours; it also made errors that humans corrected.

Evidence sources: https://www.anthropic.com/research/investigating-vault-particles-with-claude

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning guided design of strong transcriptional activation domains

The preprint abstract reports computational generation of 500,000 candidate transcriptional activation domains and selection of 18,750 high-scoring sequences for yeast experiments using sequence-to-function models including three neural networks. More than 90% of 16,770 well-measured sequences functioned as activation domains. Some designed sequences retained activity in fungi, plants and human cells. Model agreement motivated the MADRIGAL consensus predictor. Review is limited to the abstract and original-server metadata.

Technology: Neural-network sequence-to-function models (three networks); consensus predictor MADRIGAL

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: 18,750 selected sequences were tested experimentally in yeast; over 90% of 16,770 well-measured sequences functioned as activation domains, and some retained function in fungi, plants and human cells.
  • Still unknown: Review limited to the indexed abstract and publication metadata; full-paper methods and additional limitations were not assessed.
  • AI role: Neural networks scored sequences generated from known activation-domain-like sequences; high-scoring sequences were selected for experimental validation and the consensus model MADRIGAL was developed.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A%2210.64898%2F2026.10.30.763198%22&resultType=core&format=json · https://api.biorxiv.org/details/biorxiv/10.64898/2026.10.30.763198

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning guides yeast alcohol dehydrogenase specificity for trans-2-nonenal

Researchers measured a training library of yeast alcohol dehydrogenase I variants, used support-vector activity regression/classification and a thermostability model to rank combinations, then tested 22 prioritized variants. Four showed higher specific activity toward trans-2-nonenal than wild type; LWFMY also showed improved thermostability under the study’s tests. The result is laboratory enzyme engineering, not an industrial process or validated product. The primary journal version was published 2026-11-01; a search result for an earlier-hosted copy had ambiguous date/provenance, so this entry does not claim it was the first public version.

Technology: Support-vector regression/classification models and docking-informed mutation selection

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The top 22 ranked variants were experimentally evaluated; four outperformed wild type on trans-2-nonenal, with one also improved in thermal stability.
  • Still unknown: Small selected validation set; specific activity differs from catalytic efficiency across substrate concentrations; ambiguous earlier-hosted copy means first-public date is not asserted.
  • AI role: Models trained on measured variant data ranked combinations in unexplored sequence space for experimental testing.

Evidence sources: https://link.springer.com/article/10.1186/s13068-026-02839-2 · https://pubmed.ncbi.nlm.nih.gov/42972448/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary Springer full article and PubMed metadata; methods, dates, affiliations and outcomes..

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Evo 2 uses in-context learning to design de novo regulatory sequences with only tens of examples

Arc Institute reports that Evo 2 designs functional regulatory sequences from 4 to 64 in-context examples without extra training, with 30-50% yield. Of nine AI-designed enhancers delivered by AAV to mice, seven drove detectable expression and two labeled layer 5 extratelencephalic neurons with 73% and 95% specificity. Results are bioRxiv preprints; in vivo work covers only layer 5 enhancers, and other designs (proteins, CRISPR loci, antibodies, phage genomes) are computational only.

Technology: Evo 2 (genomic language model); in-context learning; reinforcement learning post-training

Family Evo: profile · models page

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: Nine AI-designed enhancers were delivered by AAV to mice: seven drove detectable expression and two labeled layer 5 extratelencephalic neurons with 73% and 95% specificity.
  • Still unknown: Only layer 5 enhancers were tested in vivo; the authors plan other cell types, and designs beyond enhancers (proteins, CRISPR loci, antibodies, phage genomes) are computational only.
  • AI role: Evo 2 generated regulatory sequences from 4 to 64 in-context examples; reinforcement learning with predictive reward models sharpened cell-type specificity.

Evidence sources: https://arcinstitute.org/news/evo-2-in-context-learning

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

A low-rank subspace of LLM representations carries most of their brain alignment

A Dartmouth preprint studied which language-model features predict human neural responses during naturalistic listening. Across three intracranial/fMRI datasets and six model conditions, reduced-rank regression recovered full linear encoding performance with about 1–3% of embedding dimensions. These directions differed from the embeddings’ principal variance axes and transferred across held-out participants, less strongly across datasets. This summary is limited to the abstract and affiliation/date metadata of the October 29, 2026 preprint. It concerns measured brain–model alignment, not proof that brains use the same representations.

Technology: Autoregressive and masked language models with ridge reduced-rank regression

Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: retrospective or benchmark
  • Tested: Linear encoding of observed intracranial/fMRI responses, reduced-rank analysis and transfer across held-out participants.
  • Still unknown: Source checking does not establish independent confirmation. Review date: 2026-11-15.
  • AI role: LLM representations are the predictors of neural activity; reduced-rank regression identifies the brain-predictive subspace.

Evidence sources: https://arxiv.org/abs/2610.40511 · https://arxiv.org/html/2610.40511v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

D2ML guided prospective ROS1 potency and selectivity optimization

Assay-aware models were used to select 84 new compounds for synthesis and prospective testing in a ROS1 kinase program, with TRKB as an anti-target. The most selective prospective compound had a measured ROS1 IC50 of 14.7 nM and 233-fold selectivity over TRKB, compared with 20.3 nM and 79-fold for the most selective training-set compound. The evidence is biochemical and does not establish therapeutic efficacy or safety.

Technology: D2ML multi-stage deep neural networks trained across direct-to-biology assay data

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: The authors prospectively synthesized and assayed 84 model-selected compounds. The most selective new compound had ROS1 IC50 14.7 nM and 233-fold selectivity over TRKB; the most selective training compound had 20.3 nM ROS1 IC50 and 79-fold selectivity.
  • Still unknown: This is an in-vitro biochemical discovery result. It does not establish cellular activity, therapeutic efficacy, safety or clinical benefit. The paper compares a historical analog but makes no claim of global chemical novelty.
  • AI role: Assay-aware models using data from multiple direct-to-biology screening stages selected compounds for prospective synthesis and ROS1/TRKB testing.

Evidence sources: https://arxiv.org/abs/2610.41070 · https://arxiv.org/html/2610.41070v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary arXiv v1 abstract, submission history and full text for date, affiliation, model methods and prospective assay result..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Healthy neural priors improve BCI decoding in patients with motor-cortex gliomas

PACT used healthy neural priors to decode hand gestures from recordings of patients with gliomas. In the reported comparison, decoding accuracy improved by more than 10%, and alignment with reference motor representations increased from r=0.66 to r=0.94. The study collected intraoperative ECoG from eight participants, four with motor-cortex gliomas and four with other gliomas. This proof of concept addresses a specific clinical population; offline decoding does not establish benefit from a deployed BCI. Review covered the abstract, affiliations and selected results.

Technology: Pathology-Aware Consensus Training (PACT), a neural decoder trained with biophysical constraints

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Offline gesture decoding on intraoperative recordings; selected results describe eight glioma participants.
  • Still unknown: Source checking does not establish independent confirmation. Review date: 2026-11-15.
  • AI role: PACT is the decoding model; it uses healthy neural recordings as priors and learns to suppress pathological noise and reweight spatial attention.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC13693691/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

CochleaNet: deep learning analysis of cochlear connectomics and gene therapy

The journal abstract describes CochleaNet, a deep-learning framework for cleared-cochlea light-sheet images, covering hair cells, spiral ganglion neurons, synapses, tonotopy and gene expression. Trained on mouse data, it was applied to gerbil and lower-resolution mouse and marmoset data. The authors say they analyzed neuron types and validated gene-therapy products but the abstract does not identify the biological or therapy-comparison findings. The date represents the 2026-10-23 journal version; an earlier preprint appeared on 2025-11-17.

Technology: CochleaNet: deep learning framework for volumetric image segmentation and analysis

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Manual-analysis and prior-segmentation comparisons; applications to neuron types, gene-therapy products, gerbil cochleae and lower-resolution mouse and marmoset imaging. The abstract does not state the specific biological outcomes.
  • Still unknown: The journal body was not accessible through the publisher (HTTP 403), and the linked preprint full-text API returned HTTP 500. Original preprint abstract adds no specific gene-therapy result. Eligibility remains unresolved until the actual application finding is read.
  • AI role: Deep learning framework that reconstructs the cochlea and segments hair cells, neurons and synapses, and quantifies tonotopy and gene expression.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A%2210.1016%2Fj.ymthe.2026.10.025%22&resultType=core&format=json · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=EXT_ID%3APPR1120449&resultType=core&format=json

Needs source check · editorial details

The source has not been fully checked or contains conflicting information. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

A language-model approach to species distribution reveals alternative stable states

In the journal abstract, a language-model-inspired generative model trained on about 27,000 reef-fish surveys produced alternative coral-associated and macroalgal-associated community configurations under identical environmental conditions. It also inferred core, bridging and satellite species positions and improved occurrence and biomass predictions over XGBoost and MLP baselines. The configurations are model-derived and consistent with previously described ecological regimes; independent field confirmation is not established by the abstract.

Technology: Language-model-inspired generative model for ecological communities

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Fitted on about 27,000 reef fish surveys; predictive skill compared with XGBoost and MLP baselines; multistability interpreted against described ecological regimes.
  • Still unknown: Review limited to the indexed abstract and publication metadata; full-paper methods and additional limitations were not assessed.
  • AI role: Generative language-model-style model that predicts each species and its biomass conditional on previously assembled species, used to generate community configurations.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A%2210.1016%2Fj.cub.2026.09.081%22&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

GPT-Rosalind agents connect genetic perturbations and neuronal signals

Analysis of a CRISPR atlas of mouse brain cells identified differences between Tsc1 and Tsc2 and a possible relationship between DCTN1 and cholesterol regulation. Experts assessed the interpretations, but the proposed mechanisms remain hypotheses to be tested in new experiments.

Technology: GPT-Rosalind and distributed agents

Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: retrospective or benchmark
  • Tested: Thousands of GPT-Rosalind agents reanalyzed an existing whole-brain CRISPR atlas; they reproduced findings from human analyses, and their novelty and value assessments aligned with blinded human experts.
  • Still unknown: The proposed biological mechanisms still need experiments to test them.
  • AI role: Each agent investigated one gene disruption, searched the literature and proposed explanations; reports were combined into a reasoning atlas.

Evidence sources: https://openai.com/index/scaling-ai-research-for-complex-biological-data/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Pinal designs functional proteins from natural-language descriptions

The 16-billion-parameter Pinal model generated protein candidates from plain-language function descriptions. In four reported design tasks—fluorescent protein, PET hydrolase, alcohol dehydrogenase and metabolic H-protein—the selected designs displayed the intended function in laboratory tests; the best H-protein raised product titre 1.7-fold relative to native E. coli H-protein in a reconstituted CO2-fixation pathway.

Technology: Pinal, a 16-billion-parameter protein foundation model with text-to-structure (T2struct) and text-conditioned sequence design (SaProt-T) components

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: In four reported design tasks—fluorescent protein, PET hydrolase, alcohol dehydrogenase and metabolic H-protein—the selected designs displayed the intended function in laboratory tests; the best H-protein raised product titre 1.7-fold relative to native E. coli H-protein in a reconstituted CO2-fixation pathway.
  • Still unknown: This is a primary-paper proof of concept with selected designs, not a broad hit-rate or industrial-performance demonstration. Experimental denominators vary substantially by task; extended data reports one active PETase among 94 screened and a fluorescent design selected from 29. The H-protein titre comparison was in a reconstituted pathway, not an organism or commercial process. No independent replication or therapeutic result is claimed.
  • AI role: Pinal maps natural-language function descriptions to candidate protein structures and sequences. The team filtered and ranked model outputs, selected candidates for synthesis, and tested intended functions. The paper reports lab-observed fluorescence, PET-hydrolysis/esterase activity, alcohol-dehydrogenase catalysis and H-protein activity; the latter improved titre in a reconstituted multi-enzyme pathway.

Evidence sources: https://www.nature.com/articles/s41586-026-11205-z

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Bolero predicts cell-state-specific gene regulation from DNA across species

Bolero predicted cell-specific responses to AP-1 activity and transferred a human-trained response model to mouse and macaque neuronal enhancers, according to Arc Institute. Its announcement also describes a computational AP-1/cofactor spacing hypothesis; these virtual edits do not establish biological causality. This summary is limited to the October 20, 2026 institutional announcement, which identifies a collaboration with Harvard Medical School and describes a 10.8-million-cell atlas. The linked paper is a preprint; full-paper methods and experimental independence were not assessed.

Technology: Bolero (cell-state-conditioned sequence-to-function Transformer)

Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: retrospective or benchmark
  • Tested: Human-trained AP-1 response predictions compared with mouse/macaque neuronal enhancers; virtual regulatory perturbations.
  • Still unknown: Source checking does not establish independent confirmation. Review date: 2026-11-15.
  • AI role: Bolero, a Transformer conditioned on cell-state embeddings, predicts cell-state-specific gene regulation directly from DNA and estimates responses in unprofiled cell states.

Evidence sources: https://arcinstitute.org/news/bolero

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ProVADA steers a protein generator to adapt proteins to non-native biological contexts

ProVADA steered a pretrained protein generator to produce active, drug-inhibitable cytosolic renin and deimmunized two murine antibody fragments while retaining antigen recognition. ESCAPE-seq measured reduced HLA-I peptide binding, and redesigned FMC63 variants retained CAR-T-cell killing. This summary is limited to the abstract of the October 18, 2026 version 2 preprint; the initial preprint dates to July 17, 2025.

Technology: ProVADA (Mixture-Adaptation Directed Annealing sampler steering a pretrained protein generator)

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: Renin activity/inhibitor response, antibody antigen recognition and HLA-I peptide binding, and FMC63 CAR-T-cell killing, as reported in the version 2 abstract.
  • Still unknown: Only the abstract, PMC version header and original-server corresponding-institution metadata were assessed; full author-affiliation list and full paper were not reviewed. Preserve ID.
  • AI role: A pretrained protein generator, steered by ProVADA's annealing sampler toward externally defined properties while preserving functional constraints.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC13742580/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A%2210.1101%2F2025.07.11.664238%22&resultType=core&format=json · https://api.biorxiv.org/details/biorxiv/10.1101/2025.07.11.664238

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google Research

AI agents in the lab: aptamers and proteins designed in a closed loop

AI agents designed DNA aptamers and protein binders, which were then tested in the lab over multiple rounds. Target success rates reached 33.5% for myoglobin and 42.1% for α-thrombin (98.5% specificity). The post summarizes three arXiv papers whose peer review is unconfirmed; it also reports initial failures.

Technology: AI agents with laboratory robots

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: Agents ran three design-test-iterate rounds of DNA aptamers against myoglobin and alpha-thrombin in the lab, reaching 33.5% and 42.1% hit rates with 98.5% specificity for thrombin; protein binders against FAP were reported at single-digit nanomolar affinity.
  • Still unknown: Peer review of the cited arXiv papers is unconfirmed; there are two protein targets and three rounds; initial rounds failed, and protein expression adds steps to each loop.
  • AI role: Agents designed and revised sequences across rounds and adjusted design rules after reading wet-lab data, under human oversight of protocols and robot runs.

Evidence sources: https://research.google/blog/grounding-ai-agents-in-biological-reality-building-blocks-for-autonomous-wet-labs/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

Searching for antimicrobial peptides with Google Antigravity

An agentic workflow searched UniProt for antimicrobial candidates in Streptococcus pseudopneumoniae and found a 97-amino-acid bacteriocin with no homologues among manually reviewed Swiss-Prot proteins. This is a computational candidate now undergoing laboratory validation; no potency or toxicity data are available.

Technology: Google Antigravity (agentic platform)

Publication: Announcement only · Test context: Test context unspecified in annotation. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: not reported
  • Tested: Nothing was tested in the lab: the candidate's novelty was checked by BLAST against Swiss-Prot, its predicted structure was compared with a known bacteriocin fold, and its genomic context was examined; the wet-lab plan is underway with no results.
  • Still unknown: No potency or toxicity results exist; the wet-lab assays outcomes are unknown; the session covered a narrow sequence window in one species.
  • AI role: An agent queried databases, filtered candidates, ran BLAST and structure comparisons, analyzed genomic neighborhoods, synthesized literature and drafted a wet-lab plan under researcher direction.

Evidence sources: https://deepmind.google/blog/accelerating-antimicrobial-discovery/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 2/5 · Validation 1/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

IRIS model reconstructs signaling histories of single cells

IRIS inferred signaling histories from gene activity after training on human stem-cell perturbations. The researchers used it to identify a signaling intervention predicted to encourage lung development and report experimental confirmation in mouse embryos. This summary is limited to MIT/Whitehead’s October 15, 2026 announcement, which explicitly dates the Nature Methods paper to September 8. The release supplies no quantitative accuracy metrics; proposed organoid and therapeutic uses are future applications.

Technology: IRIS (neural network with transfer learning)

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Inferred cell signaling histories and a lung-development prediction checked in mouse embryos, according to the institutional release.
  • Still unknown: Source checking does not establish independent confirmation. Review date: 2026-11-15.
  • AI role: IRIS learns consistent pathway fingerprints in gene activity across cell types, reconstructs signaling histories, and generated the lung-development prediction that was tested in mouse embryos.

Evidence sources: https://news.mit.edu/2026/ai-model-decodes-language-cells-use-communicate-1015

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Deep-learning design of zinc metalloproteases

RFdiffusion2 for Molecular Interfaces designs zinc-dependent proteases for peptide-bond hydrolysis. The primary paper reports cleavage at the targeted bond in 36% of 135 designs and selected designs that cleave disease-associated proteins in vitro. These biochemical results do not establish therapeutic efficacy.

Technology: RFdiffusion2 for Molecular Interfaces (RFD2-MI); AlphaFold 3

Family RFdiffusion: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The source reports the study-specific result summarized here.
  • Still unknown: Independent replication and broader deployment or clinical benefit are not established by this source review.
  • AI role: RFD2-MI generates scaffolds around zinc catalytic and peptide-binding geometries; computational filtering guided experimental testing.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC12667880/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary-source abstract and publication metadata reviewed; full text scope is specified in source_checks..

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 5/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

SPARCS screen identifies a Golgi-pH regulator of STING signaling

A genome-wide imaging and CRISPR screen of about 70 million cells identified GPR89A and GPR89B (encoding GPHR) as important for maintaining Golgi acidity and early STING trafficking and activation. The study also found a previously undescribed autophagy phenotype associated with EI24 and linked it to altered autophagosome formation.

Technology: SPARCS platform with machine-learning image analysis

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: A genome-wide imaging and CRISPR screen of about 70 million cells identified GPR89A and GPR89B (encoding GPHR) as important for maintaining Golgi acidity and early STING trafficking and activation.
  • Still unknown: The study used pools of at least 500 excised cells for sequencing, so it could not map individual images to individual guide RNAs. The paper describes GPHR as a regulator of STING trafficking and signaling; therapeutic implications remain untested.
  • AI role: Machine-learning image analysis classified phenotypes in genome-wide CRISPR screens; SPARCS recovered known autophagy genes and surfaced EI24 and GPHR-linked findings. Follow-up molecular profiling and experiments by the researchers examined these hits.

Evidence sources: https://www.helmholtz-munich.de/en/icb/news-detail/new-method-links-cell-images-to-genetic-changes

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Deep-learning prioritization of metagenomic products yielded cytotoxic compounds

MetasynBGC used gene-cluster-family-guided metagenomic mining and a Chemprop graph neural network to prioritize predicted natural-product structures for synthesis. Six selected NRPS-derived compounds showed in-vitro cytotoxic activity across seven cancer cell lines, with reported IC50 values as low as 38.04 μM; compounds D and E showed cell-line-selective activity. These are in-vitro results and do not establish in-vivo efficacy or therapeutic benefit.

Technology: Chemprop graph neural network used to prioritize predicted biosynthetic products by antitumor activity

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Six synthesized NRPS-derived compounds were tested in vitro across seven cancer cell lines; the paper reports IC50 values as low as 38.04 μM and selective activity for compounds D and E.
  • Still unknown: No in-vivo efficacy or clinical therapeutic benefit is established. The reported potency is limited to the tested in-vitro cancer cell-line assays.
  • AI role: A Chemprop graph neural network scored PRISM-predicted biosynthetic product structures for antitumor potential; the authors jointly used predicted scores and structural feasibility to prioritize compounds for synthesis and cytotoxicity testing.

Evidence sources: https://journals.asm.org/doi/10.1128/spectrum.01098-26

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary ASM article abstract, methods and author affiliation block checked for online date, model role, synthesized compounds and measured cytotoxicity..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

SynthID Bio: watermarking for proteins and synthetic structures

Methods insert a detectable signal into protein sequences and predicted structures. For binders designed against three targets (VEGF-A, SARS-CoV-2 RBD, and PD-L1), watermarked designs matched the others in success rate and affinity. This is a proof of concept that needs to be made more robust against tampering.

Technology: SynthID Bio, AlphaFold 3 (optimized component)

Family AlphaFold 3: profile · models page

Publication: Announcement only · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: direct test
  • Tested: Wet-lab tests on three targets (VEGF-A, SARS-CoV-2 RBD, PD-L1), with Adaptyv Bio helping, showed watermarked designs matched unwatermarked ones in hit rate, affinity and diversity; the phage was tested only in bacterial cultures.
  • Still unknown: The watermark must become more robust against deliberate tampering, the bacteriophage work is early-stage, and the post says no single biosecurity measure is a silver bullet.
  • AI role: SynthID Bio guided amino acid choice in a ProteinMPNN variant, adjusted predicted atomic coordinates, fine-tuned part of AlphaFold 3, and was integrated into Evo 2 to watermark a designed phage genome.

Evidence sources: https://deepmind.google/blog/introducing-synthid-bio/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Quine ranks compounds that shift pancreatic cancer cells toward a basal state

Microsoft's Quine, a multimodal biology world model, prioritized thousands of compounds for shifting tumor cell states, and a handful were tested in pancreatic ductal adenocarcinoma cell lines. The top-ranked compounds produced the largest classical-to-basal shifts, as validated by wet-lab experiments, and some moved cells toward a distinct third phenotype. The post gives no quantitative wet-lab figures and links no paper, and it measures cell-state changes rather than clinical efficacy.

Technology: Quine, multimodal biology world model with interactive harness

Publication: Announcement only · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: direct test
  • Tested: A handful of top-ranked compounds predicted by Quine were tested in patient-derived pancreatic ductal adenocarcinoma cell lines, where the largest classical-to-basal transcriptional shifts were reported.
  • Still unknown: The page calls the drug-repurposing findings early evidence and does not report how many compounds were tested or the assay statistics.
  • AI role: Quine, a multimodal biology world model with a harness, ranked thousands of compounds by their predicted ability to shift tumor cell states.

Evidence sources: https://microsoft.com/en-us/research/blog/introducing-quine-an-ai-research-system-designed-for-the-complexity-of-biology

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Anthropic

Claude identifies a new enzyme system: array-associated reverse transcriptases

About 950 Claude agents searched a DNA-sequence database for 21 hours (210 million tokens) and flagged an uncharacterized enzyme system, found mostly in bacteriophages, with a DNA repeat array similar to CRISPR. Scientists performed the laboratory work; the function remains unknown, and the work is a preprint, not a peer-reviewed publication.

Technology: Claude (agents)

Family Claude: profile · models page

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: Human scientists expressed the protein in laboratory strains and characterized it biochemically and structurally; early experiments showed the array is expressed as distinct short RNAs.
  • Still unknown: The primary function of the enzyme system is unknown; findings are preliminary, and most Claude-proposed candidates are eliminated in follow-up review.
  • AI role: About 950 agents used 210 million tokens over 21 hours to search DNA sequences, narrowing about 3,500 candidate systems to 20; one agent noticed the repeat array and filed a report.

Evidence sources: https://anthropic.com/news/claude-discovers-novel-enzyme-system

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Paper2Agent integrates perturbation datasets to prioritize GPR137 at a psoriasis risk locus

Paper2Agent-generated agents used AlphaGenome to prioritize GPR137 as the top predicted gene affected by psoriasis-associated variant rs887314 in CD4+ T cells, then integrated that prediction with existing MPRA-coupled scCRISPRi and CD4+ T-cell Perturb-seq data. A cross-screen correlation of gene-expression signatures supported GPR137 as the probable causal gene under stimulated conditions (Spearman rho 0.613 at 8 h and 0.630 at 48 h; FDR < 0.05), but not at rest. The paper presents this as activation-dependent computational evidence.

Technology: Paper2Agent-generated AlphaGenome, MPRA-coupled scCRISPRi and CD4+ T-cell Perturb-seq agents, coordinated through an AI co-scientist

Family AI Co-Scientist: profile · models page

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Paper2Agent-generated agents used AlphaGenome to prioritize GPR137 as the top predicted gene affected by psoriasis-associated variant rs887314 in CD4+ T cells, then integrated that prediction with existing MPRA-coupled scCRISPRi and CD4+ T-cell Perturb-seq data.
  • Still unknown: The analysis reuses existing published datasets and does not report new wet-lab experiments for this psoriasis case. AlphaGenome supplied a model prediction; cross-dataset signature correlation supports, but does not establish, causal biology. The GPR137 concordance was significant only after stimulation, and the authors describe it as probable/activation-dependent. A human researcher selected which AI-proposed strategy to run. Primary paper source checked; no independent replication assessed.
  • AI role: The AlphaGenome paper agent scored and ranked candidate genes at rs887314. Under human-in-the-loop supervision, the AI co-scientist inspected two other paper agents' datasets, proposed ten validation strategies, and the researcher selected signature-correlation analysis. The resulting cross-screen, cross-modality analysis found significant concordance for GPR137 only under stimulated conditions.

Evidence sources: https://www.nature.com/articles/s41586-026-11044-y

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

AlphaGenome Atlas: predictive map of human DNA variation

A precomputed database of predicted molecular effects for about 9 billion possible single-letter changes in the human genome, generated by AlphaGenome. Researchers can access it free through a portal and API. The predictions are a research tool, do not replace experiments, and are not validated for clinical use.

Technology: AlphaGenome, AlphaGenome Variant Impact (AVI)

Family AlphaGenome: profile · models page

Publication: Announcement only · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: retrospective or benchmark
  • Tested: Validation described is the company's own testing of the AlphaGenome Variant Impact score on pathogenicity benchmarks, plus case studies: a DNM1 variant prioritized with GREGoR and 22% more non-coding associations in UK Biobank.
  • Still unknown: The page does not detail how individual predictions were checked; AlphaGenome is not validated or approved for clinical use.
  • AI role: AlphaGenome predicted variant effects across cell types and tissues; these were precomputed for about 9 billion variants and combined with AlphaMissense into a single score.

Evidence sources: https://deepmind.google/blog/alphagenome-atlas-a-predictive-map-of-every-possible-dna-letter-change-in-the-human-genome/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

YOLO detector tracks endangered European hamster activity in camera traps

A YOLO object-detection model trained on 2023 camera-trap videos detected European hamsters with an F1-score of 0.93 and accuracy of 0.93. Its activity profile matched human annotation and showed peak activity at dusk and a secondary peak before sunrise. The evidence covers one season at one monitoring setup, so generalization to other sites is untested.

Technology: YOLO object detection (deep learning)

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: A YOLO detector trained on 2023 camera-trap videos was tested for European hamster detection (F1 and accuracy 0.93), and its activity profile was compared with human annotation.
  • Still unknown: The abstract states no limitations; it covers one season and does not name the sites or the number of cameras.
  • AI role: A YOLO object-detection model detected hamsters in camera-trap footage.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1002/ece3.74267&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ERTransformer-generated TCR candidates activate T cells in defined receptor contexts

ERTransformer generated epitope-conditioned TCR β-chain CDR3 candidates. Flow-cytometry experiments found selected artificial TCRs induced activation comparable to or greater than natural controls in defined TCR/pMHC contexts. This represents the September 2026 journal report; an ERTransformer precursor was posted October 20, 2023. The earlier report does not establish that these journal wet-lab results were already public.

Technology: ERTransformer; EpitopeBERT; ReceptorBERT

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Selected generated candidates tested for T-cell activation by flow cytometry.
  • Still unknown: Full experimental methods and supplements not assessed: publisher rendering exposed abstract and metadata. General specificity, safety and clinical outcomes are not established by this review.
  • AI role: Generates TCR β-chain CDR3 candidates conditioned on epitopes.

Evidence sources: https://link.springer.com/article/10.1186/s13059-026-04266-w · https://www.biorxiv.org/content/10.1101/2023.10.18.562845v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ESM-FEP combines protein-language-model fitness screening and free-energy calculations to engineer herbicide-detoxifying maize dioxygenase

ESM-FEP was applied to Zea mays dioxygenase ZmHSL1B and nominated a quadruple mutant (Q140H/Y205F/L332R/K336F) with approximately sevenfold higher catalytic efficiency for mesotrione detoxification than wild type, measured by in-vitro assays and kinetic analysis. Arabidopsis thaliana expressing the engineered variant showed enhanced herbicide tolerance in the reported transgenic tests.

Technology: ESM-FEP (fine-tuned ESM-2 protein language model plus alchemical free-energy perturbation)

Family ESM / ESMFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: ESM-FEP was applied to Zea mays dioxygenase ZmHSL1B and nominated a quadruple mutant (Q140H/Y205F/L332R/K336F) with approximately sevenfold higher catalytic efficiency for mesotrione detoxification than wild type, measured by in-vitro assays and kinetic analysis.
  • Still unknown: Reported results concern one maize dioxygenase and mesotrione in specified in-vitro and Arabidopsis tests. They do not establish crop-field performance, commercial efficacy, or generality across herbicides/enzymes. Primary PubMed abstract and publisher citation metadata checked; no independent replication assessed.
  • AI role: Fine-tuned ESM-2 screened saturation-mutagenesis variants for retained global protein fitness; free-energy perturbation then estimated substrate-binding effects and prioritized mutants. The selected M5 mutant was assayed in vitro and in transgenic Arabidopsis by the researchers.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/42675858/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Deep-learning-designed enhancers drive target-tissue reporter activity in mouse embryos

Researchers trained convolutional neural networks on E11.5 mouse heart, limb and midbrain chromatin-accessibility data, then fine-tuned the models using validated human and mouse enhancers and used model-guided sequence optimization to design 15 synthetic enhancers. In vivo transgenic reporter assays found all 15 reproducibly active in their intended tissue. Four of five heart designs and four of five CNS designs were exclusive to the target tissue; all five limb designs were strongly active in limb and had weaker activity in other TWIST1-expressing mesenchymal tissues.

Technology: Convolutional neural networks for sequence-to-accessibility and sequence-to-activity prediction, transfer learning, and Ledidi gradient-based sequence design

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Researchers trained convolutional neural networks on E11.5 mouse heart, limb and midbrain chromatin-accessibility data, then fine-tuned the models using validated human and mouse enhancers and used model-guided sequence optimization to design 15 synthetic enhancers.
  • Still unknown: The functional test was a transgenic reporter assay at embryonic day 11.5, not editing or activation of an endogenous native gene locus. Fifteen selected sequences were tested (five per target), so the reported all-positive result does not estimate success across every generated candidate. Some designs had weaker off-target activity, particularly in related mesenchymal tissues. These results do not establish therapeutic efficacy or safe gene therapy.
  • AI role: Models learned tissue-specific chromatin-accessibility patterns from genome-wide ATAC-seq, then transfer learning adapted them to validated enhancer activity. The trained models guided de novo enhancer sequence edits; the authors selected 15 candidates for transgenic mouse-embryo reporter testing, where each showed reproducible activity in its intended target tissue.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC13553301/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Anthropic

Claude designs protein binders for 15 targets

After an initial prompt, Claude selected binding sites, used design and folding models, and chose candidates, with success on 14 of 15 targets. Of 1,320 designs, 354 bound, for rates of 22–35% versus a typical 10–15%. None of the 90 designs for the maltose-binding protein worked. External labs ran the tests.

Technology: Claude (Mythos Preview and Opus 4.8)

Family Claude: profile · models page

Publication: Announcement only · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: direct test
  • Tested: External labs (Adaptyv Bio and Twist Bioscience) produced and tested the designs: 354 of 1,320 designs bound, with hit rates of 22-35% versus a typical 10-15%; none of 90 designs for maltose-binding protein bound. The collaborating external laboratories performed candidate production and binding assays within the discovery project; separate independent follow-up replication is not established by this account.
  • Still unknown: Anthropic plans further characterization to confirm hit rates and affinities; BBF-14 affinities are modest, GDF-8 data were inconclusive and excluded, and the TNFa result is unexplained.
  • AI role: Claude chose binding sites, generated candidate structures and sequences by orchestrating structure, sequence and co-folding models, ran in silico optimization and chose candidates.

Evidence sources: https://www.anthropic.com/research/Claude-accelerates-protein-design

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

FluEmbed predicts H3N2 antigenic drift from sequence alone

FluEmbed, a protein language model framework, predicted influenza A (H3N2) antigenicity from RNA sequence with Spearman correlations of 0.67 to 0.80 against hemagglutination inhibition titers. It outperformed Hamming, BLOSUM62 and phylogenetic baselines, and in-silico mutagenesis showed unconstrained mutations carry greater escape potential outside dominant lineages. Validation uses existing HI assay data and retrospective sequences; no new wet-lab tests were run.

Technology: FluEmbed (protein language model embeddings)

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: FluEmbed predictions from H3N2 sequences were correlated with existing hemagglutination inhibition titers (Spearman 0.67-0.80) and compared with Hamming, BLOSUM62 and phylogenetic baselines, with in-silico mutagenesis for escape potential.
  • Still unknown: The abstract states no limitations and does not discuss generalization beyond H3N2 or beyond the data used; no wet-lab tests were run.
  • AI role: A protein language model framework predicted H3N2 antigenicity from RNA sequence.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1371/journal.pcbi.1014628&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

MP4 text-to-protein design yields ATP-active proteins, including one with a novel fold

The transformer MP4 maps plain-language prompts to protein sequences. The study reports two experimentally validated de novo designs that were expressible and thermostable and showed ATP binding and hydrolysis in vitro. X-ray structures at 1.30 Å and 1.77 Å showed that one design adopts a fold not previously observed in the authors’ structural comparisons.

Technology: MP4, a transformer-based text-to-protein model

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The study reports two experimentally validated de novo designs that were expressible and thermostable and showed ATP binding and hydrolysis in vitro.
  • Still unknown: Only two designed proteins are described as experimentally validated in the abstract; this is an early proof of concept, not a demonstrated general success rate. ATP binding and hydrolysis are in-vitro results, not a novel physiological function or therapeutic effect. The source describes one fold as novel; no claim is made that both structures are novel.
  • AI role: MP4 converted natural-language protein specifications into candidate sequences. Researchers evaluated 96 prompts in silico, selected candidates, expressed and tested designs, and solved two crystal structures. The source attributes the designed sequences to MP4; experimental teams established expression, thermostability, ATP binding/hydrolysis and structural novelty.

Evidence sources: https://pubs.acs.org/asbcd6/article/doi/10.1021/acssynbio.6c00246/5254330/An-Artificial-Intelligence-Model-for-Translating

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-learning-guided CsrA homolog screens reveal divergent bacterial regulatory activities

A Swarm-seq assay tested more than 500 bacterial CsrA homologs for regulation of Bacillus subtilis swarming and found two functional classes: several homologs strongly inhibited swarming while others failed despite conserved sequence and canonical GGA motifs in the hag transcript. A random-forest model trained on these assays predicted homolog activity across thousands of sequences; experimental testing supported predicted Class I regulatory activity for Bdellovibrio bacteriovorus CsrA and Pseudomonas putida RsmA. The findings indicate sequence conservation and the canonical RNA motif do not alone determine these tested regulatory effects.

Technology: Random-forest machine-learning classifier trained on Swarm-seq functional assay data

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: A Swarm-seq assay tested more than 500 bacterial CsrA homologs for regulation of Bacillus subtilis swarming and found two functional classes: several homologs strongly inhibited swarming while others failed despite conserved sequence and canonical GGA motifs in the hag transcript.
  • Still unknown: Activity classifications are based on the reported B. subtilis swarming assay and tested proteins; they do not establish all native targets or regulatory functions in each source organism. The predicted function landscape across thousands of homologs is not experimentally validated wholesale. Primary publisher abstract checked; no independent replication assessed.
  • AI role: High-throughput swarming assays generated a functional dataset; machine learning was then trained to classify homolog functional activity across the bacterial domain and nominate homologs for validation. Selected CsrABb and RsmAPp predictions were experimentally validated.

Evidence sources: https://academic.oup.com/nar/article/54/14/gkag756/8753004

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning identifies the first enzyme for phenazine–thiol conjugation

A genome-informed machine-learning workflow (ML-CITO) used contrastive learning over protein-language-model representations and genomic context, starting from 14 known phenazine-modifying sequences, to identify PTC. Biochemical characterization showed PTC catalyzes glutathione-dependent phenazine modification, a reaction previously assumed to occur non-enzymatically.

Technology: ML-CITO (genome-context-informed data augmentation plus protein-language-space contrastive learning)

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: A genome-informed machine-learning workflow (ML-CITO) used contrastive learning over protein-language-model representations and genomic context, starting from 14 known phenazine-modifying sequences, to identify PTC.
  • Still unknown: The paper reports that deleting the PTC gene did not cause a strong fitness disadvantage under tested conditions. The demonstrated chemistry and substrate context should not be generalized to other phenazine reactions.
  • AI role: The model ranked candidate phenazine-interacting proteins from sparse known examples; the team then expressed and biochemically tested the candidate enzyme. AI nomination was a direct step in discovering the previously unknown enzymatic reaction.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC13419639/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

REAP-guided engineering improves P450 BM3 activity

REAP combines a protein-language-model ranking predictor with active selection and automated build-and-test cycles. The study reports a P450 BM3 quintuple variant with 57-fold higher product yield than the engineered FL#62 starting scaffold in its high-throughput assay. This is a variant-specific comparison with an already engineered scaffold, not a 57-fold gain over wild type. A separate sortase result is not included in this bounded draft.

Technology: REAP; PLM-RankReg; automated build-and-test

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Publisher indexed article reports a quintuple variant and 57-fold product-yield increase over FL#62 in the high-throughput assay; kinetic characterization is also described.
  • Still unknown: Independent replication and broader application are not established by this bounded review.
  • AI role: Learned ranking and active selection nominate variants for automated synthesis and testing; the reported gain is measured against the engineered FL#62 scaffold.

Evidence sources: https://www.nature.com/articles/s41467-026-76264-2

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Publisher indexed primary article text, title/DOI and result; PMC direct open CAPTCHA challenged. Journal/preprint chronology and sequence overlap not audited..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning and directed evolution improved an L-Dopa-producing monooxygenase

Zero-shot protein models prioritized HpaBC substitutions, which were experimentally assayed with a L-Dopa-responsive GFP biosensor. EVOLVEpro then combined experimentally validated substitutions, including directed-evolution-derived variants, and selected higher-order designs. The best reported combined variants produced up to 57–58 times the wild-type fluorescence signal under the stated assay conditions. This is a biosensor proxy for L-Dopa production, not a purified-enzyme kinetic measurement; the work also builds on prior directed-evolution mutations and does not establish clinical benefit.

Technology: MutCompute, ProteinMPNN, ESM protein language models and EVOLVEpro combinatorial design

Family ESM / ESMFold: profile · models page

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: 329 model-predicted variants were assayed; 19 showed at least 1.25-fold biosensor signal. EVOLVEpro-selected combined variants yielded up to 57–58-fold normalized fluorescence relative to wild type under the reported assay conditions.
  • Still unknown: Fluorescence from a DopA/GFP biosensor is an indirect L-Dopa-production proxy, not a purified-enzyme catalytic constant. The work includes directed-evolution-derived mutations; G295T had prior directed-evolution support. No human therapeutic efficacy is established.
  • AI role: MutCompute, ProteinMPNN and ESM prioritized single substitutions for experimental testing; EVOLVEpro combined tested mutations and iteratively selected higher-order variants, including designs incorporating directed-evolution results.

Evidence sources: https://api.biorxiv.org/details/biorxiv/10.64898/2026.07.31.742090 · https://www.biorxiv.org/content/10.64898/2026.07.31.742090v1.full

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary bioRxiv v1 API metadata and Crossref posted-date metadata checked; primary v1 full-text indexed extract reviewed for methods, biosensor-normalized outcomes, model role and affiliation. Direct full-text page open was blocked..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning prioritizes a protective Arabidopsis leaf fungus that suppresses oomycete infection

Four classifiers (random forest, support-vector machine, logistic regression and multilayer perceptron) used six years of natural Arabidopsis leaf-microbiome data to prioritize taxa associated with Albugo infection status. Follow-up plant assays found that selected health-associated microbes reduced infection; Cystofilobasidium macerans OTU126 gave the strongest individual protection, reducing Albugo infection by an average of 73%. In a 15-member synthetic leaf community, adding this fungus increased the reported infection reduction to 40.25%, compared with 19.36% for the community alone.

Technology: Random forest, support-vector machine, logistic regression and multilayer perceptron classifiers with cross-validated feature selection

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Follow-up plant assays found that selected health-associated microbes reduced infection; Cystofilobasidium macerans OTU126 gave the strongest individual protection, reducing Albugo infection by an average of 73%.
  • Still unknown: This is a controlled Arabidopsis–Albugo plant-pathogen model, not a field crop efficacy trial. ML features are associations with infection status; they do not establish the microbes caused naturally observed health status. Isolates were matched to predicted OTUs using marker sequences, which the authors note do not fully establish strain-level functional identity. Genomic terpene-cluster hypotheses were not directly tested. No commercial or field-ready biocontrol performance is established.
  • AI role: Models classified infected versus uninfected plants from microbial community profiles and prioritized disease-associated and health-associated taxa for follow-up. Researchers then tested selected cultured isolates in planta and tested Cystofilobasidium in a synthetic community.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/42521687/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning enriched MthUPO variants for beta-damascone hydroxylation

Researchers combined experimental activity measurements with machine learning to enrich a second MthUPO peroxygenase library for substrate-specific activity. Across 506 distinct variants screened and sequenced, 97 of 108 variants in the ML-enriched library outperformed wild type in the screening assay, compared with 14% in the initial htFuncLib library. The authors report an 11.8-fold turnover number for the best variant producing 4-hydroxy-beta-damascone, with greater than 99% regioisomeric excess. Separate 5–12-fold enrichment and recovery figures in the paper are simulation benchmarks and are not treated as experimental outcomes.

Technology: Machine-learning enrichment of experimental enzyme-activity data following htFuncLib design

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: The preprint reports screening and sequencing 506 distinct MthUPO variants. In the ML-enriched library, 97 of 108 variants outperformed wild type in the screening assay, compared with 14% in the initial htFuncLib library. The best reported variant had an 11.8-fold turnover number for 4-hydroxy-beta-damascone with greater than 99% regioisomeric excess.
  • Still unknown: The preprint is not peer reviewed. The paper separately reports simulation-only MLEE/htFuncLib comparisons, which are excluded from the experimental result summary. Assay outcomes do not establish process-scale productivity or commercial suitability.
  • AI role: Machine learning used experimental enzyme-activity measurements to enrich a library of MthUPO variants for testing, following an earlier htFuncLib design stage.

Evidence sources: https://api.biorxiv.org/details/biorxiv/10.64898/2026.07.20.739108 · https://www.biorxiv.org/content/10.64898/2026.07.20.739108v1.full

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Official bioRxiv API v1 metadata/abstract and primary indexed v1 full-text excerpts; simulation and experimental results kept separate..

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

EnzGFM-Agent experimentally prioritizes activity-enhancing secreted alkaline phosphatase variants

EnzGFM-Agent scored all possible single amino-acid substitutions in secreted alkaline phosphatase (SEAP) and selected top candidates from EnzGFM-1.5B and ESM2-3B for testing in transfected Chinese hamster ovary cells. Among the five variants uniquely prioritized by EnzGFM, all five had significantly higher measured activity than wild type in the reported assay; three exceeded 1.5-fold, and A46I reached 6.52-fold. This is an experimentally measured enzyme-engineering result for the tested SEAP variants.

Technology: EnzGFM-Agent with EnzGFM-1.5B and ESM2-3B

Family ESM / ESMFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: EnzGFM-Agent scored all possible single amino-acid substitutions in secreted alkaline phosphatase (SEAP) and selected top candidates from EnzGFM-1.5B and ESM2-3B for testing in transfected Chinese hamster ovary cells.
  • Still unknown: The evidence is limited to single substitutions in one SEAP reporter protein tested after transfection in CHO-S cells, with three independent biological replicates; it does not establish improved activity for other enzymes, industrial process performance, or a naturally occurring enzyme function. The five EnzGFM-specific variants all differed significantly from wild type, but the study also evaluated ESM2-ranked variants and five variants shared by both models. This is a paper-specific experimental validation, not independent replication.
  • AI role: The workflow scored every possible SEAP single-point substitution, ranked candidates using EnzGFM-1.5B and ESM2-3B, and selected 15 unique variants for experimental validation. The distinct EnzGFM-prioritized group comprised five variants; the paper reports significantly increased SEAP activity for all five, with A46I at 6.52-fold wild type.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC13494029/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Deep learning identified minimal mutations for cross-reactive anti-PD-L1 antibodies

A deep-learning model trained on deep-mutational-scanning sequence–binding data selected minimal mutation combinations for a human anti-PD-L1 antibody with mouse cross-reactivity. Variants with four to five substitutions showed in-vitro and cellular binding comparable to conventionally recombined variants with 13–15 substitutions; the paper also reports blockade assays and therapeutic activity in a mouse tumor model. These are preclinical antibody-engineering results, not evidence of human efficacy. The preprint reports patent applications related to the subject matter.

Technology: Deep-learning model trained on deep-mutational-scanning sequence–binding data

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: The preprint reports in-vitro and cellular binding properties of selected variants, PD-1/PD-L1 competition/blockade assays, epitope mapping and mouse tumor-model assessment. The 4–5 substitution variants are compared with highly mutated variants requiring 13–15 substitutions.
  • Still unknown: Mouse-model activity is preclinical and does not establish human therapeutic efficacy. The preprint reports patent applications relevant to the work; no freedom-to-operate or patent-priority conclusion is made.
  • AI role: A deep-learning model trained on DMS-derived sequence-binding data selected minimal mutation combinations for experimental antibody characterization.

Evidence sources: https://www.biorxiv.org/content/10.64898/2026.07.15.738714v1.full · https://www.biorxiv.org/content/10.64898/2026.07.15.738714v1 · https://api.biorxiv.org/details/biorxiv/10.64898/2026.07.15.738714

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary bioRxiv v1 indexed full-text/abstract, author-affiliation metadata and official API metadata; direct page open returned 403..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

UniKP prioritizes a sulfotransferase for sulfotyrosine production

UniKP was used to rank sulfotransferase homologues by predicted kinetics. The study reports experimental testing of five prioritized candidates and identifies EfST1A1, with about 6.5-fold higher in-vitro activity than the initial hit under the reported assay. Later directed evolution is not attributed to the model. This is a bounded enzyme-discovery result, not proof of industrial production or broad enzyme generality.

Technology: UniKP kinetic prediction

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Publisher abstract/indexed article text reports five tested AI-ranked enzymes and identifies EfST1A1; institute research announcement summarizes experimental activity and later evolution.
  • Still unknown: Independent replication and broader application are not established by this bounded review.
  • AI role: UniKP ranks enzyme homologues by predicted kinetic properties; researchers tested top candidates and separately applied directed evolution to the selected hit.

Evidence sources: https://www.sciencedirect.com/science/article/abs/pii/S1385894726047108

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Publisher article preview/indexed text and institute source; full publisher text inaccessible. Issue date is recorded; exact online-first and earlier version not established..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Diffusion-guided peptide design yields antimicrobial candidate Arcinin

ARCADIAMP combines iterative discrete diffusion generation, ESM2-based activity classifiers and sequence novelty filtering. Eight of ten experimentally screened candidates met the paper’s MIC threshold; Arcinin showed activity against tested pathogens and reduced bacterial burden in a mouse wound infection model. These are selected preclinical findings, not evidence of human safety or treatment efficacy; formal resistance-evolution testing remains future work.

Technology: Iterative discrete diffusion model with ESM2 classifiers and novelty filter

Family ESM / ESMFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: As described in the primary source
  • Still unknown: Independent replication and global priority not established by this review.
  • AI role: Iterative discrete diffusion model with ESM2 classifiers and novelty filter

Evidence sources: https://www.nature.com/articles/s41467-026-75030-8

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary publisher/preprint material inspected for identity, AI role and bounded result..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AI-prioritized CRISPR screen identifies ALOX5 and OXTR as regulators of keratinocyte IL17RA in psoriasis

A genome-wide CRISPR knockout screen in primary human epidermal keratinocytes, combined with VirtualCRISPR language-model prioritization, identified ALOX5 and OXTR as previously underappreciated cell-intrinsic regulators of surface IL17RA. Targeted knockout in keratinocytes from two adult donors reduced IL17RA and inflammatory cytokine secretion. Topical inhibition of ALOX5 or OXTR also reduced disease signs in an imiquimod mouse model of psoriasiform dermatitis.

Technology: VirtualCRISPR, a functional-genomics language-model framework trained on BioGRID-ORCS screens

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: A genome-wide CRISPR knockout screen in primary human epidermal keratinocytes, combined with VirtualCRISPR language-model prioritization, identified ALOX5 and OXTR as previously underappreciated cell-intrinsic regulators of surface IL17RA.
  • Still unknown: The genetic validation used primary cells from two donors, while therapeutic effects were tested in an imiquimod-induced mouse model, not patients. The screen provided the initial unbiased hits; the AI supplied an orthogonal novelty-prioritization layer and did not itself measure biological effects. The paper reports a provisional patent application covering methods and findings. No clinical efficacy, safety or topical treatment benefit in humans is established.
  • AI role: After the unbiased genome-wide CRISPR screen, VirtualCRISPR scored whether perturbing each gene in primary adult keratinocytes would alter IL17RA expression. The authors combined screen enrichment with the model’s lower prior probability for noncanonical hits to prioritize high-novelty regulators; targeted CRISPR knockout and pharmacologic mouse experiments tested the resulting biology.

Evidence sources: https://www.nature.com/articles/s41467-026-75249-5

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Active learning improved whole-cell antibiotic screening and validation

A closed-loop active-learning campaign for Borrelia burgdorferi whole-cell screening increased the experimental hit rate five-fold, from 0.2% in investigator-selected plates to 1.0% in the AL-guided campaign. The same study separately reports prospective in-silico selection of diverse compounds across bacterial species, with 11.0% experimental validation of predicted hits and 53-fold enrichment over investigator-directed screening; those validated compounds showed the intended narrow-spectrum B. burgdorferi activity. These are assay and screening results, not clinical efficacy.

Technology: Active learning using a directed message-passing neural network and evidential Dirichlet head

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: In a closed-loop Borrelia burgdorferi campaign, active-learning-guided selection produced a 1.0% experimental screening hit rate versus 0.2% in investigator-selected plates. A separate prospective in-silico selection across bacterial species achieved 11.0% experimental validation of predicted hits (53-fold enrichment over investigator-directed screening); the paper reports intended narrow-spectrum activity for B. burgdorferi.
  • Still unknown: Preprint evidence is not independent replication or clinical testing. The cross-species hit-validation experiment is distinct from the closed-loop Borrelia hit-rate comparison. Direct full page retrieval was blocked; checked material was the primary indexed preprint and archived primary article text.
  • AI role: A directed message-passing neural network with an evidential Dirichlet head supported active-learning compound selection for whole-cell inhibition testing; selected compounds were then experimentally screened.

Evidence sources: https://www.biorxiv.org/content/10.64898/2026.07.04.736489v1 · https://pmc.ncbi.nlm.nih.gov/articles/PMC13344992/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary bioRxiv indexed text, PubMed metadata/abstract and PMC indexed full-text excerpt; direct bioRxiv and PMC page loads were blocked..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AI-guided mining identifies and engineers an alkaline xylanase

A fine-tuned sequence model and generative workflow guided identification of xylanase 8E20 and mutant 8E20-178. The primary abstract reports an optimum-pH shift from 8 to 10 and a 1.9-fold improvement, but calls the metric catalytic activity while indexed conclusion text calls it catalytic efficiency; this draft preserves the ambiguity. It does not infer unreported assay conditions or industrial-scale performance.

Technology: AAEPre sequence model; xynODM generative model

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Primary abstract reports 8E20-178, pH optimum shift and 1.9-fold improvement.
  • Still unknown: Publisher full text inaccessible; exact improvement metric and assay conditions unresolved. Prior sequence characterization and independent replication not assessed.
  • AI role: Sequence models and composite selection index prioritized candidate enzyme; generative model guided variant engineering.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/42398548/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary source checked for identity, publication date, AI workflow and the bounded experimental result stated..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning guided a PET-binding module into a reusable whole-cell catalyst

An HMM and supervised PET hydrolase model screened sequence candidates and prioritized the PET-binding module tCBM13-1. Experiments found greater PET adsorption and retention than the CBM2 comparator; when displayed with FAST-PETase on E. coli, the system increased PET-film depolymerization by about 43% and retained over 64% of initial activity after ten cycles. These are laboratory assays, not evidence of industrial-scale recycling.

Technology: Profile-HMM homology search combined with supervised PET hydrolase machine-learning model PETML

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The authors experimentally measured tCBM13-1 adsorption and retention on PET and tested whole-cell co-display with FAST-PETase. The publisher abstract reports about 43% greater PET-film depolymerization (2793 μg/(d·cm²)) and over 64% residual activity after ten cycles.
  • Still unknown: Results are laboratory PET assays; they do not establish industrial-scale recycling, environmental deployment, or lifecycle benefit.
  • AI role: A supervised PET hydrolase machine-learning model within an HMM-based sequence-mining workflow prioritized tCBM13-1 from candidate modules for experimental adsorption and whole-cell depolymerization tests.

Evidence sources: https://www.sciencedirect.com/science/article/pii/S0960852426013611 · https://pubmed.ncbi.nlm.nih.gov/42385911/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary publisher-indexed full text and primary NCBI/PubMed metadata/abstract, including official article date and affiliations..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

BirdNET embeddings distinguish individual North Atlantic right whales

Audio embeddings from the BirdNET model robustly distinguished individual North Atlantic right whales in upcall recordings from 234 samples across 11 individuals at 3 sites. Simulations showed that an extra 7 dB or more of signal excess is needed to identify individuals under ambient noise. The study uses on-animal archival tags on an endangered species, but the sample is small and the model was not trained for whales.

Technology: BirdNET audio embeddings (deep learning)

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: BirdNET audio embeddings were tested for distinguishing 11 individual right whales in 234 upcall samples, and noise simulations estimated the extra signal excess needed for individual identification compared with detection.
  • Still unknown: The abstract lists no explicit limitations; noise effects are simulated rather than field-measured, and the dataset is small (11 individuals at 3 sites).
  • AI role: The BirdNET deep-learning model generated audio embeddings used to classify individual identity.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1371/journal.pcbi.1013321&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Few-shot protein-language-model design yields high-output bicyclogermacrene synthase variants

After experimentally screening activity-enhancing single mutants of LdTPS5, the researchers used PIFS-PLM, a few-shot model that selects features from protein-language-model embeddings using active-site and conserved-residue information, to propose mutation combinations. All 10 tested model-predicted combinations were active; nine produced more than 100 mg/L bicyclogermacrene in yeast, and the top variant produced 213.59 mg/L versus 2.93 mg/L for wild type (about 73-fold higher yield). A related selectivity-optimized combination increased bicyclogermacrene selectivity by 15%.

Technology: PIFS-PLM (Physics-Inspired Feature Selection of Protein Language Models), using ESM-2 embeddings with a few-shot mutation-combination classifier

Family ESM / ESMFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: After experimentally screening activity-enhancing single mutants of LdTPS5, the researchers used PIFS-PLM, a few-shot model that selects features from protein-language-model embeddings using active-site and conserved-residue information, to propose mutation combinations.
  • Still unknown: The 73-fold result is product yield in the reported yeast system, reflecting effects of enzyme activity and expression; it is not a 73-fold increase in catalytic rate. The selectivity improvement was 15%. The model relies on a defined active site/conserved-site set and a small, enzyme-specific experimental dataset. Product profiles differed between cell-free extracts and yeast; these results do not establish universal performance across enzymes or production settings.
  • AI role: After single substitutions were screened and activity-enhancing mutations experimentally validated, PIFS-PLM used 60–100 measured combinations to learn from selected ESM-2 features at active-site and conserved residues and predict beneficial combinations. Ten predicted LdTPS5 variants were tested; all were active, and the top variants substantially increased measured yeast product yield. The model also guided a selectivity optimization test. Initial single-mutation nomination used MM/GBSA or sequence conservation analysis, separate from PIFS-PLM combinatorial predictions.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC13502649/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

MutexaGPT-guided cold-adapted amylase variants

The MutexaGPT multi-agent large language model system translated a cold-adaptation design request into molecular simulations and linker-variant proposals for bidomain amylase. The authors report experimentally validated variants with 1.7-fold and 3.7-fold activity enhancement at 0 °C. The preprint appeared on 19 June 2026; the peer-reviewed journal version followed on 10 September 2026.

Technology: Multi-agent large language model system coordinating high-throughput molecular dynamics simulations and mutation design

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: The primary article reports two MutexaGPT-guided bidomain amylase linker variants with experimentally measured 1.7-fold and 3.7-fold activity enhancement at 0 °C. The Methods define relative activity using the 0 °C to 45 °C activity ratio.
  • Still unknown: The experimental result is specific to the tested bidomain amylase and variants; broader enzyme-family transfer and independent reproduction are not established.
  • AI role: LLM agents clarify the design request, translate it into computational plans and mutation patterns, and explain simulation outputs. High-throughput molecular dynamics metrics guide the proposed variants; experimental assays test their activity.

Evidence sources: https://assets-eu.researchsquare.com/files/rs-10072540/v1_covered_65be6e89-30da-488a-a7d7-0310d3a07d5f.pdf · https://www.nature.com/articles/s43588-026-01049-y

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary Research Square v1 full-text indexed content and metadata; primary Nature Computational Science indexed article text, including abstract, Methods, experimental result, date and affiliation block..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Deep learning identifies antimicrobial peptides hidden in prion-related proteins

APEX 1.1 screened 19.3 million peptide fragments from 2,897 prion-related proteins and identified 1,179 candidate antimicrobial peptides, termed prionins. Among 75 synthesized candidates, 59 inhibited at least one tested bacterial strain; two selected peptides reduced Acinetobacter baumannii burden in a mouse skin-abscess model. The study shows that prion-related sequences can yield experimentally active antimicrobial fragments, but does not show that these peptides are naturally released or function as host-defense agents in their source organisms.

Technology: APEX 1.1 deep-learning antimicrobial-peptide prediction and prioritization

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Of 75 synthesized APEX-selected prionins, 59 inhibited at least one bacterial strain in vitro; two selected peptides reduced A. baumannii burden in a mouse skin-abscess model.
  • Still unknown: No evidence shows these fragments are naturally released from their source proteins or physiologically function as host-defense peptides in native organisms. Mouse results do not establish human efficacy or safety.
  • AI role: APEX 1.1 scored peptide fragments for predicted antimicrobial activity and prioritized candidates from a search across prion-related proteins for synthesis and experimental screening.

Evidence sources: https://www.nature.com/articles/s41564-026-02408-1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary Nature Microbiology full text as indexed, including methods, results, author contributions and affiliations..

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Deep learning guides higher-activity SpuFz1 nuclease variants

A DeepSequence-based predictor ranked SpuFz1 mutations without experimental activity labels; experimentally confirmed substitutions were combined into SpuFz1-v3. At the B2M locus in HEK293T cells, the article reports 35.52% indel formation versus 3.07% for wild type. Comparative sequence analysis, not this AI model, guided the separate ultrashort RNA design. The reported activity is assay- and locus-specific; off-target and clinical safety are not established.

Technology: DeepSequence-based Fanzor-Fitness Predictor

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The primary source reports the study-specific measured result summarized here.
  • Still unknown: Independent replication and clinical or industrial efficacy are not established.
  • AI role: A learned sequence fitness predictor ranks candidate mutations; researchers assay single substitutions and combine successful variants.

Evidence sources: https://www.nature.com/articles/s41467-026-74624-6

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary source indexed full text or abstract and publication metadata reviewed; scope stated in source_checks..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Fung-AI prioritizes peptides with measured antifungal activity

Fung-AI combines a generative adversarial network with learned antifungal and hemolysis classifiers to select peptides for testing. The study first tested five peptides from separate clusters; activity of peptide 40 prompted testing of eight more members from cluster 17. Five of the nine tested cluster-17 peptides inhibited Fusarium graminearum, a post-selection cluster-specific outcome—not a general 5/13 model hit rate. Four peptides showed activity against Candida albicans, while none inhibited C. auris up to 500 μg/mL. The paper’s HepG2 cytotoxicity statements are internally inconsistent for peptide 48 and peptide 65, and peptide 12 and 40 have reported cytotoxicity; these assays establish neither human safety nor clinical efficacy.

Technology: Fung-AI GAN; learned antifungal and hemolysis classifiers

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Primary article reports staged testing, fungal MICs and cell-line cytotoxicity; adaptive denominator and conflicting LC50 statements were reconciled in the continuation audit.
  • Still unknown: Sequence-level prior art and independent replication not assessed; HepG2 assays do not establish human safety.
  • AI role: GAN generates candidates; antifungal and hemolysis classifiers narrow the candidate set before synthesis and fungal assays.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC13289945/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary PLOS article and indexed PMC full text checked; follow-up audit read full article and resolves staged denominator and LC50 inconsistency, not by remeasurement..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-learning screening identifies bacterial-selective membrane-active compounds

The June 2026 bioRxiv version describes an initial machine-learning-guided screen of ZINC20 compounds balancing predicted antibacterial activity, mammalian-cell toxicity, and mechanism-of-action novelty. Among 61 purchased candidates, six showed partial growth inhibition; 49 were non-toxic to HepG2 cells at 50 μM. Flow cytometry indicated membrane disruption for four of the six initial active hits and membrane depolarization for two. A September version reports additional screening and revised models; its expanded results are not attributed to the June version here. The study does not report a novel protein target or clinical antibiotic.

Technology: Graph-convolutional neural networks for antibacterial activity/toxicity with XGBoost mechanism-of-action classification and novelty filtering

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: The June 2026 v1 reports that 61 purchased candidates were tested, six showed partial growth inhibition, and 49/61 were non-toxic to HepG2 cells at 50 μM. Flow cytometry indicated membrane integrity disruption for four initial active hits and membrane depolarization for two.
  • Still unknown: The initial work does not report a novel protein target and is limited to the stated bacterial assays and HepG2 toxicity concentration. The September version adds expanded screening and revised analyses; those later results are not attributed to the June v1.
  • AI role: Learned activity and toxicity models, together with a mechanism-novelty classifier, prioritized compounds for purchase and experimental testing.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC13278111/ · https://pubmed.ncbi.nlm.nih.gov/42327090/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary PMC-indexed full text for the June bioRxiv v1, primary PubMed preprint abstract/metadata, and primary PMC/PubMed metadata for the later September version to distinguish version-specific results..

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AI-designed bacterial anti-defense proteins enable phage infection and plasmid retention

RFdiffusion-based de novo binder design and experimental screening produced synthetic proteins that inhibited three bacterial defense systems: Thoeris and Avs1 in Bacillus subtilis assays, and DdmDE in Vibrio cholerae. Five designed proteins inhibited Thoeris and nine inhibited Avs1 in the reported screens. An engineered SBSphiJ phage carrying both anti-Thoeris and anti-Avs1 proteins replicated in B. subtilis cells expressing both defenses. In native V. cholerae O1 El Tor, two of four tested DdmDE binders improved plasmid resistance to clearance.

Technology: RFdiffusion de novo binder design, ProteinMPNN sequence design and AlphaFold2/3 structure-guided candidate scoring

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: RFdiffusion-based de novo binder design and experimental screening produced synthetic proteins that inhibited three bacterial defense systems: Thoeris and Avs1 in Bacillus subtilis assays, and DdmDE in Vibrio cholerae.
  • Still unknown: These are laboratory proof-of-concept results in three bacterial defense systems and specific strains; they do not demonstrate therapeutic phage efficacy in animals or humans. Binder success was low relative to the computationally designed candidate sets, and the authors note that a verified binder may not work against homologous defense proteins. The native V. cholerae result was tested for plasmid retention, not general transformation across clinical isolates. The earlier bioRxiv version is the same study and is not a separate record.
  • AI role: RFdiffusion generated candidate binder backbones for selected target sites in the Thoeris, Avs1 and DdmDE defense proteins; ProteinMPNN designed sequences and AlphaFold-based co-folding/scoring prioritized candidates for experiments. The selected synthetic binders were tested for defense inhibition, direct interactions and ability to restore phage infection or plasmid retention. The measured anti-defense functions and engineered-phage outcomes are the paper-specific results enabled by AI-guided protein design.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC13402655/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Family-specific machine learning prioritizes enzyme variants for prospective testing

Researchers screened 1,513 natural SGNH-hydrolase homologues and trained task-specific sequence CNNs on the resulting activity, thermostability, and substrate-preference data. The models helped select candidates for a prospective 230-sequence assay panel, where CNN-selected candidates were enriched for target traits relative to generalist pretrained and physics-based baselines; the panel also included other selection strategies and random controls. Assays used clarified E. coli lysates and measured relative maximum velocity, integrating expression and catalysis. The first result-bearing version was posted to bioRxiv on 2026-06-03; a Research Square version followed on 2026-06-05.

Technology: Task-specific convolutional neural networks trained on amino-acid sequences; CNN-based HEV and UCB selection strategies

Publication: Preprint · Test context: prospective or independent check. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: prospective or independent check
  • Tested: A prospective panel of 230 previously untested sequences was assayed; CNN selection strategies were enriched for the predicted target traits compared with generalist pretrained and physics-based approaches. The panel pooled sequences selected by multiple strategies and included random controls.
  • Still unknown: Clarified lysate MaxV integrates enzyme expression and catalytic activity and is not a purified-enzyme turnover measurement. The result is scoped to this enzyme family and the reported assay setup; prospective panel composition was mixed across selection strategies.
  • AI role: Supervised family-specific CNNs trained on sequence and assay data ranked untested homologues for activity, thermostability, and substrate preference, and guided candidate selection for prospective experimental testing.

Evidence sources: https://www.biorxiv.org/content/10.64898/2026.06.02.729712v1.full

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary bioRxiv v1 full text as indexed, including title, posting date, affiliations, methods, prospective assay results and limitations; primary Research Square v1 identity/date record..

Editorial assessment: Impact 3/5 · Validation 4/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Satellite model maps collapse and surviving Acropora corals in Belize

A supervised Random Tree classifier on satellite images predicted a drop of over 90% in live Acropora cover at Belize from 2011 to 2023 across 25 km². Field searches in June 2025 found live or recently dead colonies at 28 of 42 predicted sites and live colonies at 16 previously undocumented locations. The field check was a search for survivors at sites the model flagged, not a blind test.

Technology: Random Tree supervised classification of satellite imagery (machine learning)

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: A Random Tree classifier's predicted over-90% decline in Acropora cover (2011-2023) was compared with field data from 2012-2025, and field searches at 42 predicted sites found live or recently dead colonies at 28 of them.
  • Still unknown: The abstract states no limitations; the search was a targeted check of flagged sites rather than a blind test, and 14 of 42 predicted sites had no colonies found.
  • AI role: A supervised Random Tree classifier identified Acropora in satellite images and generated the cover predictions that guided the field searches.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41598-026-47081-w&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ML ranking identifies active sesquiterpene-synthase variants

A composite score combines Boltz2-predicted binding and DLKcat-predicted turnover to rank variants in a focused Agr5 library. Of 15 top-ranked variants tested, four produced more viridiflorol than wild-type Agr5; the paper reports 11.05-fold higher production for Y374F and a 952.79 mg/L titer in its yeast assay. The initial V314G design involved mechanism-guided human analysis, and the measured values are assay-specific.

Technology: DeEnzyme_score combining Boltz2 and DLKcat predictions

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The article reports four activity-enhancing variants among 15 tested and specifies the Y374F production ratio and titer in a yeast assay.
  • Still unknown: Prior sequence-level overlap, earlier disclosure and independent replication remain unestablished; the proposed allosteric explanation partly relies on simulation.
  • AI role: Boltz2 and DLKcat predictions are combined to rank variants selected for experimental testing; the initial V314G design is attributed to mechanism-guided analysis, not AI.

Evidence sources: https://pubs.acs.org/accacs/article/16/14/13022/5114435/Customizing-Natural-Products-of-Sesquiterpene

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Publisher indexed full article text including title, methods, results, article history and author-affiliation block; direct ACS endpoint returned 403..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Spatial-DC maps a candidate fibroblast–cancer signaling interaction in pancreatic tumors

Spatial-DC reconstructed cell-type-resolved spatial proteomes and suggested COL8A1–ITGA2 communication between cancer-associated fibroblasts and pancreatic cancer cells. Multiplexed immunohistochemistry supported the spatial predictions; human protein co-expression was associated with survival. The pairing was already known: the result concerns its inferred spatial deployment, not a new receptor pair or established causal mechanism. This record uses the May 27, 2026 journal version; the result already appeared in the December 20, 2024 preprint.

Technology: Spatial-DC; graph-enhanced transfer learning

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Retrospective spatial proteomic reconstruction and human survival association. Same-study multiplexed immunohistochemistry supports spatial distributions and co-expression, not causal COL8A1–ITGA2 communication.
  • Still unknown: No causal signaling perturbation, therapeutic benefit or independent replication is established in the reviewed material. Current review read indexed publisher introduction and NLM abstract/metadata; full journal methods were inaccessible. Earlier primary preprint XML was read in the project audit on November 16, 2026; revised journal quantitative details were not assessed.
  • AI role: Graph-enhanced transfer learning deconvolved mixed spots and reconstructed cell-type-resolved protein profiles used to infer spatial interactions.

Evidence sources: https://www.sciencedirect.com/science/article/abs/pii/S2405471226000943 · https://pubmed.ncbi.nlm.nih.gov/42202779/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Co-Scientist and aging: metabolism and the integrated stress response

Calico Life Sciences used Co-Scientist to formulate the hypothesis that metabolism regulates the cellular integrated stress response, refined experiments with the tool, and reports new findings. The post gives no numerical data, and the results have not yet been published or peer reviewed.

Technology: Co-Scientist

Family AI Co-Scientist: profile · models page

Publication: Announcement only · Test context: Test context unspecified in annotation. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: not reported
  • Tested: Calico reports that experiments were run with the Co-Scientist hypothesis and produced new findings; the post does not describe the experiments, assays or results.
  • Still unknown: Not reported (the results are not yet published; the team plans to publish them).
  • AI role: Co-Scientist generated the hypothesis that metabolism regulates the integrated stress response, and was used to refine experimental design.

Evidence sources: https://deepmind.google/blog/opening-new-paths-in-aging-research/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 2/5 · Validation 1/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

Co-Scientist: genetic factors to rejuvenate cells

Omar Abudayyeh and Jonathan Gootenberg used Co-Scientist to search tens of thousands of papers, obtaining over 20 proposed genetic factors for reversing cellular aging, a couple of which were validated in the lab, and reducing analysis of genetic-screening data from six months to a few days. The post gives no sample size or peer review status.

Technology: Co-Scientist

Family AI Co-Scientist: profile · models page

Publication: Announcement only · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: direct test
  • Tested: Lab tests validated a couple of the more than 20 proposed genetic factors, which pushed cells toward a younger state; the experiments and exact number are not described.
  • Still unknown: Not reported (the post gives no sample size or peer-review status and contains only the investigators' quotes).
  • AI role: Co-Scientist scanned tens of thousands of papers, proposed genetic factors for reversing cellular aging and analysed screening data alongside the literature.

Evidence sources: https://deepmind.google/blog/fast-tracking-genetic-leads-to-reverse-cellular-aging/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-learning metabolomics identifies widespread promiscuous activities in Escherichia coli enzymes

In vitro metabolomics of 667 purified E. coli enzymes, analyzed with a machine-learning-derived multivariate classifier, indicated previously unrecognized promiscuous activities in 11% of enzymes tested. Follow-up reconstruction assigned 22 reaction stoichiometries to 11 enzymes, with four reactions experimentally validated; examples include expanded nucleotide-substrate use by DeoA, cofactor and sugar promiscuity in NanK, and FMN dephosphorylation by CobC. Follow-up assays also found that pyrimidine-specific ribonucleoside hydrolase RihB acted on thymidine, guanosine and deoxyuridine, further extending its measured substrate range.

Technology: Machine-learning-derived multivariate classifier applied to enzyme metabolomics ion-trace data

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: In vitro metabolomics of 667 purified E. coli enzymes, analyzed with a machine-learning-derived multivariate classifier, indicated previously unrecognized promiscuous activities in 11% of enzymes tested.
  • Still unknown: The classifier operated at a 33% false-discovery rate; the 135 ion traces were putatively annotated, and only four of 22 reconstructed reaction stoichiometries were experimentally validated. Therefore the full 11% result is a set of candidate activities, not 11% independently confirmed enzyme functions. The RihB activities were prompted by a low-confidence potassium-thymidine-adduct trace and then tested directly with purified enzyme; they should be distinguished from the classifier’s higher-confidence hit set.
  • AI role: A machine-learning-derived multivariate classifier, with a reported false-discovery rate of 33%, detected unexpected changes in metabolite ion traces in the enzyme assays and helped nominate candidate promiscuous activities. The researchers then reconstructed candidate reaction stoichiometries and experimentally validated four reactions.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC13389410/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Fragment-guided AI protein redesign yields a Kemp eliminase and fluorophore binder

Crystallographic fragment screening of a designed binding protein guided redesign into a fluorophore binder and a Kemp eliminase. LigandMPNN and structure-prediction models supported the design workflow; crystal structures and enzyme assays checked the resulting functions. The paper reports catalytic efficiency of 3.2 million M⁻¹ s⁻¹ for the eliminase. This entry tracks the 4 May 2026 journal version; a preprint preceded it.

Technology: LigandMPNN with Rosetta relaxation for sequence design; RaptorX and later AlphaFold 3 for structure checking

Family AlphaFold 3: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Fragment screening, crystallography and enzyme assays of redesigned proteins; reported Kemp eliminase catalytic efficiency 3,200,000 M⁻¹ s⁻¹.
  • Still unknown: The reported kinetics and structural tests are within the study; independent replication was not assessed. The journal version follows an earlier preprint.
  • AI role: LigandMPNN with Rosetta relaxation designed enzyme sequences; RaptorX and later AlphaFold 3 predictions checked candidate structures.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC13245809/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A%2210.1038%2Fs41557-026-02125-6%22&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AI-guided design enables E. coli ribosomal proteins without isoleucine

Researchers used computational protein design and generative AI to recode ribosomal proteins in E. coli so they function without isoleucine, then assembled an engineered strain that remained viable and genomically stable over more than 450 laboratory generations. The result shows a reduced amino-acid alphabet can support core translation machinery in this bacterium; the organism still uses isoleucine elsewhere.

Technology: Protein design models including ESM2, MSATr, ProteinMPNN and AfDesign

Family ESM / ESMFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Researchers used computational protein design and generative AI to recode ribosomal proteins in E. coli so they function without isoleucine, then assembled an engineered strain that remained viable and genomically stable over more than 450 laboratory generations.
  • Still unknown: This is not a whole 19-amino-acid organism: isoleucine was eliminated from the engineered ribosomal proteins, while the rest of the cell still uses it. Growth was slower than wild type according to the paper.
  • AI role: The researchers used AI-guided computational design to propose amino-acid substitutions that allowed selected ribosomal proteins to function after removing isoleucine; the strain was assembled and experimentally tested by the team.

Evidence sources: https://magazine.columbia.edu/article/could-life-be-made-simpler

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: other. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

CellOracle-guided pancreatic islet differentiation experiments reveal STAT1 and ZEB1 lineage roles

A digital model built from 400,603 cells across 61 single-cell datasets used CellOracle gene-regulatory-network inference and in-silico transcription-factor perturbations to nominate regulators of stem-cell-derived human islet differentiation. CRISPRi experiments supported a previously unreported role for STAT1 in exocrine specification: transient STAT1 knockdown reduced exocrine markers, and JAK–STAT inhibition preserved beta-cell markers and improved glucose-stimulated insulin secretion. CRISPRi also supported a stage-dependent role for ZEB1, whose early knockdown impaired endocrine specification and later knockdown reduced off-target serotonergic cells while increasing beta-cell proportion. The date is the first preprint posting; this entry summarizes the later version reviewed in the research record, which reports the specific STAT1 and ZEB1 experiments.

Technology: CellOracle v0.12.0 within a digital model integrating single-cell transcriptomic and chromatin-accessibility data

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: CRISPRi experiments supported a previously unreported role for STAT1 in exocrine specification: transient STAT1 knockdown reduced exocrine markers, and JAK–STAT inhibition preserved beta-cell markers and improved glucose-stimulated insulin secretion.
  • Still unknown: This is a bioRxiv preprint (first posted 30 April 2026; the source lists a later version on 17 September 2026), not peer-reviewed at this source date. It integrates existing datasets and reports laboratory validation in stem-cell-derived islet cultures; it does not establish human therapeutic benefit or mature clinical transplantation outcomes. Regulatory-model predictions remain context- and stage-dependent.
  • AI role: The integrated cell-state model inferred regulatory networks and performed in-silico transcription-factor knockout simulations to identify candidate lineage regulators and intervention windows. These model predictions guided subsequent CRISPRi and pharmacologic perturbations in stem-cell-derived islet differentiation cultures.

Evidence sources: https://www.biorxiv.org/content/10.64898/2026.04.27.721124v2.full

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

PEPIC identifies a BHET-hydrolyzing enzyme candidate

PEPIC is a sequence-based convolutional model that predicts PET-hydrolyzing potential. The study reports prioritizing W2061_PET9 and experimentally showing that purified protein converts soluble BHET into MHET, confirmed by HPLC and mass spectrometry. The experiment used a soluble PET intermediate; it does not establish degradation of polymer PET. Homology screening and human selection also contributed.

Technology: PEPIC sequence convolutional neural network

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Purified W2061_PET9 converted BHET to MHET in an in-vitro assay confirmed chromatographically and by mass spectrometry.
  • Still unknown: Polymer PET degradation, exact sequence-level prior characterization and independent replication are not established.
  • AI role: PEPIC scores sequences for predicted PET-hydrolytic activity to prioritize a candidate after genomic/homology selection; direct assay validates soluble BHET turnover.

Evidence sources: https://journals.asm.org/doi/pdf/10.1128/aem.01586-25

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary publisher PDF indexed full text confirms title/DOI, publication date, AI role and Fig. 6 experimental validation..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AggreProt-guided mutations improve LinB solubility and yield

AggreProt is a deep neural-network predictor of residue-level protein aggregation propensity. Researchers used its predicted aggregation-prone regions to guide mutations in haloalkane dehalogenase LinB and report reduced aggregation, improved soluble yield and up to a 100% increase in yield over wild type under their assay. This is one engineered-enzyme result; model validation on selected peptides/proteins is ancillary, and the result does not establish industrial biodegradation performance.

Technology: AggreProt deep neural-network predictor

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The paper reports the specific in-vitro protein assay summarized here.
  • Still unknown: Clinical or diagnostic deployment and independent replication are not established.
  • AI role: The model identifies residue-level aggregation regions; predictions inform mutations that are experimentally assessed for aggregation, solubility and yield.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC13328301/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary paper indexed full text and publication metadata reviewed..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

DefensePredictor uncovers dozens of previously unknown bacterial anti-phage systems

DefensePredictor, a machine-learning classifier using protein-language-model embeddings, nominated candidate defense genes across diverse E. coli strains. Of 94 predicted gene clusters selected for testing, experiments validated 42 previously unknown systems; analyses of 1,000 prokaryotic genomes also revealed thousands of uncharacterized protein clusters without homology to known systems.

Technology: DefensePredictor, gradient-boosted classification over protein-language-model embeddings and genomic context

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Of 94 predicted gene clusters selected for testing, experiments validated 42 previously unknown systems; analyses of 1,000 prokaryotic genomes also revealed thousands of uncharacterized protein clusters without homology to known systems.
  • Still unknown: The 42 validated systems came from a selected subset of predictions, not all computationally proposed systems; genome-wide counts are predicted protein clusters, not experimentally tested systems. The source reports a 45% validation rate for 94 tested units.
  • AI role: The machine-learning model identified likely defense proteins beyond conventional defense-island co-localization methods, guiding experimental cloning and phage-challenge validation. Validation confirmed 42 new systems among 94 tested candidates.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC13092281/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AI-designed protein family binds and senses small molecules, including cortisol

The study combines trRosetta-based scaffold hallucination and Rosetta design with ProteinMPNN/LigandMPNN sequence design and AlphaFold structural filtering to create NTF2-like binders for six small-molecule targets. An optimized cortisol binder had a measured dissociation constant of 68 nM. A purified-protein, split-luciferase cortisol sensor showed an estimated response EC50 of 72 nM; this is a sensor response measurement, not the binding-affinity value or a clinical detection limit. The sensor also responded to cortisone and 17-alpha-hydroxyprogesterone, so specificity needs improvement. The March 2026 peer-reviewed paper is the source of record.

Technology: trRosetta-based hallucination; Rosetta; ProteinMPNN/LigandMPNN sequence design; AlphaFold structural filtering

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Designed binders for six small-molecule targets were measured for binding affinity (nanomolar to low micromolar), and a cortisol binder was used to build a biosensor.
  • Still unknown: The abstract states no specific limitations.
  • AI role: Deep learning combined with physics-based methods generated a protein family with varied designable pocket geometries.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41467-026-70953-8&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ORI engineers active lysozymes, heat-tolerant chitinases and dual-activity enzymes

Tencent researchers used ontology-conditioned generation and experimental feedback to engineer enzymes. In the reported assays, lysozyme TX-RL15 exceeded the earlier TX-L6 design by 66-fold and hen egg-white lysozyme by over two orders of magnitude. Chitinase TX-SC2 retained activity after incubation at 85 °C; selected proteins combined lysozyme and chitinase activity. These are bounded biochemical results. The date represents initial journal publication; the version of record followed on May 7, 2026.

Technology: ORI; protein generation model (PGM); unified sequence model (USM)

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Expression, enzyme-activity and heat-treatment assays on selected generated proteins.
  • Still unknown: Separate supplements and independent replication not assessed. Assay-specific advantages do not establish industrial or clinical outcomes.
  • AI role: Ontology-conditioned generation and reinforcement learning from experimental feedback.

Evidence sources: https://www.nature.com/articles/s41467-026-69855-6?error=cookies_not_supported

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Horizyn-1 experimentally identifies enzymes for orphan and nonnatural reactions

Researchers used the reaction–protein dual-encoder model Horizyn-1 to prioritize enzyme candidates, then experimentally validated activity in orphan-reaction deorphaning, enzyme-promiscuity tests and nonnatural transformations. For an uncharacterized Thermus aquaticus protein, LC–MS confirmed dCMP kinase activity; a selected transaminase produced noncanonical amino acids from lysine-derived donors, with isotope tracing supporting predominant alpha-amine transfer.

Technology: Horizyn-1

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Researchers used the reaction–protein dual-encoder model Horizyn-1 to prioritize enzyme candidates, then experimentally validated activity in orphan-reaction deorphaning, enzyme-promiscuity tests and nonnatural transformations.
  • Still unknown: Evidence is from specific in-vitro assays and tested sequences/reactions; model retrieval performance and the tested enzyme activities do not establish broad generalization. The paper reports that the predicted CMP kinase activity for one protein was not confirmed, while dCMP activity was.
  • AI role: A contrastive model jointly embedded reaction fingerprints and ProtT5 protein-sequence representations, ranked candidate enzyme–reaction pairs, and informed experimental selection. The authors assayed predictions with biochemical assays, LC–MS and isotope tracing; some fine-tuning results are predictive-performance findings rather than biological discoveries.

Evidence sources: https://doi.org/10.1073/pnas.2520070123

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

M3-CAD designs antimicrobial QLX-3DV peptides

M3-CAD combines sequence, predicted structural and physicochemical features in a multimodal peptide-design pipeline. The journal study reports in-vitro activity for selected QLX-3DV peptides, including untemplated candidates, against tested bacterial strains; selected leads also received additional host-toxicity and mouse wound-infection assays. These are selected preclinical results, not evidence of clinical efficacy or safety. The article cites an earlier M3-CAD/QL-AMP-1 preprint; this entry is restricted to the reported QLX-3DV study result and does not claim the pipeline itself as new.

Technology: M3-CAD multimodal deep-learning pipeline

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Selected designed peptides were synthesized and tested in vitro; selected leads were also examined in mouse wound infection and host-toxicity assays.
  • Still unknown: Clinical efficacy and safety, independent replication and full sequence-level comparison with the earlier M3-CAD preprint are not established here.
  • AI role: M3-CAD uses multimodal learned sequence and predicted-structure representations with activity-related objectives to generate and rank peptides for experimental testing.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC13185819/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary PMC full text including publication metadata, methods and experimental results..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning prioritizes microbiome metabolites that stimulate interleukin-8 secretion

Researchers trained ML models on public drug-development data to prioritize biological properties of microbiome metabolites, then tested predictions prospectively. In vitro experiments identified previously unreported IL-8 secretion stimulation by spermine and spermidine, metabolites previously regarded as anti-inflammatory. The paper reports average accuracy of 92.86% in prospective experiments, while cautioning that sample sizes were small.

Technology: Machine-learning property prediction models implemented in KNIME and trained on public drug-development data

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Researchers trained ML models on public drug-development data to prioritize biological properties of microbiome metabolites, then tested predictions prospectively. The selected metabolites were tested in new in-house assays; this is direct experimental testing within the discovery study.
  • Still unknown: The unexpected activity is demonstrated in vitro and does not establish in-vivo host effects. The article explicitly cautions that prospective validation sample sizes were small; reported model metrics vary by task.
  • AI role: Predicted and triaged microbiome metabolites for target assays, including IL-8 activation; prospective laboratory assays then established the unexpected cytokine-stimulatory activity of spermine and spermidine.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC13059116/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-learning analysis separates genetic from environmental drivers of stimulated cytokine responses in healthy adults

In 406 healthy GEOCODE cohort participants, whole-blood cytokine responses to Toll-like-receptor (TLR) and T-cell-receptor (TCR) stimulation varied across people, including high- and low-IL-6-response profiles associated with monocyte counts, sex and sampling season. Feature-importance analysis of machine-learning models found TCR-induced cytokine predictions were mainly driven by genetic features, whereas TLR-induced predictions were driven more by environmental and biological features. Combining genetic with biological and environmental data improved prediction by about 0.2 on average, and the reported findings were evaluated in the independent Milieu Intérieur cohort.

Technology: Machine-learning models including random forest, gradient-boosted trees, artificial neural networks and elastic net; random forest feature-importance analysis

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: In 406 healthy GEOCODE cohort participants, whole-blood cytokine responses to Toll-like-receptor (TLR) and T-cell-receptor (TCR) stimulation varied across people, including high- and low-IL-6-response profiles associated with monocyte counts, sex and sampling season.
  • Still unknown: The discovery cohort was a selected healthy-adult sample (406 participants), and the correlations for TLR-induced response prediction were modest (average 0.27; TCR-induced 0.53). Genetic-only models may overestimate predictive accuracy, as the authors report. Feature importance identifies predictive contribution, not causal biological effects; results do not establish disease-specific immune mechanisms or treatment response.
  • AI role: Models predicted measured cytokine levels following TLR or TCR stimulation from genetic, biological and environmental participant features. Random-forest permutation importance identified which input categories contributed to predictions; the study also compared predictions across cross-validation and a separate Milieu Intérieur cohort. Feature importance is associational and does not establish that a feature causally regulates cytokine production.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/41735307/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ETopt-guided screening identifies thermophilic beta-agarases

ETopt combines sequence, structural and protein-language-model features in an ensemble XGBoost model to predict enzyme optimum temperatures. The authors report wet-lab screening that identified two novel thermophilic beta-agarases. The draft is limited to those source-identified enzyme findings; model benchmarks and other screened enzyme classes are not counted as separate discoveries. Industrial process performance and independent replication are not established.

Technology: ETopt multimodal model; ensemble XGBoost; ProtBERT and Prime embeddings

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The source highlights experimental identification of two novel thermophilic beta-agarases; detailed enzyme activity and comparator data were not available in the reviewed abstract.
  • Still unknown: Full article, exact assay results, sequences, prior enzyme overlap and independent replication were not reviewed.
  • AI role: ETopt integrates learned sequence and structural representations with selected features to predict optimum temperature and prioritize candidate enzymes for wet-lab validation.

Evidence sources: https://www.sciencedirect.com/science/article/pii/S0141813026009943 · https://pubmed.ncbi.nlm.nih.gov/41747980/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Publisher highlights and abstract; PubMed primary metadata and abstract..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning screening identifies GW3965 activity against invasive group A Streptococcus

A neural network trained on measured growth responses from 2,560 compounds screened 6,111 FDA-evaluated compounds for activity against multidrug-resistant emm92 group A Streptococcus. Of nine selected compounds tested, GW3965 reduced bacterial survival by 99% at an MIC of 6.25 μM. The authors also report improved wound closure in an infected human skin-equivalent model and reduced lesion area and bacterial burden in mice. This is preclinical evidence and does not establish efficacy or safety in people. The peer-reviewed journal article followed on 2026-05-30 and reports the same core GW3965 result.

Technology: Deep neural network trained on compound SMILES and measured bacterial growth responses; model screened an FDA-evaluated compound library

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: The model was trained on 2,560 compounds and screened 6,111. Nine selected compounds were validated; GW3965 had MIC 6.25 μM and reduced viable emm92 iGAS by 99%. In a human skin-equivalent model infected tissues treated with GW3965 showed wound closure; in a mouse SSTI model 30 mg/kg reduced lesion size and day-6 wound bacterial burden by about 2 logs on average.
  • Still unknown: Results come from in-vitro assays and animal/tissue models; the preprint reports no synergy with ampicillin. Human clinical efficacy, dosing, safety and mechanism against GAS remain unestablished.
  • AI role: A neural network learned structure–growth-inhibition patterns from measured compound screening data, prioritized GW3965 for follow-up, and the selected candidate was experimentally tested in bacterial, human skin-equivalent, and mouse infection models.

Evidence sources: https://assets-eu.researchsquare.com/files/rs-8779898/v1/a183f808-a19a-41be-8b56-ad4009293888.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary Research Square v1 full-text PDF, including posted date, affiliations, Methods and experimental outcomes; primary Nature publisher indexed article text and journal-version abstract/results..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Pichia-CLM improves recombinant protein expression in Komagataella phaffii

A GRU encoder-decoder language model learns host-specific amino-acid-to-codon patterns and designs synonymous sequences. Across six proteins, the study reports the best titer for five against four commercial optimization tools; human serum albumin reached about threefold the native-sequence titer. Results are plate-scale expression measurements, not manufacturing economics or clinical product outcomes.

Technology: Pichia-CLM (GRU encoder-decoder)

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The source reports the study-specific result summarized here.
  • Still unknown: Independent replication and broader deployment or clinical benefit are not established by this source review.
  • AI role: The learned host codon language model generates optimized coding sequences; researchers experimentally compare protein titers with native sequences and four commercial tools.

Evidence sources: https://www.pnas.org/doi/10.1073/pnas.2522052123

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary-source abstract and publication metadata reviewed; full text scope is specified in source_checks..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

De novo designed proteins modulate GPCR signalling in the dopamine D1 receptor

Researchers designed four protein exoframe modulators targeting the transmembrane region of the dopamine D1 receptor. Structural studies and functional assays showed agonist-positive, negative and biased allosteric modulation; an agonist-positive modulator restored activity in loss-of-function receptor mutants. The journal abstract supports these receptor-level experiments, rather than a clinical treatment claim.

Technology: Computational de novo protein design (AI-based protein design pipeline)

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Structural and functional investigation of four designed modulators of the D1 receptor, including rescue of loss-of-function mutants.
  • Still unknown: Review limited to abstract, online date and affiliations. Therapeutic efficacy in patients is not established by these receptor experiments.
  • AI role: A deep-learning, hallucination-like protein-design workflow used structural prompts to generate transmembrane binding modes.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A%2210.1038%2Fs41586-025-09957-1%22&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

EnzySFC discovers nitrile-hydratase activity in nitrilases and evolves amide-selective variants

The AI-assisted EnzySFC workflow experimentally validated 10 previously uncharacterized nitrilases selected from 1,113 candidates. A wild-type Phototrophicales nitrilase showed exclusive nitrile hydratase activity; a separate Actinomycetia nitrilase was AI-evolved, with four tested top variants reaching 100% amide yield and complete substrate conversion.

Technology: EnzySFC AI-assisted enzyme discovery and functional-evolution platform

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The AI-assisted EnzySFC workflow experimentally validated 10 previously uncharacterized nitrilases selected from 1,113 candidates.
  • Still unknown: The reported outcomes are enzyme and substrate assays; the abstract does not establish industrial scale performance or application in a living system. The 10 discovery candidates and 16 mutants were experimental follow-ups to model prioritization, not comprehensive validation of all 1,113 sequences.
  • AI role: EnzySFC combined de novo enzyme discovery and multiobjective activity/specificity optimization to rank candidate nitrilases and propose evolved mutants. The study experimentally tested ten discovery candidates and sixteen mutants from one evolution prediction cycle; the named activity switch and product-selective variants were measured in assays.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/41645784/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

GPT-5 reduces the cost of cell-free protein synthesis

A cycle between GPT-5 and a robotic lab explored over 36,000 compositions and reduced protein-production cost by 40% compared with the previous baseline. This is an experimental result in one CFPS setup; it does not establish the same reduction for other proteins, scales, or facilities.

Technology: GPT-5

Family GPT-5: profile · models page

Publication: Announcement only · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: direct test
  • Tested: Over six rounds in a cloud laboratory, the system tested more than 36,000 reaction compositions on 580 plates and reached a 40% lower protein production cost (57% lower reagent cost).
  • Still unknown: Not stated in the reviewed material
  • AI role: GPT-5 proposed and analyzed experiments in a closed loop with Ginkgo Bioworks automation.

Evidence sources: https://openai.com/index/gpt-5-lowers-protein-synthesis-cost/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Iterative discovery of potent polymeric antibiotics via multi-stage and multi-task learning against antimicrobial resistance

PolyCLOVER integrated multi-stage learning, active learning and high-throughput experiments across a ~100,000-member poly(β-amino ester) space, identifying three lead nanoparticles with measured antibacterial activity and in-vivo mouse outcomes. Mouse groups were small; this does not establish human efficacy or safety.

Technology: Multi-stage self-supervised learning and active learning coupled to high-throughput experimentation

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Primary Nature abstract/full text reports experimental MICs of 4 and 8 μg/mL against named resistant bacteria and mouse infection models; this is one study, not three discoveries.
  • Still unknown: Independent replication and exhaustive prior-art review are not established.
  • AI role: Multi-stage self-supervised learning and active learning coupled to high-throughput experimentation

Evidence sources: https://www.nature.com/articles/s41467-026-68682-z

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary source checked for study identity, stated AI contribution, date and bounded result..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Anthropic

Claude in biology labs: Biomni, MozzareLLM, and others

Anthropic describes three cases: Biomni (Stanford) completed a GWAS analysis in 20 minutes in an initial trial and suggested new transcription factors in embryonic development; MozzareLLM (MIT and Whitehead) interprets gene clusters from CRISPR screens; a third lab is still testing. This is an Anthropic account without independent evaluation; some figures are estimates.

Technology: Claude (Biomni, MozzareLLM)

Family Claude: profile · models page

Publication: Announcement only · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: retrospective or benchmark
  • Tested: Biomni's case studies (GWAS analysis, cloning design blind-evaluated against a postdoc, embryonic-tissue cell analysis) were checked by the Biomni team, citing a bioRxiv preprint; MozzareLLM was assessed qualitatively by its lab, with no quantitative benchmark.
  • Still unknown: The article says the validation is not independent, some figures are estimates (such as the three-week human comparison), and Biomni cannot yet do everything out of the box.
  • AI role: A Claude-powered agent (Biomni) analysed GWAS and embryonic cell data and suggested transcription factors; a Claude-powered system (MozzareLLM) interprets gene clusters from CRISPR screens.

Evidence sources: https://www.anthropic.com/news/accelerating-scientific-research

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AI and developmental profiling reveal conserved rules for microRNA strand selection

Using a high-throughput qPCR assay, Arizona State University researchers measured microRNA 5p/3p strand use across six Caenorhabditis elegans developmental stages and selected tissues. A machine-learning model trained on experimentally measured strand usage identified compositional and structural sequence features associated with strand preference; the learned relationships predicted strand usage in other vertebrates, including humans, and the authors report that the rules are conserved but context-dependent.

Technology: Predictive machine-learning model using 77 biologically informed features, trained on high-throughput qPCR measurements from the HiTmiSS platform

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Using a high-throughput qPCR assay, Arizona State University researchers measured microRNA 5p/3p strand use across six Caenorhabditis elegans developmental stages and selected tissues.
  • Still unknown: The dataset measured strand usage in C. elegans across six developmental stages and specific tissues; cross-species including human results are model predictions/validation against available strand-usage data, not direct tests of gene-regulatory outcomes in human cells. The result concerns strand-selection patterns and predictors, not disease causality or therapeutic efficacy.
  • AI role: The model learned strand-preference relationships from the experimental strand-usage dataset, quantified feature contributions, and predicted preference across additional nematode microRNAs and vertebrate species. Experimental qPCR measurements established the C. elegans profiles; predictions across species extend the analysis but are not themselves experimental confirmation of function in each organism.

Evidence sources: https://academic.oup.com/nar/article/54/2/gkaf1510/8425383

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Self-iterative learning finds RNA-targeting prokaryotic Argonautes and an RNA-modification-sensing activity

The self-iterative hierarchical ensemble model (SIM) nominated six high-confidence prokaryotic Argonaute candidates; five were experimentally validated as RNA-targeting proteins. The study identified three uncharacterized Agos with a novel N-terminal domain, defining a previously unrecognized subclass. Biochemical and in-vivo experiments on Haloferax profundi Ago (HpAgo) established RNA cleavage and a distinctive RNA-modification-sensing capability, which the authors used to build a method for quantifying modified RNA.

Technology: Self-Iterative hierarchical ensemble Model (SIM), an ensemble-learning and self-training pipeline

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The self-iterative hierarchical ensemble model (SIM) nominated six high-confidence prokaryotic Argonaute candidates; five were experimentally validated as RNA-targeting proteins.
  • Still unknown: Five of six selected candidates were validated; the broader predicted sequence space was not experimentally tested. The specific RNA modification-sensing mechanism and assay performance are supported for HpAgo and tested conditions, not all Argonautes. No therapeutic or universal detection claim is made. Primary abstract/citation record checked; no independent replication assessed.
  • AI role: SIM iteratively expanded its training labels to rank prokaryotic Argonautes for RNA targeting from sparse initial data; it prioritized candidates and identified a previously unrecognized domain-containing subclass. The researchers validated RNA cleavage and RNA-modification sensing biochemically and in vivo.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/42039271/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

CleaveNet designs peptide substrates with MMP13 cleavage profiles

CleaveNet combines a generative model conditioned on protease cleavage profiles with a predictor that ranks candidate sequences. In an in-vitro screen of 95 peptide substrates against 12 matrix metalloproteinases, the authors report that designed substrates were cleaved by MMP13 and identify substrates with different efficiency/selectivity trade-offs. The panel includes training controls and baseline candidates; it is not 95 novel AI discoveries. These are laboratory substrates, not validated diagnostic or therapeutic products.

Technology: CleaveNet generator and predictor; deep learning

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The paper reports the specific in-vitro protein assay summarized here.
  • Still unknown: Clinical or diagnostic deployment and independent replication are not established.
  • AI role: A generator proposes peptide sequences conditioned on cleavage profiles; a model ranks candidates and researchers test synthetic substrates against recombinant enzymes.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC12775482/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Full primary journal article text, results and experimental assays reviewed. The preprint version-specific claim correspondence was not checked..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

SMRTnet predicts small molecules that bind RNA

The January 2026 Nature Biotechnology abstract reports 40 small-molecule hits across ten disease-associated RNA targets, with nanomolar-to-micromolar binding. One predicted compound reduced MYC expression, inhibited proliferation and promoted apoptosis in three cancer cell lines. SMRTnet combines language models with convolutional and graph attention networks, using RNA secondary structure. These are laboratory results.

Technology: SMRTnet (language-model-based)

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: RNA-binding assays and cancer-cell experiments.
  • Still unknown: Not assessed: full-paper validation rates and therapeutic performance.
  • AI role: Multimodal deep learning predicted small molecule-RNA interactions.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/41482542/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Knowledge-graph AI prioritizes TPP1 as a Parkinson’s disease target linked to α-synuclein homeostasis

Standigm ASK’s knowledge-graph model ranked Parkinson’s disease target associations; subgraph enrichment and expression analyses narrowed 2,527 predicted targets to five candidates, including TPP1. In a human cell-based α-synuclein seeding model, TPP1 knockdown increased aggregate accumulation and insoluble α-synuclein, experimentally supporting a TPP1–α-synuclein relationship and a possible protective role.

Technology: Standigm ASK knowledge-graph path-based deep-learning model with subgraph-level over-representation re-ranking

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: In a human cell-based α-synuclein seeding model, TPP1 knockdown increased aggregate accumulation and insoluble α-synuclein, experimentally supporting a TPP1–α-synuclein relationship and a possible protective role.
  • Still unknown: The functional experiment is a human neuroblastoma cell model exposed to α-synuclein preformed fibrils, not a clinical or in-vivo therapeutic demonstration. Knockdown increasing aggregation supports a protective association; direct enzymatic cleavage of α-synuclein is proposed from structural modeling, not established by the reported work. TPP1 is a candidate therapeutic target, not a validated treatment.
  • AI role: The model integrated heterogeneous biomedical databases, encoded relational paths using a convolutional network and attention aggregation, and produced disease-target scores. Researchers applied enrichment and transcriptomic filters to prioritize TPP1, then validated TPP1’s relationship to α-synuclein aggregation via siRNA knockdown and cellular assays; AlphaFold-Multimer modeling suggested, but did not prove, direct proteolysis.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC12814083/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-learning metabolomics prioritizes succinate in sepsis-associated encephalopathy

In plasma samples from 29 controls, 32 sepsis patients and 27 patients with sepsis-associated encephalopathy (SAE), LC–MS/MS plus machine-learning feature selection identified a 12-metabolite profile. Succinate rose stepwise from controls to sepsis to SAE, ranked among the five leading XGBoost features, and correlated with APACHE II, SOFA and lactate. In a cecal-ligation-and-puncture mouse model, dietary succinate supplementation before sepsis worsened cognitive-test performance, neuronal injury and microglial activation; succinate also increased inflammatory gene expression in cultured microglia.

Technology: Random forest and support-vector-machine feature ranking with XGBoost classification

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: In a cecal-ligation-and-puncture mouse model, dietary succinate supplementation before sepsis worsened cognitive-test performance, neuronal injury and microglial activation; succinate also increased inflammatory gene expression in cultured microglia.
  • Still unknown: The human discovery cohort comprised 88 participants and the held-out model test set had 18 participants, so diagnostic performance and metabolite ranking need larger external cohorts. Succinate supplementation was given before CLP in mice, and cell experiments used a microglial cell line; these are preclinical perturbations and do not prove that succinate causes human SAE or that targeting it improves patient outcomes. Associations with severity scores are observational.
  • AI role: Machine-learning feature selection and classifier comparison were applied to the measured plasma metabolomics to identify discriminatory features; succinate ranked among the top five XGBoost features. Researchers then tested its functional relevance through supplementation in a mouse sepsis model and in cultured microglia. The animal/cell experiments were human-directed follow-up, not model-generated predictions.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC12809278/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-learning prioritization and CRISPR knockout identify OsWOX13 as a rice heat-tolerance regulator

Meta-transcriptomic comparison of heat-tolerant and susceptible rice genotypes yielded 409 meta-differentially expressed genes. Random-forest and XGBoost feature ranking with SHAP independently prioritized OsWOX13; CRISPR/Cas9 knockout lines showed lower survival under heat stress (~20% versus ~60% in wild type) and delayed flowering, with accompanying changes in ABA, oxidative-stress markers and stress/flowering-associated proteins.

Technology: Random forest and XGBoost models with SHAP feature attribution

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Meta-transcriptomic comparison of heat-tolerant and susceptible rice genotypes yielded 409 meta-differentially expressed genes.
  • Still unknown: The PubMed primary abstract supports gene prioritization and knockout evidence, but does not establish detailed direct molecular targets or field-level yield effects. Reported survival figures concern the study's heat-stress condition and experimental lines; this is not evidence of breeding performance or resilience in field environments. Mechanistic pathway interpretations remain limited to associated physiological/proteomic measurements.
  • AI role: The models prioritized stress-associated genes from meta-transcriptomic differences; OsWOX13 was among the top-ranked genes in both models. The team tested its predicted positive role using CRISPR/Cas9 knockout, followed by heat-stress survival, flowering, physiology and proteomic assays.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/41424172/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 88 of works of the same type, year and subfield (OpenAlex work metrics, 6 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

CycleRFdiffusion-designed TP4 is a reproducible cyclic peptide binder to TREM2

A structure-conditioned workflow combining CycleRFdiffusion, ProteinMPNN and HighFold generated and screened 1,500 cyclic peptide–TREM2 complexes and prioritized four candidates. The lead TP4 showed weak but reproducible TREM2 binding in spectral-shift, microscale-thermophoresis and surface-plasmon-resonance assays. The authors report favorable plasma stability and moderate metabolic stability, presenting the work as an initial proof of concept for TREM2-targeted cyclic binders.

Technology: CycleRFdiffusion backbone design, ProteinMPNN sequence design and HighFold structural prediction/screening

Family RFdiffusion: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The lead TP4 showed weak but reproducible TREM2 binding in spectral-shift, microscale-thermophoresis and surface-plasmon-resonance assays.
  • Still unknown: The only reported lead, TP4, is described as a weak binder. The evidence is biophysical binding and stability profiling, not demonstrated modulation of TREM2 signaling, microglial function, disease pathology or therapeutic efficacy. Treat this as proof-of-concept binder discovery.
  • AI role: The AI pipeline generated and ranked 1,500 candidate cyclic peptide–TREM2 complexes based on structure, sequence and predicted energetics, from which four were selected for experiments; TP4 was then assessed with multiple orthogonal binding assays.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/41435973/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

OpenAI

GPT-5 and molecular cloning: 79-fold greater efficiency

In a cycle with Red Queen Bio, GPT-5 optimized a cloning protocol, introducing RecA and gp32 as a new mechanism: 2.6-fold improvement with the enzymatic reaction, 36-fold with transformation, and 79-fold when combined. This applies to the model system used, with scientists carrying out the protocols; reported by OpenAI.

Technology: GPT-5

Family GPT-5: profile · models page

Publication: Announcement only · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: direct test
  • Tested: Scientists carried out GPT-5-modified cloning protocols over several rounds; the final protocol recovered 79 times more sequence-verified clones for a fixed amount of input DNA than the baseline.
  • Still unknown: The work used a benign experimental system with a limited task scope.
  • AI role: GPT-5 reasoned about the protocol and proposed modifications, including a RecA and gp32 mechanism; humans only ran the protocol and uploaded data.

Evidence sources: https://openai.com/index/accelerating-biological-research-in-the-wet-lab/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-learning screening identified experimentally active pore-forming peptides

Machine-learning models and membrane-related filters were used to prioritize natural peptide sequences for testing. The study reports seven experimentally active peptides; three showed broad-spectrum activity and pore-forming evidence in the reported membrane assays. Peptide 615 had already been disclosed in a Peking University patent published in March 2025, before the related preprint and journal article. This entry represents one study campaign and makes no first-disclosure or sequence-novelty claim. Assay results do not establish clinical safety or efficacy.

Technology: ProtBERT-BFD sequence embeddings with a support-vector classifier and membrane-interaction, secondary-structure and contact-map predictions

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The paper reports seven experimentally active selected peptides and pore-forming measurements for three broad-spectrum candidates.
  • Still unknown: One peptide sequence, 615, was disclosed in an earlier university patent. The article does not establish clinical safety or efficacy, and no universal mechanism is claimed.
  • AI role: Machine-learning sequence and membrane-related models prioritized peptide candidates for experimental testing.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC12904036/ · https://patents.google.com/patent/CN119591677A/zh · https://doi.org/10.1101/2025.05.17.654650

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary PMC full text, metadata and affiliation block; published patent text for earlier public disclosure; primary preprint record identified in the frozen source review..

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ML-guided CRISPRi combinations improve isoprenol production

An ensemble of machine-learning regressors recommended combinations of up to four CRISPRi guide RNAs during six design–build–test–learn cycles in engineered Pseudomonas putida. The paper reports an approximately fivefold isoprenol-titer improvement over its control for highlighted strains. It reports 347 unique arrays for the highlighted optimization and 472 unique strains overall; these are different denominators. A 2024 DOE meeting abstract describes an earlier campaign but does not quantify the later result; the journal article date is used without claiming first public disclosure. The finding is specific to this strain and laboratory workflow, not industrial fuel production.

Technology: Automated Recommendation Tool ensemble of machine-learning regressors

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The article reports six optimization cycles and about fivefold higher isoprenol titer versus the stated control, with distinct counts of 347 unique arrays and 472 unique strains overall.
  • Still unknown: The result is study-specific; industrial scale-up and independent replication are not established. The earlier DOE abstract establishes a related campaign, not the later quantitative result.
  • AI role: The Automated Recommendation Tool, an ensemble of ML regressors, recommended guide-RNA combinations after initial genes and training data were supplied by researchers.

Evidence sources: https://www.nature.com/articles/s41467-025-66304-8_reference.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary publisher article/PDF and institutional preprint metadata reviewed for identity, bounded outcomes, AI role, affiliations and chronology..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

tFold designs binding antibodies against four defined antigens

Tencent researchers generated antibodies against influenza hemagglutinin, PD-1, PD-L1 and SARS-CoV-2 RBD, followed by phage-display screening and SPR characterization. Anti-SC2RBD-H330 showed neutralization IC50 of 0.42 nM. Screening remained extensive, and some confident epitope predictions failed competition tests. The date represents initial journal publication, preceding the January 19, 2026 version of record; a tFold precursor was posted February 8, 2024.

Technology: tFold System; tFold-Ab; tFold-Ag; AlphaFold-Multimer

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Binding and competition assays on generated antibodies; RBD-antibody neutralization.
  • Still unknown: Separate supplements and independent replication not assessed. First-public dates of individual journal experiments were not established by precursor version comparison.
  • AI role: Generates CDR sequences and predicts complexes for structure-guided filtering.

Evidence sources: https://www.nature.com/articles/s41467-025-67361-9?error=cookies_not_supported · https://www.biorxiv.org/content/10.1101/2024.02.05.578892v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

CACLENS predicts and validates a zearalenone-degrading enzyme

The multitask deep-learning enzyme-screening framework CACLENS selected ten candidate enzymes for zearalenone (ZEN) degradation, which were expressed and tested. One candidate degraded more than 90% of both ZEN and its analogue α-zearalenol (α-ZOL) in the reported assay.

Technology: CACLENS (Cross-Attention & Contrastive Learning-enabled Enzyme Selection), a multimodal multitask deep-learning enzyme classifier

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The multitask deep-learning enzyme-screening framework CACLENS selected ten candidate enzymes for zearalenone (ZEN) degradation, which were expressed and tested.
  • Still unknown: This is an in-vitro enzyme degradation finding; it does not establish safety of degradation products, remediation performance at environmental scale, or food/feed safety. The abstract does not name the successful enzyme or its degradation-product identity.
  • AI role: CACLENS combined reaction-type classification, EC-number prediction and reaction-feasibility estimates to screen candidate enzymes. The paper then expressed and assayed ten selected enzyme candidates against ZEN; one showed >90% degradation of ZEN and α-ZOL.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC12904076/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Heat-resistant crops: a hybrid photosynthesis enzyme

At Michigan State University, AlphaFold predicted the 3D structure of glycerate kinase in a crop and a heat-loving alga. Replacing three flexible loops with rigid loops from the alga produced a hybrid that remained stable up to 65 °C. The enzyme has not yet been tested in whole plants.

Technology: AlphaFold

Family AlphaFold: profile · models page

Publication: Announcement only · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: direct test
  • Tested: Hybrid glycerate kinase enzymes with three loops replaced by rigid loops from a heat-loving alga were tested in the lab, and one stayed stable up to 65 °C; whole-plant testing has not been done.
  • Still unknown: It is not yet known whether the hybrid enzyme improves heat resilience in a whole plant; the page does not say whether the AlphaFold-predicted structures were experimentally validated.
  • AI role: AlphaFold predicted the enzyme's 3D structures, which were then used in molecular simulations to identify loops to modify.

Evidence sources: https://deepmind.google/blog/engineering-more-resilient-crops-for-a-warming-climate/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

RFdiffusion2 scaffolds enzyme active sites at atomic level

RFdiffusion2 scaffolded all 41 active sites in a diverse benchmark, compared with 16 for previous methods. Designs were tested experimentally, with fewer than 96 sequences needed per case to find active enzymes across three catalytic mechanisms. The journal abstract's three-mechanism count conflicts with a five-reaction claim in the Institute for Protein Design announcement, so the abstract should be treated as authoritative.

Technology: RFdiffusion2

Family RFdiffusion: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Designed enzymes were expressed and screened experimentally for retroaldolase, cysteine hydrolase and zinc hydrolase activity, with kinetic constants measured for the best variants.
  • Still unknown: Authors state the designs are not as active as native enzymes, the theozymes may miss interactions needed for high activity, and scaffolding catalytic atoms does not guarantee proficient enzymes.
  • AI role: RFdiffusion2 generated scaffolds from atom-level active-site descriptions, LigandMPNN assigned sequences, and structure predictors filtered designs before testing.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC12791007/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Computational design of zinc metallohydrolases with RFdiffusion2

RFdiffusion2 designed zinc metallohydrolases from quantum-chemistry-derived active-site geometries. The first round of 96 designs produced an enzyme with catalytic efficiency of 16,000 M-1 s-1, and a second round of 96 yielded three more highly active enzymes, with k/K up to 53,000 M-1 s-1. Crystal structures (PDB 9PYJ, 9PYL) were determined, but the evidence covers one designed enzyme family.

Technology: RFdiffusion2 with quantum-chemistry active-site models

Family RFdiffusion: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: From 96 designs, the best had kcat/KM of 16,000 M^-1 s^-1; a second round of 96 gave three more active enzymes, up to 53,000 M^-1 s^-1 and kcat up to 1.5 s^-1. The crystal structure of the most active design is very close to the model.
  • Still unknown: Use in practical applications is not reported in the abstract.
  • AI role: RFdiffusion2 designed zinc metallohydrolases from quantum-chemistry active sites; PLACER and Chai-1 analysed the designs.

Evidence sources: https://doi.org/10.1038/s41586-025-09746-w

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: other. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Riff-Diff: computational enzyme design by catalytic motif scaffolding

The December 2025 Nature version reports Riff-Diff designs for retro-aldol and Morita-Baylis-Hillman reactions, with activities comparable to enzymes optimized by experimental evolution. Six high-resolution structures support near-atomic active-site placement. The abstract describes a hybrid of machine learning and atomistic modelling. The work also has a bioRxiv preprint first indexed in August 2024; the January 2026 journal issue is later than online publication.

Technology: Riff-Diff (RFdiffusion with rotamer-based motif placement)

Family RFdiffusion: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Designed-enzyme activity and six experimental high-resolution structures.
  • Still unknown: Not assessed: full-paper success rates, thermal assays and supplementary data.
  • AI role: Riff-Diff combines learned protein generation and atomistic catalytic-motif scaffolding.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A%2210.1038%2Fs41586-025-09747-9%22&resultType=core&format=json · https://www.tugraz.at/en/news/article/enzyme-design-ai-riff-diff

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AI-guided subtilisin variants improve aqueous amide formation

The study reports that wild-type subtilisin BPN′ directly catalyzed condensation of free carboxylic acids and amines in aqueous PBS (pH 8, 40 °C), forming a peptide-bond amide without stoichiometric coupling reagents. The authors combined AI-guided mutation ranking with a FRET assay, reducing 7,999 in-silico variants at three active-site positions to five experimental candidates. The top triple mutant N155G/N218D/M222Q had reported relative activity of 753 ± 14% of wild type and also coupled biotin to a peptide fragment in water, extending beyond peptide–peptide ligation.

Technology: AI-guided enzyme mutation prediction combining binding-free-energy analysis of active-site substitutions with an experimental FRET-based rapid product-detection platform

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The study reports that wild-type subtilisin BPN′ directly catalyzed condensation of free carboxylic acids and amines in aqueous PBS (pH 8, 40 °C), forming a peptide-bond amide without stoichiometric coupling reagents.
  • Still unknown: The reported chemistry is demonstrated for specific peptide fragments and a biotin–peptide coupling example under the stated aqueous conditions; the paper does not establish broad substrate generality or production-scale utility. The 753% figure is reported relative catalytic activity with wild type as the reference, not a yield increase or 753-fold improvement. The reaction requires follow-up for broader substrate scope and process optimization.
  • AI role: The AI-guided pipeline evaluated combinations at residues N155, N218 and M222, narrowing 7,999 in-silico variants to five prioritized mutants for FRET-based experimental validation. This reduced the candidate set and enabled selection of the N155G/N218D/M222Q variant; the AI component was applied after the paper established the wild-type aqueous amidation reaction.

Evidence sources: https://www.sciencedirect.com/science/article/pii/S1385894725106992

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 47 of works of the same type, year and subfield (OpenAlex work metrics, 2 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Healthier bees: AlphaFold predicts vitellogenin's structure

With the Norwegian University of Life Sciences, predicted structures of honeybee vitellogenin were confirmed by electron microscopy in July 2025; AlphaFold 3 helped study its interactions with other molecules. Early evidence suggests breeding times for resilient bees could fall from months to weeks, a still preliminary claim.

Technology: AlphaFold, AlphaFold 3

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: AlphaFold-predicted vitellogenin structures were confirmed by electron microscopy in July 2025, per the post; the linked paper reports a cryo-EM structure of native honey bee vitellogenin, and the breeding-time claim comes from early trials.
  • Still unknown: The breeding-time benefit (months to weeks) is preliminary and not cited to a publication, and implications for other egg-laying species are possibilities, not established findings.
  • AI role: AlphaFold 3 predicted the vitellogenin structures and helped study how it binds other molecules inside the bee.

Evidence sources: https://deepmind.google/blog/breeding-healthier-and-stronger-honeybees/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41467-025-58575-y&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 12 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

RFD2-MI designs sequence-specific cysteine proteases

RFD2-MI designs cysteine-nucleophile proteases that cleave peptide backbones in a sequence-dependent manner. The preprint abstract reports rate enhancements up to 3 × 10^7 and crystal structures close to design models. The journal version reports additional experiments; this draft is bounded to the preprint abstract result and does not transfer later results across versions.

Technology: RFD2-MI; ProteinMPNN; AlphaFold 3

Family ProteinMPNN: profile · models page

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: The source reports the study-specific result summarized here.
  • Still unknown: Independent replication and broader deployment or clinical benefit are not established by this source review.
  • AI role: RFD2-MI generates catalytic protein structures; sequence design and structure prediction support experimental selection.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC12668006/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary-source abstract and publication metadata reviewed; full text scope is specified in source_checks..

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning-guided discovery of thermophilic carbonic anhydrases

AdaBoost and LightGBM classifiers applied to sequence descriptors prioritized three gamma carbonic anhydrases from Fang Hot Spring metagenomic data. The recombinant enzymes showed carbonic anhydrase activity and melting temperatures from 97.0 to 109.1 °C in the reported assays. Some candidates are similar to annotated proteins; the contribution is discovery and testing of thermostable enzymes, not new protein families or demonstrated industrial performance.

Technology: AdaBoost; LightGBM

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The primary source reports the study-specific measured result summarized here.
  • Still unknown: Independent replication and clinical or industrial efficacy are not established.
  • AI role: Classifiers rank candidate metagenomic enzyme sequences for expression; researchers express, purify and assay carbonic anhydrase activity and thermal stability.

Evidence sources: https://www.nature.com/articles/s41598-025-24713-1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary source indexed full text or abstract and publication metadata reviewed; scope stated in source_checks..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ML prioritization identifies two antimicrobial peptide candidates

CalcAMP ranked 16,384 14-residue sequences in a researcher-defined LL-37-derived design space bounded by patent WO2015088344. Researchers selected GDST-038 and GDST-045 for laboratory tests; the paper reports activity against tested bacteria and elimination of detectable adherent S. aureus in a 3D human epidermal infection model at concentrations of at least 15 µM. Retro-inverso variants were prepared and tested after selection. These results do not establish human safety or clinical efficacy, and exact sequence-level patent priority remains open.

Technology: CalcAMP machine-learning activity ranking

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The primary article reports in-vitro antimicrobial testing, biofilm assays and a 3D human epidermal infection model for selected parent/retro-inverso peptides.
  • Still unknown: These are in-vitro and tissue-model findings, not clinical efficacy or human safety. Sequence-level patent scope and independent replication are not established.
  • AI role: CalcAMP ranked peptide sequences by predicted antimicrobial activity; researchers selected and experimentally tested two lead sequences and retro-inverso variants.

Evidence sources: https://www.mdpi.com/2079-6382/14/11/1172

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary MDPI full article, abstract/results and author-affiliation block reviewed..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Evo genomic language model designs functional toxin-antitoxin and anti-CRISPR systems

Evo used genomic context to generate functional bacterial defence-system components. The paper reports an AI-generated toxin and low-sequence-similarity antitoxins that neutralized it in growth assays, including antitoxins that also acted against multiple natural toxins; it also reports five generated anti-CRISPR proteins that protected bacteria against Cas9 activity in cell and phage assays. Two generated anti-CRISPR proteins lacked significant sequence similarity to known proteins. These experiments demonstrate function-guided genomic design beyond simple copying of natural sequences.

Technology: Evo 1.5, a long-context generative genomic language model trained on prokaryotic DNA sequences

Family Evo: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The paper reports an AI-generated toxin and low-sequence-similarity antitoxins that neutralized it in growth assays, including antitoxins that also acted against multiple natural toxins; it also reports five generated anti-CRISPR proteins that protected bacteria against Cas9 activity in cell and phage assays.
  • Still unknown: This is a set of functional in-vitro bacterial assays, not evidence of therapeutic use, broad host range or ecological performance. One generated toxin had 71% sequence identity to a known RelE toxin; novelty is especially supported by low-similarity antitoxins and anti-CRISPRs, two of which lacked significant sequence identity to known proteins. Earlier Evo phage-genome design and Evo 2 regulatory-sequence work are different papers and different functional results; this row is limited to the toxin-antitoxin/anti-CRISPR findings in the cited Nature article.
  • AI role: Researchers prompted Evo 1.5 with genomic contexts that associate functionally related bacterial genes and used it to generate candidate toxin-antitoxin and anti-CRISPR sequences. Candidates were filtered computationally, synthesized/expressed and tested in bacterial growth, Cas9-protection and phage assays. AI generated candidates; the biological activity was established experimentally.

Evidence sources: https://www.nature.com/articles/s41586-025-09749-7.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-learning microscopy isolates micronucleated cells and reveals chromosome-abnormality formation

The MAGIC platform used an XGBoost image classifier to identify and photolabel micronucleated cells for single-cell genome sequencing. In near-diploid human cell models, the selected cells exposed new chromosome-abnormality formation patterns, including dicentric-associated breakage-fusion-bridge events; measured baseline abnormalities approximately doubled in TP53-deficient cells, and chromosome losses outnumbered gains.

Technology: MAGIC adaptive imaging platform with an XGBoost micronuclei classifier

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: In near-diploid human cell models, the selected cells exposed new chromosome-abnormality formation patterns, including dicentric-associated breakage-fusion-bridge events; measured baseline abnormalities approximately doubled in TP53-deficient cells, and chromosome losses outnumbered gains.
  • Still unknown: Findings come from near-diploid non-transformed cell lines and sequencing limited mainly to chromosome abnormalities larger than 200 kb in dividing cells. Micronucleated RPE-1 cells did not show significant CA enrichment, unlike MCF10A; spontaneous cell-line results do not establish cancer outcomes in patients.
  • AI role: Machine learning examined live confocal images, identified micronucleated cells and drove targeted photolabelling so selected cells could be sorted for downstream single-cell sequencing. The model enabled targeted sampling; it did not itself infer chromosome-abnormality mechanisms from sequence data.

Evidence sources: https://www.nature.com/articles/s41586-025-09632-5

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 28 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Explainable neural-network analysis nominated GPC6 as a Parkinson disease dopaminergic-neuron marker

An explainable machine-learning workflow applied to four post-mortem midbrain single-nucleus RNA-sequencing datasets identified cell-type-specific gene signatures that distinguished Parkinson disease from control cells across datasets (mean balanced accuracy 0.92) and nominated ten candidate genes. GPC6 emerged as a marker in dopaminergic neurons; rare GPC6 variants were enriched among Parkinson disease cases in analyses spanning three case-control cohorts.

Technology: Explainable neural-network classifiers with LIME feature attribution applied to single-nucleus RNA-sequencing data

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: GPC6 emerged as a marker in dopaminergic neurons; rare GPC6 variants were enriched among Parkinson disease cases in analyses spanning three case-control cohorts.
  • Still unknown: The expression analysis used post-mortem datasets and the rare-variant association is not a functional perturbation or proof of causality. GPC6 remains a candidate marker/association requiring further validation. The reported balanced accuracy is cell-level classifier performance, not a clinical diagnostic test. Source checked; independent replication beyond the reported cohorts was not assessed.
  • AI role: The neural-network classifiers identified cell-type-specific genes important for distinguishing Parkinson disease from control nuclei across four datasets; LIME feature attribution prioritized candidate genes. The researchers then used external rare-variant case-control analyses to validate GPC6 enrichment. The model nominates markers and candidates; it does not establish that GPC6 causes Parkinson disease.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/41120302/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 88 of works of the same type, year and subfield (OpenAlex work metrics, 5 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Deep-learning feature-guided design yields antimicrobial peptides D1 and D2

DLFea4AMPGen uses deep-learning activity models and SHAP feature attribution to define peptide fragments used in a candidate sequence space. Sixteen selected peptides were synthesized; the primary article reports 12 with at least two tested activities and identifies D1 and D2 as active against drug-resistant bacterial isolates. D1 also reduced bacterial load in a reported mouse infection experiment. This is preclinical evidence, not human treatment or safety; the study does not establish that every generated sequence or all three activity classes were validated.

Technology: Deep-learning activity prediction and SHAP-guided feature-fragment generation

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The primary source reports the described assays.
  • Still unknown: Independent replication and exhaustive sequence-priority review not established.
  • AI role: Deep-learning activity prediction and SHAP-guided feature-fragment generation

Evidence sources: https://www.nature.com/articles/s41467-025-64378-y.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary publisher article/indexed full text and bibliographic metadata reviewed for identity, AI role, dates, affiliations, and bounded result..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

CellTransformer maps previously uncatalogued regions of the mouse brain

A transformer model applied to spatial transcriptomics mapped fine-grained mouse-brain tissue domains and revealed hundreds of regions absent from existing anatomical atlases. The study evaluated the data-driven domains against the Allen Mouse Brain Common Coordinate Framework and reported strong alignment alongside the novel subdivisions.

Technology: CellTransformer

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: The study evaluated the data-driven domains against the Allen Mouse Brain Common Coordinate Framework and reported strong alignment alongside the novel subdivisions.
  • Still unknown: The new domains are data-driven subdivisions; their biological meaning and generalization beyond the analyzed mouse-brain dataset require further study. Nature's paper page was accessible in search and as a PDF abstract, but its HTML reader view redirected to a cookie wall.
  • AI role: The transformer analyzed cell-neighborhood patterns in spatial transcriptomics and assigned tissue domains without manual annotation; human researchers compared outputs with the Allen reference atlas.

Evidence sources: https://www.nature.com/articles/s41467-025-64259-4

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 16 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Evozyne; NVIDIA

ProT-VAE designs synthetic enzymes with measured activity and high-temperature stability

The transformer–variational-autoencoder ProT-VAE generated synthetic enzyme sequences that were tested experimentally. Among 190 designed human phenylalanine-hydroxylase sequences, 69 had measurable activity and 19 exceeded wild-type activity, with the best at 2.5× wild type. A designed γ-carbonic anhydrase had a measured melting temperature of 116 °C (61 °C above the most thermostable sequence previously reported by the authors) and remained active in 23% methyl diethanolamine at pH 11.25 and 93 °C.

Technology: Protein Transformer Variational AutoEncoder (ProT-VAE), using ProtT5nv transformer representations and a family-specific variational autoencoder

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The transformer–variational-autoencoder ProT-VAE generated synthetic enzyme sequences that were tested experimentally.
  • Still unknown: The reported outcomes are laboratory expression/activity/stability tests on selected synthetic sequences; most generated candidates were not active or did not express. The 0.8× activity comparison is relative to bovine α-carbonic anhydrase under the stated assay, and the 61 °C comparison is against the previously reported thermostable γ-CA referenced by the authors. Industrial relevance is a proposed application, not demonstrated process performance.
  • AI role: The generative model learned protein sequence–function patterns, produced candidate enzyme sequences conditioned in a learned latent space, and guided selection for synthesis. The authors measured enzyme function and stability experimentally; AlphaFold was used only for structure predictions of selected designs.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC12541330/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 62 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

HuDiff humanizes antibodies and nanobodies while retaining measured binding

HuDiff generated humanized variants of a murine SARS-CoV-2 RBD antibody and two alpaca nanobodies. Bio-layer interferometry measured 0.15 nM affinity for the best humanized antibody versus 0.12 nM for its parent; a humanized nanobody measured 2.52 nM versus 5.47 nM. The journal abstract also reports viral neutralization. The October 2025 journal date follows an October 25, 2024 precursor describing one antibody and one nanobody.

Technology: HuDiff-Ab; HuDiff-Nb

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Newly humanized molecules underwent binding assays; journal abstract reports neutralization.
  • Still unknown: Main journal methods are subscription content and were not assessed. Humanness scores do not establish reduced human immunogenicity or clinical safety.
  • AI role: Adaptive autoregressive diffusion generates humanized framework sequences from CDR inputs.

Evidence sources: https://www.nature.com/articles/s42256-025-01120-9?error=cookies_not_supported · https://www.biorxiv.org/content/10.1101/2024.10.22.619416v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Arc Institute

First bacteriophage genomes designed by a genome language model

The Arc Institute group used Evo to design genomes resembling that of phage ΦX174: of 285 designs, 16 yielded functional phages (67–392 mutations from the closest natural one; 13 had never-seen mutations). They infect only two E. coli strains; a cocktail overcame resistance in three evolved strains. bioRxiv preprint; peer review is unconfirmed.

Technology: Evo (genome language model)

Family Evo: profile · models page

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: Of 285 Evo-designed genomes, 16 were synthesized and gave functional phages in E. coli, and these were tested for fitness, host range and, for three resistant strains, a phage cocktail.
  • Still unknown: The source states host range was limited to E. coli C and W, all 16 phages were based on the phiX174 template, and clinical use is not shown.
  • AI role: Evo, fine-tuned on Microviridae sequences, generated whole phage genomes resembling phiX174.

Evidence sources: https://arcinstitute.org/news/hie-king-first-synthetic-phage · https://www.biorxiv.org/content/10.1101/2025.09.12.675911

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 5/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning selects an N-TIMP2 variant with altered MMP specificity

Models trained on binding data for MMP-1, MMP-3 and MMP-9 guided selection of an N-TIMP2 variant. The paper reports high affinity for MMP-9, moderate affinity for MMP-3 and low affinity for MMP-1 relative to wild type. This is an in-vitro protein specificity result, not clinical efficacy. Earlier N-TIMP2 specificity engineering exists; this draft claims only the study’s selected variant and assay result.

Technology: Machine-learning binding predictors for MMP-1, MMP-3 and MMP-9

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Primary article reports an experimentally tested selected N-TIMP2 variant with differential MMP binding profile.
  • Still unknown: Conference abstract posting/result chronology and exact sequence-level overlap remain unresolved; no clinical claim.
  • AI role: Binding predictors scored mutation candidates and guided selection of variant for experimental binding assays.

Evidence sources: https://febs.onlinelibrary.wiley.com/doi/10.1111/febs.70249 · https://pubmed.ncbi.nlm.nih.gov/40912922/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary source checked for identity, date, model role and bounded reported result..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning reveals broad and previously uncharacterized fungal polyketide-synthase activity

A genome-mining and cell-free enzyme study characterized fungal type III polyketide synthases (T3PKSs), revealing broad and varied substrate/product profiles. An ML model trained on 341 enzyme–substrate pairs predicted substrate specificity and guided tests that found two fungal enzymes, AcaePKS and PansPKS, active on all ten tested substrates, including non-natural substrates absent from training data.

Technology: Feedforward neural-network multilayer perceptron trained on protein-language-model embeddings and substrate fingerprints

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: An ML model trained on 341 enzyme–substrate pairs predicted substrate specificity and guided tests that found two fungal enzymes, AcaePKS and PansPKS, active on all ten tested substrates, including non-natural substrates absent from training data.
  • Still unknown: The enzyme products were characterized in vitro; the study examines product scope and biocatalytic potential, not native ecological roles or medical efficacy. Two candidate proteins expressed poorly, and the model's transfer accuracy was lower on non-fungal enzyme classes. The reported five enzyme sample includes only three well-expressing candidates, of which two showed the broadest tested activity.
  • AI role: The model learned fungal T3PKS enzyme–substrate activity from measured data, was tested on unseen substrates, and predicted candidate promiscuous fungal enzymes. Guided experiments confirmed broad substrate activity for AcaePKS and PansPKS, including substrate pairs outside training data; cell-free assays and LC–MS characterized enzyme products.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC12559476/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 13 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-learning-guided mining identifies 91 previously unknown PET hydrolases, including acid-active enzymes

Researchers iteratively combined sequence mining, machine-learning activity/property predictors and high-throughput assays to screen natural enzyme homologs. They experimentally identified 91 previously unknown PET hydrolases among the 212 candidates assayed; 35 had activity on crystalline PET at pH 4.5, and four matched or exceeded a benchmark enzyme under that screening condition.

Technology: Three-round iterative machine-learning pipeline (including VAE/rank model, supervised activity predictors, ESM embeddings and property predictors)

Family ESM / ESMFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Researchers iteratively combined sequence mining, machine-learning activity/property predictors and high-throughput assays to screen natural enzyme homologs.
  • Still unknown: The article defines PET hydrolase by demonstrated PET-depolymerizing activity and notes that activity may be promiscuous relative to the enzyme’s native role. Assays are small-scale and 48 hours; the authors say further enzyme engineering is needed for meaningful bioreactor studies. This is distinct from the catalogued AI-engineered FAST-PETase: it discovers natural homologs and diverse acid/temperature activity profiles rather than optimizing one known PETase scaffold.
  • AI role: Machine-learning activity, thermal-stability and acid-tolerance predictions prioritized natural homologs across successive mining rounds; new assay results updated later predictors. Researchers expressed/purified over 200 candidates and measured PET hydrolysis to verify activity.

Evidence sources: https://doi.org/10.1021/acscatal.5c03460

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 39 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

BindCraft designs protein binders in one shot

An open AlphaFold2-based hallucination pipeline designs binders with reported experimental success rates of 10-100% across 12 targets, with affinities predominantly in the nanomolar range. The authors also demonstrated effects in cell assays and patient-derived samples; these are experimental results, not established clinical treatments. This entry uses the Nature publication date (27 August 2025); the preprint appeared in 2024.

Technology: BindCraft (AlphaFold2-based design)

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Binders tested experimentally across 12 targets (receptors, allergens, de novo proteins, nucleases) with reported success rates of 10-100%, plus functional assays in patient-derived samples and cell assays.
  • Still unknown: The abstract gives no number of designs per target or per-target success rates, and states no limitations.
  • AI role: Open-source automated pipeline using AlphaFold2 weights to design binders in one shot, without high-throughput screening.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-025-09429-6&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 454 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ESM-guided metagenome mining identifies seven new Cas12a subtypes with distinct activities

AIL-Scan fine-tuned an evolutionary-scale protein language model (ESM) on Cas proteins to search metagenomic sequences and identify previously undocumented Cas12a lineages. The study classified seven previously unreported Cas12a subtypes with distinct CRISPR-locus organization, then biochemically characterized 16 Cas12a proteins from the candidate clades, documenting different DNA-cleavage and PAM-recognition preferences. The authors also demonstrated a specific assay strategy for distinguishing the KRAS G12C variant using a Cas12a variant with broader PAM recognition.

Technology: AIL-Scan, using fine-tuned ESM-2 protein language-model embeddings and a machine-learning classifier to identify Cas proteins and predict Cas12a trans-cleavage activity

Family ESM / ESMFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The study classified seven previously unreported Cas12a subtypes with distinct CRISPR-locus organization, then biochemically characterized 16 Cas12a proteins from the candidate clades, documenting different DNA-cleavage and PAM-recognition preferences.
  • Still unknown: The paper reports metagenomic classification and biochemical assays, not a clinical genome-editing or diagnostic application. The PAM/SNP detection demonstration is an experimental proof of concept. The trans-cleavage prediction model was trained on a small labeled set (69 proteins), with two unreported candidates used as an explicit prediction check; CRISPR loci and small-gene annotation in metagenomes remain incomplete. Distinct from other Cas-protein discovery papers by target family, candidate set and measured result.
  • AI role: The fine-tuned ESM model screened metagenomic proteins for Cas identity and generated embeddings; supervised models on small experimental labels predicted trans-cleavage activity to help select uncharacterized candidates. Researchers then manually inspected candidate loci and experimentally measured binding, cleavage and PAM preferences. AI supported sequence-family identification and candidate prioritization; laboratory assays established the reported biochemical properties.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC12375029/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 27 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

GPT-4b micro: Yamanaka factor variants with over 50-fold higher expression

OpenAI and Retro Biosciences used a specialized model to design Yamanaka factor variants that produced over 50-fold higher expression of reprogramming markers in vitro than controls, with results replicated across multiple donors, cell types, and delivery methods, and pluripotency confirmed. Reported by OpenAI; the model is not publicly available.

Technology: GPT-4b micro

Publication: Announcement only · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: direct test
  • Tested: Designed Yamanaka factor variants were tested in vitro across multiple donors, cell types and delivery methods, with over 50-fold higher expression of reprogramming markers and pluripotency confirmed relative to controls.
  • Still unknown: The page gives no detailed protocol or data, the model is not publicly available, and the results are reported by OpenAI only.
  • AI role: A specialized GPT-4b micro model designed Yamanaka factor variants.

Evidence sources: https://openai.com/index/accelerating-life-sciences-research-with-retro-biosciences/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AlphaCD maps experimentally measured catalytic properties across cytidine deaminases and supports a high-fidelity editor

The researchers measured catalytic efficiency, editing windows, motif preferences and off-target activity for 1,100 APOBEC-like cytidine deaminases fused to nCas9 in HEK293T cells, then trained AlphaCD to predict these properties. They applied it to 21,335 UniProt deaminases and experimentally checked a 28-protein subsample. As an application, the team used AlphaCD to identify a high-efficiency deaminase and alanine scanning to produce the E100A variant, which retained editing activity while reducing off-target activity; cell experiments tested editing at endogenous loci and a PCSK9 splice site.

Technology: AlphaCD, a machine-learning model trained on sequence, predicted structure and related protein features

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The researchers measured catalytic efficiency, editing windows, motif preferences and off-target activity for 1,100 APOBEC-like cytidine deaminases fused to nCas9 in HEK293T cells, then trained AlphaCD to predict these properties.
  • Still unknown: The large functional dataset was measured in HEK293T reporter assays. The 21,335-protein landscape is model-predicted, with a 28-protein validation subsample. The engineered E100A editor was tested in cells; these experiments do not establish therapeutic safety, in-vivo efficacy or clinical benefit. This is distinct from the earlier AlphaFold2 structural-clustering papers: the named discovery and assay here concern the 1,100-enzyme dataset, AlphaCD predictions and A0A2R2Z4E4 E100A variant, not CD0208^P52A or the 2023 DddA-like clade findings.
  • AI role: AlphaCD learned experimentally measured activity, editing-window, motif and off-target properties across 1,100 deaminases, predicted properties for 21,335 database sequences, and was applied to select a high-efficiency candidate for subsequent alanine-scanning engineering. The authors report the model's predictions validated on 28 sampled deaminases; the E100A mutation itself was experimentally produced by alanine scanning, not generated by the model.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/40825882/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 80 of works of the same type, year and subfield (OpenAlex work metrics, 7 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

PepMLM designs experimentally binding target-conditioned peptide ligands, including peptides used for targeted protein degradation

PepMLM fine-tunes ESM-2 using masked-language modeling conditioned on a target protein sequence to generate de novo linear peptide binders without structural input. In the reported experiments, selected peptides showed sequence-specific binding to targets including NCAM1 and AMHR2; incorporation into ubiquitin-based antibody constructs enabled targeted degradation of disease-relevant proteins, including Huntington’s disease drivers and viral phosphoproteins.

Technology: PepMLM, ESM-2 fine-tuned with target-sequence-conditioned masked language modeling

Family ESM / ESMFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: In the reported experiments, selected peptides showed sequence-specific binding to targets including NCAM1 and AMHR2; incorporation into ubiquitin-based antibody constructs enabled targeted degradation of disease-relevant proteins, including Huntington’s disease drivers and viral phosphoproteins.
  • Still unknown: The reported evidence is preclinical binding/degradation assays, not therapeutic efficacy, pharmacokinetics or safety. The article notes that experimental optimization/stabilization of its selected peptides was not performed and further validation is needed for viral-target applications. Primary publisher full text/abstract checked; no independent replication assessed.
  • AI role: Given a target protein sequence, PepMLM generates candidate linear binder peptide sequences by reconstructing masked peptide regions. The authors tested selected designs in binding and degradation assays and reported target-specific activity in the stated systems.

Evidence sources: https://www.nature.com/articles/s41587-025-02761-2

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 103 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

HydrAMP-guided design of short tryptophan-rich antimicrobial peptides

HydrAMP generated analogues from a short Arg/Trp prototype, with AMPlify scoring and empirical filters used to select candidates. The authors synthesized three peptides and measured antimicrobial activity across bacterial strains; AMP3 showed the broadest activity among those tested, including MIC50 values of 15 μM against E. coli and 125 μM against P. aeruginosa. This is a small, model-assisted peptide panel built on a known sequence motif, not evidence of therapeutic efficacy.

Technology: HydrAMP deep-learning peptide sequence generator and AMPlify deep-learning antimicrobial-activity scorer

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Three peptides were synthesized and tested in vitro. The lead AMP3 had reported MIC50 values of 15 μM against E. coli and 125 μM against P. aeruginosa, with broader spectrum than AMP1/AMP2 in the tested panel.
  • Still unknown: The design starts from a known short Arg/Trp peptide motif and tests a very small selected set; clinical efficacy, safety, and broad-spectrum performance beyond the strains tested are unknown. The accessible source comparison establishes it is not the Pexiganan/Temporin-A wet-lab result reported in the earlier HydrAMP paper, but does not assert exhaustive sequence-prior-art completeness.
  • AI role: HydrAMP generated candidate analogues from the AMP1 prototype; the AMPlify learned model scored antimicrobial propensity, followed by empirical candidate filtering and wet-lab testing.

Evidence sources: https://chemistry-europe.onlinelibrary.wiley.com/doi/pdfdirect/10.1002/chem.202501918 · https://pubmed.ncbi.nlm.nih.gov/40801152/ · https://www.nature.com/articles/s41467-023-36994-z

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary Wiley publisher full-text indexed article including sequences, Methods, MIC results and publication date; PubMed metadata/abstract/affiliation; separate primary Nature Communications article reviewed to compare the prior HydrAMP wet-lab prototypes and results..

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Archaeasins: deep learning finds antibiotic candidates in the archaeal proteome

The APEX 1.1 deep learning platform scanned 233 archaeal proteomes and flagged 12,623 candidate antimicrobial molecules. Of 80 synthesized candidates, 93% showed in vitro activity against key pathogens, and three selected compounds reduced a drug-resistant bacterium in animal models after a single dose. Candidate and synthesis counts differ between press coverage and secondary reports, and the mouse details come from press coverage rather than the full text, so the paper's own figures should be checked.

Technology: APEX 1.1 deep learning antimicrobial peptide predictor

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: 80 top-ranked APEX-predicted peptides were synthesized and tested by broth microdilution MIC assays, then in mouse skin-abscess and thigh-infection models (six mice per group).
  • Still unknown: Authors state the model is sequence-based without structural information, pseudo-MIC labels may add noise, screening used only reviewed UniProt sequences, the thigh model did not prevent infection, efficacy declined between days 2 and 4 in the skin model, and only female mice were used.
  • AI role: APEX 1.1, a bacterial-strain-specific antimicrobial activity predictor, scored 193,331,608 peptides from the archaeome and selected candidates for synthesis.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC12408343/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 87 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

BEAUT uncovers new gut-microbial bile-acid enzymes and the novel bile acid 3-acetoDCA

The BEAUT workflow combines ESM-2 protein-language-model embeddings with a neural classifier and active-site-pocket-similarity data augmentation to search gut microbial proteins for bile-acid enzymes. The authors predicted over 600,000 candidate enzymes and experimentally tested selected predictions; assays identified new bile-acid activities including MABH and ADS. ADS produces 3-acetoDCA, a previously undescribed bile-acid carbon skeleton, reported in human feces and shown to alter microbial composition and promote Lactobacillus-associated effects in experimental studies.

Technology: BEAUT, ESM-2 protein-language-model embeddings plus dense neural-network classifier, augmented using active-site-pocket similarity

Family ESM / ESMFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The authors predicted over 600,000 candidate enzymes and experimentally tested selected predictions; assays identified new bile-acid activities including MABH and ADS.
  • Still unknown: The 600,000 figure is a prediction set, not experimentally characterized enzymes; authors tested selected subsets. Physiological claims are based on study assays and animal/microbial experiments, not demonstrated human clinical effects. The journal full text was inaccessible in this reader; the Peking University research team's detailed release summarizes its Cell paper, methods and experiments.
  • AI role: Screened microbial protein sequences for likely bile-acid-metabolizing function and prioritized candidate enzymes; experimental characterization of selected AI predictions identified enzyme activities and the ADS/3-acetoDCA pathway.

Evidence sources: https://cqb.pku.edu.cn/info/1101/3119.htm

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 58 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Virtual Lab: AI agents design nanobodies against SARS-CoV-2

A group of AI agents led by a principal agent designed 92 nanobodies: two bound to spike protein from recent variants in lab tests, and humans wrote about 1% of the conversation. Only 2 of 92 bound; the date is the Nature publication date given on the page.

Technology: Multiple AI agents (Virtual Lab)

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The Virtual Lab designed 92 nanobodies; experiments (multiplexed ELISA) showed a range of functional binders. Two showed improved binding to JN.1 or KP.3 while keeping strong binding to the ancestral spike protein.
  • Still unknown: The two are called candidates for further investigation. Neutralization and clinical use not reported in the abstract. Full text is paywalled; counts of tested designs beyond 92 and binding values were not read.
  • AI role: An LLM principal investigator agent guided LLM scientist agents through meetings, with human feedback; they built a pipeline using ESM, AlphaFold-Multimer and Rosetta. Humans ran the experiments.

Evidence sources: https://doi.org/10.1038/s41586-025-09442-9

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 388 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

SONAR identifies 3′UTR sequence features that raise T-cell receptor protein expression and CD8-cell killing

The XGBoost-based SONAR pipeline modeled protein abundance from thousands of sequence features across human immune cells and cell lines, identifying cell-type- and activation-dependent regulatory patterns. Reporter assays found that CT-rich 3′UTR motifs increased protein output while AG-rich motifs were generally repressive. SONAR-informed synthetic 3′UTRs raised CDK4-specific T-cell-receptor protein without increasing its mRNA; engineered CD8+ T cells carrying the CT-rich constructs killed CDK4-expressing melanoma cells more effectively in a coculture assay.

Technology: SONAR, an XGBoost machine-learning pipeline using thousands of sequence features to predict protein abundance

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Reporter assays found that CT-rich 3′UTR motifs increased protein output while AG-rich motifs were generally repressive.
  • Still unknown: The target-cell killing experiment used engineered CD8+ cells in vitro and three donors; no animal or patient treatment outcome was tested. Protein abundance prediction used limited cell and immune-subset datasets, and model feature importance is not itself causal evidence. The assay supports effects for the tested synthetic 3′UTRs and CDK4-specific TCR in the reported system, not all proteins or therapeutic settings.
  • AI role: SONAR learned sequence-feature patterns associated with protein abundance in immune cells, including cell-type- and activation-dependent feature importance. Its outputs guided selection of 3′UTR motifs and synthetic sequence design; reporter assays and a CDK4-specific TCR construct then tested their effects on protein output and CD8+ T-cell cytotoxicity. The model generated design hypotheses; the downstream experiments established the measured motif and T-cell results.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC12285716/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 66 of works of the same type, year and subfield (OpenAlex work metrics, 5 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

VenusMine identifies PET hydrolase KbPETase

VenusMine combines structure-based search with protein-language-model representations and learned property filters to prioritize candidate PET hydrolases from large sequence collections. The authors tested 34 selected proteins and report PET-hydrolyzing activity for 14; KbPETase had an optimum activity temperature 32 °C above the IsPETase comparator. This is laboratory enzyme evidence, not demonstrated waste-processing performance. A 2024 preprint reportedly contains an earlier version; its exact version-to-result chronology was not freshly resolved, so this draft uses the journal version date and does not claim first discovery priority.

Technology: VenusMine; ProstT5; ESM2

Family ESM / ESMFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: 34 selected proteins were tested; the source reports PET-hydrolyzing activity for 14 and characterizes KbPETase experimentally.
  • Still unknown: Scale-up, real-world plastic-waste processing, independent replication and exact relation of the earlier preprint’s experiments remain outside this bounded review.
  • AI role: VenusMine uses structure search, protein-language-model embeddings and learned property predictions to select enzyme candidates for experimental testing.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC12228687/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Indexed primary article text including title, methods, results and publication metadata; direct PMC open triggered CAPTCHA..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-learning screening selects compounds active against Candida

Random-forest prediction and chemical filters helped narrow commercial compounds for experimental testing. Of 17 selected compounds, the paper reports 11 with inhibitory activity against Candida albicans at tested concentrations. Compound 17 had an IC50 of 0.071 μM and MIC of 4.5 μM, compared with fluconazole values of 0.39 μM and 0.78 μM, respectively; its lower IC50 but higher MIC does not establish superiority. The reported selectivity-index definition appears inconsistent with tabulated values, so this draft omits that index. Cell assays do not establish clinical efficacy or safety.

Technology: Random-forest classifiers and regressors; chemical similarity filtering

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Primary full text reports 17 experimentally tested compounds and 11 active; compound-specific IC50/MIC values are stated with different fluconazole comparators.
  • Still unknown: Molecule-level previous activity, exact compound identities and independent replication remain to review.
  • AI role: Random-forest models predict activity and toxicity to prioritize commercially available compounds; selected compounds are experimentally tested.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC12320878/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary full-text article on PMC indexed text and methods/results read; online date, title and DOI verified..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

A closed-loop platform engineers enzymes with minimal human input

The July 2025 Nature Communications paper reports AI-guided biofoundry engineering of two enzymes in four rounds over four weeks, testing fewer than 500 variants per enzyme. Different AtHMT variants achieved about 90-fold greater ethyl-iodide substrate preference and 16-fold ethyltransferase activity; YmPhytase activity improved 26-fold at neutral pH. The system combines ML variant selection, language-model access and automated experiments; the authors describe conditional-to-high automation.

Technology: Machine learning and LLM-guided variant design with biofoundry automation

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Enzyme activity and purified-variant kinetic measurements in two campaigns.
  • Still unknown: Not assessed: supplements; broad applicability beyond two enzymes. Authors classify automation between levels 3 and 4.
  • AI role: Protein models and regression selected variants; the biofoundry built and tested them.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC12215622/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 140 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

BroadAMP-GPT generated and experimentally tested broad-spectrum AMP_S13

BroadAMP-GPT combined transformer-based peptide generation, computational screening and experimental testing. The paper reports that 57% of AI-generated candidates showed potent activity against ESKAPE pathogens. AMP_S13 showed broad-spectrum antimicrobial activity, stability under tested conditions, low cytotoxicity and hemolysis, and efficacy in Galleria mellonella and a murine MRSA wound model. These preclinical findings do not establish human safety or efficacy; the 57% abstract metric is kept distinct from the separately reported assays of 14 synthesized candidates.

Technology: Transformer-based peptide generation with deep-learning-guided screening

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The paper reports a 57% potent-activity rate among AI-generated candidates and describes AMP_S13 broad-spectrum activity, stability, low cytotoxicity/hemolysis, Galleria mellonella infection testing and a murine MRSA wound-infection model. A separate cohort statement reports six of 14 synthesized candidates inhibiting both E. coli and S. aureus at MIC ≤16 μg/mL; these are distinct denominators/outcomes.
  • Still unknown: The results are laboratory and animal preclinical findings, not human efficacy or safety. The entry makes no claim that the peptide sequence was first disclosed here or is globally novel.
  • AI role: BroadAMP-GPT generated peptide candidates with a transformer and applied deep-learning-guided screening before experimental validation; AMP_S13 is the named lead candidate in the reported assays.

Evidence sources: https://www.tandfonline.com/doi/full/10.1080/19490976.2025.2523811 · https://www.tandfonline.com/doi/pdf/10.1080/19490976.2025.2523811

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary publisher full article/PDF, publisher publication history and NCBI/PubMed metadata/abstract; first-public preprint and sequence priority are not claimed..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

AlphaGenome: predicting the effects of DNA variants

The model reads sequences up to 1 million base pairs and predicts thousands of gene-regulation properties at single-letter resolution. It outperformed the best external models in 22 of 24 sequence evaluations and matched or outperformed the best in 24 of 26 variant evaluations. Stated limits: distant regulatory elements; not validated for personal genomes or clinical use.

Technology: AlphaGenome

Family AlphaGenome: profile · models page

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: AlphaGenome was evaluated on sequence and variant-effect benchmarks against external models, outperforming in 22 of 24 and matching or exceeding in 24 of 26 evaluations, plus a TAL1 case study.
  • Still unknown: The source states it is not validated for personal genomes or clinical use, distant regulatory elements remain a challenge, and cell-specific patterns need improvement.
  • AI role: A model takes up to 1 million DNA base pairs and predicts thousands of genomic tracks and variant effects.

Evidence sources: https://deepmind.google/blog/alphagenome-ai-for-better-understanding-the-genome/ · https://www.biorxiv.org/content/10.1101/2025.06.25.661532

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Deep learning identifies MYT1L as a regulator that restricts glioblastoma plasticity

scDORI inferred enhancer-linked regulatory networks from single-cell multiome data and prioritized regulators for perturbation. MYT1L overexpression restricted alternative cell states and reduced tumor growth and invasion in mice; loss of MYT1L restored malignant features. This is preclinical evidence, not an established treatment. The result was public in a May 13, 2025 preprint; the peer-reviewed paper appeared November 4, 2026.

Technology: scDORI; deep learning for enhancer-linked gene-regulatory network inference

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Patient-derived cell perturbations and mouse xenograft experiments tested inferred regulators.
  • Still unknown: Clinical benefit and independent follow-up replication were not established by this review.
  • AI role: Deep network inference nominated regulatory programs and MYT1L for experimental testing.

Evidence sources: https://www.nature.com/articles/s41586-026-11177-0 · https://www.biorxiv.org/content/10.1101/2025.05.13.653733v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AI-designed enhancers control gene expression in healthy mammalian cells

CRG researchers measured more than 64,000 synthetic enhancers across seven stages of blood-cell development and found that enhancer combinations can activate genes in one cell type but repress them in another. The AI model used those measurements to design previously unseen regulatory DNA sequences that drove the intended cell-selective reporter expression in mouse blood cells.

Technology: Machine-learning enhancer design model developed by Lars Velten's group

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: CRG researchers measured more than 64,000 synthetic enhancers across seven stages of blood-cell development and found that enhancer combinations can activate genes in one cell type but repress them in another.
  • Still unknown: The enhancer designs are proof-of-concept reporter regulators, tested in mouse blood cells; this is not evidence of a safe or effective gene therapy. The paper's Cell page redirected to a cookie wall, so detailed findings rely on the CRG's linked official summary.
  • AI role: Researchers used large-scale enhancer activity measurements in healthy blood-cell models to train the model, which generated synthetic DNA enhancers for specified cell-type expression patterns; experiments tested the designs in mouse blood cells.

Evidence sources: https://www.crg.eu/en/news/ai-designed-dna-controls-genes-healthy-mammalian-cells-first-time

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 37 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Experimental feedback guides ESM-2 to brighter CreiLOV variants

Reinforcement Learning from eXperimental Feedback (RLXF) used a learned fluorescence reward model to align ESM-2 with measured CreiLOV sequence-function data. In the peer-reviewed study, the authors expressed 60 generated variants in E. coli and measured cellular fluorescence; six of ten aligned ESM-2 five-mutant designs exceeded CreiLOV fluorescence, and three aligned designs exceeded CreiLOV-T7S. The first bioRxiv version (posted 2025-05-08) already described experimentally enhanced fluorescence from RLXF-aligned designs; detailed measurements here are from the 2026 journal version. The earlier ICLR workshop comparison was computational and did not report new wet-lab fluorescence assays.

Technology: 650-million-parameter ESM-2 protein language model aligned by reinforcement learning from experimental feedback using a learned sequence-fluorescence reward model

Family ESM / ESMFold: profile · models page

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: The journal article reports 60 CreiLOV designs expressed in E. coli BL21(DE3) and measured for total cellular fluorescence. Six of ten aligned ESM-2 five-mutant designs exceeded CreiLOV fluorescence, and three aligned designs exceeded the prior CreiLOV-T7S control; the brightest reported design was 1.2-fold above T7S.
  • Still unknown: The measurements use total cellular fluorescence and are specific to tested designs/conditions; broader cellular contexts, protein function beyond fluorescence, and independent reproduction are not established. The qualitative enhanced-fluorescence result is present in the 2025 preprint abstract; detailed design-set metrics are verified from the later journal version.
  • AI role: A learned sequence-fluorescence reward model trained on measured CreiLOV variant data guides supervised fine-tuning and PPO alignment of ESM-2. The aligned model generates the candidate sequences that were expressed and assayed.

Evidence sources: https://www.biorxiv.org/content/10.1101/2025.05.02.651993v1 · https://www.nature.com/articles/s41467-026-77557-2 · https://openreview.net/pdf?id=yuiQMh52Bm

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary bioRxiv v1 abstract, posting-date and affiliation metadata; primary GEM@ICLR workshop paper full text for methods and experimental outcomes; primary Nature Communications indexed text for later journal-version comparison..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AlphaFold structure analysis helps reveal PHGDH transcriptional activity linked to Alzheimer’s amyloid pathology

AlphaFold-based structural analysis highlighted a helix-helix-turn-helix subdomain in PHGDH that resembles a TALE homeodomain, prompting tests of a previously unrecognized DNA-binding and transcriptional function. Experiments supported PHGDH binding to promoters and activation of IKKα and HMGB1 in astrocytes, independently of its metabolic enzyme activity. In mouse models and human brain organoids, PHGDH expression promoted amyloid pathology; a blood-brain-barrier-permeable inhibitor of its transcriptional function reduced amyloid pathology and improved measured behavioral deficits in mice.

Technology: AlphaFold structure prediction, followed by structural comparison of the predicted PHGDH fold

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: AlphaFold-based structural analysis highlighted a helix-helix-turn-helix subdomain in PHGDH that resembles a TALE homeodomain, prompting tests of a previously unrecognized DNA-binding and transcriptional function.
  • Still unknown: The AlphaFold-derived structural resemblance generated the mechanistic hypothesis; experiments establish the reported PHGDH transcriptional and pathology effects in the tested models. The inhibitor and behavioral effects were tested in mice, not people. Human brain organoid findings are in vitro and do not establish clinical disease causation or treatment efficacy. PHGDH overexpression alone did not reproduce all Alzheimer’s disease features, and no human therapeutic outcome is reported.
  • AI role: The researchers inspected AlphaFold-predicted PHGDH structure and identified a helix-helix-turn-helix subdomain resembling a TALE homeodomain, which motivated the hypothesis and targeted mutation tests for DNA binding. AlphaFold did not establish PHGDH function or disease causality; the transcriptional activity and pathology effects were tested through biochemical, cellular, organoid and mouse experiments.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC12204802/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 20 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Serine hydrolases designed from scratch

The April 2025 Science version describes RFdiffusion enzyme backbones and PLACER assessment of catalytic-state ensembles. Laboratory characterization found catalytic efficiency up to 2.2 × 10^5 M^-1 s^-1, crystal structures with C-alpha RMSD below 1 angstrom, and catalysts in five folds different from natural serine hydrolases. This summary follows the abstract and figure descriptions; an earlier bioRxiv version was posted in August 2024.

Technology: RFdiffusion and PLACER

Family RFdiffusion: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Enzyme activity assays and crystallographic comparison of designed catalysts.
  • Still unknown: Not assessed: complete design success rates and supplementary data.
  • AI role: RFdiffusion generated backbones; PLACER assessed catalytic-state geometry.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/39946508/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 275 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Active learning identifies salt-rich media for flaviolin production

An automated recommendation tool proposed culture-medium mixtures for engineered Pseudomonas putida across three semi-automated campaigns. The paper reports direct flaviolin-titer improvements of about 70% and 60% in two campaigns and a 350% process-yield improvement in a third, each versus its campaign starting medium. The spectroscopic proxy changed by different amounts, so proxy values are not substituted for direct quantification. Results are specific to the tested strain, media and objective and do not establish industrial performance. A preprint preceded the journal article, but its exact posting date and version-specific outcome set are unresolved; the journal date is used without asserting first disclosure.

Technology: Automated Recommendation Tool for active-learning medium optimization

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Primary article reports direct titer changes of about 70% and 60% in two campaigns and a 350% yield change in a third, with each comparison tied to its own starting medium.
  • Still unknown: Outcomes are specific to the engineered strain and the tested campaigns; the earlier preprint date/version contents and independent replication are not established.
  • AI role: The Automated Recommendation Tool proposed media formulations iteratively from experimental measurements; feature-importance analysis highlighted salt effects, but does not prove a unique causal mechanism.

Evidence sources: https://www.nature.com/articles/s42003-025-08039-2.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary publisher article/PDF and institutional preprint metadata reviewed for identity, bounded outcomes, AI role, affiliations and chronology..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

MICrONS: a map of one cubic millimeter of mouse visual cortex

A map with over half a billion connections and about 75,000 reconstructed neurons from roughly 28,000 sections, produced over nine years by more than 150 researchers. AI traced every neuron outline, and the result requires human checking, much of which is already done. This is about one-thousandth of a mouse connectome, and cell counts vary within the text.

Technology: AI neuron segmentation

Publication: Peer-reviewed paper · Test context: Test context unspecified in annotation. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: not reported
  • Tested: Not reported
  • Still unknown: The source states the map is about one-thousandth of a mouse connectome, the result requires human proofreading (much already done), and the connectome is not necessarily the neural code.
  • AI role: AI traced the outline of every neuron through the electron microscopy slices.

Evidence sources: https://engineering.princeton.edu/news/2025/04/09/scientists-map-vision-centers-mammalian-brain · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-025-08790-w&resultType=core&format=json

The legacy label “Not reported” does not distinguish information absent from reviewed material from information not assessed. It is not a scientific finding of absence.

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 367 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 5/5 · Validation 1/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ESM-Ezy mining identifies multicopper oxidases and asparaginases with experimentally improved catalytic properties

The ESM-1b semantic-embedding mining strategy nominated low-sequence-similarity enzyme candidates. In the reported experimental assessment, 44% of tested multicopper oxidases outperformed their query enzymes in at least one measured property (including catalytic efficiency, heat/solvent tolerance or pH stability); 51% were reported to excel in environmental-remediation applications. Among mined L-asparaginases, 40% had higher specific activity and catalytic efficiency than query enzymes.

Technology: ESM-Ezy, using ESM-1b protein-language-model embeddings and semantic-space similarity

Family ESM / ESMFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: In the reported experimental assessment, 44% of tested multicopper oxidases outperformed their query enzymes in at least one measured property (including catalytic efficiency, heat/solvent tolerance or pH stability); 51% were reported to excel in environmental-remediation applications.
  • Still unknown: Percentages describe the selected candidates and tested property comparisons in this study, not the fraction of all database enzymes that will be superior. The reported remediation suitability is assay-specific; environmental or industrial deployment is not established. Primary abstract/full-text search result checked; no independent replication assessed.
  • AI role: ESM-1b sequence embeddings and semantic-space similarity ranked low-similarity candidate enzymes from sequence databases for experimental follow-up; assays assessed catalytic properties of selected multicopper oxidases and L-asparaginases.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/40188191/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 38 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

CataPro-assisted enzyme mining identifies a high-activity Sphingobium enzyme for vanillin production

CataPro, a deep-learning model using pretrained protein-language representations and molecular fingerprints to predict enzyme kinetics, was combined with sequence and structure-based filtering to screen candidate enzymes for converting 4-vinylguaiacol to vanillin. The team experimentally identified Sphingobium sp. CSO (SsCSO), whose measured activity was 19.53 times that of the initial enzyme CSO2, and then used CataPro-guided engineering to increase SsCSO activity a further 3.34-fold.

Technology: CataPro

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The team experimentally identified Sphingobium sp.
  • Still unknown: The reported 19.53× and 3.34× figures are relative activities for the tested 4-vinylguaiacol oxidation assay; they do not establish industrial-scale yield or process performance. Model predictions were used alongside conventional sequence/structure methods, not as the sole discovery method.
  • AI role: CataPro predicted kinetic parameters to prioritize candidate carotenoid-cleavage-oxygenase family enzymes and rank variants; researchers combined its predictions with BLAST, structural similarity and laboratory enzyme assays. It materially narrowed the enzyme search and guided mutations, while experiments established measured activity.

Evidence sources: https://doi.org/10.1038/s41467-025-58038-4

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 140 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

TeleProt-guided engineering yields a nuclease variant with improved neutral-pH biofilm degradation

A machine-learning-guided design and high-throughput screening campaign produced a Bacillus licheniformis NucB variant with 19-fold higher measured specific activity than wild type at pH 7. At 1 μM in a laboratory biofilm assay at neutral pH, the selected variant degraded 71% ± 2% of the biofilm versus 13% ± 3% for wild-type NucB.

Technology: TeleProt machine-learning-guided protein-library design framework; best-performing ML3 variant selected by model-based optimization with a deep neural network (MBO-DNN)

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: A machine-learning-guided design and high-throughput screening campaign produced a Bacillus licheniformis NucB variant with 19-fold higher measured specific activity than wild type at pH 7.
  • Still unknown: The evidence is from enzyme and in-vitro B. licheniformis biofilm assays; it does not show wound healing, animal efficacy or clinical benefit. At the 1 μM dose, the paper reports the stated biofilm comparison; conditions and estimates are study-specific. This is engineering of a known nuclease, not discovery of a naturally new enzyme or organismal mechanism.
  • AI role: TeleProt combined sequence-evolutionary information with experimental genotype–phenotype data to propose diverse NucB variants for high-throughput microfluidic activity screening. The selected ML-guided variant was then purified and assessed in a lower-throughput nuclease assay and a neutral-pH biofilm degradation assay.

Evidence sources: https://doi.org/10.1016/j.cels.2025.101236

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 45 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning and peptide-array experiments reveal new substrates for protein-modifying enzymes

A hybrid workflow coupled machine-learning models with enzyme-treated peptide arrays to predict previously unexplored protein post-translational-modification sites. Experimental validation confirmed 37–43% of proposed sites across SET8 and SIRT1–7; mass spectrometry confirmed dynamic methylation for several SET8 substrates and deacetylation at 64 unique SIRT2 sites. The work also found altered SET8 substrate networks associated with breast-cancer missense mutations.

Technology: Enzyme-specific ML-hybrid substrate prediction models trained from high-throughput peptide-array assays

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Experimental validation confirmed 37–43% of proposed sites across SET8 and SIRT1–7; mass spectrometry confirmed dynamic methylation for several SET8 substrates and deacetylation at 64 unique SIRT2 sites.
  • Still unknown: Reported validation rate is 37–43% of proposed sites, so predictions are not uniformly confirmed. Results concern in-vitro enzyme assays and selected sites; disease relevance of the altered SET8 network is not established as a clinical mechanism.
  • AI role: Learned enzyme-specific substrate preferences from experimental array measurements and prioritized novel PTM sites for experimental follow-up; this enabled the reported new SET8/SIRT substrate assignments.

Evidence sources: https://www.nature.com/articles/s42004-025-01717-6.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 77 of works of the same type, year and subfield (OpenAlex work metrics, 6 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AI-guided promoter editing raises maize ZmVTE4 expression and kernel alpha-tocopherol

Basenji2-based deep-learning models and editing-plasticity analysis guided CRISPR editing of maize ZmVTE4, a gene involved in vitamin E biosynthesis. Researchers isolated a previously unobserved four-base deletion in the gene’s 5′ untranslated region; the edited allele significantly increased ZmVTE4 expression and α-tocopherol content in kernels. A separate promoter-variant assay supported the four-base deletion, rather than a co-occurring 82-base deletion, as the key change associated with increased reporter activity.

Technology: Basenji2-long and Basenji2-3K-NAM sequence-to-expression deep-learning models with editing-plasticity analysis

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: A separate promoter-variant assay supported the four-base deletion, rather than a co-occurring 82-base deletion, as the key change associated with increased reporter activity.
  • Still unknown: The demonstrated editing result concerns one ZmVTE4 allele and maize material in the reported experiments; it does not establish effects on yield, nutritional outcomes in consumers, or performance across environments and genetic backgrounds. The four-base deletion is newly engineered and was not found in the 721 accessions examined, rather than a naturally observed allele. The paper reports limitations in prediction for rice and tomato and incomplete modeling of environmental influences; claims here are restricted to the maize ZmVTE4 test.
  • AI role: Deep-learning models predicted plant gene expression from regulatory DNA and used interpretability/occlusion and in-silico editing to identify promoter/5′UTR sequences likely to alter expression. The editing-plasticity predictions guided CRISPR-Cas9 targeting of ZmVTE4; the resulting four-base 5′UTR deletion was experimentally linked to increased gene expression and maize-kernel α-tocopherol. This is a measured crop-specific molecular/metabolite outcome, not merely an in-silico editing proposal.

Evidence sources: https://link.springer.com/article/10.1186/s13059-025-03516-7

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 23 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AI structural modeling helps map nearly 300 metabolite-binding interactions in E. coli

A cross-institutional study combined chemoproteomics with AI-driven structural modeling to identify nearly 300 small-molecule ligands and their binding sites on proteins in E. coli. The map adds experimentally grounded candidate links between bacterial metabolites and proteins involved in core cellular processes.

Technology: AI-driven structural modeling (Kozakov laboratory; specific model not named in institutional announcement)

Publication: Peer-reviewed paper · Test context: Test context unspecified in annotation. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: not reported
  • Tested: The map adds experimentally grounded candidate links between bacterial metabolites and proteins involved in core cellular processes.
  • Still unknown: The cited count is nearly 300 ligands and binding sites; the announcement does not establish that every interaction is a newly discovered physiological mechanism. No named AI model was verified.
  • AI role: AI structural modeling contributed to identifying and localizing metabolite-protein interactions alongside chemoproteomic experiments; the institutional source does not name the exact AI model.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/39862855/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 14 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 1/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Canopy traits across tropical forests mapped with deep learning and satellite data

Field data from more than 1,800 vegetation plots, combined with satellite, terrain, climate and soil data and convolutional neural networks, predicted variation in 13 tree functional traits. Tropical American forests show about 40% greater functional richness than African and Asian forests, and African forests show 32% higher functional divergence than American forests. The authors caution that accuracy depends on data quality and coverage, and that AI should support rather than replace field sampling.

Technology: Convolutional neural networks combined with satellite spectral data

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Trait predictions from field data on more than 1,800 vegetation plots, combined with satellite, terrain, climate and soil data, were checked with an uncertainty analysis.
  • Still unknown: Accuracy depends on data quality and coverage, and further data collection would refine the maps.
  • AI role: Not stated in the page read; the page does not mention artificial intelligence or machine learning.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-025-08663-2&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 55 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

BioEmu-1: ensembles of protein structures with deep learning

An open-source model that generates thousands of protein structures per hour on a GPU and reproduces the equilibrium distribution of a molecular-dynamics simulation using 10,000 to 100,000 times fewer GPU-hours. Predicted folding free energies correlate well with experimental values. The authors call it ‘a first step.’

Technology: BioEmu-1

Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: retrospective or benchmark
  • Tested: Generated ensembles were compared with a molecular dynamics simulation of protein G using 10,000 to 100,000 times fewer GPU hours, and predicted folding free energies were correlated with experimental values.
  • Still unknown: The authors state they are aware of limitations but the post does not list them, and it is presented as an early step.
  • AI role: A generative deep learning model produces thousands of protein structures per hour on a GPU.

Evidence sources: https://www.microsoft.com/en-us/research/blog/exploring-the-structural-changes-driving-protein-function-with-bioemu-1/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

FragFold identifies inhibitory fragments within bacterial proteins

MIT researchers used FragFold, an AlphaFold-based approach, to predict fragments that bind to and inhibit full-length proteins in E. coli. The researchers experimentally confirmed more than half of its binding or inhibition predictions, including cases without prior structural interaction data.

Technology: FragFold, using AlphaFold

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The researchers experimentally confirmed more than half of its binding or inhibition predictions, including cases without prior structural interaction data.
  • Still unknown: The institutional report says more than half of predictions were accurate; this is a study-specific validation result, not a general success rate across proteins or organisms.
  • AI role: Researchers computationally fragmented bacterial proteins and used AlphaFold predictions to nominate inhibitory fragments; experiments tested binding and inhibition. The AI supplied structural predictions, not experimental verification.

Evidence sources: https://news.mit.edu/2025/ai-system-fragfold-predicts-protein-fragments-0220

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 15 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Arc Institute

Evo 2: the largest AI model for biology

Trained on more than 9.3 trillion nucleotides from over 128,000 genomes, it reads sequences up to 1 million nucleotides, flags harmful mutations, and designs genetic elements and genomes as long as those of simple bacteria. In BRCA1 tests it exceeds 90% accuracy in distinguishing benign from potentially pathogenic variants (a result limited to that gene). Preprint; human pathogens were excluded from training.

Technology: Evo 2 (genome language model)

Family Evo: profile · models page

Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: retrospective or benchmark
  • Tested: The model was tested on BRCA1 variants, where it reached over 90% accuracy distinguishing benign from potentially pathogenic mutations, and on identifying genetic changes affecting protein function and fitness.
  • Still unknown: The source reports the BRCA1 result for one gene only and states no other limitations beyond human pathogens being excluded from training.
  • AI role: A genome language model trained on 9.3 trillion nucleotides flags harmful mutations and designs genetic elements.

Evidence sources: https://arcinstitute.org/news/evo2

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-learning-assisted taxonomy identifies Limnognathia desmeti

Learned analyses of transcriptomic data contributed to an integrative description of Limnognathia desmeti from the Crozet Islands, the second species described in Micrognathozoa. Phylogenomic and molecular evidence also contributed; machine learning alone did not establish the species. Greenland and Pyrenean populations were not definitively split. Limited specimen and population sampling constrains the result.

Technology: Variational autoencoder and random-forest species delimitation

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Molecular analyses, microscopy and multiple delimitation approaches supported the formal description.
  • Still unknown: Broader geographic diversity and dispersal history remain uncertain.
  • AI role: Variational autoencoder and random-forest species delimitation supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11836703/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary full-text species-delimitation Methods, taxonomic Results, date and affiliations; supplements not reviewed..

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-learning network analysis identifies a conserved cell-mechanics gene module

Researchers combined transcriptomic profiles with single-cell mechanical measurements from mouse and human systems and used machine-learning-based discriminative network analysis to infer a conserved five-gene module associated with cell mechanical properties. In perturbation experiments, changing CAV1 expression by silencing or overexpression altered measured cell stiffness, providing experimental support for a selected marker’s role.

Technology: PC-corr machine-learning-based discriminative network analysis

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Researchers combined transcriptomic profiles with single-cell mechanical measurements from mouse and human systems and used machine-learning-based discriminative network analysis to infer a conserved five-gene module associated with cell mechanical properties.
  • Still unknown: Only CAV1 was experimentally tested as a selected target in the reported perturbation experiments; the other genes are candidate regulators. Experiments show changes in measured stiffness in specific cell models, not broad physiological or clinical effects. Paper’s eLife assessment notes further validation, especially of genes with previously unknown roles, would strengthen the network.
  • AI role: PC-corr analyzed transcriptomic data linked to measured mechanical phenotypes to identify a conserved five-gene signature and prioritize candidate regulators; selected genes were computationally assessed across six validation datasets, and CAV1 was experimentally perturbed to test effects on cell mechanics.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11832173/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 53 of works of the same type, year and subfield (OpenAlex work metrics, 2 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ProtGPS reveals sequence signals that direct proteins to subcellular compartments

ProtGPS, a protein language model and generative sequence-design system, helped identify amino-acid sequence features that guide proteins to subcellular compartments. The paper applies the model to predict disease-mutation-associated localization shifts and reports cellular testing of 20 examples.

Technology: ProtGPS protein language model and generative design algorithm

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The paper applies the model to predict disease-mutation-associated localization shifts and reports cellular testing of 20 examples.
  • Still unknown: The 20 tested mutations are a sample of a much larger computational screen; the predictions do not establish causal disease mechanisms in patients.
  • AI role: ProtGPS predicted compartment localization and mutation-associated shifts, and its generative component designed sequences for target localization. Researchers experimentally tested 20 mutation predictions in cells and reported confirmation.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC12034300/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 109 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

EVOLVEpro improves proteins up to 100-fold with few experiments

For the January 2025 journal version indexed by PubMed and Europe PMC, the Science abstract reports that EVOLVEpro combines protein language models, regression and experimental feedback to improve six proteins used in RNA production, genome editing and antibody binding, with gains up to 100-fold in desired properties. A July 2024 preprint described five proteins. The date identifies the indexed journal version, not the first public announcement.

Technology: EVOLVEpro (protein language model with active learning)

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Experimental engineering of proteins for RNA production, genome editing and antibody binding.
  • Still unknown: Not assessed: full-paper per-protein data and generalization beyond the studied proteins.
  • AI role: Active learning used protein-language-model representations and regression to select variants.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/39571002/ · https://news.massgeneralbrigham.org/en/mgb-bidmc-unveil-ai-protein-engineer

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 264 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning-guided cell-free engineering creates specialized amide synthetases for pharmaceutical molecules

A cell-free design-build-test-learn platform generated sequence-function data from 1,217 McbA variants across 10,953 reactions. Ridge-regression ML guided enzyme variants that produced nine small-molecule pharmaceuticals with activity 1.6- to 42-fold above the parent; initial substrate-promiscuity assays found McbA could make 11 pharmaceutical compounds plus hybrid molecules. This is experimental enzyme engineering, not evidence of therapeutic efficacy of the products.

Technology: Augmented ridge-regression ML models within a cell-free sequence-function mapping and enzyme-engineering workflow

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: A cell-free design-build-test-learn platform generated sequence-function data from 1,217 McbA variants across 10,953 reactions.
  • Still unknown: The platform optimizes in-vitro enzyme production of molecules and does not establish drug efficacy, manufacturing readiness, or broad transferability beyond tested reactions. The nine-product gains are relative to the parent enzyme under the reported assay conditions.
  • AI role: Used measured variant fitness data to predict and select higher-activity McbA variants for distinct amide-forming reactions; selected variants were synthesized and assayed.

Evidence sources: https://www.nature.com/articles/s41467-024-55399-0.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 153 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

MaSIF designs binders for protein–drug complexes

MaSIF designed proteins that recognize three complexes containing progesterone, venetoclax, or actinonin, then validated them in the lab. A venetoclax-inducible system activated the antitumor activity of CAR T cells in vitro; this is a proof of concept without clinical validation.

Technology: MaSIF (deep learning on molecular surfaces)

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: MaSIF-designed binders were validated in the lab against three drug-bound complexes, and a venetoclax-inducible system switched on CAR-T tumor-killing activity in vitro.
  • Still unknown: The source states the CAR-T work is in vitro only, with no clinical validation, and binding sites are hard to predict.
  • AI role: MaSIF generated surface fingerprints and searched for complementary protein surfaces to design binders.

Evidence sources: https://actu.epfl.ch/news/a-deep-learning-pipeline-for-controlling-protein-3/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 93 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

EvoGradient evolves antimicrobial peptides active against resistant pathogens

EvoGradient uses an explainable deep-learning potency model and virtual sequence evolution to optimize candidate antimicrobial peptides. Applied to peptides encoded by low-abundance human oral bacteria, it virtually evolved 32 candidates; the six most effective were synthesized and tested against multidrug-resistant pathogens. The authors report activity across carbapenem-resistant E. coli, K. pneumoniae and A. baumannii and vancomycin-resistant E. faecium; pep-19-mod reduced bacterial loads by more than 95% in mouse thigh-infection models under the reported systemic and local administrations. These are preclinical findings and do not establish human efficacy or safety.

Technology: Explainable deep-learning antimicrobial-peptide potency model with gradient-based virtual sequence evolution

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Six synthesized peptides were tested against multidrug-resistant pathogens; the authors also report pep-19-mod in-vivo testing in mouse thigh-infection models, with over 95% bacterial-load reduction under the described systemic and local administration.
  • Still unknown: Mouse infection outcomes do not establish human safety, dosing, efficacy or clinical translation; only six candidates from the 32 virtual evolutions were synthesized and tested.
  • AI role: EvoGradient predicts peptide potency and virtually modifies peptide sequences through gradient-based evolution; the workflow selected optimized candidates for synthesis and experimental testing.

Evidence sources: https://www.nature.com/articles/s41564-024-01907-3 · https://pubmed.ncbi.nlm.nih.gov/39825096/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary Nature Microbiology indexed article text and primary PubMed abstract/metadata/affiliations; conclusions limited to reported preclinical assays..

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Active learning identifies a protoglobin variant with 99% yield and cis-selective cyclopropanation

An iterative machine-learning campaign engineered the protoglobin from Pyrobaculum arsenaticum (ParPgb) for non-native cyclopropanation of 4-vinylanisole. After two model-guided rounds following an initial library screen, variant MPFDY gave 99% total cyclopropane yield and a 14:1 cis:trans ratio in the reported assay, compared with a parent variant that produced about 40% yield and favored trans product. The result demonstrates a specific sequence combination with improved yield and stereoselectivity in this enzyme scaffold.

Technology: ALDE (Active Learning-assisted Directed Evolution), using a bootstrapped deep-neural-network ensemble and Thompson sampling

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: After two model-guided rounds following an initial library screen, variant MPFDY gave 99% total cyclopropane yield and a 14:1 cis:trans ratio in the reported assay, compared with a parent variant that produced about 40% yield and favored trans product.
  • Still unknown: Cyclopropanation is an established non-native enzyme reaction class; the study's distinct result is the ALDE-guided ParPgb sequence and its measured product yield/selectivity. The 99% yield and 14:1 ratio are study assay results, not evidence of industrial process readiness. The study is a single enzyme-engineering campaign.
  • AI role: ALDE trained on measured sequence–yield data, quantified model uncertainty, and selected variants across two iterative rounds in a five-residue, 3.2-million-sequence design space. Experimental screening measured product yield and stereoselectivity; the model-guided variants included MPFDY, whose combined mutations improved the targeted cis-product objective where tested single mutations and their naive recombinations did not.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11739421/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 186 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Deep learning enables multi-site bone marrow adiposity genetics in UK Biobank

Deep learning measured bone marrow fat fraction at four skeletal sites in UK Biobank MRI data; after quality control, 48,608 participants contributed to genome-wide analyses. In the white-participant meta-GWAS, the study reported 67, 147, 134 and 174 independent significant variants for femoral head, total hip, femoral diaphysis and spine, mapping to 54, 90, 43 and 100 genes, respectively. Analyses also identified site- and sex-specific genetic associations and added signals in multi-ancestry analyses.

Technology: Previously validated deep-learning models for MRI segmentation and bone marrow fat-fraction measurement

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Deep learning measured bone marrow fat fraction at four skeletal sites in UK Biobank MRI data; after quality control, 48,608 participants contributed to genome-wide analyses.
  • Still unknown: The GWAS, transcriptome-wide and colocalization results are statistical associations and candidate gene mappings, not proof of causal mechanisms. Ancestry subgroup sizes differ, and results do not establish clinical utility or disease prevention. The deep-learning measurement method was established in an earlier study; this record concerns the distinct population-scale phenotype and genetic findings. A 2025 author correction fixed a misspelling of AKNA as ANKA; it did not alter the reported analysis or findings.
  • AI role: The study applied deep-learning models to segment four skeletal regions and derive bone marrow fat-fraction phenotypes at UK Biobank scale. Those AI-derived measurements were the phenotypes used in the study's subsequent genetic association analyses; the paper's genetic findings are not themselves outputs of a generative model.

Evidence sources: https://www.nature.com/articles/s41467-024-55422-4

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 30 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Protein language model guides creatinase stability optimization

A deep protein language model used prior single- and multiple-mutant measurements to guide combinations across two design/test rounds. The 13-substitution variant 13M4 reportedly retained nearly wild-type catalytic activity and increased apparent melting temperature by 10.19 °C. This assay does not establish long-term stability across production conditions, and the model recombined previously measured mutations.

Technology: Deep protein language model

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Publisher article and PubMed metadata report 13M4, near-WT catalytic activity and apparent melting-temperature increase.
  • Still unknown: Prior exact-sequence reporting and independent replication not established.
  • AI role: Protein language model guided combinations of previously tested mutations for experimental stability and activity assays.

Evidence sources: https://doi.org/10.1002/mlf2.12151

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary source checked for identity, date, model role and bounded reported result..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

PICNIC machine learning predictions revealed previously poorly characterized human proteins in cellular condensates

The PICNIC gradient-boosting model ranked proteins for membership in biomolecular condensates using sequence-pattern and AlphaFold2-derived structural features. To test its predictions, researchers expressed 24 poorly characterized human proteins in U2OS cells: 21 formed mesoscale fluorescent foci, and 18 met the paper’s high-confidence condensate threshold. The experiments mapped examples to cytoplasmic bodies, nuclear bodies, nucleoli or P-bodies; for example, TYW5 and RAD51AP1 showed distinct localization patterns. Across the tested set, the authors found condensate localization was not limited to highly disordered proteins.

Technology: PICNIC, a gradient-boosting machine-learning classifier using protein sequence co-occurrence patterns, sequence complexity/disorder features and AlphaFold2-derived structural features

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: To test its predictions, researchers expressed 24 poorly characterized human proteins in U2OS cells: 21 formed mesoscale fluorescent foci, and 18 met the paper’s high-confidence condensate threshold.
  • Still unknown: The experimental test used fluorescently tagged proteins expressed in U2OS cells, so observed localization may depend on expression level, tag, cell type or condition. The study did not establish each protein’s endogenous localization or biological function; the authors explicitly state that condensate localization alone makes no functional claim. Three of 24 proteins did not form detectable mesoscale foci, and three additional results were low-confidence under the paper’s size threshold. This is targeted experimental validation of model-selected candidates, not an independently replicated atlas.
  • AI role: PICNIC scored human proteins for likely cellular-condensate membership and selected poorly characterized positive predictions for experimental testing. Its rankings directly guided the 24-protein U2OS localization validation; the model was built to predict condensate members (including clients), not only proteins that self-drive phase separation.

Evidence sources: https://www.nature.com/articles/s41467-024-55089-x.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 52 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

PlantRNA-FM identifies translation-associated RNA motifs and structure motifs in plants

The PlantRNA-FM model, pretrained on sequence and predicted RNA structure from 1,124 plant transcriptomes, identified 5′-UTR motifs associated with translation efficiency across plant transcriptomes. Reporter assays in plants tested selected predictions: mutating a high-translation-associated RNA secondary-structure motif reduced measured translation efficiency by up to 5.3-fold, while disrupting a model-prioritized low-translation-associated RNA G-quadruplex motif increased it by up to 5.8-fold. The paper reports the rG4 motifs as translation-repressive, consistent with prior evidence on individual rG4s; the transcriptome-wide model prioritization and tested instances are the scoped findings here.

Technology: PlantRNA-FM, an interpretable RNA foundation model integrating RNA sequence and predicted structure

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Reporter assays in plants tested selected predictions: mutating a high-translation-associated RNA secondary-structure motif reduced measured translation efficiency by up to 5.3-fold, while disrupting a model-prioritized low-translation-associated RNA G-quadruplex motif increased it by up to 5.8-fold.
  • Still unknown: Experimental reporter validation was limited to selected motifs and had n=3, as reported in the figure caption. The model’s transcriptome-wide associations are not individually causal validations. The repressive role of rG4 motifs was previously known for individual cases; the paper’s contribution includes model-prioritized plant 5′-UTR motifs and experimental effects for selected constructs, not first discovery of rG4 biology. Reporter assays do not establish whole-plant phenotype or crop benefit.
  • AI role: PlantRNA-FM was pretrained with plant RNA sequence and structure data, then fine-tuned on translation-efficiency data. An attention-contrast interpretation framework prioritized 5′-UTR sequence/structure motifs associated with translation; researchers tested selected motif predictions by mutating reporter constructs and measuring translation efficiency in plants.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11652376/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 65 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ApexGO optimizes antimicrobial peptide activity

ApexGO uses Bayesian optimization with a peptide variational autoencoder and an activity-prediction oracle to optimize derivatives of ten peptide templates. The primary article reports 100 synthesized sequences tested in vitro and activity improvements; selected candidates showed activity in mouse A. baumannii infection models. These results are preclinical and do not establish human efficacy.

Technology: ApexGO; peptide variational autoencoder; Bayesian optimization; APEX predictor

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The primary source reports the study-specific result summarized here.
  • Still unknown: Independent replication and broad deployment are not established by this review.
  • AI role: A predictor scores generated peptide derivatives and Bayesian optimization uses feedback to propose variants; researchers synthesize and test them in vitro and in mouse models.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC13201158/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary-source indexed article or abstract and bibliographic metadata reviewed; complete source scope is given in source_checks..

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

CHOOSER uses protein foundation models to discover functional Casλ homologs

The CHOOSER framework fine-tuned ESM-2 protein-language-model representations for alignment-free CRISPR-Cas discovery and pre-crRNA self-processing predictions. It nominated 11 Casλ homologs, nearly doubling the known Casλ set. Wet-lab work on selected homologs confirmed pre-crRNA processing for five of six tested proteins; EphcCasλ also cleaved target DNA and showed trans-cleavage activity, with a TTR PAM reported.

Technology: CHOOSER, a protein foundation-model framework fine-tuning ESM-2 representations for Cas homolog identification and Cas12 pre-crRNA-processing classification

Family ESM / ESMFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Wet-lab work on selected homologs confirmed pre-crRNA processing for five of six tested proteins; EphcCasλ also cleaved target DNA and showed trans-cleavage activity, with a TTR PAM reported.
  • Still unknown: This is a specific functional genomics result, not evidence of a ready-to-use diagnostic or gene-editing tool. Only selected candidates were purified and tested; the paper reports pre-crRNA activity in five of six tested Casλ proteins and DNA-cleavage validation for a subset. EphcCasλ activity and PAM characterization are in-vitro findings. The article has a publisher-listed author correction dated 2025-01-09; this record relies on the version-of-record text and retains limited scope.
  • AI role: CHOOSER mined candidate proteins adjacent to CRISPR arrays using learned sequence representations rather than alignment-only detection and prioritized Cas12 homologs by predicted pre-crRNA processing. The study reported 11 Casλ homologs and experimentally tested selected predictions; purified EphcCasλ was validated for self-processing pre-crRNA, DNA cleavage and trans-cleavage.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11576732/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 26 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Evo: a foundation model for DNA, from molecules to genomes

Published in Science, Evo (Arc Institute, Stanford, and UC Berkeley) designed EvoCas9-1, a CRISPR system sharing only about 73% of its sequence with CRISPR-Cas9 but with comparable activity after eleven designs, and new IS200/IS605 transposons that cut and paste DNA. The team calls this the beginning and conducted an ethics and safety study.

Technology: Evo (DNA foundation model)

Family Evo: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Evo was prompted to generate CRISPR systems; EvoCas9-1 was obtained after testing eleven designs, validated in the lab by an experimental biology subteam.
  • Still unknown: The page gives no success rates or error analysis, and an ethics and safety investigation addressed potential risks rather than efficacy.
  • AI role: Evo, a DNA-trained foundation model, designed the protein and RNA components of EvoCas9-1 and new IS200/IS605 transposons.

Evidence sources: https://arcinstitute.org/news/evo-science

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 737 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

GRACE designs two carbonic-anhydrase-like enzymes with measured catalytic activity

The GRACE workflow combined RFdiffusion structure generation, ProteinMPNN sequence design, CLEAN enzyme classification and solubility/structure filtering to select two carbonic-anhydrase-like sequences from 10,000 candidates. The authors report experimental expression and carbonic anhydrase activity: dCA12_2 at 400 WAU/mL and dCA23_1 at 260 WAU/mL in the reported assay.

Technology: GRACE (RFdiffusion, ProteinMPNN and CLEAN, with computational solubility/structural filters)

Family RFdiffusion: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The authors report experimental expression and carbonic anhydrase activity: dCA12_2 at 400 WAU/mL and dCA23_1 at 260 WAU/mL in the reported assay.
  • Still unknown: The reported activity values are from the study's carbonic-anhydrase assay and do not establish purified-enzyme kinetics, physiological CO2-fixation performance, or industrial utility. The designs and validation are a two-protein proof of concept; no independent replication assessed.
  • AI role: The generative-design workflow produced structures/sequences and machine-learning function classification prioritized two candidates from a 10,000-sequence pool; researchers expressed both and measured carbonic-anhydrase activity.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11669173/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 89 of works of the same type, year and subfield (OpenAlex work metrics, 20 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

HiDDEN finds malignant-like cell signals in early multiple-myeloma precursor samples

HiDDEN refines sample-level case labels at single-cell resolution. In human bone-marrow single-cell RNA-seq data, its labels tracked manual neoplastic-cell annotations in mixed precursor samples and inferred malignant-like cells in three low-purity MGUS samples that the original manual analysis had labeled as containing only healthy cells. The inferred cells were enriched for previously described malignancy-associated gene signatures. In a separate mouse demyelination dataset, HiDDEN identified an early endothelial subpopulation with expression changes associated with angiogenesis, blood–brain-barrier breakdown and extracellular-matrix responses.

Technology: HiDDEN, a machine-learning framework for refining disease/condition labels at single-cell resolution

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: HiDDEN refines sample-level case labels at single-cell resolution.
  • Still unknown: The MGUS finding is a transcriptomic inference in previously collected samples, computationally supported by known gene signatures and sample-purity estimates; the paper does not report independent pathology confirmation, clinical follow-up, or validation as a diagnostic test. The endothelial subpopulation was found in a mouse lesion model; its association with blood–brain-barrier dysfunction does not establish causality.
  • AI role: The model inferred which individual cells in case samples were affected rather than assigning the sample-level case label to all cells. This enabled the paper’s reanalysis of human myeloma precursor single-cell data and mouse demyelination single-nucleus data, yielding the specified malignancy and endothelial-population findings.

Evidence sources: https://www.nature.com/articles/s41467-024-53666-8.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 24 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 2/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Transcriptome-based machine learning reveals an MCL1 dependency in alveolar soft part sarcoma

A support-vector-regression model trained to infer gene dependencies from RNA-sequencing profiles predicted selective MCL1 dependence in alveolar soft part sarcoma (ASPS), distinct from TFE3-translocation renal cell carcinoma. The authors tested the prediction with CRISPR knockout and the MCL1 inhibitor murizatoclax in two ASPS cell lines; the reported viability effects were greater in ASPS than comparator tRCC lines.

Technology: Transcriptome-based support-vector regression model (RBF kernel; 657 highly predictable gene dependencies)

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: The authors tested the prediction with CRISPR knockout and the MCL1 inhibitor murizatoclax in two ASPS cell lines; the reported viability effects were greater in ASPS than comparator tRCC lines.
  • Still unknown: bioRxiv preprint, not peer reviewed as stated by PMC. Validation was in two ASPS cell lines with CRISPR and inhibitor assays; it does not establish patient benefit or clinical efficacy. The model nominates dependencies and is not a functional screen itself.
  • AI role: The model predicted per-gene dependency scores from tumor/cell-line expression profiles and prioritized selective vulnerabilities. The study then tested MCL1 dependence in ASPS using three distinct sgRNAs and murizatoclax; those experiments, rather than the prediction alone, provide the reported validation.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11527139/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-guided synthetic regulatory DNA drives cell-type-specific expression across cell lines and vertebrates

The CODA platform used the Malinois deep-learning model and sequence-search algorithms to design synthetic cis-regulatory elements (CREs). In massively parallel reporter assays, 94.1% of the synthetic sequences met the authors’ cell-type-specificity criterion, versus 73.6% of model-selected natural sequences and 40.6% of chromatin-mark-selected natural sequences; synthetic sequences also had higher measured target/off-target separation across K562, HepG2 and SK-N-SH cells. In transgenic zebrafish, two of three liver-directed and two of three neuron-directed CREs showed the intended activity. One neuron-directed sequence, synN1, retained cortical and neuronal expression in mice, including postnatal neocortical layer 6 and subplate neurons.

Technology: CODA (Computational Optimization of DNA Activity) with Malinois deep neural network and AdaLead, simulated annealing, and Fast SeqProp sequence search/design algorithms

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: In massively parallel reporter assays, 94.1% of the synthetic sequences met the authors’ cell-type-specificity criterion, versus 73.6% of model-selected natural sequences and 40.6% of chromatin-mark-selected natural sequences; synthetic sequences also had higher measured target/off-target separation across K562, HepG2 and SK-N-SH cells.
  • Still unknown: The in-vitro work used three transformed human cell lines and reporter assays. In-vivo testing covered a small selected set: two of three liver-directed and two of three neuron-directed elements were active in zebrafish; mouse validation focused on neuronal synN1, with n=3 synN1 mice for postnatal cell-type analysis and n=6 for the 5-week cortical reporter assay. These are transgenic reporter results, not therapeutic delivery or evidence of clinical utility. Earlier synthetic enhancer studies existed; this paper’s counted result is its CODA/Malinois comparison against natural sequences plus selected vertebrate validation, not the general idea of AI-designed enhancers.
  • AI role: Malinois predicted MPRA activity of candidate DNA sequences in K562, HepG2 and SK-N-SH cells; CODA iteratively optimized sequences for specified cell-type specificity. Researchers synthesized and experimentally measured the designs by MPRA, then tested selected CREs with reporter constructs in zebrafish and mice. The paper attributes the novel sequences and their measured expression specificity to this model-guided design loop.

Evidence sources: https://www.nature.com/articles/s41586-024-08070-z

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 180 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

CRISPR perturbation and Bayesian network inference map immune-disease regulatory cascades in primary CD4+ T cells

The study used CRISPR knockouts of 84 genes in primary human CD4+ T cells and a new linear latent causal Bayes (LLCB) network-inference method to estimate 211 regulatory connections. Integrating the inferred network with immune GWAS analyses identified an interconnected module in which KMT2A was inferred upstream of JAK-STAT and IL-2 signaling regulators including IRF4, STAT5A/B and IL2RA, linking an epigenetic regulator to immune-disease-associated pathways.

Technology: Linear latent causal Bayes (LLCB), a Bayesian extension of linear latent causal network inference, applied to CRISPR knockout and RNA-sequencing data

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The study used CRISPR knockouts of 84 genes in primary human CD4+ T cells and a new linear latent causal Bayes (LLCB) network-inference method to estimate 211 regulatory connections.
  • Still unknown: The study perturbed 84 selected genes in primary CD4+ T cells from three donors, so this is not a complete T-cell regulatory map. LLCB “direct effects” are statistical effects adjusted for other perturbed genes and are not evidence of physical binding; effects from unperturbed genes can be attributed to modeled edges. The GWAS integration identifies pathway links, not proof that the inferred cascade causes clinical disease. A prior bioRxiv version (10.1101/2023.09.17.557749) is the same study and is not counted separately.
  • AI role: LLCB inferred direct and indirect regulatory relationships from perturbation-associated expression effects across 84 targeted genes. The model-derived network and gene-program analysis, combined with immune GWAS statistics, supported the specific KMT2A-to-JAK-STAT/IL-2 regulatory finding. CRISPR perturbation supplied the experimental perturbation data; the network algorithm converted those measurements into the reported regulatory map.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11605694/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 90 of works of the same type, year and subfield (OpenAlex work metrics, 21 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning-guided amidase engineering improves enantioselectivity

Random-forest models using chemistry and geometry descriptors were trained on 240 prior amidase datasets and used to guide substrate and variant selection. The primary article reports that variant I198F/W328F achieved an enantioselectivity E-value of 303.5, compared with 5.7 for wild type, for the reported substrate assay. Most experimental substrate predictions agreed, with identified exceptions. This bounded laboratory result does not establish broad performance across substrates or industrial application.

Technology: Random-forest classification models and descriptor-guided variant design

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Primary paper reports the stated assay results.
  • Still unknown: Independent replication and exhaustive prior-art review not established.
  • AI role: Random-forest classification models and descriptor-guided variant design

Evidence sources: https://www.nature.com/articles/s41467-024-53048-0

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary publisher full text and bibliographic metadata reviewed for identity, AI role, dates, affiliations and bounded results..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Iterative Gaussian process screening discovers selective aurein-derived antifungal peptides

An uncertainty-aware Gaussian process regression workflow iteratively selected 23 synthetic α/β-peptide analogues for synthesis and measurements of fungal inhibition and human-red-blood-cell hemolysis. One lead, peptide 6-4, was a measured variant of an existing aurein-derived template and achieved a selectivity index of 57.1 (60.9 using the study’s finer MIC interval) versus 1.2 for aurein 1.2. Its high selectivity was driven by low hemolysis; the result is a specific peptide-design and assay contribution, not a new peptide scaffold or evidence of therapeutic efficacy. The ChemRxiv v1 result was posted on 2024-10-04; the peer-reviewed journal article followed in 2025.

Technology: Iterative uncertainty-aware Gaussian process regression using molecular descriptors and measured antifungal/hemolysis labels

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: Across six iterative rounds, uncertainty-aware Gaussian process regression selected 23 peptide variants for synthesis and antifungal/hemolysis assays. Lead 6-4 had a selectivity index of 57.1 (60.9 using the study’s finer MIC interval), versus 1.2 for aurein 1.2; the authors attribute the selectivity to reduced hemolysis.
  • Still unknown: Peptide 6-4 modifies a known aurein-derived template. The result is a measured sequence and assay outcome, not evidence of a new scaffold or therapeutic efficacy. The version-specific first public result is the 2024 ChemRxiv preprint; the journal article followed in 2025.
  • AI role: An uncertainty-aware Gaussian process regression model selected peptide variants for synthesis and assays, and measured outcomes were fed back into subsequent selection rounds.

Evidence sources: https://chemrxiv.org/engage/chemrxiv/article-details/67a7e4386dde43c9081bdfc1 · https://pubs.rsc.org/en/content/articlepdf/2025/sc/d4sc06689h · https://pubmed.ncbi.nlm.nih.gov/40028619/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary ChemRxiv version-history record and v2 manuscript; primary RSC indexed full article PDF, including sequence tables, Results, Methods and publication metadata; PubMed primary abstract/affiliation metadata..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

FlyWire: the complete map of neurons in the Drosophila brain

Connectome of the adult Drosophila brain: nearly 140,000 neurons, over 15 million synaptic connections, and over 8,000 cell types. The page says deep neural networks segmented images and extracted neurons, but automated segmentation had millions of errors and required about 33 years of human correction across over 50 labs and citizen scientists.

Technology: Deep neural networks for segmentation

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: The connectome was checked against an earlier partial map, and connectomes were found to be highly consistent between animals.
  • Still unknown: About 0.5% of neurons have developmental wiring defects, connection signs were predicted rather than measured, and the dataset was not built for a specific medical application.
  • AI role: Deep neural networks segmented electron microscopy images and extracted neurons; the first automated pass had millions of errors and was corrected by humans.

Evidence sources: https://mrclmb.ac.uk/?p=4875

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 758 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 5/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AI-assisted cerebellar atlases reveal primate-specific cellular organization

Deep-learning segmentation and spatial annotation supported cerebellar atlases of macaques, marmosets and mice, revealing primate-specific cell subtypes and regional gene-expression differences. Transcriptomic gradients corresponded with awake functional-connectivity patterns. This represents the September 27, 2024 journal version; an institutional announcement gives September 26, and the observed associations do not establish causal functions.

Technology: Deep-learning cell segmentation; STAGATE; DestVI

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: New spatial tissue measurements, biological replicates and comparisons with awake fMRI supported the reported anatomical associations within this study.
  • Still unknown: Not assessed: exhaustive supplements, earlier-preprint chronology and independent confirmation. Functional mechanisms are not established by spatial correlations.
  • AI role: Deep-learning models segment cells; STAGATE clusters spatial domains and DestVI assigns cell-type probabilities.

Evidence sources: https://pure.eur.nl/ws/files/169697439/Cross-species_single-cell_spatial_transcriptomic_atlases_of_the_cerebellar_cortex.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Deep learning identifies previously unreported protein interactions in human bacterial pathogens

A RoseTTAFold2-Lite pipeline combined residue coevolution, deep-learning complex prediction and AlphaFold modelling to screen protein pairs across 19 human bacterial pathogens. In a validation set selected using the deep-learning pipeline, co-immunoprecipitation detected four of six predicted interactions. Three pairs lacked prior direct interaction evidence: E. coli UbiE–YcaR, P. aeruginosa PA4105–PA4106, and P. aeruginosa WbpG–HisF2. The PA4105–PA4106 structural similarity to proteins in 3-thiaglutamate biosynthesis and the WbpG–HisF2 result provide functional hypotheses, not established native-pathogen mechanisms.

Technology: RoseTTAFold2-Lite integrated with direct coupling analysis and AlphaFold complex modelling

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: In a validation set selected using the deep-learning pipeline, co-immunoprecipitation detected four of six predicted interactions.
  • Still unknown: The positive interaction evidence comes from heterologous co-immunoprecipitation/pull-down assays, with two independent experiments per positive pair, rather than native-pathogen systems. The validation subset was small (four positives among six AI-selected candidates); most of the thousands of predicted interactions remain untested, and false positives and false negatives are expected. Proposed roles for PA4105–PA4106 and UbiE–YcaR remain hypotheses; this paper does not establish their physiological function or a therapeutic target.
  • AI role: RoseTTAFold2-Lite used residue–residue coevolution and predicted complex structures to rank candidate protein pairs in a proteome-wide screen. The authors selected six pairs from the final deep-learning pipeline for co-immunoprecipitation tests; four yielded positive interaction signals. Three were described as lacking previously reported direct interaction evidence, including UbiE–YcaR, PA4105–PA4106 and WbpG–HisF2. The models enabled the experimental target selection; they did not themselves establish physical interactions.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11445079/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 52 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

MILTON-augmented disease cohorts surface putative new gene–disease associations in UK Biobank data

MILTON uses machine learning on routine clinical biomarkers, quantitative traits and plasma proteomics to predict disease phenotypes and augment case labels in UK Biobank. In a phenome-wide analysis of 484,230 genome-sequenced participants, the authors report 182 putative novel gene–disease associations that were not genome-wide significant in baseline cohorts or explained by biomarker associations; 38.9% of the 167 mappable signals were nominally supported with the same direction in FinnGen.

Technology: MILTON (Machine Learning with Phenotype Associations), an ensemble machine-learning framework for disease prediction from clinical and molecular biomarkers

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: In a phenome-wide analysis of 484,230 genome-sequenced participants, the authors report 182 putative novel gene–disease associations that were not genome-wide significant in baseline cohorts or explained by biomarker associations; 38.9% of the 167 mappable signals were nominally supported with the same direction in FinnGen.
  • Still unknown: The authors classify the 182 associations as putative novel; they are computational associations, not experimentally established causal genes. FinnGen nominal same-direction support applied to 38.92% (65/167) of the mappable putative novel PheWAS associations, while the paper also reports other aggregate exome-wide results. A separate example, APOB–thyrotoxicosis, showed FinnGen support. These signals require further biological validation.
  • AI role: MILTON predicted cases from clinical biomarkers and other quantitative traits, augmenting phenotype labels for genetic association analyses. The AI-enabled cohort enrichment yielded putative novel gene–disease signals; the study compared them with baseline analyses, orthogonal prioritization tools and FinnGen.

Evidence sources: https://www.nature.com/articles/s41588-024-01898-1.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 137 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning identifies alkaline laccases from Lepista nuda

Machine-learning pH-optimum predictors screened metagenomic sequences, and two predicted Lepista nuda laccases were produced and profiled. The authors report LnLccA and LnLccB optimum activity at pH 10 and 9 for substrate 2,6-DMP, respectively, while both had pH 4 optima for ABTS. The substrate-dependent contrast is retained; it is not a general alkaline optimum or industrial demonstration.

Technology: Supervised, semi-supervised and unsupervised pH-optimum prediction models

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Open primary full article reports recombinant LnLccA and LnLccB pH profiles, substrate-specific optima and validation of ML predictions.
  • Still unknown: Prior characterization and novelty of sequence/activity remain open; industrial applicability not demonstrated.
  • AI role: Models trained on curated laccase pH data predicted candidates from metagenomic sequences that were expressed and profiled.

Evidence sources: https://link.springer.com/article/10.1186/s13068-024-02566-6

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary source checked for identity, date, model role and bounded reported result..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

AlphaProteo: AI-designed protein binders

A system for designing high-affinity protein binders, tested against seven targets, achieved 88% laboratory success on BHRF1 and bindings 3 to 300 times stronger than the best existing methods. It produced no valid binders for TNFα; strong binding is only a first step.

Technology: AlphaProteo

Family AlphaProteo: profile · models page

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: Binders were designed against seven target proteins and tested in the lab, with 3 to 300 times stronger binding and higher experimental success rates than the best existing methods.
  • Still unknown: Per-target success rates are not given on the abstract page and no limitations are listed.
  • AI role: A family of machine learning models for protein design that generates the binders.

Evidence sources: https://arxiv.org/abs/2409.08022

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Deep learning maps fine-scale plant distributions and fire-associated change

DeepBiosphere combines a residual CNN trained on citizen-science observations with aerial imagery to map California distributions for 2,221 plant species. The authors report median AUC 0.95 versus 0.88 for climate-only models and use repeat imagery to map plant-community change associated with the 2013 Rim Fire. These are California-specific ecological maps and observational associations; the study does not establish causal effects of fire or a new species discovery.

Technology: DeepBiosphere residual convolutional neural network for species-distribution mapping

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: The primary article reports held-out mapping median AUC 0.95 versus 0.88 for climate-only models and uses repeat imagery to map community change following the 2013 Rim Fire.
  • Still unknown: The fire-severity relationship is observational and does not establish causation. Mapping performance is California-specific and weaker for rare species.
  • AI role: A residual CNN trained with citizen-science observations and aerial imagery produced fine-scale species-distribution maps that were then used to estimate rapid plant-community change.

Evidence sources: https://www.pnas.org/doi/10.1073/pnas.2318296121 · https://pmc.ncbi.nlm.nih.gov/articles/PMC11406280/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary PNAS indexed article and PMC full-text indexed content, including abstract, Methods, results and plant/fire case studies; PubMed primary metadata/affiliations and first-online date..

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Learned call analysis identifies receiver-specific marmoset vocal patterns

Classifiers identified receiver-related structure in calls from ten captive marmosets. Playback experiments also found stronger responses to calls originally directed at the listener. The authors interpret these patterns as vocal labels; a later commentary disputes whether the evidence establishes name-like reference, and the authors defend that interpretation. The bounded finding is receiver-specific acoustic and behavioral differentiation, not decoded human-like names.

Technology: Random-forest acoustic classifiers

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Recorded calls and playback responses were analyzed.
  • Still unknown: Name-like reference and alternative explanations remain debated; the captive sample is small.
  • AI role: Random-forest acoustic classifiers supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/39208084/ · https://zoology.ubc.ca/edg/pdfs/Oren2024.pdf · https://pubmed.ncbi.nlm.nih.gov/39747802/ · https://pubmed.ncbi.nlm.nih.gov/40866605/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Indexed primary published-paper Methods and abstract, PubMed metadata, critical commentary abstract and authors response abstract..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-learning Treg framework identifies SMARCC1 as a previously unrecognized regulator

A framework combining single-cell ATAC-seq, motif analysis, topic modeling, machine learning and genetic variation identified SMARCC1 as a previously unrecognized controller of regulatory T-cell chromatin programs. CRISPR-Cas9 deletion of Smarcc1 in Tregs produced the predicted loss of accessibility in topic-3 regulatory regions, supporting the computationally inferred connection.

Technology: Integrated machine-learning framework using topic modeling and high-density genetic variation over single-cell ATAC-seq profiles

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: A framework combining single-cell ATAC-seq, motif analysis, topic modeling, machine learning and genetic variation identified SMARCC1 as a previously unrecognized controller of regulatory T-cell chromatin programs.
  • Still unknown: CRISPR validation supported the predicted SMARCC1 effect for topic 3; topic 10 was not represented and the authors note possible redundancy with SMARCC2. The result concerns regulatory chromatin programs in Tregs, not clinical or therapeutic outcomes.
  • AI role: The framework organized co-varying open chromatin regions into Treg regulatory modules, inferred transcription-factor effects on those programs, and nominated SMARCC1 as a regulator of topic-3/10 accessibility. The team tested the prediction by CRISPR-Cas9 editing Tregs and profiling chromatin accessibility.

Evidence sources: https://cbdm.hms.harvard.edu/assets/Publications/2024pub/chowdharyPNAS.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 32 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Generative antibody language model yields a pseudovirus-neutralizing XBB spike binder

Tencent AI Lab and collaborators developed PALM-H3, a pretrained antibody language model that generates heavy-chain CDRH3 sequences, and A2binder to rank antigen-antibody binding. For SARS-CoV-2 XBB, which the paper says was not present during training, the selected generated antibody (Artificial 1) bound spike and neutralized pseudovirus in vitro, with reported KD 0.13 nM and IC50 0.00301 μg/mL. The study also assayed generated antibodies against wild-type, Alpha and Delta; performance varied by antibody and variant.

Technology: PALM-H3 pretrained antibody generative language model and A2binder antigen-antibody affinity predictor

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: For SARS-CoV-2 XBB, which the paper says was not present during training, the selected generated antibody (Artificial 1) bound spike and neutralized pseudovirus in vitro, with reported KD 0.13 nM and IC50 0.00301 μg/mL.
  • Still unknown: The functional evidence is in vitro, using SARS-CoV-2 pseudovirus rather than live-virus or clinical outcomes. The study tested selected antibody candidates; several generated candidates performed substantially worse, so the reported top result is not representative of all generated sequences. The stated novelty is experimental neutralization against the XBB variant by a generated antibody, not proof of a treatment or protection in people.
  • AI role: PALM-H3 generated CDRH3 sequences conditioned on antigen specificity using a RoFormer-based antibody model; A2binder scored antibody-antigen pairs to select candidates for binding, surface-plasmon-resonance and pseudovirus neutralization tests. The paper reports it generated and selected an XBB spike-binding antibody despite XBB being absent from its training set.

Evidence sources: https://www.nature.com/articles/s41467-024-50903-y

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 80 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

MODIFY-guided library screening identifies a cytochrome c variant active in both carbon–boron and carbon–silicon bond formation

Researchers applied MODIFY, an ensemble model combining protein language models and sequence-density models, to design a diverse high-fitness library for Rhodothermus marinus cytochrome c. Experimental screening identified variant MLYPPT, which catalyzed both non-natural borylation and silylation: reported total turnover numbers were 2,880 and 2,740, with enantiomeric ratios of 94:6 and 97:3, respectively. The same library yielded distinct high-performing variants for each reaction.

Technology: MODIFY (ML-optimized library design with improved fitness and diversity)

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Experimental screening identified variant MLYPPT, which catalyzed both non-natural borylation and silylation: reported total turnover numbers were 2,880 and 2,740, with enantiomeric ratios of 94:6 and 97:3, respectively.
  • Still unknown: The C–B and C–Si reactions had previously been demonstrated with separately evolved cytochrome c variants; this finding is the discovery of additional, structurally distinct variants and a generalist MLYPPT catalyst. The reported yields and enantiomeric ratios are from the study's assays; industrial scale-up was not shown.
  • AI role: MODIFY used evolutionary sequence models and protein language models for zero-shot fitness estimates, then co-optimized predicted fitness and residue-level diversity to propose a 1,000-variant library. Researchers synthesized and assayed sampled library members; the model-guided library contained experimentally measured new cytochrome c variants, including MLYPPT with activity in both reactions.

Evidence sources: https://par.nsf.gov/servlets/purl/10529357

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 99 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning uncovers four new extracellular contractile injection-system toxins

An XGBoost classifier trained on genomic and biochemical features prioritized eCIS-associated toxins from bacterial and archaeal genomes. Four candidates, EAT14–17, were experimentally shown to be toxic in bacterial and/or yeast cells; EAT14 also inhibited Rac1/Cdc42-mediated mitogenic signaling in human cells.

Technology: XGBoost classifier for extracellular contractile injection-system-associated toxins (EATs)

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Four candidates, EAT14–17, were experimentally shown to be toxic in bacterial and/or yeast cells; EAT14 also inhibited Rac1/Cdc42-mediated mitogenic signaling in human cells.
  • Still unknown: Four of eight selected candidates were toxic under the tested heterologous-expression assays; the authors report an empirical positive predictive value of 50% for this sample and discuss possible host/compartment mismatch. Toxin delivery by native eCIS and physiological target organisms remain incompletely established. EAT14’s signaling effect was shown in human cells, not as a therapeutic effect.
  • AI role: The classifier used genomic and protein biochemical features to score candidate EAT genes across eCIS-positive microbial genomes. The model narrowed candidates for functional tests; heterologous expression assays validated four previously uncharacterized toxins, with follow-up active-site mutations and a human-cell signaling assay.

Evidence sources: https://link.springer.com/article/10.1038/s44320-024-00053-6

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 83 of works of the same type, year and subfield (OpenAlex work metrics, 9 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AI-guided HuJ3 mutation improves HIV-1 gp120 binding and neutralization in assays

The IsAb2.0 workflow, combining AlphaFold-Multimer complex modeling with FlexddG mutation scoring, proposed five substitutions to improve the binding of humanized J3 nanobody HuJ3 to HIV-1 gp120. Of these, E44R experimentally improved binding (ELISA EC50 1.696 nM) and neutralization relative to HuJ3; the other proposed substitutions did not consistently improve binding, and T99R failed to express. The gain did not restore neutralization to the parental llama J3 level.

Technology: IsAb2.0 antibody-design protocol integrating AlphaFold-Multimer 2.3/3.0, SnugDock, alanine scanning and FlexddG

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Of these, E44R experimentally improved binding (ELISA EC50 1.696 nM) and neutralization relative to HuJ3; the other proposed substitutions did not consistently improve binding, and T99R failed to express.
  • Still unknown: This is a single, previously known J3 nanobody humanization/optimization case, not de novo antibody discovery, animal efficacy or clinical evidence. Only one of five proposed mutants improved binding; several predictions failed and the authors note limitations in prediction accuracy and manual steps. E44R remained less potent than parental J3.
  • AI role: The protocol modeled the HuJ3–gp120 complex, nominated interface hotspots and computationally ranked five single-residue mutations. Researchers synthesized/tested variants in binding and HIV-1 neutralization assays; the measured E44R improvement is the experimental outcome, while most predictions did not improve performance.

Evidence sources: https://academic.oup.com/bib/article/25/5/bbae445/7758758

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 81 of works of the same type, year and subfield (OpenAlex work metrics, 11 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AI helps resolve an unknown regulatory mechanism in bacterial acetyl-CoA synthetase

Researchers at the University of Greifswald resolved how a post-translational modification regulates Bacillus subtilis acetyl-CoA synthetase, a central metabolic enzyme. The study used AI-based structural prediction to explain the molecular mechanism at near-atomic detail.

Technology: AlphaFold2 and AI-assisted structural analysis

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: Test context unspecified in annotation. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: not reported
  • Tested: Researchers at the University of Greifswald resolved how a post-translational modification regulates Bacillus subtilis acetyl-CoA synthetase, a central metabolic enzyme.
  • Still unknown: The institutional announcement says the regulatory mechanism is evolutionarily conserved and may matter in higher organisms; that implication is not itself a demonstration in those organisms.
  • AI role: AlphaFold2 structural predictions supported the researchers' mechanistic interpretation of how the enzyme is regulated; the source describes a collaboration with biochemical and structural experiments, not an AI-only result.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11255334/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 27 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 1/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Gubra

Machine-Learning-Guided Peptide Drug Discovery: Development of GLP-1 Receptor Agonists with Improved Drug Properties

A machine-learning-guided peptide optimization workflow selected GUB021794, reported at 0.018 nM GLP-1R potency in the paper assay and dose-dependent weight loss in diet-induced obese mice. These preclinical results do not establish human efficacy, safety or a dosing regimen.

Technology: streaMLine machine-learning QSAR analysis integrated with peptide synthesis and screening

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Primary abstract reports the named candidate and mouse outcomes; no human clinical result is established.
  • Still unknown: Independent replication and exhaustive prior-art review are not established.
  • AI role: streaMLine machine-learning QSAR analysis integrated with peptide synthesis and screening

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/38977267/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary source checked for study identity, AI operation, dates, affiliations and the bounded result. PubMed affiliation block names Gubra..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ESM3 and esmGFP: a fluorescent protein designed by a language model

ESM3 designed esmGFP, a fluorescent protein of 229 amino acids with 96 mutations relative to the closest known protein (about 58% similarity). In a second round of 96 designs, some proteins glowed as brightly as natural GFPs. The comparison with ‘more than 500 million years of evolution’ is an analogy-based estimate.

Technology: ESM3

Family ESM / ESMFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: In two rounds of 96 designs, lab measurements of fluorescence found some proteins as bright as natural GFPs; esmGFP differs by 96 mutations and is about 58% similar to the closest natural protein.
  • Still unknown: The page lists no formal limitations; the designed B8 was about 50 times dimmer and slower to mature than natural GFPs.
  • AI role: ESM3 generated new GFP candidates prompted with structural residues from natural GFP.

Evidence sources: https://www.evolutionaryscale.ai/blog/esm3-release

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Deep-learning design produced soluble analogues of membrane-protein folds with preserved binding functions

A deep-learning pipeline based on inversion of AlphaFold2 and ProteinMPNN designed soluble proteins with claudin, rhomboid-protease and GPCR-like folds. Experiments reported stable, folded soluble designs and structural agreement with target folds. The researchers grafted functional motifs, including a ghrelin-receptor epitope that bound its antibody and active/inactive A2A-receptor-like designs that differed in mini-Gs binding; a binding-site knockout served as a control.

Technology: AF2seq deep-learning backbone design with ProteinMPNN sequence design

Family ProteinMPNN: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Experiments reported stable, folded soluble designs and structural agreement with target folds.
  • Still unknown: This is a protein-design proof of concept with in-vitro biophysical and binding assays; the analogues are not native membrane receptors and do not demonstrate signaling, drug activity or therapeutic efficacy. The GPCR analogues tested binding to mini-Gs in conformation-designed constructs. The separate live-catalog record on designed proteins modulating dopamine D1 receptor signaling reports a different binder/target and signaling result; this paper reports soluble receptor-fold analogues and binding assays, not D1 signaling modulation.
  • AI role: The deep-learning design pipeline generated soluble analogues of complex membrane-protein folds and the scaffolds into which native functional motifs were grafted. The study tested solubility, thermal stability, structural accuracy and selected binding functions experimentally, including conformation-specific A2A-receptor-like constructs.

Evidence sources: https://www.nature.com/articles/s41586-024-07601-y

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 184 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

African elephants address one another with calls resembling names

About 470 calls from 101 callers to 117 recipients in Kenya: a model trained only on acoustic structure identifies the intended recipient, and playback experiments show elephants respond more to calls addressed to them. The model confirms a previously suspected pattern; more data are needed to isolate the ‘names’.

Technology: Acoustic classification model

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: About 470 calls from 101 callers to 117 receivers were analyzed, and playback experiments showed elephants respond more to calls addressed to them.
  • Still unknown: More data are needed to isolate the names, it is unknown whether elephants name other things, and naming was not universal.
  • AI role: A model trained only on acoustic structure identified the intended recipient of each call, supporting the name-like finding.

Evidence sources: https://newsmediarelations.colostate.edu/2024/06/10/elephants-have-names-for-each-other-like-people-do-new-study-shows/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 77 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-learning-guided lysine dioxygenase variant improves turnover

A structure-based machine-learning mutation-ranking method, combined with molecular dynamics and human rational design, guided variant selection for a lysine dioxygenase. The authors report a five-substitution variant with total turnover number 218 ± 3, compared with 160 ± 2 for wild type in the reported assay. The improvement is specific to this enzyme and assay, not a claim of industrial or therapeutic performance. The source uses conflicting variant abbreviations: its abstract says LPNYI and its Results say LPYNI.

Technology: MutComputeX; molecular dynamics

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: Primary abstract reports the designed pentamutant’s TTN and wild-type comparator.
  • Still unknown: Independent replication and broader application are not established by this bounded review. The source uses conflicting abbreviations for the five-substitution variant.
  • AI role: Structure-based ML prioritized amino-acid substitutions; MD and rational selection also contributed to the final combined variant.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/38895203/ · https://www.biorxiv.org/content/10.1101/2024.06.04.597480v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary preprint version and journal methods/results; the variant abbreviation differs between abstract and Results..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AlphaFold2-guided mining identifies context-independent cytidine deaminases for efficient base editing

Researchers used AlphaFold2 to predict structures for 1,483 cytidine deaminases, clustered the predictions and experimentally characterized representatives. The screen identified previously uncharacterized enzymes with high editing efficiency, and CD0208^P52A showed little sequence-context preference across AC/GC/TC/CC motifs, a narrow editing window and relatively low off-target activity. In mammalian cells, its base editor introduced stop codons in single- and multicopy genes without DNA double-strand breaks.

Technology: AlphaFold2 protein-structure prediction, followed by structural clustering of candidate cytidine deaminases

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Researchers used AlphaFold2 to predict structures for 1,483 cytidine deaminases, clustered the predictions and experimentally characterized representatives.
  • Still unknown: The base-editing experiments were preclinical cell experiments, not animal or clinical efficacy tests. AlphaFold2 supplied predicted rather than experimentally solved structures. The paper reports multiple promising deaminases; CD0208^P52A is the specific context-independent editor highlighted here. No therapeutic safety or in-vivo delivery claim is implied.
  • AI role: AlphaFold2-predicted structures at scale were used to mine and cluster 1,483 putative deaminases; representatives were then selected for experimental characterization in cytosine base editors. The model enabled structure-guided candidate selection; editing outcomes were measured experimentally in mammalian cells.

Evidence sources: https://www.nature.com/articles/s41551-024-01220-8.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 37 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-learning analysis linked two sepsis-associated neutrophil clusters to five mouse-validated marker genes

Reanalysis of single-cell profiles from four sepsis patients and two healthy controls identified two neutrophil clusters enriched in the sepsis samples and a set of cluster-associated gene modules. LASSO and seven additional machine-learning classifiers prioritized eight hub genes, including ALPL, CD177, S100A8, S100A9 and STXBP2. Expression analyses in separate public sepsis transcriptome cohorts supported several markers, and qRT-PCR in cecal-ligation/puncture and LPS mouse models found increased neutrophil expression of those five genes. The result links sepsis-associated neutrophil states to a small marker panel, with experimental expression validation in mice.

Technology: LASSO feature selection followed by seven machine-learning classifiers (k-nearest neighbors, linear discriminant analysis, logistic regression, naive Bayes, random forest, recursive partitioning and support-vector machine), combined with hdWGCNA and single-cell transcriptomic analysis

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Expression analyses in separate public sepsis transcriptome cohorts supported several markers, and qRT-PCR in cecal-ligation/puncture and LPS mouse models found increased neutrophil expression of those five genes.
  • Still unknown: The single-cell discovery comparison was very small (four sepsis cases and two controls) and the two clusters are analysis-derived states in this dataset; the study does not establish that they are distinct universal neutrophil lineages. The public bulk cohorts validate expression/classification associations, not an independent single-cell subtype. qRT-PCR confirmed expression changes for five genes in mouse models; it did not test whether the genes cause sepsis phenotypes or are therapeutic targets. The article itself notes earlier reports of other sepsis-related neutrophil subtypes, so this record is bounded to its specific clusters/markers and assays.
  • AI role: Machine-learning feature selection prioritized eight hub genes from expression modules associated with the two sepsis-enriched neutrophil clusters. The models were assessed on separate public training and validation transcriptome cohorts; the resulting marker set guided qRT-PCR validation of five genes in two sepsis mouse models.

Evidence sources: https://www.frontiersin.org/journals/immunology/articles/10.3389/fimmu.2024.1377817/full

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 19 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Active learning identifies a higher-activity streptavidin variant for gold-catalyzed hydroamination

An iterative active-learning campaign engineered a streptavidin artificial metalloenzyme for the known-to-the-authors gold-catalyzed hydroamination of 2-ethynylaniline to indole. Gaussian-process models guided screening across five residues in a 3.2-million-variant sequence space. The final-round library contained 38% variants more active than the prior reference, compared with 3% in the initial library. The leading variant, Sav CTNLV, showed 18-fold higher cell-specific activity than wild-type Sav and threefold higher activity than the previous reference variant in replicate testing.

Technology: Active-learning Gaussian-process regression model with optimized sequence descriptors and experimental-noise estimates

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The leading variant, Sav CTNLV, showed 18-fold higher cell-specific activity than wild-type Sav and threefold higher activity than the previous reference variant in replicate testing.
  • Still unknown: The hydroamination reaction and streptavidin–gold artificial metalloenzyme were established in earlier work by the same group; this record is specifically the later AI-guided identification of a higher-activity five-residue variant, not discovery of a new reaction. Activity was measured in an E. coli periplasmic whole-cell assay; purified variants did not retain the same rank order in vitro. Industrial utility or catalytic scale-up was not demonstrated.
  • AI role: The model used measured sequence-activity data from thousands of variants to select exploratory and exploitation libraries across five residues, then prioritized variants for experimental testing. It enabled identification of a high-activity sequence in a large combinatorial space; the paper reports considering experimental uncertainty in its rankings.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11273458/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 37 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-learning-designed peptides show dual GCGR and GLP-1R activity

An ensemble of multitask convolutional neural networks trained on 125 peptide analogues guided optimization of 15 synthesized variants. The study reports three potent dual agonists; P3 improved potency at both glucagon and GLP-1 receptors compared with the best dual agonist in the training set. These are cell-based receptor assays only. The paper reports weak performance for designing GCGR-selective peptides and establishes no clinical safety or therapeutic efficacy.

Technology: Ensemble multitask convolutional neural networks for peptide sequence optimization

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Primary paper reports the stated assay results.
  • Still unknown: Independent replication and exhaustive prior-art review not established.
  • AI role: Ensemble multitask convolutional neural networks for peptide sequence optimization

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11374683/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary publisher/PMC article and journal metadata reviewed; author affiliation confirmed from publisher article PDF. The PMC full-text open displayed a reCAPTCHA..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Ensemble machine learning identified CDK8 inhibitor hits

A six-model ensemble classified 1.6 million commercial molecules and reduced them to a 1,672-compound CDK8-focused screening library. Docking and visual inspection selected 25 candidates for enzymatic assays; six showed CDK8 inhibition. The most potent hit, T479-0984, had a reported IC50 of 99.9 nM. The paper reports low structural similarity (Tanimoto <0.300) to a comparison set of 30 diverse known CDK8 inhibitors. These are assay hits and starting points, not validated drugs or evidence of therapeutic efficacy.

Technology: Six-model ensemble classifier (logistic regression, k-nearest neighbors, support vector classification, random forest, XGBoost and multilayer perceptron)

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Twenty-five candidates from the focused library underwent CDK8 enzymatic testing; six showed inhibition. T479-0984 was reported with IC50 99.9 nM. The paper compared hits against 30 diverse known CDK8 inhibitors and reports Tanimoto similarity below 0.300.
  • Still unknown: The assay establishes in-vitro enzyme inhibition only, not cellular target engagement, animal efficacy, clinical utility, or exhaustive prior patent/publication status for the compounds.
  • AI role: The six-model ensemble filtered a 1.6-million-molecule commercial library into a 1,672-molecule CDK8-focused library; docking and visual inspection selected candidates that were then tested enzymatically.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11081523/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary PMC full text, article metadata, methods/results, author affiliations and compound-level structural comparison reviewed..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AlphaFold 3: structure and interactions of life’s molecules

It predicts joint structures of proteins, nucleic acids, small molecules and ions. DeepMind and Isomorphic Labs report at least 50% greater accuracy than existing methods for protein interactions with other molecules. Free for non-commercial research through the AlphaFold Server.

Technology: AlphaFold 3

Family AlphaFold 3: profile · models page

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Higher accuracy than docking tools for protein–ligand interactions, than nucleic-acid-specific predictors for protein–nucleic acid interactions, and than AlphaFold-Multimer v2.3 for antibody–antigen prediction.
  • Still unknown: Accuracy on new experimental targets beyond the benchmarks is not described in the abstract.
  • AI role: A diffusion-based architecture predicting joint structures of proteins, nucleic acids, small molecules, ions and modified residues.

Evidence sources: https://doi.org/10.1038/s41586-024-07487-w

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 17949 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 5/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Sperm whales: a possible ‘phonetic alphabet’ in their calls

Analysis of nearly 9,000 codas from the eastern Caribbean clan, collected over more than ten years, proposes a combinatorial structure involving rhythm, tempo, rubato and ornamentation. Published in Nature Communications; communicative function remains unknown, data cover one clan, and the page does not specify machine-learning methods.

Technology: Machine-learning analysis of whale calls

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Nearly 9,000 codas from the eastern Caribbean clan, collected over more than ten years, were analyzed for context-sensitive features (rubato, ornamentation) and categorical rhythm and tempo.
  • Still unknown: The communicative function of many codas is unknown, whether the features encode intent needs future field research, and the duality-of-patterning link is only a possibility.
  • AI role: Not reported; the page does not specify what machine learning did in the analysis.

Evidence sources: https://www.projectceti.org/blog-posts/sperm-whale-phonetic-alphabet-proposed-for-the-first-time

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 96 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

HydrogelFinder: A Foundation Model for Efficient Self-Assembling Peptide Discovery Guided by Non-Peptidal Small Molecules

HydrogelFinder used a generative model and learned prediction to prioritize 111 candidate peptides; 17 were synthesized and nine self-assembled under tested conditions. Cell assays covered two selected materials and do not establish general safety or clinical suitability.

Technology: HydrogelFinder generative language model and learned gelation prediction

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Primary abstract reports 111 candidates, 17 synthesized and nine experimentally self-assembling; cell compatibility tests were limited to two selected materials.
  • Still unknown: Independent replication and exhaustive prior-art review are not established.
  • AI role: HydrogelFinder generative language model and learned gelation prediction

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/38704695/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary source checked for study identity, AI operation, dates, affiliations and the bounded result..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

GPT-3.5 reasoning helps connect MRE11 to social-context sleep regulation in Drosophila

A study combined GPT-3.5-generated gene-function context with a genome-wide Drosophila RNAi behavioral screen. It identified MRE11 knockdown as reducing sleep in group-housed flies, a result the authors report as previously unrecognized in this social context. GPT-3.5 reasoning proposed a link from MRE11 to Dop1R1/dopaminergic signaling; treatment with a Dop1R1 antagonist partially rescued sleep, locomotor and social-activity phenotypes in MRE11-RNAi flies.

Technology: GPT-3.5, prompted for genome-wide Drosophila gene-function interpretation and chain-of-thought relationship hypotheses

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: It identified MRE11 knockdown as reducing sleep in group-housed flies, a result the authors report as previously unrecognized in this social context.
  • Still unknown: The experiments are in Drosophila and concern sleep in the presence of conspecifics. Dop1R1 antagonist treatment only partially rescued phenotypes; this does not establish direct MRE11 regulation of the receptor or dopamine pathway. GPT-3.5 had high false-negative rates in the authors’ known-gene evaluation and contributed knowledge summarization/hypothesis generation, while the broad RNAi screen itself was experimental. No mammalian or human result is implied.
  • AI role: GPT-3.5 was queried about fly gene functions and used to construct hypothesized relationships among behavioral-screen candidates; it proposed a MRE11–Dop1R1 link that researchers tested pharmacologically. The RNAi screen and behavioral measurements were experimental, while direct molecular regulation was not demonstrated.

Evidence sources: https://www.nature.com/articles/s41467-024-48005-w

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 26 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Profluent

OpenCRISPR-1: a genome editor designed by language models

Profluent built an atlas of 5.1 million CRISPR-Cas proteins and generated a Cas9-like editor with 403 mutations relative to SpCas9. Of 48 generated sequences, the best (OpenCRISPR-1) achieved 55.7% on-target editing versus 48.3% for SpCas9, and 0.32% off-target editing versus 6.1%. Human cells in the laboratory only; company data; the page does not indicate peer review.

Technology: Protein language models

Family OpenCRISPR: profile · models page

Publication: Publication status unspecified in annotation · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Not reported
  • Test type: direct test
  • Tested: Of 48 selected generated sequences, the best (OpenCRISPR-1) was tested in human cells: 55.7% on-target editing vs 48.3% for SpCas9, and 0.32% off-target vs 6.1%.
  • Still unknown: The company article has no explicit limitations section and does not state whether a preprint or peer review exists.
  • AI role: A language model pretrained on protein data and steered to CRISPR families using a 5.1-million-protein atlas generated Cas9-like nucleases and compatible sgRNAs.

Evidence sources: https://www.profluent.bio/media/editing-the-human-genome-with-ai

The legacy label “Not reported” does not distinguish information absent from reviewed material from information not assessed. It is not a scientific finding of absence.

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 5/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Structure-based virtual screening identifies mammalian enzymes catalyzing the second step of carnitine biosynthesis

The OSMES workflow screened AlphaFold-modeled vitamin-B6-dependent enzymes by docking the substrate into active sites and ranking catalytically favorable conformations. It identified human SHMT1/2 as enzymes with hydroxytrimethyllysine-aldolase activity, validated in vitro; mouse Tha1 showed much higher catalytic efficiency, supporting a species difference in the pathway. The authors infer that humans may have lost a dedicated gene, with SHMT activity partly compensating.

Technology: OSMES (One Substrate–Many Enzymes Screening), automated structure-based docking/ranking workflow using AlphaFold models

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: It identified human SHMT1/2 as enzymes with hydroxytrimethyllysine-aldolase activity, validated in vitro; mouse Tha1 showed much higher catalytic efficiency, supporting a species difference in the pathway.
  • Still unknown: Enzyme activity was validated in vitro. The claim about loss of a dedicated human gene is an evolutionary inference, not a direct gene-knockout demonstration; the authors note unresolved pathway questions.
  • AI role: Ranked candidate enzymes for a specified metabolic reaction using docked substrate conformations and catalytically favorable active-site geometry; guided biochemical testing that verified SHMT1/2 activity.

Evidence sources: https://www.nature.com/articles/s41467-024-47466-3.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 87 of works of the same type, year and subfield (OpenAlex work metrics, 13 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Progeni predicts candidate melanoma and colorectal-cancer targets validated in cell models

Progeni combines a probabilistic knowledge graph with graph neural networks to rank disease-associated molecular targets. The authors selected six candidates lacking literature support and reported that knockdown of HSP90AB1 and MME reduced proliferation in a murine melanoma line, while knockdown of ADCY5, ADRA2A and EEF2 reduced proliferation in a murine colorectal-cancer line.

Technology: Progeni probabilistic knowledge graph and graph neural network

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The authors selected six candidates lacking literature support and reported that knockdown of HSP90AB1 and MME reduced proliferation in a murine melanoma line, while knockdown of ADCY5, ADRA2A and EEF2 reduced proliferation in a murine colorectal-cancer line.
  • Still unknown: The functional validations were in B16F10 and MC38 murine cell lines, not human tumor samples or clinical settings. The paper evaluates a model and validates selected candidates; it does not establish these as clinically effective therapeutic targets.
  • AI role: The graph neural network ranked new target hypotheses from heterogeneous biological networks and literature co-occurrence evidence. Researchers then tested selected predictions by shRNA knockdown in representative murine cancer cell lines.

Evidence sources: https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1011945

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 81 of works of the same type, year and subfield (OpenAlex work metrics, 11 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

DeepEvo yields heat-tolerant G3PDH variants

A generative model and protein-language-model-based temperature selector were combined in an iterative design procedure. The article reports eight heat-tolerant glyceraldehyde-3-phosphate dehydrogenase variants among 30 generated and tested sequences. This is an experimentally screened enzyme set, not proof of industrial deployment.

Technology: DeepEvo generative model; protein language model temperature selector

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Publisher article reports eight heat-tolerant variants among 30 generated sequences tested.
  • Still unknown: Exact assay definition, earlier disclosure and sequence-level prior overlap remain open.
  • AI role: Generative model proposes functional sequences and PLM-based selector ranks high-temperature candidates in iterative design/test loop.

Evidence sources: https://spj.science.org/doi/10.34133/bdr.0031

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary source checked for identity, date, model role and bounded reported result..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Tree-based machine learning nominates Parabacteroides merdae and N-acetyl-D-mannosamine as ulcerative-colitis alleviators in a mouse model

The researchers combined personalized gut-microbiome community metabolic models from human iHMP data with tree-based machine-learning feature selection to identify microbial strains and metabolites associated with ulcerative-colitis samples. Parabacteroides merdae ATCC 43,184 and N-acetyl-D-mannosamine were among the most predictive features; the strain was also the main modeled microbial contributor to the metabolite. In a DSS-induced mouse colitis experiment, oral administration of the bacterium or metabolite improved reported colitis measures.

Technology: Tree-based supervised classifiers including random forest, gradient-boosted regression trees, XGBoost, XGBoost-random forest and LightGBM, combined with individualized genome-scale microbiome community models

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: In a DSS-induced mouse colitis experiment, oral administration of the bacterium or metabolite improved reported colitis measures.
  • Still unknown: The human data are observational microbiome samples, while intervention was tested in a DSS-induced mouse model, not in people. The animal groups contained eight mice each; findings do not establish treatment efficacy or safety in human ulcerative colitis. Metabolite-production associations came from community-model simulations and were not direct evidence of individual strain production in each person. The model provides candidate prioritization, not a causal human mechanism.
  • AI role: Personalized community metabolic models estimated strain-level contributions to secreted metabolites in human iHMP samples. Tree-based classifiers selected discriminative strain and metabolite features; XGBRF feature attribution ranked P. merdae ATCC 43,184 and N-acetyl-D-mannosamine highly, motivating the mouse validation. The AI analysis nominated these targets; it did not establish human therapeutic efficacy.

Evidence sources: https://www.tandfonline.com/doi/full/10.1080/19490976.2024.2336877

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 88 of works of the same type, year and subfield (OpenAlex work metrics, 19 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Deep-learning-assisted DIP-MS identifies POLR2E as a constitutive prefoldin-like-complex subunit

DIP-MS combines deep-learning-based signal processing with native-gel fractionation and mass spectrometry to resolve complex isoforms. In human prefoldin-like-complex experiments, repeated coelution and affinity-purification data supported POLR2E as a constitutive PFDL component, updating the view that it was merely an associated protein.

Technology: Deep Interactome Profiling by Mass Spectrometry (DIP-MS), including deep-learning-based signal processing

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: In human prefoldin-like-complex experiments, repeated coelution and affinity-purification data supported POLR2E as a constitutive PFDL component, updating the view that it was merely an associated protein.
  • Still unknown: The paper proposes POLR2E as a constitutive PFDL subunit based on coelution and affinity-purification in the tested human-cell system; it does not establish that every cellular PFDL complex contains POLR2E or fully characterize its functional contribution. DIP-MS also recapitulated known ASDURF association, which is not counted here as novel because ASDURF’s PAQosome/PFDL role had been reported earlier.
  • AI role: Deep-learning signal processing helped resolve coexisting protein-complex isoforms from AP-MS/native PAGE/DIA-MS data. DIP-MS on UXT and PFDN2 baits repeatedly detected PFDL coelution with POLR2E; affinity purification of four tagged PFDL-like subunits also recovered POLR2E, supporting the proposed constitutive association.

Evidence sources: https://www.nature.com/articles/s41592-024-02211-y

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 32 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

A biophysics-informed protein language model designs functional GFP variants

Mutational Effect Transfer Learning (METL) combines transformer pretraining on simulated protein biophysics with fine-tuning on experimental sequence–function data. Using only 64 experimental GFP examples, the authors designed 20 previously untested variants and measured their fluorescence; 16 showed measurable fluorescence, although none was brighter than wild-type GFP. The result demonstrates functional design under a small-data setup, not improved brightness over wild type. The result was present in the bioRxiv version posted 2024-03-17; the GEM@ICLR 2024 workshop paper provides the detailed 20-variant assay report, and the later journal version is the same study.

Technology: Mutational Effect Transfer Learning (METL), a transformer protein language model pretrained on simulated biophysical features and fine-tuned on experimental data

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: Twenty METL-designed GFP variants were synthesized and experimentally measured; 16 showed measurable fluorescence, but none exceeded wild-type brightness.
  • Still unknown: The panel contains 20 designed variants in a specific GFP task; the result does not show improvement over wild-type brightness or establish general performance across protein families.
  • AI role: METL-Local was pretrained on simulated GFP sequence/biophysical data, fine-tuned using 64 experimental examples, and used with sequence optimization to select variants for wet-lab testing.

Evidence sources: https://www.biorxiv.org/content/10.1101/2024.03.15.585128v1 · https://openreview.net/pdf?id=hWISqnHBr7

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary bioRxiv v1 indexed abstract, posting date and affiliation metadata; primary GEM@ICLR workshop paper full text with methods and experimental outcomes..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

MolE prioritizes existing drugs with measured Staphylococcus aureus growth effects

MolE-XGBoost prioritized existing compounds. Of six tested commercially available compounds, elvitegravir, opicapone and ebastine showed measured growth-inhibitory effects on S. aureus; visomitin was a known-antimicrobial positive control. Effects varied: elvitegravir inhibited growth at 8 μg/mL, while the other two limited or delayed growth. This is an in-vitro finding, not a clinical antibiotic.

Technology: MolE molecular representation learning; XGBoost

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: The primary source reports the study-specific result summarized here.
  • Still unknown: Independent replication and broad deployment are not established by this review.
  • AI role: A learned molecular representation and XGBoost rank candidates for experimental screening; selected compounds undergo MIC and growth-curve assays.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11986102/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary-source indexed article or abstract and bibliographic metadata reviewed; complete source scope is given in source_checks..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning guides engineering of an alkaloid-biosynthesis enzyme

Researchers developed MutComputeX, a structure-based residual neural network for assessing amino-acid substitutions in protein–ligand environments, and used it with a 4′-O-methylnorbelladine biosensor to engineer the plant methyltransferase Nb4OMT in Escherichia coli. The paper reports engineered variants with 60% higher product titer, twofold higher catalytic activity, threefold lower off-product regioisomer formation and 5 °C higher thermostability; ligand-aware predictions helped identify the beneficial A53M substitution. These enzyme and product results do not demonstrate complete microbial production of galantamine.

Technology: MutComputeX structure-based residual neural network for protein–ligand mutation design

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: MutComputeX-guided, manually curated Nb4OMT variants were assayed using a 4NB biosensor and biochemical methods. The article reports 60% higher product titer, twofold higher catalytic activity, threefold lower off-product formation and 5 °C higher thermostability for engineered variants.
  • Still unknown: The result concerns a pathway enzyme and branchpoint intermediate in the reported microbial/biochemical systems; full pathway production of galantamine and industrial manufacturing were not demonstrated.
  • AI role: MutComputeX predicted sequence substitutions in a modeled protein–ligand microenvironment and guided selection of Nb4OMT variants for biosensor screening and enzyme assays; ligand-aware predictions supported A53M.

Evidence sources: https://www.nature.com/articles/s41467-024-46356-y

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary Nature Communications full text as indexed, including methods/results, author contributions and affiliation block..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

An ML-generated MinE homolog restores bacterial cell division patterning in vivo

A machine-learning-generated MinE homolog, synMinEv25, restored Min oscillations and wild-type-like cell growth and morphology when substituted for the native minE gene in an E. coli strain lacking minDE. Researchers generated variants with a multiple-sequence-alignment variational autoencoder, used AlphaFold2-based structural scoring to select candidates, and screened them in synthetic-cell mimics before in-vivo testing. The validated variant is a distant homolog with less than 50% sequence identity to E. coli MinE.

Technology: Multiple-sequence-alignment variational autoencoder (MSA-VAE), AlphaFold2 Multimer, and a structure-based divide-and-conquer scoring pipeline

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Researchers generated variants with a multiple-sequence-alignment variational autoencoder, used AlphaFold2-based structural scoring to select candidates, and screened them in synthetic-cell mimics before in-vivo testing.
  • Still unknown: This is a proof-of-principle result in one engineered E. coli MinDE system, not a general demonstration that generated proteins can replace natural genes. The paper notes weak correlation between several in-silico scores and individual measured protein properties; it reports one best-performing variant and does not establish applications outside this assay.
  • AI role: The MSA-VAE generated MinE homolog sequences; structure-based scoring estimated membrane binding, MinD complex formation and dimerization to prioritize 48 candidates from 4,000 generated sequences. Cell-free synthetic-cell assays identified 14 variants with Min oscillations, and researchers tested these in E. coli; synMinEv25 restored oscillations, growth and morphology. AI generated and prioritized candidates, while wet-lab experiments established function.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC10914801/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 23 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Enzymit Ltd.

CoSaNN designs and tests an induced-fit carbohydrate kinase

CoSaNN combined neural structure prediction, learned sequence optimization and solubility ranking to design variants of an induced-fit ROK carbohydrate kinase. The authors measured expression, thermal stability and activity in vitro. The paper reports measurable expression for 189 of 348 designs in one table, while the strategy-level counts do not sum to that total; no success-rate claim is made here.

Technology: CoSaNN: AlphaFold2-based conformation modeling, ProteinMPNN sequence design and SolvIT graph neural network ranking

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The authors applied CoSaNN to redesign the Arthrobacter polyphosphate glucokinase ROK kinase and performed expression, thermal-stability and enzyme-activity assays on designed variants. The primary Results table reports 189 designs with detectable soluble expression among 348 listed designs; its strategy counts (119+95+94 plus 40 in another strategy; rows in the prose/table are internally inconsistent) do not reconcile cleanly, so the record does not use a single total as a success-rate denominator.
  • Still unknown: The study tests a company-designed enzyme set in vitro; it does not establish industrial deployment. The primary paper’s count inconsistency remains explicitly disclosed.
  • AI role: CoSaNN uses deep-learning structure prediction (AlphaFold2), sequence optimization with ProteinMPNN and a learned solubility classifier (SolvIT) to rank enzyme designs for experimental characterization.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC10945820/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary PMC full text and PNAS DOI record for CoSaNN-guided glucokinase design, Enzymit affiliations, experimental activity results and the discrepancy in reported design totals..

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning and genetic search design antimicrobial peptides

Regression and k-nearest-neighbor models with a genetic algorithm guided optimization from a shrimp lipopolysaccharide-binding peptide to LBD_A–D. The article reports in-vitro tests, NMR structure for LBD_B and improved survival in a kuruma prawn co-injection experiment. This is not human therapy and does not establish treatment efficacy after infection.

Technology: Regression and k-nearest-neighbor models; genetic algorithm

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Article reports synthesized/tested peptide variants, NMR characterization and a prawn co-injection assay.
  • Still unknown: Full exact assay/organism scope and peptide sequence novelty/independent replication not exhaustively compared.
  • AI role: Activity models and genetic search guided sequence variants which were synthesized and biologically tested.

Evidence sources: https://doi.org/10.1016/j.jare.2024.02.016 · https://pubmed.ncbi.nlm.nih.gov/38431124/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary source checked for identity, date, model role and bounded reported result..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Gaussian-process-assisted engineering of a ketoreductase for an ipatasertib precursor

A six-round engineering campaign combined mutational scans, structure-guided libraries and Gaussian-process models. The model-aided library yielded M3, reported with 22-fold higher apparent activity than wild type in the UV assay. Three subsequent conventional mutagenesis rounds produced M6, a 10-substitution descendant with 64-fold higher apparent kcat, but also about 12-fold higher apparent KM; do not attribute M6 as a direct model output or treat kcat as production yield. The authors report a 100 g/L laboratory process test with ≥98% conversion and 99.7% diastereomeric excess after 30 hours. These are process-relevant laboratory results, not a commercial production run or drug efficacy.

Technology: Gaussian-process regression

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: M3 and later M6 were tested for activity; article reports preparative reactions at 100 g/L substrate and conversion/stereoselectivity under specified laboratory conditions.
  • Still unknown: M6 follows conventional site-mutagenesis rounds after the AI-informed step; apparent kcat differs from process yield and KM changed; no commercial production/clinical efficacy claim.
  • AI role: Gaussian-process predictions informed smaller follow-up variant libraries within a broader workflow that also used mutational scanning, rational design and conventional evolution.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC10902378/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary PMC full article, methods/results, publication history, affiliations and chronology of model and conventional rounds..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ModelAngelo: automated atomic model building in cryo-EM maps

ModelAngelo combines cryo-EM density, protein sequence and local geometry to build atomic models. Beyond method benchmarks, the paper identifies six previously unassigned proteins in a photosynthetic supercomplex and four additional radial-spoke proteins plus two central-apparatus proteins in Chlamydomonas reinhardtii. These assignments use existing experimental cryo-EM maps and sequence searches; the paper leaves the newly assigned ciliary proteins' functional roles for further investigation.

Technology: ModelAngelo graph neural network with hidden Markov model sequence search

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Models built by ModelAngelo were compared with those of human experts, and its protein identification was compared with human experts on proteins of unknown sequence.
  • Still unknown: Not reported in the abstract read; the nucleotide result is limited to backbone accuracy.
  • AI role: A graph neural network combining cryo-EM density, protein sequence and local geometry to build atomic models and identify proteins.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-024-07215-4&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 713 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

SLIDE uncovers a keratinocyte-centered signal in systemic sclerosis

An interpretable machine-learning analysis of multi-omic systemic-sclerosis data identified an unexplored keratinocyte-centered signature, validated by protein staining, and a previously unreported interaction between myeloid-cell and fibroblast states and macrophage HLA signaling. The same study also analyzed CD4 T-cell clonal expansion, but this record treats the paper's disease-mechanism findings as one discovery.

Technology: SLIDE (Significant Latent Factor Interaction Discovery and Exploration)

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: An interpretable machine-learning analysis of multi-omic systemic-sclerosis data identified an unexplored keratinocyte-centered signature, validated by protein staining, and a previously unreported interaction between myeloid-cell and fibroblast states and macrophage HLA signaling.
  • Still unknown: The paper reports computationally identified relationships and validation of one signature by protein staining; these findings do not by themselves establish causal mechanisms or clinical utility.
  • AI role: SLIDE integrated single-cell and spatial multi-omic data, identified interacting latent factors associated with systemic-sclerosis disease severity, and surfaced candidate mechanisms; the keratinocyte signal was checked with protein staining by the study team.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11588359/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 37 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-learning models identify carbon-source-dependent metabolic reactions that control Escherichia coli growth

Elastic-net and multilayer-perceptron models integrated genome-wide gene-deletion data and simulated fluxes to predict beneficial and detrimental reactions for E. coli growth on 30 carbon sources. Beyond training-set examples, the authors experimentally validated model predictions about the glyoxylate shunt, pyruvate dehydrogenase, and redundant purine-biosynthesis reactions. The findings showed that effects of energy-generating pathways depend on carbon source, while biosynthetic pathways were more generally growth-promoting across the tested conditions.

Technology: Elastic-net regression and multilayer-perceptron models

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Beyond training-set examples, the authors experimentally validated model predictions about the glyoxylate shunt, pyruvate dehydrogenase, and redundant purine-biosynthesis reactions.
  • Still unknown: Reported reaction effects are for E. coli under the tested carbon-source and genetic-perturbation conditions; they do not establish universal metabolic-control rules. The results should not be interpreted as direct clinical or industrial outcomes. Primary PubMed abstract and PMC article record checked; no independent replication assessed.
  • AI role: The models combined gene-deletion phenotypes with simulated metabolic fluxes to prioritize reactions whose perturbation would benefit or harm growth across carbon conditions; selected out-of-training-set predictions were experimentally tested.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC10912204/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 85 of works of the same type, year and subfield (OpenAlex work metrics, 16 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AlphaFold-Multimer-guided experiments corroborate six novel human protein interfaces

Researchers developed a fragmentation strategy for AlphaFold-Multimer and applied it to 62 human protein interactions associated with neurological disorders. Eighteen predicted structures were judged correct or likely correct, and experiments corroborated six previously unreported interfaces: FBXO23–STX1B, STX1B–VAMP2, ESRRG–PSMC5, PEX3–PEX19, PEX3–PEX16 and SNRPB–GIGYF1.

Technology: AlphaFold-Multimer with a protein-fragmentation strategy

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Eighteen predicted structures were judged correct or likely correct, and experiments corroborated six previously unreported interfaces: FBXO23–STX1B, STX1B–VAMP2, ESRRG–PSMC5, PEX3–PEX19, PEX3–PEX16 and SNRPB–GIGYF1.
  • Still unknown: The article reports 18 correct or likely correct models among 62 interactions and experimental support for six interfaces; these are not a comprehensive interactome map. AlphaFold-Multimer showed limited specificity and was sensitive to input-fragment length, as the source explicitly notes.
  • AI role: AlphaFold-Multimer predicted domain–motif interface structures; researchers tested predictions and used the model-informed fragmentation strategy to prioritize interaction interfaces. Wet-lab assays corroborated six novel contacts, while the tool had limited specificity and performance declined with longer/full-length input sequences.

Evidence sources: https://link.springer.com/article/10.1038/s44320-023-00005-6

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 119 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Designed proteins bind peptides with the highest affinity reported

Using RFdiffusion and ProteinMPNN, researchers designed proteins that bind helical peptides such as glucagon, neuropeptide Y and parathyroid hormone, with what the authors consider the highest reported affinity for a computer-generated biomolecule. They detect low concentrations of peptides in human serum and withstand heat; a parathyroid-hormone biosensor produced a 21-fold increase in bioluminescence. The press release gives no affinity values.

Technology: RFdiffusion and ProteinMPNN

Family RFdiffusion: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Designed binders were tested in the lab for peptide detection (mass spectrometry, including low-concentration peptide detection) and a bioluminescence biosensor produced a 21-fold signal increase with parathyroid hormone; the press release gives no affinity values.
  • Still unknown: The abstract does not describe how affinity was measured, does not mention serum testing, and states no limitations; the press release states no limitations.
  • AI role: RFdiffusion, extended for flexible targets and partial diffusion, designed binders to helical peptides, de novo or by refining other designs.

Evidence sources: https://www.ipd.uw.edu/2023/12/ai-generates-proteins-with-exceptional-binding-strengths/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-023-06953-1&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 290 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AI-assisted hippocampal atlas maps Alzheimer-associated cellular changes around amyloid plaques

In a six-donor postmortem hippocampal study, neural cell segmentation and Spatial-ID annotation helped map Alzheimer-associated cell distributions and gene-expression changes. Immunofluorescence supported the inferred accumulation patterns of microglia and astrocytes around amyloid-beta plaques. These are observational associations, not causal effects or clinical benefit. This record describes the December 7, 2023 preprint; the later Neuron paper was published online March 31, 2025 with an expanded cohort, whose counts and final methods are not pooled here.

Technology: StereoCell/ESPANet cell segmentation; Spatial-ID deep neural network and graph convolutional network

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: Spatial transcriptomics and single-nucleus sequencing in three AD and three control donor brains; immunofluorescence checked the plaque-associated glial distribution pattern within the same study.
  • Still unknown: Postmortem observational results do not establish disease causation. No independent replication was assessed. Preprint methods retain an unfinished training-epoch placeholder. The reviewed claims and AI architectures refer to the primary preprint, not authenticated final journal methods; candidate blood biomarkers and diagnostic efficacy are outside this record's represented claim.
  • AI role: Trained neural segmentation delineated cells; Spatial-ID used a DNN and GCN to assign spatial cell types, enabling plaque-associated cell-distribution analysis.

Evidence sources: https://www.medrxiv.org/content/10.1101/2023.12.06.23299593v1.full.pdf · https://www.medrxiv.org/content/10.1101/2023.12.06.23299593v1.full · https://pubmed.ncbi.nlm.nih.gov/40168986/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

3D-HiChAT predicts embryonic enhancer hubs validated by CRISPR interference

A random-forest model integrating ATAC-seq, H3K27ac ChIP-seq and HiChIP data predicted enhancer effects across mouse trophoblast stem cells, embryonic stem cells and extraembryonic endoderm cells. CRISPRi tests supported a shared Enh14 hub regulating both Tfcp2l1 and Gli2 in embryonic stem cells, and an Enh4 hub linked to seven genes in extraembryonic endoderm; four of those seven showed expression reductions as predicted. Across 40 tested enhancer-promoter pairs, experiments yielded 12 true positives and 13 true negatives, with overall prediction accuracy 0.71.

Technology: 3D-HiChAT, an optimized Random Forest model combining chromatin accessibility, H3K27ac and 3D HiChIP connectivity features

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: CRISPRi tests supported a shared Enh14 hub regulating both Tfcp2l1 and Gli2 in embryonic stem cells, and an Enh4 hub linked to seven genes in extraembryonic endoderm; four of those seven showed expression reductions as predicted.
  • Still unknown: The experiments were performed in mouse-derived cell lines, not intact embryos or human cells. Only selected predictions were tested; the 40-pair validation had 12 true positives, 13 true negatives and an overall reported accuracy of 0.71, leaving discordant/uncertain pairs. The Enh14 and Enh4 results support specific cell-type-dependent regulatory hubs but do not establish the full genome-wide predicted map as experimentally validated or prove developmental phenotypes in vivo.
  • AI role: The model used 1D and 3D chromatin features to rank genome-wide enhancer-promoter pairs by predicted expression impact after in-silico enhancer perturbation. Researchers selected predictions for CRISPRi validation and measured target-gene expression by RT-qPCR; the paper reports specific validated multi-gene enhancer hubs and positive/negative prediction rates across 40 pairs.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC10897904/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 51 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

A recurrent neural network uncovers a non-canonical human GRHL1 DNA-binding site

A convolutional recurrent neural network found 46 GRHL1 ChIP-seq sites missed by a conventional position-weight matrix, including sequences lacking the previously considered obligatory CNNG motif. Isothermal titration calorimetry confirmed GRHL1 binding to a selected non-canonical sequence lacking that core motif; nucleotide mutagenesis experiments also showed that model scores tracked measured affinity across tested variants.

Technology: Convolutional recurrent neural network trained on GRHL1 HT-SELEX binding data

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Isothermal titration calorimetry confirmed GRHL1 binding to a selected non-canonical sequence lacking that core motif; nucleotide mutagenesis experiments also showed that model scores tracked measured affinity across tested variants.
  • Still unknown: Only one of the 46 predicted non-canonical sites was directly validated by ITC in this study; 46 is a model-nominated set, not 46 experimentally confirmed bindings. Biochemical affinity was tested with the isolated GRHL1 DNA-binding domain and synthetic DNA, not in vivo transcriptional effects for the novel site.
  • AI role: The CNN/RNN learned GRHL1 sequence-binding patterns from HT-SELEX and scored human-cell ChIP-seq peaks. It identified 46 sites not detected by FIMO/PWM; one selected high-scoring non-canonical site was experimentally confirmed by ITC, and predicted variant affinity changes were compared with ITC results.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC10696883/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 31 of works of the same type, year and subfield (OpenAlex work metrics, 2 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning identifies bacterial strains and combinations that reduce pathogen colonization in Arabidopsis

Random forest and elastic-net models analyzed 136 synthetic communities, each composed of five bacterial strains, and identified strain identity as the strongest predictor of protection against Pseudomonas syringae DC3000 in Arabidopsis. Independent test communities confirmed predictive performance, and follow-up plant assays showed that Acidovorax Leaf76, Rhizobium Leaf68 and Pseudomonas Leaf15 each reduced pathogen colonization by about two orders of magnitude versus axenic controls; a Rhizobium Leaf371–Arthrobacter Leaf337 combination also reduced colonization.

Technology: Random forest and elastic-net generalized linear models (GLMNet)

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Independent test communities confirmed predictive performance, and follow-up plant assays showed that Acidovorax Leaf76, Rhizobium Leaf68 and Pseudomonas Leaf15 each reduced pathogen colonization by about two orders of magnitude versus axenic controls; a Rhizobium Leaf371–Arthrobacter Leaf337 combination also reduced colonization.
  • Still unknown: The experiments used gnotobiotic Arabidopsis thaliana and P. syringae DC3000, not field crops. Validation was in the same study, with 24 plants per treatment in the key individual-strain validation assay. The work establishes protection in this model system; it does not establish the underlying mechanisms, efficacy in other plants or field performance.
  • AI role: Models predicted pathogen outcomes from strain presence or abundance across the 136 screened communities and ranked strains by feature importance. Researchers tested the leading strains and combinations in additional plant experiments, validating the model-prioritized protection results.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/38042924/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 131 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Chroma: a generative model for protein design

Generate:Biomedicines published the Chroma model in Nature; it designs new proteins with requested biophysical, biological and therapeutic properties. The weights are freely available to academic and nonprofit researchers. The press release gives no experimental figures, and performance claims come from the company.

Technology: Chroma (generative model)

Family Chroma: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The Nature paper reports experimental characterization of 310 designed proteins and crystal structures of two designs matching the predicted structures (backbone RMSD about 1.0 Å); the press release itself describes no experiments and states no limitations.
  • Still unknown: The paper abstract states no limitations; the press release gives no experimental figures and its performance claims come from the company.
  • AI role: A generative model samples protein structures and sequences and can be conditioned on specified properties or functions.

Evidence sources: https://generatebiomedicines.com/news-releases/generatebiomedicines-announces-publication-of-its-chroma-model-in-nature/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-023-06728-8&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 585 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Multi-omics models guide experimental validation of flowering-time genes

Researchers trained models on genomic, transcriptomic and methylomic data for Arabidopsis traits and used feature interpretation to select flowering-time gene candidates. The preprint reports nine experimentally validated genes inferred to contribute to flowering time, tested using mutant phenotypes. Random-forest and ridge-regression models were used, so do not assign every gene to the random forest alone. Count one study, not nine discoveries. This draft follows the primary 2023 preprint version; the journal paper appeared in 2024.

Technology: Random-forest models; ridge regression; multi-omics integration

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: The preprint reports experimental validation of nine flowering-time gene candidates using single- and/or double-mutant phenotypes.
  • Still unknown: Individual candidate-selection routes are not all attributed to one model; the result is from a preprint version and does not establish independent replication.
  • AI role: Random-forest and ridge-regression models used multi-omics data to predict plant traits and identify features associated with flowering-time predictions.

Evidence sources: https://storage.prod.researchhub.com/uploads/papers/2023/11/15/2023.11.14.566971.full.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary bioRxiv v1 full text, posted-date line, abstract, methods and author affiliations..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Deep-learning-guided cell-free screening identifies antimicrobial peptides

Generative and predictive deep-learning models proposed and prioritized peptide sequences for cell-free synthesis and screening. The authors report 30 functional antimicrobial peptides among 500 candidates tested across five design–build–test rounds. The assay tested peptides in unpurified cell-free mixtures, a limitation for interpreting activity and potency. The reported effects do not establish clinical efficacy or antibiotic approval.

Technology: Variational autoencoders; convolutional and recurrent neural-network regressors

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: 500 model-prioritized candidates were produced and screened; 30 showed functional antimicrobial activity in the reported assays.
  • Still unknown: Unpurified CFPS assay mixtures, animal/human efficacy and independent replication limit interpretation.
  • AI role: Generative models proposed peptide sequences and learned regressors prioritized candidates before cell-free synthesis and activity tests.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/37938588/ · https://pmc.ncbi.nlm.nih.gov/articles/PMC10632401/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary PubMed record, abstract and figure captions; indexed full-text results for assay conditions..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Meta AI

Meta AI: near-real-time image decoding from brain activity

A three-part system (image encoder, brain encoder, decoder) reconstructs viewed images from magnetoencephalography signals at each time point, with signals better aligned to DINOv2 than the other models compared. Images remain imperfect; fMRI decodes better; the page gives no numerical metrics; preprint.

Technology: Magnetoencephalography decoder

Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: retrospective or benchmark
  • Tested: MEG recordings from healthy volunteers viewing images were decoded at each time point; several pretrained image models were compared (signals aligned best with DINOv2), and MEG decoding was compared with fMRI. Participant and image counts not stated on the blog.
  • Still unknown: Generated images are imperfect: high-level features are often kept but low-level features are often wrong. MEG decoding is less precise than fMRI decoding.
  • AI role: AI models built the image representations and generated the output images; volunteers supplied the brain recordings.

Evidence sources: https://ai.meta.com/blog/brain-ai-image-decoding-meg-magnetoencephalography/ · https://arxiv.org/abs/2310.19812

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-learning-guided screening discovers potent innate immune signaling modulators

At the University of Chicago, an active-learning pipeline iteratively selected molecules from a 139,998-compound library for in-vitro assays of NF-κB and interferon regulatory factor (IRF) signaling. After screening about 2% of the library, the study reported compounds suppressing NF-κB up to 15-fold or elevating NF-κB and IRF up to 5-fold and 6-fold, respectively. Follow-up cytokine assays found PME-4007 increased IFN-β secretion about threefold when co-delivered with the STING agonist cGAMP.

Technology: Machine-learning active-learning pipeline with variational-autoencoder molecular representations, Gaussian-process regression and Bayesian optimization

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: At the University of Chicago, an active-learning pipeline iteratively selected molecules from a 139,998-compound library for in-vitro assays of NF-κB and interferon regulatory factor (IRF) signaling.
  • Still unknown: The work reports in-vitro pathway modulation and primary-cell cytokine assays, not vaccination, immunotherapy efficacy, animal outcomes or clinical benefit. Fold changes are study-specific relative to agonist-only controls. A subset of candidates was further characterized; four tested compounds were flagged as non-viable by the cytostasis/cytotoxicity confluency mask despite strong pathway effects.
  • AI role: Models were iteratively trained on screening measurements and used Bayesian optimization to select subsequent compounds for in-vitro testing. This model-guided selection led to the reported NF-κB/IRF modulators; follow-up cytokine release assays tested a subset.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/38020385/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 86 of works of the same type, year and subfield (OpenAlex work metrics, 21 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

EVEscape: predicting viral variants before they emerge

Starting from January 2020, the tool predicted SARS-CoV-2 mutations with accuracy similar to experimental antibody-binding assays and ranked their prevalence better; it generalizes to HIV and influenza. Retrospective validation; prospective performance is not reported.

Technology: EVEscape

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: EVEscape was trained only on pre-2020 sequences and compared against high-throughput experimental scans of SARS-CoV-2 variation that later emerged, a retrospective validation; prospective performance is not reported.
  • Still unknown: The abstract states no limitations; the page says testing on Lassa and Nipah is ongoing and vaccine or therapy use against future variants is a future application.
  • AI role: A deep learning model of historical viral sequences, combined with biophysical and structural information, scored the escape potential of mutations.

Evidence sources: https://hms.harvard.edu/news/ai-tool-can-help-forecast-viral-outbreaks · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-023-06617-0&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 245 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

AlphaMissense: cataloging missense genetic mutations

It classified 89% of the 71 million possible missense variants: 32% as likely pathogenic and 57% as likely benign, with 90% precision on a database of known variants. Predictions are not intended for direct clinical use.

Technology: AlphaMissense

Family AlphaMissense: profile · models page

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: The model was validated on genetic and experimental benchmarks, and Genomics England cross-referenced its predictions with previously aggregated human variant pathogenicity data and found them consistent. The reported comparison uses previously collected pathogenicity and variant data; an independent data source is distinct from a new independent experimental replication.
  • Still unknown: The model does not predict structural or stability effects, the blog says predictions are not designed for direct clinical use, and the abstract describes no clinical validation.
  • AI role: AlphaMissense, an AlphaFold adaptation fine-tuned on human and primate variant population frequency data, classified missense variants and released predictions for all possible human substitutions.

Evidence sources: https://deepmind.google/blog/a-catalogue-of-genetic-mutations-to-help-pinpoint-the-cause-of-diseases/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1126/science.adg7492&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 2381 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AI-assisted protein-family analysis identifies a tested toxin-antitoxin superfamily

AI-predicted structural information helped researchers prioritize previously poorly characterized protein families. Experiments supported a translation-targeting toxin-antitoxin superfamily, TumE–TumA. This is one tested family-level result; the broader proposed families and predicted fold are not hundreds of separately validated discoveries. Native-organism function and general experimental validation across the predicted families remain unestablished.

Technology: AlphaFold structural models and DeepFRI functional prediction

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Selected proteins received laboratory functional validation.
  • Still unknown: Independent replication and function of the broader computational candidate set were not assessed.
  • AI role: AlphaFold structural models and DeepFRI functional prediction supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://www.nature.com/articles/s41586-023-06622-3 · https://edoc.unibas.ch/entities/publication/5b5fe8a1-910e-4fa9-81a1-c5bceb0297be · https://www.sib.swiss/news/artificial-intelligence-tools-shed-light-on-millions-of-proteins

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Indexed primary publisher Methods and publication metadata, institutional abstract and announcing-institution text; full PDF and supplements not freshly recovered..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Deep-learning knee imaging phenotypes reveal osteoarthritis genetic and fracture associations

Using knee DXA scans from the UK Biobank, a ResNet-101 model identified 1,931 more probable knee-osteoarthritis cases than health-record coding and a U-Net segmentation model estimated minimum joint-space width (mJSW). In 29,257 participants, the quantitative mJSW phenotype had substantially higher SNP heritability than the binary phenotypes and yielded 18 genome-wide-significant loci, compared with one for the image-based binary classifier and six for ICD-10 coding; only one mJSW locus was already significant in a large prior knee-OA GWAS. The AI-derived mJSW phenotype was also associated with fracture history in an epidemiological analysis, a signal not seen with the binary case labels.

Technology: ResNet-101 convolutional neural network for knee-OA classification and U-Net image segmentation for joint-space measurement

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Using knee DXA scans from the UK Biobank, a ResNet-101 model identified 1,931 more probable knee-osteoarthritis cases than health-record coding and a U-Net segmentation model estimated minimum joint-space width (mJSW).
  • Still unknown: The GWAS used 29,257 quality-controlled UK Biobank participants with knee DXA images, all from the analyzed white-British ancestry subset; ancestry transferability is unestablished. DXA images were non-weight-bearing. The 18 loci are genome-wide associations for a model-derived quantitative endophenotype, not 18 proven causal OA genes; the paper notes only one also reached significance in a previous large knee-OA GWAS and many loci overlap broad anthropometric associations. The fracture result is observational and does not show that OA or joint-space narrowing causes fractures. The medRxiv preprint DOI 10.1101/2023.03.07.23286909 is an earlier version of this same study and is not counted separately.
  • AI role: The neural networks converted routine knee DXA scans into two disease phenotypes: a case–control OA classifier and a quantitative minimum joint-space-width measurement. The researchers used these AI-derived phenotypes in GWAS and downstream epidemiological association analyses, increasing statistical power and revealing genetic and fracture associations that were not significant with ICD-10 case coding or the binary classifier.

Evidence sources: https://www.nature.com/articles/s41746-023-00903-x

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 71 of works of the same type, year and subfield (OpenAlex work metrics, 17 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning guides engineering of fluorescent calcium indicators

An ensemble of three regression models trained on experimentally derived GCaMP mutation libraries screened 1,423 uncharacterized variants. The study reports eGCaMP and eGCaMP+ with faster kinetics and larger fluorescence responses than previously published fast variants, and eGCaMP2+ with greater dynamic range than the tested sixth-, seventh- and eighth-generation indicators. The preprint reports cell and neuron assays. These are laboratory sensor results, not human or clinical validation. This is one study; the journal article is a later version.

Technology: Ensemble of three regression models

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: The study reports experimental evaluation in vitro, HEK293 cells and primary neurons for selected GCaMP variants.
  • Still unknown: Results are laboratory sensor assays, not evidence of human or clinical use; the record uses the 2023 preprint version while the journal paper appeared in 2024.
  • AI role: An ensemble of three regression models predicted mutation effects and screened 1,423 GCaMP variants for experimental follow-up.

Evidence sources: https://assets-eu.researchsquare.com/files/rs-3146778/v1_covered_50aaf3d0-2f80-4222-91d1-d14c7e34a5f6.pdf?c=1711091168

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary Research Square v1 full text, posted-date header, abstract/results, author affiliations and version-of-record note..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

RFdiffusion: designing new proteins with a diffusion model

A guided diffusion model generates tailored protein structures, such as proteins that bind tightly to another molecule, and outperforms previous methods across a wide range of problems. The post says only that ‘a useful fraction’ of generated designs work as intended, without figures; the software was released in March 2023.

Technology: RFdiffusion

Family RFdiffusion: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Designs were characterized experimentally (symmetric assemblies, metal-binding proteins, protein binders), and a cryo-EM structure of a designed binder was nearly identical to the design model; the abstract reports no overall success rates.
  • Still unknown: The abstract states no explicit limitations; the press post gives no success rates, sample sizes or benchmark numbers and says only that a useful fraction of designs work.
  • AI role: A generative diffusion model fine-tuned from RoseTTAFold generated protein backbones for monomer, binder, oligomer, enzyme scaffold and motif designs.

Evidence sources: https://www.ipd.uw.edu/2023/07/ai-tools-are-designing-entirely-new-proteins-that-could-transform-medicine/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-023-06415-8&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 2729 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AI-assisted macaque cortex atlas reveals regional cell-type organization

AI-assisted segmentation and Spatial-ID annotation helped map 264 transcriptome-defined cell types across the macaque cortex. Regional cell-type composition tracked cortical hierarchy, and cross-species comparisons identified primate-specific types enriched in layer 4. This is the July 12, 2023 journal version; the atlas is observational, uses male monkeys, and has limited transcript capture per spatial cell.

Technology: AI-assisted cell segmentation; Spatial-ID (deep neural network and graph convolutional network)

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: New macaque tissue measurements and spatial maps were compared across sections and biological replicates; this supports a bounded anatomical result, not causal brain function.
  • Still unknown: Not assessed: independent replication and exhaustive supplementary material. The paper identifies limited transcript sensitivity and absence of female subjects.
  • AI role: Image segmentation identifies cells; Spatial-ID trains a DNN on nucleus profiles and a GCN incorporating spatial neighborhoods to assign spatial cell labels.

Evidence sources: https://www.sciencedirect.com/science/article/pii/S0092867423006797 · https://github.com/TencentAILabHealthcare/Cell_Segmentation_Fluorescence · https://bcas.cas.cn/issue/2023_3/202311/P020231112626745733563.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AlphaFold-assisted structural clustering discovers new deaminase functions and compact base-editing enzymes

The team used AlphaFold2 structure predictions to cluster the deaminase family and found that most proteins in a DddA-like clade were not double-stranded DNA deaminases. The work characterized smaller single-strand-specific deaminases, developed a compact cytidine deaminase suitable for packaging a base editor into one AAV, and demonstrated robust editing by a clade member in soybean.

Technology: AlphaFold2-assisted structure-based protein-family clustering

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The team used AlphaFold2 structure predictions to cluster the deaminase family and found that most proteins in a DddA-like clade were not double-stranded DNA deaminases.
  • Still unknown: This paper is an AI-assisted protein-discovery and engineering study; experimental base editing demonstrates the utility of selected deaminases, while therapeutic application remains prospective. The study was followed by a 2024 correction/erratum; the record references the original research article, DOI 10.1016/j.cell.2023.05.041, not the erratum.
  • AI role: AlphaFold2 predicted protein structures that enabled structure-similarity clustering of a full deaminase family. The clustering guided discovery of functional diversity and selection of enzymes, which the researchers characterized experimentally and used in cell, animal and plant base-editing tests.

Evidence sources: https://doi.org/10.1016/j.cell.2023.05.041

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 254 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Organ-specific learned networks prioritize tested seed-lipid regulators

Learned organ classification enabled regulatory-network inference that nominated seed-lipid regulators for genetic and physiological testing in Arabidopsis. The result was public in a June 2023 preprint before the April 2024 journal article. Candidate counts differ between versions, so the entry does not transfer final-version counts or mechanistic detail to the preprint. Crop-yield benefit was not established.

Technology: K-nearest-neighbor organ classification and regulatory-network inference

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: Predicted regulators received genetic and physiological experimental follow-up.
  • Still unknown: Cross-species utility, crop performance and version-specific numerical correspondence remain unestablished.
  • AI role: K-nearest-neighbor organ classification and regulatory-network inference supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://www.biorxiv.org/content/10.1101/2023.06.17.542927v1 · https://api.biorxiv.org/details/biorxiv/10.1101/2023.06.17.542927 · https://www.pnas.org/doi/10.1073/pnas.2322751121

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary bioRxiv API v1 abstract/date plus journal abstract, selected Methods and affiliations; no full preprint numerical endpoints transferred..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

SIMMER-guided experiments identify gut bacterial strains that deplete methotrexate

SIMMER combines reaction-chemistry and protein-sequence similarity to predict gut microbial species and enzymes responsible for drug biotransformations. For methotrexate (MTX), the model predicted candidate strains and enzyme sequences; researchers tested 42 gut bacterial isolates by HPLC. SIMMER predicted 13 isolates as MTX-metabolizing, and 10 of those tested isolates showed MTX depletion; the association was significant in the study (Fisher exact p=0.046). Predicted MTX-hydrolysis enzyme signatures were also enriched in stool samples of MTX nonresponders in an analyzed rheumatoid-arthritis cohort.

Technology: SIMMER, a computational reaction-similarity and protein-homology pipeline using chemical fingerprints, MetaCyc reactions and gut metagenomic sequences

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: For methotrexate (MTX), the model predicted candidate strains and enzyme sequences; researchers tested 42 gut bacterial isolates by HPLC.
  • Still unknown: The study reports species-level evidence for bacterial MTX depletion, but the enzyme predictions themselves were not individually purified or functionally validated. Only 13 of 42 tested isolates were predicted positive and 10 depleted MTX; predictions may have false positives, as the authors note. The RA association is observational and does not establish that microbial metabolism caused treatment nonresponse. The underlying fact that gut bacteria can metabolize MTX was already known; this result concerns species-level prediction and validation, not discovery of MTX metabolism itself.
  • AI role: SIMMER proposed specific bacterial species and enzyme sequences for known drug transformations. The team tested its MTX predictions against an independent panel of 42 gut bacterial isolates and HPLC-measured drug depletion, and compared predicted enzyme presence with baseline stool metagenomes and later MTX response in an RA dataset.

Evidence sources: https://elifesciences.org/articles/82401

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 86 of works of the same type, year and subfield (OpenAlex work metrics, 25 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

First draft of the human pangenome reference

A consortium of 119 scientists from 60 institutions published in Nature a pangenome combining assembled genomes from 47 people of diverse ancestries, rather than one person. Google adapted DeepVariant to identify variants and DeepConsensus to correct sequencing errors; a long-read assembler achieved over 99.999% final accuracy. This is a draft: benefits for diagnosis and treatment remain prospective.

Technology: DeepVariant and DeepConsensus

Family DeepConsensus: profile · models page

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: The draft pangenome from 47 assemblies was evaluated for accuracy at structural and base-pair levels, and short-read analysis was compared with GRCh38-based workflows (34% fewer small-variant errors).
  • Still unknown: The abstract states no explicit limitations; the blog says this is a draft and benefits for diagnosis and treatment remain prospective.
  • AI role: Google adapted DeepVariant to call variants and DeepConsensus to correct sequencing errors in the pangenome assembly.

Evidence sources: https://blog.google/innovation-and-ai/technology/health/first-pangenome-reference-nature-paper-ai/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-023-05896-x&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 1258 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

BacterAI experimentally maps amino-acid growth requirements of two oral Streptococcus species

BacterAI autonomously designed and ran growth experiments with robotic liquid handling to learn amino-acid requirements for the oral commensals Streptococcus gordonii and S. sanguinis. It learned separate growth rules for each organism and showed that the closely related species have different amino-acid auxotrophies. For S. gordonii, its neural-network growth predictions exceeded 90% accuracy after nine days and 3,024 experiments; the article also reports transfer-learning tests in larger media.

Technology: BacterAI, a reinforcement-learning agent coupled to iterative neural-network growth models and automated laboratory robots

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: BacterAI autonomously designed and ran growth experiments with robotic liquid handling to learn amino-acid requirements for the oral commensals Streptococcus gordonii and S. sanguinis.
  • Still unknown: The measured outcomes are in defined growth media under the tested laboratory conditions, not direct evidence about nutrient exchange or health effects in a human oral microbiome. The article establishes that the two species have different amino-acid requirements; transfer learning and high growth-prediction accuracy describe the method’s performance, not separate biological discoveries. Replicate growth assays showed high agreement, but inferred rules may not cover every environmental condition.
  • AI role: The agent selected successive combinations of amino acids to test based on prior growth results, executed those tests through an automated pipeline, and distilled results into interpretable growth rules. It enabled an unbiased experimental mapping of nutrient requirements from a blank slate for both species.

Evidence sources: https://www.nature.com/articles/s41564-023-01376-0

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 52 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

An MRI decoder turns brain activity into text

A noninvasive decoder reconstructs continuous text from fMRI that closely, and sometimes precisely, matches the original meaning in about half of trials (after dozens of hours of training with podcasts; it also works on silent videos). Published in Nature Neuroscience; it works only with cooperative participants and is not word-for-word transcription.

Technology: Semantic decoder from fMRI

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: A non-invasive fMRI decoder generated continuous text that conveyed the meaning of perceived speech, imagined speech and silent videos; the press release says it matched the original meaning in about half of trials and works only with cooperative participants.
  • Still unknown: The abstract does not report participant numbers or semantic accuracy measures and states no explicit limitations; the press release says it is not word-for-word transcription.
  • AI role: A decoder trained on fMRI recordings of cortical semantic representations generated word sequences matching the meaning of brain activity.

Evidence sources: https://news.utexas.edu/2023/05/01/brain-activity-decoder-can-reveal-stories-in-peoples-minds/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41593-023-01304-9&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 519 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 5/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Single-cell imaging and machine learning reveal a Cdk2–COP9 interaction that induces a senescence-like immune state in Drosophila cells

A synthetic genetic-interaction and cell-morphology map covering more than 6,800 Drosophila genes supported machine-learning-based gene-function assignments across 47 modules. In follow-up, Cytoclass single-cell classification revealed a state-specific interaction between Cdk2 and COP9 signalosome knockdown: co-depletion arrested cell-cycle progression and produced a senescence-like phenotype with increased antimicrobial-peptide expression in cultured cells and hemocytes, including without bacterial infection.

Technology: Single-cell morphology classifier ensemble and machine-learning analysis of genetic-interaction profiles (Cytoclass; support-vector machines and random forest)

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: A synthetic genetic-interaction and cell-morphology map covering more than 6,800 Drosophila genes supported machine-learning-based gene-function assignments across 47 modules.
  • Still unknown: The findings are from cultured Drosophila cells and hemocytes; they do not establish mammalian or human disease mechanisms. The senescence-like phenotype and antimicrobial-peptide induction were measured after experimental RNAi co-depletion. The global module assignments and many network interactions are computational inferences; only selected interactions received targeted follow-up.
  • AI role: The authors generated single-cell morphology and synthetic genetic-interaction profiles, used machine-learning-based function discovery to assign functions to genes in 47 modules, and applied Cytoclass state classification to reveal interactions hidden in bulk-average profiles. The state-specific Cdk2–COP9 interaction was followed up by double-RNAi assays, single-cell transcriptomics, beta-galactosidase and antimicrobial-peptide measurements.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/37116498/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 72 of works of the same type, year and subfield (OpenAlex work metrics, 14 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Learned promoter models identify human-specific transcription motifs

Models trained on promoter activity measurements helped characterize a human-specific downstream promoter variant and select synthetic motifs with species-specific activity. In-vitro transcription assays tested the predicted preferences with human and Drosophila extracts. These are gene-regulation and construct-assay findings, not demonstrated effects in organisms or disease.

Technology: HARPE-trained support-vector regression

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Predicted promoter preferences were compared with measured transcription in extracts.
  • Still unknown: Native-chromatin regulatory effects and organism-level outcomes were not established.
  • AI role: HARPE-trained support-vector regression supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://escholarship.org/uc/item/1rt3r053 · https://escholarship.org/content/qt1rt3r053/qt1rt3r053.pdf · https://pubmed.ncbi.nlm.nih.gov/37163335/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Institutional manuscript abstract and selected SVR Methods; primary PubMed early-online date, affiliations and experimental figure captions..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Language models guide improvements to seven antibodies

Two rounds of laboratory evolution, using at most 20 variants per antibody, improved the affinity of four mature antibodies by up to sevenfold and three immature ones by up to 160-fold. Biochemical and pseudovirus tests; no demonstration of therapeutic benefit in humans.

Technology: Protein language models

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Language-model-guided variants of seven antibodies were made in two lab evolution rounds (at most 20 variants per antibody) and binding was measured by biolayer interferometry; four antibodies were also checked for thermostability, polyspecificity and pseudovirus neutralization.
  • Still unknown: The authors state that affinity gains are smaller than in vivo somatic hypermutation, the method improves existing binding rather than creating new binders, and it may work less well under other selection pressures; no therapeutic benefit in humans is shown.
  • AI role: General protein language models guided affinity maturation by choosing which antibody variants to test.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC10869273/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41587-023-01763-2&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 481 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Learned community models prioritize a tested bacterial colonization inhibitor

Models trained on community composition helped identify Enterococcus faecalis as an inhibitor of Enterococcus faecium colonization in human stool-derived laboratory communities. Follow-up supported the predicted interaction under the tested conditions. The result was public in an April 2023 preprint before the March 2024 journal article. It does not demonstrate prevention of infection or beneficial intervention in humans.

Technology: Random-forest and neural-ODE community models

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: Model-prioritized interactions received in-vitro validation.
  • Still unknown: Host effects, clinical benefit and transfer beyond the culture conditions remain unestablished.
  • AI role: Random-forest and neural-ODE community models supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://www.biorxiv.org/content/10.1101/2023.04.19.537502v1 · https://api.biorxiv.org/details/biorxiv/10.1101/2023.04.19.537502 · https://pmc.ncbi.nlm.nih.gov/articles/PMC10944475/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary bioRxiv API v1 abstract/date; journal article selected model and intervention Results, metadata and affiliations. Detailed final-version counts not transferred..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ExoGRU-guided experiments support RBM24 involvement in small-RNA secretion

A recurrent model of RNA-sequence features helped nominate RNA-binding proteins involved in secretion. Perturbation and binding experiments supported RBM24 involvement in the tested cell system, alongside previously implicated factors. This mechanism study was public in an April 2023 preprint before the 2024 journal version. It does not establish a universal secretion mechanism, in-vivo benefit or a therapeutic result.

Technology: ExoGRU gated recurrent neural network

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: Selected factors received perturbation and extracellular RNA-binding follow-up.
  • Still unknown: Generality across cell types and effects in vivo remain unestablished.
  • AI role: ExoGRU gated recurrent neural network supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://www.biorxiv.org/content/10.1101/2023.04.04.535452v1 · https://api.biorxiv.org/details/biorxiv/10.1101/2023.04.04.535452 · https://pmc.ncbi.nlm.nih.gov/articles/PMC11019361/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary bioRxiv API v1 metadata/abstract and journal PMC selected model, validation and limitations text; supplement not reviewed..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

CLEAN experimentally identifies SAM hydrolase activity in previously uncharacterized proteins

Researchers at the University of Illinois Urbana–Champaign developed CLEAN, a contrastive-learning enzyme-function predictor, and used it to identify the SAM hydrolase activity of the previously uncharacterized protein TTHA0338. In vitro HPLC–MS and enzyme-kinetic experiments also corrected the annotation of MJ1651 and confirmed the multiple activities of SsFIA.

Technology: CLEAN (contrastive learning-based enzyme annotation model)

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: In vitro HPLC–MS and enzyme-kinetic experiments also corrected the annotation of MJ1651 and confirmed the multiple activities of SsFIA.
  • Still unknown: The study validated a small, selected set of proteins, and also showed CLEAN can assign functions to non-enzymes; the demonstrated result is specific to the tested proteins. TTHA0338 is one of multiple results reported by this paper; record describes the previously unknown function rather than treating every model benchmark as a separate discovery.
  • AI role: CLEAN predicted enzyme commission functions from protein sequences; researchers then validated the predicted activity experimentally. It was the model that nominated TTHA0338 as a SAM hydrolase, not an autonomous discovery or validation system.

Evidence sources: https://doi.org/10.1126/science.adf2465

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 482 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Acoustic learning maps spider-monkey associations with forest disturbance

A learned call detector, manual confirmation and occupancy modeling analyzed recordings from 341 Costa Rican sites. In this sample, Geoffroy's spider monkey was detected only above 80% forest cover and away from nearby primary paved roads. These are regional observational associations, not causal or universal habitat thresholds; imperfect detection and the focus on one call type limit inference.

Technology: Attention-equipped residual convolutional call detector

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Model positives from field recordings were manually checked and linked to landscape observations.
  • Still unknown: Causality, independent ground-truth survey agreement and transfer to other regions were not established.
  • AI role: Attention-equipped residual convolutional call detector supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC10015327/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary article dates, affiliations and selected detector Methods, Results and limitations; supplement not reviewed..

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

LuxSit: de novo luciferases active in the laboratory

The group designed new luciferases and identified three active enzymes among the initial designs. The 117-residue LuxSit catalyzes blue light with the DTZ substrate; its function is verified in the laboratory, while therapeutic or industrial applications remain prospective.

Technology: Family-wide hallucination and deep learning

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Designed luciferases were tested in activity assays with synthetic luciferin substrates; the most active design had catalytic efficiency comparable to native luciferases and high substrate selectivity.
  • Still unknown: The abstract does not state how many designs were tested or how many were active, and states no explicit limitations; therapeutic or industrial applications remain untested.
  • AI role: Deep learning generated idealized scaffolds with binding pockets and matching sequences used to design artificial luciferases.

Evidence sources: https://www.ipd.uw.edu/2023/02/machine-learning-generates-custom-enzymes/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-023-05696-3&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 470 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Learned morphology guides targeted isolation of gut bacteria

A learned morphology model selected likely target genera for robotic isolation, followed by sequencing to confirm cultured organisms. This supports a bounded, experimentally tested recovery result. The platform's broader diversity-selection component uses feature distances and is not itself described as learned. Neither every isolate nor every genome is counted as a separate discovery, and clinical benefit was not established.

Technology: CAMII random-forest colony taxonomy prediction

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: New colonies selected by the taxonomy model were isolated and sequenced.
  • Still unknown: Generality beyond the studied samples and culture conditions remains unestablished.
  • AI role: CAMII random-forest colony taxonomy prediction supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://www.nature.com/articles/s41587-023-01674-2 · https://www.nature.com/articles/s41587-023-01674-2.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Indexed primary publisher PDF targeted-isolation Methods, publisher abstract/date and affiliations; prior experimental Results review retained..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

CellOracle-guided zebrafish perturbations identify axial-mesoderm effects of noto and lhx1a loss

The CellOracle machine-learning workflow used single-cell gene-regulatory networks to prioritize transcription factors for perturbation in zebrafish development. CRISPR/Cas9 tests confirmed a model-predicted, previously unreported increase in prechordal-plate differentiation after loss of noto, alongside the known loss of notochord. Model-prioritized lhx1a loss also inhibited axial-mesoderm differentiation: single-cell RNA-seq found expansion of early notochord cells and reductions in later notochord and prechordal-plate populations; orthogonal RNA assays supported reduced notochord/prechordal markers.

Technology: CellOracle, a machine-learning gene-regulatory-network model with in-silico transcription-factor perturbation

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: CRISPR/Cas9 tests confirmed a model-predicted, previously unreported increase in prechordal-plate differentiation after loss of noto, alongside the known loss of notochord.
  • Still unknown: Evidence is from zebrafish crispants during embryonic development, not stable mutants or mammalian/human systems; the authors report no gross morphological change despite transcriptomic/cell-population effects. The CellOracle model depends on inferred networks and makes cell-state-shift predictions, not direct causal proof of all inferred edges. A later unreviewed 2024 preprint raises implementation and benchmarking concerns, particularly about its use of ATAC-seq for distal regulatory interactions and comparisons with other GRN tools; this record relies only on the specific perturbation predictions tested experimentally and does not adopt the paper's broader benchmarking claims.
  • AI role: The model built context-specific networks from single-cell data, simulated transcription-factor loss, and prioritized noto and lhx1a for experimental follow-up. Researchers used CRISPR/Cas9 crispants, single-cell RNA-seq and marker assays to test the predicted zebrafish cell-fate effects; the prechordal-plate increase after noto loss was reported as previously unrecognized.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC9946838/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 766 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning guided XylM variant improved biosensor-indicated productivity

Two machine-learning-assisted directed-evolution cycles produced XylM-D140E-V144K-F243L-N244S, reported to have 15-fold higher 3-methylsalicylic acid productivity than wild-type XylM. Productivity was estimated using XylS-dependent fluorescence in a sensor strain; this is a biosensor-based productivity readout, not a purified-enzyme kinetic constant.

Technology: Machine-learning-assisted directed evolution using biosensor-derived activity data

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The publisher reports XylM-D140E-V144K-F243L-N244S with 15-fold greater 3-methylsalicylic-acid productivity than wild type after two ML-assisted directed-evolution cycles, using XylS-dependent fluorescence from a sensor strain as the measured indicator.
  • Still unknown: The signal is an indirect biosensor-based productivity proxy; the reported value is not a purified-enzyme kinetic constant.
  • AI role: Machine learning used sensor-derived data to guide two rounds of directed evolution and prioritize XylM variants.

Evidence sources: https://pubs.acs.org/doi/abs/10.1021/acssynbio.2c00577 · https://pubmed.ncbi.nlm.nih.gov/36734676/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary ACS publisher indexed abstract and affiliation/article-history metadata, cross-checked against primary PubMed abstract and metadata..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-learning pipeline identifies antimicrobial hexapeptides

A sequential machine-learning pipeline screened virtual libraries of 6–9-residue peptides and identified three leading antimicrobial hexapeptides. The paper reports activity against multidrug-resistant clinical isolates and therapeutic effects in a mouse bacterial-pneumonia model. These are laboratory and animal results, not clinical efficacy.

Technology: Sequential machine-learning pipeline for peptide selection, classification, ranking and activity regression

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The authors report a sequential machine-learning pipeline over virtual 6–9 amino-acid sequence space; the leading three identified hexapeptides were tested against clinical multidrug-resistant isolates and in a mouse bacterial-pneumonia model.
  • Still unknown: The mouse model and reported comparison with penicillin do not establish human efficacy. The PubMed primary metadata gives journal online date 2023-01-12; no earlier pre-2020 version is identified in the reviewed primary material.
  • AI role: Sequential learned modules selected, classified, ranked and regressed peptide candidates from the virtual sequence library; the reported three lead hexapeptides are outputs of that workflow.

Evidence sources: https://www.nature.com/articles/s41551-022-00991-2 · https://pubmed.ncbi.nlm.nih.gov/36635418/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary Nature indexed article text and PubMed abstract/metadata/affiliation record for the ML hexapeptide screening study, its January 2023 publication date and reported mouse pneumonia assays..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Meta AI

ESM Metagenomic Atlas: over 600 million protein structures

Meta AI published predicted structures for almost the entire MGnify90 database (over 600 million), generated in about two weeks on roughly 2,000 GPUs with a 15-billion-parameter language model, up to 60 times faster than previous methods. These are predictions, and high-confidence structures are a much smaller subset.

Technology: ESMFold (15-billion-parameter language model)

Family ESM / ESMFold: profile · models page

Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: retrospective or benchmark
  • Tested: The preprint reports structure prediction up to 60 times faster while maintaining accuracy versus state-of-the-art methods and an atlas of 617 million structures; the blog describes no experimental checks.
  • Still unknown: The blog lists no limitations and notes some sequences fold into disordered forms; high-confidence predictions are a subset.
  • AI role: Language model (15B parameters) predicted protein structures from sequence.

Evidence sources: https://ai.meta.com/blog/protein-folding-esmfold-metagenomics/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1101/2022.07.20.500902&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Ligand-based discovery of coronavirus main protease inhibitors using MACAW molecular embeddings

MACAW molecular embeddings and ligand-based models prioritized purchased compounds for assays. The paper reports inhibition of recombinant coronavirus main protease for eight of ten selected molecules; compound 7 also reduced hCoV-229E infectivity in a cell assay. This is not a clinical COVID-19 result or proof of SARS-CoV-2 patient benefit.

Technology: MACAW molecular embeddings used in ligand-based machine learning virtual screening

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Primary article reports 8/10 selected compounds inhibiting recombinant Mpro and follow-up cellular testing of compound 7 against HCoV-229E.
  • Still unknown: Independent replication and exhaustive prior-art review are not established.
  • AI role: MACAW molecular embeddings used in ligand-based machine learning virtual screening

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC9621234/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary source checked for study identity, stated AI contribution, date and bounded result..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Novo Nordisk A/S

Mass spectrometry and machine learning identify bioactive peptide NHPD-50

A predicted-peptide-variant model trained using mass-spectrometry observations from seven mouse tissues helped prioritize candidate peptides. The authors report that NHPD-50, a secretogranin-1-derived peptide, lowered blood glucose and increased plasma insulin in specified diabetic-mouse experiments. The authors state that the data do not establish the peptide’s exact mechanism. These are early preclinical findings, not a human diabetes treatment; other predicted peptides had different or absent outcomes.

Technology: Predicted-peptide-variant machine-learning prioritization from mass-spectrometry data

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Primary paper reports the stated mouse assay results.
  • Still unknown: Independent replication and exhaustive prior-art review are not established; the paper states that the mechanism is undetermined.
  • AI role: A predicted-peptide-variant model prioritized candidate sequences from mass-spectrometry data; biological activity was established through subsequent assays.

Evidence sources: https://www.nature.com/articles/s41467-022-34031-z

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary Nature full article and bibliographic metadata reviewed for identity, AI role, date, affiliation and bounded mouse results..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Bayesian language-model design improved measured scFv binding

An end-to-end Bayesian language-model workflow designed diverse single-chain antibody fragment libraries that were measured for binding. In the 2023 journal version, the best generated scFv bound 28.7-fold better than the best directed-evolution comparator, and 99% of designs in the most successful library improved on the initial candidate. The 2022 bioRxiv v1 abstract already reported this result chain, giving 28.8-fold for the best generated scFv; the small difference is version-specific. These are in-vitro binding results, not clinical efficacy.

Technology: Bayesian optimization with a language-model-based antibody sequence representation

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: Designed scFvs were empirically measured. The publisher abstract reports the best generated scFv had 28.7-fold greater binding than the best directed-evolution scFv and 99% of the designed scFvs in the most successful library improved on the initial candidate.
  • Still unknown: The journal version reports the best generated scFv at 28.7-fold better binding while the v1 abstract reports 28.8-fold; this minor version discrepancy is retained. The measured binding result does not establish clinical efficacy.
  • AI role: A Bayesian, language-model-based workflow designed candidate scFv libraries, whose binding was experimentally measured.

Evidence sources: https://api.biorxiv.org/details/biorxiv/10.1101/2022.10.07.502662 · https://www.nature.com/articles/s41467-023-39022-2.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Official bioRxiv API v1 date/abstract and primary Nature journal PDF for full result and affiliations..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AlphaFold helps interpret activation of PINK1, a Parkinson’s-linked protein

After experimentally determining different forms of PINK1, the WEHI group used AlphaFold to model the human protein and its dimer, obtaining structures consistent with experimental results. The predictions also helped interpret how mutations associated with early-onset Parkinson’s may affect dimerization; this does not amount to a validated therapy.

Technology: AlphaFold

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The WEHI group compared an AlphaFold model of human PINK1 dimer with its own experimental structure work and used it to model disease mutations, which it says confirmed their suspicions.
  • Still unknown: The article describes no specific limitations of the predictions and says plenty of questions remain.
  • AI role: AlphaFold modelled human PINK1 and its dimer and the effect of mutations on dimerisation.

Evidence sources: https://deepmind.google/blog/targeting-early-onset-parkinsons-with-ai/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ProteinMPNN: protein sequences designed in about one second

A neural network generates the amino-acid sequence for a given protein shape in about one second, over 200 times faster than the best previous software, without specialist tuning. A related study shows that proteins designed with ProteinMPNN fold correctly much more often, including nanoscale rings and complex assemblies. Researchers describe it as the beginning of machine learning in protein design.

Technology: ProteinMPNN

Family ProteinMPNN: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Speed is compared with prior software; a related Science study tested lab-made designs, which folded as intended more often, including nanoscale rings imaged by electron microscopy.
  • Still unknown: The article gives no quantitative success rates and notes protein data is scarce, requiring trial and error.
  • AI role: Neural network generated amino-acid sequences for given protein shapes.

Evidence sources: https://www.ipd.uw.edu/2022/09/proteinmpnn-excels-at-creating-new-proteins/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1126/science.add2187&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 2354 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

MIMaL predicts and experimentally validates five yeast gene functions from protein–metabolite profiles

MIMaL used tree-based regression and SHAP interpretation on yeast proteome/metabolome measurements, then clustered protein-control profiles across single-gene knockouts. It predicted functions for five genes, with experimental validation reported for all five. Two previously uncharacterized genes, YJR120W and YDL157C, were predicted to modulate mitochondrial translation; functions were also predicted and validated for SDH9, ISC1 and FMP52. The study additionally validated previously unrecognized protein links to citrate.

Technology: MIMaL: ExtraTrees regression of metabolite abundance from proteomic profiles, SHAP interpretation, and UMAP/OPTICS clustering of protein-control profiles

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: MIMaL used tree-based regression and SHAP interpretation on yeast proteome/metabolome measurements, then clustered protein-control profiles across single-gene knockouts.
  • Still unknown: This was a yeast study using a prior 174-single-gene-knockout multi-omics dataset, with 348 condition profiles; it does not establish corresponding human gene functions. Five gene functions are reported validated, and the article specifically states YJR120W and YDL157C modulate mitochondrial translation. Some experimental validation used existing knockout strains and measurements; no independent replication is implied.
  • AI role: A tree-based model predicted metabolite profiles from proteomic measurements; SHAP-derived protein-control values and clustering across knockout conditions were used to infer new protein–metabolite connections and gene functions. The researchers experimentally validated predicted citrate-control relationships and five gene-function predictions.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC9801967/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 81 of works of the same type, year and subfield (OpenAlex work metrics, 25 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Deep learning helps experimentally map a ribozyme neutral network

After earlier experimental evolutionary rounds, the authors trained a deep classifier to guide the final in-silico generation and experimentally tested candidates. They measured all 2^16 intermediates between two active RNA ligase ribozymes separated by 16 mutations and report an extensive network of neutral paths. The result applies to this ribozyme and assay, not all RNA evolution. The paper also notes possible model overfitting and the restricted classifier’s accuracy.

Technology: Deep-learning classifier; evolutionary algorithm

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The study reports experimental activity measurements for all 2^16 connecting sequence variants and a screen of more than 120,000 ribozyme sequences.
  • Still unknown: The neutral network is bounded to one RNA ligase ribozyme and assay; generality across other landscapes is not established, and the paper notes potential classifier overfitting.
  • AI role: A deep-learning classifier trained on prior experimental generations guided in-silico selection, mutation and recombination in the final generation.

Evidence sources: https://www.nature.com/articles/s41467-022-32538-z

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary Nature Communications indexed full text, abstract, main text, publication metadata and author affiliation..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AlphaFold Database: over 200 million structures

The predicted-protein-structure archive, maintained with EMBL-EBI, grew from nearly 1 million to over 200 million structures, covering practically all cataloged proteins, with free access and bulk download from Google Cloud. More than 500,000 researchers in 190 countries had already used it. These are predictions, not experimental measurements.

Technology: AlphaFold 2

Family AlphaFold: profile · models page

Publication: Announcement only · Test context: Test context unspecified in annotation. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: not reported
  • Tested: The post gives no accuracy metrics for the new structures; it cites two Nature AlphaFold papers and uses in experiments as evidence.
  • Still unknown: The post states no explicit limitations or caveats on prediction accuracy.
  • AI role: AlphaFold predicted protein structures added to the AlphaFold Database.

Evidence sources: https://deepmind.google/discover/blog/alphafold-reveals-the-structure-of-the-protein-universe/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 5/5 · Validation 1/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ProtGPT2: a language model generating novel protein sequences

ProtGPT2, a 738-million-parameter autoregressive transformer trained on UniRef50, generates de novo protein sequences in seconds. The authors report the generated sequences are distantly related to natural ones, sample unexplored regions of protein space, and match natural sequences in amino acid and disorder propensities. Caveat: the evaluation is computational (sequence searches and similarity networks), with no experimental validation in the sources checked.

Technology: ProtGPT2 (transformer protein language model)

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Generated sequences were assessed computationally: composition and disorder propensities, sequence searches, and AlphaFold-predicted folds; no experiments are described in the abstract read.
  • Still unknown: Not reported in the abstract read; no experimental validation is described.
  • AI role: ProtGPT2, a 738-million-parameter autoregressive transformer trained on UniRef50, generated de novo protein sequences.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41467-022-32007-7&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 883 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Active learning improves a cell-free carbon-fixation cycle

An ensemble learned from experimental feedback to select compositions for an engineered carbon-fixation cycle. Laboratory measurements supported improved glycolate production compared with the study's reference cycle. This is an in-vitro pathway result, not carbon capture by living organisms or a demonstrated industrial process. Other applications in the same paper are not additional catalog entries.

Technology: METIS XGBoost active learning

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Selected pathway compositions received biochemical measurements.
  • Still unknown: Independent replication, in-vivo function and scale-up remain unestablished.
  • AI role: METIS XGBoost active learning supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://www.nature.com/articles/s41467-022-31245-z · https://www.nature.com/articles/s41467-022-31245-z.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Indexed primary publisher PDF model Methods, date and affiliations; prior primary experimental Results review retained. Direct PDF access blocked..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

BirdNET: citizen science through bird-song identification

The app recognizes over 3,000 species from their calls, with more than 2.2 million people contributing since 2018; in four test cases, its data reproduced known patterns such as song distribution and the migration ranges of two species. Published in PLOS Biology; only four cases, with no identification-error figures.

Technology: BirdNET (bioacoustics)

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: In four test cases where conventional research had robust answers, BirdNET data replicated known patterns such as song-type distribution and migration ranges.
  • Still unknown: No identification-error rates are reported and the authors describe the validation as the first step.
  • AI role: Bird-song recognition app identified species from calls and produced citizen-science observations.

Evidence sources: https://news.cornell.edu/stories/2022/06/ai-powered-birdnet-app-makes-citizen-science-easier · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1371/journal.pbio.3001670&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 88 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

EMBL

Model of the human nuclear pore with AlphaFold2 and electron microscopy

AlphaFold2 and ColabFold, combined with cryo-tomography and cryo-electron microscopy, produced a model of the human nuclear pore complex (about 30 proteins, about 1,000 copies, 120 MDa) covering over 90% of the complex, versus 46% for previous models. Flexible parts remain unresolved; a complete dynamic model is future work.

Technology: AlphaFold2 and ColabFold

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: The integrative model (AlphaFold2 with cryo-ET and cryo-EM data) was validated against experimental data and covers more than 90% of the NPC core versus 46% for previous models.
  • Still unknown: Flexible peripheral parts are not modelled, no quantitative confidence scores are given, and a complete dynamic model is future work.
  • AI role: AlphaFold2 and ColabFold predicted protein structures that were assembled into the nuclear pore model.

Evidence sources: https://www.embl.org/news/science/puzzling-out-the-structure-of-a-molecular-giant/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1126/science.abm9506&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 360 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ColabFold: free protein folding with AlphaFold2 or RoseTTAFold

ColabFold pairs AlphaFold2 or RoseTTAFold with fast MMseqs2 homology search, running free on Google Colab. Beyond benchmark comparisons, the paper reports predicting structures for 1,762 Methanocaldococcus jannaschii proteins in 48 hours on one GPU and using predicted complexes to help determine the 120 MDa human nuclear-pore structure with cryo-electron microscopy. The proteome structures are computational predictions, not experimentally validated structures. The nuclear-pore application overlaps the separate nuclear-pore entry; this is a methodological contribution, not an additional independent biological finding.

Technology: ColabFold (AlphaFold2, RoseTTAFold, MMseqs2)

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Throughput was measured (close to 1,000 structures per day on one GPU) and predicted complexes were used alongside cryo-EM to help determine the human nuclear-pore structure.
  • Still unknown: Not reported in the abstract read.
  • AI role: AlphaFold2 or RoseTTAFold, paired with MMseqs2 homology search, run free on Google Colaboratory.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41592-022-01488-1&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 9430 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Knowledge-graph-guided discovery of antibiotic-resistance associations

KIDS combined an expert-curated knowledge graph with learned link-prediction models to rank gene–antibiotic resistance hypotheses for experiments. The authors report 93 validated associations involving 83 E. coli genes and 15 antibiotics, including six genes they describe as previously unknown resistance genes; five tested Salmonella homologs also showed resistance. These are laboratory gene–drug associations, not clinical validation. The article’s second-round percentage conflicts with its counts (29/90 is 32.2%, although the text states 28.8%). Count one study.

Technology: Knowledge graph with PRA, multilayer perceptron, AdaBoost-stacked model, TransE and TransD

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The authors report wet-lab validation of 93 gene–antibiotic associations, including five tested Salmonella homologs.
  • Still unknown: Reported percentages are internally inconsistent for 29 of 90 second-round tests; these assays do not establish clinical relevance or treatment outcomes.
  • AI role: KIDS ranked gene–antibiotic links from an expert-curated graph with learned link-prediction models, directing experiments to candidate resistance associations.

Evidence sources: https://www.nature.com/articles/s41467-022-29993-z

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary Nature indexed full text, article dates, affiliations, methods/results, contributions and stated counting discrepancy..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

FAST-PETase: an AI-engineered enzyme degrades PET waste

A machine-learning model suggested PETase mutations, then tested on 51 post-consumer containers, fibers and fabrics. FAST-PETase degrades PET below 50 °C, in some cases down to monomers within 24 hours; production at scale and environmental use still needed development.

Technology: MutCompute / 3D convolutional network

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: FAST-PETase was tested on 51 post-consumer containers, fibers and fabrics, degrading PET below 50 °C, in some cases to monomers within 24 hours.
  • Still unknown: Scale-up, field use at polluted sites and cost are not demonstrated, and the work covers only PET.
  • AI role: Machine-learning model suggested enzyme mutations.

Evidence sources: https://news.utexas.edu/2022/04/27/plastic-eating-enzyme-could-eliminate-billions-of-tons-of-landfill-waste/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1038/s41586-022-04599-z&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 1271 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-learning selection uncovers plant enzymes for alkaloid pathway branches

Support-vector-machine ranking selected poppy homologs whose experimental characterization identified aromatic acetaldehyde synthase activity in PsTyDC6 and phenylpyruvate-decarboxylase activity in PsPDC1. In vitro and engineered E. coli metabolomics detected pathway products including norcoclaurine and reticuline, demonstrating alternative routes from tyrosine into benzylisoquinoline alkaloids.

Technology: High-dimensional support vector machine (SVM) enzyme-sequence classifiers

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Support-vector-machine ranking selected poppy homologs whose experimental characterization identified aromatic acetaldehyde synthase activity in PsTyDC6 and phenylpyruvate-decarboxylase activity in PsPDC1.
  • Still unknown: The study demonstrates enzyme activity and pathway products in vitro and engineered E. coli; it does not establish the exact physiological roles of these enzymes in living poppy plants. The paper describes several enzyme activities as potential and says natural pathway roles need further elucidation.
  • AI role: SVM models ranked more than 100 homologous candidate sequences for specialized aromatic-acetaldehyde-synthase and phenylpyruvate-decarboxylase functions. The team selected high-scoring sequences for in-vitro enzyme assays and metabolomics of reconstructed E. coli pathways.

Evidence sources: https://www.nature.com/articles/s41467-022-28883-8

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 89 of works of the same type, year and subfield (OpenAlex work metrics, 41 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Learned activity models improve viral diagnostic assay designs

Learned diagnostic-activity scores guided assay design across viral variation. Experiments on selected synthetic targets supported improved detection compared with the studied conservation-based designs. The larger cross-species design catalog is computational; it is not thousands of validated diagnostics. Patient-specimen sensitivity, clinical specificity and field performance were not established.

Technology: ADAPT learned sequence-activity classifier and regressor

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Selected assay designs were tested on synthetic viral targets.
  • Still unknown: Clinical and environmental sample performance require further validation.
  • AI role: ADAPT learned sequence-activity classifier and regressor supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC9287178/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary article metadata, affiliations, selected design Methods, experimental Results and limitations; supplements not reviewed..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

HydrAMP generates antimicrobial peptide analogues validated in bacterial assays

HydrAMP, a conditional variational autoencoder with antimicrobial activity classifiers, generated peptide analogues and de novo candidates. The 2022 preprint reports wet-lab confirmation of increased activity for a generated Pexiganan-based analogue; the 2023 journal version reports testing across five bacterial strains, including nine analogues of clinical prototypes and six analogues of an inactive peptide. The tested panel was small and model/human filtered; it does not establish therapeutic efficacy.

Technology: Conditional variational autoencoder with classifier-guided peptide generation and molecular-dynamics-based candidate preselection

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: Generated peptide analogues were experimentally evaluated for antimicrobial activity across five bacterial strains; the journal abstract reports high activity for nine generated analogues of clinical prototypes and six analogues of an inactive peptide.
  • Still unknown: The tested panel is selected and small; these are in-vitro assays and do not establish safety, therapeutic efficacy or clinical benefit. The 2023 correction replaced a duplicated generation-statistics figure, not the wet-lab result stated here.
  • AI role: A conditional VAE generated peptide analogues and unconstrained candidates from antimicrobial activity conditions; ranking and molecular-dynamics preselection helped choose candidates for experimental validation.

Evidence sources: https://www.biorxiv.org/content/10.1101/2022.01.27.478054v1 · https://www.nature.com/articles/s41467-023-36994-z · https://www.nature.com/articles/s41467-023-40879-6.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary bioRxiv v1 abstract/metadata; primary Nature Communications indexed full article including methods/results; primary publisher author correction..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Cell2location: mapping fine-grained cell types in spatial transcriptomics

Cell2location is a Bayesian model that resolves fine-grained cell types in spatial transcriptomics data by accounting for technical variation and sharing information across locations. Applied to three tissues, it revealed regional astrocyte subtypes in mouse brain, a rare pre-germinal center B cell population in human lymph node, and fine immune populations in human gut lymphoid follicles. Caveat: the reported results are biological findings from the method's application, with no single quantitative accuracy figure confirmed in the abstract.

Technology: Cell2location (Bayesian deep probabilistic model)

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Cell2location was assessed in three tissues (mouse brain, human lymph node, human gut) and reported to give higher sensitivity and resolution than existing tools.
  • Still unknown: The abstract names no conditions under which the claimed higher sensitivity holds and no limitations.
  • AI role: A Bayesian model that resolves fine-grained cell types in spatial transcriptomics data.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41587-021-01139-4&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 1414 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-learned multiscale simulations reveal lipid-dependent RAS organization at membranes

The MuMMI machine-learning multiscale simulation ensemble found that local membrane lipid fingerprints correlate with wild-type KRAS clustering and orientation, with many simulated RAS–RAS interfaces appearing nonspecific. Complementary liposome and diffusion experiments supported lipid-dependent RAS adhesion and dynamics, motivating the authors' conclusion that lipids can mediate RAS multimer organization.

Technology: MuMMI (Multiscale Machine-learned Modeling Infrastructure)

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Complementary liposome and diffusion experiments supported lipid-dependent RAS adhesion and dynamics, motivating the authors' conclusion that lipids can mediate RAS multimer organization.
  • Still unknown: Most specific orientation and interface conclusions are simulation-derived; experimental comparisons provide validation of selected membrane adhesion/diffusion behaviors, not direct demonstration of these mechanisms in intact cells or cancer outcomes. Authors describe absolute diffusion values as semiquantitative and dependent on methods/force fields.
  • AI role: Machine learning selects diverse local membrane patches from a macroscale lipid model to launch and couple atomistic/coarse-grained simulations, producing an ensemble of more than 100,000 simulations. The resulting simulations identified lipid-composition relationships with RAS orientation and multimer organization; selected predictions were compared with surface-plasmon resonance, fluorescence-correlation and other biophysical experiments.

Evidence sources: https://www.pnas.org/content/pnas/119/1/e2113297119.full.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 97 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Learned soundscape features predict tropical-forest species occurrence

Learned soundscape features predicted occurrence for 34 of 39 surveyed bird and herpetofaunal species across a forest-degradation gradient in Sabah. Predictions used the broader acoustic environment rather than direct identification of every species call. Site-level evaluation supports a bounded monitoring result, not proof of causal habitat quality or a replacement for species surveys.

Technology: Convolutional audio embeddings and Gaussian-mixture occurrence models

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Occurrence predictions were compared with field point counts using cross-validation.
  • Still unknown: Transfer to other ecosystems and causal interpretation are not established by this study.
  • AI role: Convolutional audio embeddings and Gaussian-mixture occurrence models supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://nsojournals.onlinelibrary.wiley.com/doi/10.1111/oik.08525

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary publisher metadata, affiliations, abstract and selected Methods/Results; supplements not reviewed..

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

A neural network “hallucinates” new protein structures

The network predicted about 2,000 new sequences as foldable; over 100 were produced in the lab, and for three, nuclear magnetic resonance and X-ray crystallography confirmed the predicted shapes. Only three proteins were characterized in detail; using the method to create useful functions (drugs, enzymes) is a future goal.

Technology: Neural network for structure prediction (hallucination)

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: 129 hallucinated sequences were synthesised, 27 gave monodisperse proteins, and three structures determined by X-ray crystallography or NMR matched the hallucinated models.
  • Still unknown: Only three proteins were characterised in detail, and using the method for useful functions is a future goal.
  • AI role: Neural network hallucinated new protein sequences predicted to fold.

Evidence sources: https://newsroom.uw.edu/news-releases/deep-learning-dreams-new-protein-structures · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1038/s41586-021-04184-w&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 653 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-learning-guided evolution identifies higher-activity sortase A variants

Two machine-learning-guided directed-evolution series for sortase A were run with and without the known high-activity 5M variant in the initial training set. The authors report new variants with higher activity than 5M in a GFP/YFP cross-linking assay. This is not a 2.2–2.5-fold improvement over wild type; the known 5M comparator is already engineered. Feature selection used ability to find 5M and may bias exploration. The related GB1 analysis uses an existing dataset, not a second laboratory discovery. Sequence-level comparison with later sortase work remains open.

Technology: Gaussian-process Bayesian selection model

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Constructed variants were assayed for SrtA-catalyzed GFP/YFP cross-linking; resulting variants exceeded the 5M comparator.
  • Still unknown: The assay is specific; 5M is an already high-activity engineered comparator; feature selection may bias to 5M; independent sequence-level overlap/replication open.
  • AI role: A Gaussian-process ML model ranked variants from initial and second-round data to design subsequent SrtA mutant libraries.

Evidence sources: https://pubs.acs.org/accacs/article/11/23/14615/417029/Machine-Learning-Guided-Library-Design-Cycle-for · https://www.eng.tohoku.ac.jp/news/detail-%2C-id%2C2033.html

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary ACS indexed full article and Tohoku research announcement; methods/results and journal timing..

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

MutCompute guides thermostable Bst DNA polymerase variants

MutCompute predicted substitutions in Bst DNA polymerase; the authors constructed and tested variants in loop-mediated isothermal amplification assays. The Mut235 triple variant showed the highest activity among four tested triple-mutant variants and remained active at 73 °C; its reported melting temperature was 78.1 °C. The evidence is in vitro and does not establish clinical diagnostic performance. The journal article cites a bioRxiv v1 posted 2021-04-15, but that version’s full text was inaccessible in this review; this entry uses the verified journal result and does not claim when the Mut235 result first appeared.

Technology: MutCompute machine-learning amino-acid environment scoring

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The authors assayed ML-predicted Bst polymerase substitutions in LAMP/LAMP-OSD reactions. Mut235 showed the highest activity among four triple-mutant variants tested, supported amplification at 73 °C, and had a measured melting temperature of 78.1 °C versus 75.5 °C for parental Bst-LF in the reported thermal-shift assay.
  • Still unknown: The article cites an earlier bioRxiv preprint dated 2021-04-15, but this check did not retrieve its full v1 text; no claim is made that every final-version assay/result was present in v1. Results are in-vitro amplification assays, not clinical diagnostic validation.
  • AI role: MutCompute machine learning predicted amino-acid substitutions in Bst DNA polymerase; the authors constructed and assayed variants including Mut235.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC9514386/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary PMC full journal text for MutCompute, LAMP assays, University of Texas affiliation and measured variants; primary bioRxiv v1 endpoint was attempted but blocked, so earlier-version result details remain bounded to what the journal and cited metadata support..

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

EVE: evolutionary model predicts disease-variant pathogenicity without labels

EVE, an unsupervised deep generative model trained on evolutionary sequence variation across about 140,000 species, predicted pathogenicity for more than 36 million variants in 3,219 disease genes and classified over 256,000 variants of unknown significance. The authors report it outperforms methods trained on labelled disease data and performs on par with high-throughput experimental predictions. Caveat: the authors state EVE is not a diagnostic test.

Technology: EVE (deep generative evolutionary model, variational autoencoder)

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: EVE's predictions were compared with methods trained on labelled disease data and with high-throughput experimental predictions.
  • Still unknown: The authors state EVE is not a diagnostic test.
  • AI role: An unsupervised deep generative model trained on evolutionary sequence variation across about 140,000 species, predicting pathogenicity for over 36 million variants.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-021-04043-8&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 855 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning prioritizes agonists for two human olfactory receptors

A random-forest classifier trained from molecular descriptors screened five databases containing about 23 million compounds for agonists of human olfactory receptors OR1A1 and OR2W1. Following docking and researcher selection, the paper reports in-vitro activation by ethyl 2-(3-bromophenyl)acetate and methyl 2-(2-methylphenyl)acetate at OR1A1 and 2-ethoxynaphthalene at OR2W1. These are cell-assay results; they do not establish human perception, tissue safety or clinical benefit. The compounds were selected after multiple filters, so the results are not a database-wide hit rate. Prior ML work on related receptors exists; this record is limited to the reported candidates and assays.

Technology: Random-forest classifier; molecular descriptors; homology-model docking

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Selected candidates were tested in vitro; the paper reports two active OR1A1 ligands and one active OR2W1 ligand.
  • Still unknown: No in-vivo human olfaction or clinical effect is established. A 2018 study used ML for related human receptors; candidate-level overlap and priority remain bounded follow-up.
  • AI role: The random-forest classifier screened chemical databases to rank candidates; docking and researcher selection further narrowed compounds for in-vitro validation.

Evidence sources: https://www.mdpi.com/1422-0067/22/21/11546

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary publisher article, article metadata and author-affiliation block reviewed; summary is limited to reported assay results..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning guides fatty acyl-ACP reductase engineering

A machine-learning-guided iterative design–test–learn workflow engineered fatty acyl-ACP reductase variants for microbial fatty alcohol production. Across ten rounds, the study reports over twice the production of the starting natural sequences under tested conditions; the top sequence also had enhanced catalytic rate on palmitoyl-ACP. These are laboratory results, not industrial-scale production. The journal article date is used here; the earlier preprint’s exact posting and result-version differences were not checked.

Technology: Machine-learning-guided protein engineering; Gaussian-process optimization

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The paper reports ten design–test–learn rounds, in-vivo microbial fatty-alcohol production and catalytic-rate characterization on palmitoyl-ACP.
  • Still unknown: Results are under the study’s microbial test conditions and do not establish industrial production; exact preprint-version results remain unchecked.
  • AI role: Machine-learning models iteratively selected reductase sequences for experimental testing and subsequent model updates.

Evidence sources: https://www.nature.com/articles/s41467-021-25831-w

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary Nature Communications indexed full text, abstract, article date, author affiliations; preprint chronology not fully checked..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

Enformer: gene expression predicted from DNA sequence

A Transformer-based model predicts gene expression from 200,000 base pairs of input DNA, over five times more than earlier methods, and is much more accurate at predicting variant effects. Enhancer attribution scores are comparable to methods using experimental data. The post gives no numerical gain and acknowledges that linking enhancers to the genes they regulate remains an open problem.

Technology: Enformer (Transformer)

Family Enformer: profile · models page

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Enformer is compared with prior models on variant-effect prediction and enhancer contribution scores against methods using experimental data, with a rs11644125 case study.
  • Still unknown: Linking enhancers to the genes they regulate is described as a major unsolved problem, and many disease variants are spuriously correlated.
  • AI role: Transformer model predicted gene expression from 200,000 base pairs of DNA.

Evidence sources: https://deepmind.google/blog/predicting-gene-expression-with-ai/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1038/s41592-021-01252-x&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 1533 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ECNet-guided selection of higher-order TEM-1 variants

ECNet integrated evolutionary context with learned sequence representations to prioritize previously untested higher-order TEM-1 variants. The authors constructed and measured 37 selected variants under ampicillin selection, using relative resistance as an enzyme-fitness proxy. The assays do not imply clinical treatment outcomes, and the reported group of 37 should not be read as showing that every variant improved.

Technology: ECNet evolutionary-context-integrated deep-learning framework

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Thirty-seven previously untested higher-order TEM-1 variants selected with ECNet were constructed and experimentally measured under ampicillin selection against wild-type and study controls. The study uses relative resistance as an enzyme-fitness proxy.
  • Still unknown: The bounded entry does not imply that all 37 variants improved equally, does not establish clinical outcomes, and makes no claim of independent sequence novelty beyond the authors’ screened literature.
  • AI role: ECNet integrated evolutionary context and a learned representation to prioritize higher-order TEM-1 variants for experimental construction and ampicillin-selection measurements.

Evidence sources: https://www.nature.com/articles/s41467-021-25976-8 · https://zhaogroup.web.illinois.edu/publications/HZ370.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary Nature indexed article and author-hosted full text; official bioRxiv API checked and identified the previously associated preprint DOI as a different paper..

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Cross-species machine learning identifies transcription factors that negatively regulate plant nitrogen-use efficiency

Evolutionarily informed XGBoost and GENIE3 analyses of conserved nitrogen-response expression patterns prioritized transcription factors associated with nitrogen-use efficiency in Arabidopsis and maize. Loss-of-function tests found higher NUE in Arabidopsis myb33 and tcp2 mutants under both low- and high-nitrogen conditions, and in six other tested TF mutants specifically under low nitrogen; a maize nfya3 mutant also showed higher nitrogen-use efficiency under sufficient nitrogen in field tests.

Technology: Evolutionarily informed XGBoost models and GENIE3 random-forest gene-regulatory-network inference

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Loss-of-function tests found higher NUE in Arabidopsis myb33 and tcp2 mutants under both low- and high-nitrogen conditions, and in six other tested TF mutants specifically under low nitrogen; a maize nfya3 mutant also showed higher nitrogen-use efficiency under sufficient nitrogen in field tests.
  • Still unknown: The machine-learning prediction scores are not themselves causal evidence; this paper adds functional evidence from specific mutants. Some tested maize candidate roles were assessed through Arabidopsis homologs, and the maize NFYA3 phenotype described here was under sufficient-N field conditions, not a general yield benefit across conditions.
  • AI role: The models ranked conserved nitrogen-responsive genes and transcription factors for contribution or network connectivity to NUE; researchers selected the candidates and tested loss-of-function mutants in Arabidopsis and maize, including field-grown maize.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC8463701/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 109 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning identifies moth-receptor agonists that attract larvae

Random-forest and k-nearest-neighbors QSAR models trained on known odorant responses prioritized natural-product candidates for two Spodoptera littoralis odorant receptors. The paper reports receptor responses for selected compounds and Y-maze preference for seven strong agonists; all seven were preferred at 10⁻² dilution, while three retained preference at 10⁻³. The percentages of receptor-active candidates apply to selected test sets and are not field trap accuracy. The fly-neuron assay is heterologous, and larval choice is laboratory evidence, not crop-protection efficacy or ecological safety. This extends the same group’s 2020 receptor-ligand study; count the 2021 behavior result as one study, not each molecule or a separate model discovery.

Technology: Random-forest QSAR and k-nearest-neighbors classifier

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Selected ML-prioritized receptor ligands were tested electrophysiologically and in larval Y-maze assays; the paper reports the stated dose-dependent preference outcomes.
  • Still unknown: No field efficacy, human health benefit, broad ecological selectivity or global priority is established. The study extends earlier same-group work; its compounds and behavior result are bounded to the reported assays.
  • AI role: The models ranked candidate natural-product ligands for SlitOR24/SlitOR25; researchers selected the targets and candidates and performed the receptor and larval behavioral experiments.

Evidence sources: https://link.springer.com/article/10.1007/s00018-021-03919-2

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary publisher article, article metadata and author-affiliation block reviewed; summary is limited to reported assay results..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning-assisted directed evolution with in silico combinatorial screening

A path-independent machine-learning protocol screens full combinatorial protein libraries in silico, with training-set design that avoids near-zero-fitness variants. On a four-site epistatic fitness landscape the optimized protocol reached the global fitness maximum up to 81-fold more often than single-step greedy optimization. Caveat: validation is on a combinatorial fitness landscape dataset, not a new wet-lab evolution campaign.

Technology: Machine-learning-assisted directed evolution (in silico combinatorial screening)

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: The ML-assisted directed evolution protocol was compared with single-step greedy optimization on a four-site epistatic fitness landscape.
  • Still unknown: Not reported in the abstract read.
  • AI role: A path-independent machine-learning protocol that screens full combinatorial libraries in silico, with training-set design avoiding near-zero-fitness variants.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1016/j.cels.2021.07.008&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 220 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Learned models reveal nonmonotonic bacterial patterns in boreal lakes

Models fitted to observations from 70 Scandinavian lakes identified nonmonotonic associations between browning indicators and bacterial community diversity. XGBoost explained up to about 45% of community-composition variation in the analyzed data. The fitted ridges are not demonstrated causal tipping points; translating this spatial gradient into future temporal trajectories requires further validation.

Technology: XGBoost, random forests and feed-forward neural networks

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Models analyzed field-sampled microbial communities with held-out evaluation.
  • Still unknown: Covariation, causal mechanisms and long-term temporal predictions remain unresolved.
  • AI role: XGBoost, random forests and feed-forward neural networks supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://www.nature.com/articles/s43705-021-00038-8 · https://www.nature.com/articles/s43705-021-00038-8.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary publisher PDF abstract and selected Methods/Results/limitations; indexed publisher publication metadata. Supplements not reviewed..

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AlphaFold Protein Structure Database: first 350,000 models

Predicted structures for the human proteome (about 20,000 proteins) and 20 other organisms were published, over 350,000 models in total, hosted free by EMBL-EBI. DeepMind announced an expansion to over 100 million structures. These are predictions that complement experimental methods. Note: the page header says July 22, 2022, but the text calls CASP14 “last December” and the 2022 post refers to it as a 2021 release; date corrected to 2021.

Technology: AlphaFold 2

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Predictions cover 98.5% of human proteins; 58% of residues have a confident prediction and 36% very high confidence. Case studies illustrate possible hypotheses.
  • Still unknown: The structures are predictions; the abstract does not report experimental confirmation of new structures.
  • AI role: AlphaFold 2 applied at proteome scale, with confidence metrics built on the model.

Evidence sources: https://doi.org/10.1038/s41586-021-03828-1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 3058 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 5/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ProGen generates artificial proteins with enzyme activity

A protein language model generated synthetic lysozymes: two of the five characterized in cells degraded bacterial cell walls with activity comparable to egg-white lysozyme. Laboratory tests and crystallography support the result, but do not demonstrate clinical applications. The preprint dates to July 18, 2021; the Nature Biotechnology publication and UCSF announcement are from January 2023.

Technology: ProGen

Family ProGen: profile · models page

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: Five model-generated lysozyme families were experimentally characterised, with activity and catalytic efficiency compared with natural lysozymes, and crystal structures reported.
  • Still unknown: Not reported in the abstract read.
  • AI role: Protein language model trained on over 280 million sequences generated artificial lysozymes.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1101/2021.07.18.452833&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

RoseTTAFold: accessible prediction of protein structures

A three-track neural network (sequence, residue interactions, and 3D structure) computes a structure in about ten minutes on a single gaming computer. Baker’s group predicted hundreds of new structures, including little-known proteins in the human genome. The web server received over 4,500 requests in the previous month. The page does not state limitations.

Technology: RoseTTAFold

Family RoseTTAFold: profile · models page

Publication: Peer-reviewed paper · Test context: Test context unspecified in annotation. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: not reported
  • Tested: The page describes the tool's speed and hundreds of new structures but does not describe how predictions were checked.
  • Still unknown: Not reported.
  • AI role: Three-track neural network predicted protein structures.

Evidence sources: https://www.ipd.uw.edu/2021/07/rosettafold-accurate-protein-structure-prediction-accessible-to-all/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1126/science.abj8754&resultType=core&format=json

The legacy label “Not reported” does not distinguish information absent from reviewed material from information not assessed. It is not a scientific finding of absence.

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 5653 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 1/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning reveals a nitrogen-assimilation strategy in Bacteroides

A microbiome metabolomics study used random-forest models to associate bacterial metabolite profiles with taxonomy and found a previously undescribed nitrogen-assimilation pattern in Bacteroides/Bacteroidetes. Follow-up growth experiments tested 60 Bacteroidetes strains with asparagine or glutamine as the sole nitrogen source and supported the predicted phenotype; the study also reports a distinct reduced-sulfur utilization result in B. thetaiotaomicron.

Technology: Random-forest models over metabolomics profiles

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Follow-up growth experiments tested 60 Bacteroidetes strains with asparagine or glutamine as the sole nitrogen source and supported the predicted phenotype; the study also reports a distinct reduced-sulfur utilization result in B. thetaiotaomicron.
  • Still unknown: The study's abstract reports machine-learning discovery of a new metabolism and the figure captions document follow-up culture tests; the article's full text is paywalled in this session, so the exact mechanistic pathway is not asserted. Mouse and culture evidence do not establish a human health effect.
  • AI role: The models identified metabolite features that distinguished taxa and highlighted a previously undescribed Bacteroides metabolic phenotype; researchers then tested bacterial growth under defined nitrogen sources. Comparative-genomics pathways remained candidate explanations rather than proven pathways.

Evidence sources: https://www.nature.com/articles/s41586-021-03707-9

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 417 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

iTARGEX analysis identifies ELG1 as a regulator of yeast transcription buffering and four protein-turnover regulators

iTARGEX applied a mixture-regression model to gene-expression profiles from 1,484 single-gene deletion yeast mutants, associating transcriptomic changes with S-phase transcriptional buffering and global protein turnover. Experiments supported ELG1 as a previously unrecognized regulator of transcriptional buffering during S phase; the phenotype was reproduced in a second genetic background and rescued by a PCNA-D150E mutation. Follow-up protein-turnover measurements showed altered turnover in MRN1, UPF3, RAM1 and SSN3 deletion strains.

Technology: iTARGEX, a finite-mixture regression framework fitted by expectation maximization to single-gene-deletion transcriptomic data

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Experiments supported ELG1 as a previously unrecognized regulator of transcriptional buffering during S phase; the phenotype was reproduced in a second genetic background and rescued by a PCNA-D150E mutation.
  • Still unknown: The study used budding yeast and pre-existing deletion-mutant transcriptome profiles, followed by experiments on selected candidates. ELG1's role in S-phase transcription buffering was reproduced in a second yeast genetic background and rescued by PCNA-D150E; the protein-turnover effects concern yeast deletion strains. These results do not establish corresponding human functions.
  • AI role: iTARGEX associated gene-expression profiles from deletion strains with biological traits without requiring prior gene annotations and nominated candidate regulators. The researchers then performed S-phase transcription and protein-turnover experiments to test the model-prioritized candidates, supporting ELG1's role in transcriptional buffering and altered turnover for four deletion strains.

Evidence sources: https://academic.oup.com/nar/article/49/13/7318/6312755

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 32 of works of the same type, year and subfield (OpenAlex work metrics, 2 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Satellite-image clustering guides discovery of four late Miocene fossil sites in Mozambique’s Urema Rift

Unsupervised k-means clustering of seven Landsat 8 spectral bands highlighted a small set of priority areas for paleontological field survey in Gorongosa National Park. The 2018 campaign documented four new late Miocene fossil localities; three lay within the model’s target cluster and the fourth was less than 90 m away in the same gully system. Together, the sites expose the first reported estuarine coastal-forest fossil setting in the East African Rift System, filling a regional paleoenvironmental gap.

Technology: Unsupervised k-means clustering of Landsat 8 imagery

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Unsupervised k-means clustering of seven Landsat 8 spectral bands highlighted a small set of priority areas for paleontological field survey in Gorongosa National Park.
  • Still unknown: This was a pilot study in a 36 km² area and the model had notable false positives: reported precision was 55.56%, despite 84.6% overall accuracy. Three of four new localities overlapped the selected cluster; the fourth was less than 90 m away and was found by surveying the same gully system, so model localization for that site is indirect. The findings are four new fossil localities, not four new hominin taxa; the reported estuarine coastal-forest interpretation is a paleoenvironmental inference.
  • AI role: The authors clustered seven spectral bands without pre-labeled land-cover classes, compared clusters with known fossil localities, and selected the cluster containing most known sites as a predictive target for field prospecting. This decision-support workflow directed the 2018 field survey that documented the new localities; a random forest was separately used to estimate spectral-band importance.

Evidence sources: https://peerj.com/articles/11573/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 71 of works of the same type, year and subfield (OpenAlex work metrics, 14 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

H01: reconstruction of one cubic millimeter of human cerebral cortex

A sample of about 1 mm³ of human cortex was imaged at 4 nm across about 5,300 sections and automatically reconstructed: 1.4 petabytes, tens of thousands of neurons, 130 million annotated synapses, and 104 manually corrected cells. The sample is about one-millionth of a brain; reconstruction is imperfect and initial findings are preliminary.

Technology: Automated segmentation of electron microscopy images

Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: retrospective or benchmark
  • Tested: The automated reconstruction was checked by manual proofreading of about 104 cells; the abstract says connections were validated.
  • Still unknown: Automated reconstruction is imperfect, the sample is about one-millionth of a brain, and findings are preliminary.
  • AI role: Automated machine-learning pipeline reconstructed neurons and synapses from electron microscopy images.

Evidence sources: https://research.google/blog/a-browsable-petascale-reconstruction-of-the-human-cortex/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1101/2021.05.29.446289&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AI-generated cell-penetrating peptide selected with molecular-dynamics simulations

An RNN generated peptide sequences and molecular-dynamics-based scoring prioritized Pep-MD for synthesis. In cell assays, the authors report less toxicity than TAT and stronger cell permeability under the tested conditions. The paper proposes a membrane-entry mechanism based partly on simulated structures; these cell results do not establish therapeutic efficacy.

Technology: Recurrent neural network sequence generator with molecular-dynamics-based selection score

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The selected synthesized peptide was tested in cell-viability, confocal-imaging and FACS permeability assays against TAT.
  • Still unknown: Cell-based assays and a single comparator; no clinical or therapeutic efficacy evidence; mechanistic interpretation partly depends on modeled structures.
  • AI role: An RNN generated candidate peptides; a simulation-derived score prioritized one sequence for synthesis and experimental cell testing.

Evidence sources: https://www.nature.com/articles/s41598-021-90245-z.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary publisher full PDF, methods/results, publication information and author affiliations..

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning guides refinement of DNA aptamers binding NGAL

Neural models trained on particle-display affinity-screening data guided sequence mutation and candidate nomination, followed by synthesis and experimental affinity testing. The authors report a 23-nucleotide truncation of candidate G13 with measured dissociation constant 1.5 nM, compared with about 8 nM for the full-length parent under their assay. This is a binding result, not demonstrated diagnostic or therapeutic efficacy. The screening set is not 187,499 independently validated agents; sequence-level prior art remains open.

Technology: Neural-network models trained on particle-display affinity data

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The study reports particle-display screening and measured NGAL binding, including full Kd for truncated and parent G13.
  • Still unknown: Binding assay only; no diagnostic/therapeutic outcome; one holdout set used for tuning and sequence novelty requires further review.
  • AI role: Trained neural models extrapolated from experimentally measured sequences to nominate and iteratively refine higher-affinity DNA aptamers and truncations.

Evidence sources: https://www.nature.com/articles/s41467-021-22555-9

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary publisher indexed full text, article dates, methods/results, affiliations and disclosed limitations..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

An integrated Arabidopsis regulatory network predicts 13 previously unlinked reactive-oxygen-stress regulators

Researchers integrated DNA-motif, chromatin-accessibility, TF-binding and expression evidence with supervised learning to build an Arabidopsis gene-regulatory network of 1,491 transcription factors and 1.7 million TF–target interactions. The network predicted candidate reactive-oxygen-species (ROS) regulators; loss- and gain-of-function phenotyping supported 13 regulators not previously linked to ROS or stress responses. Newly tested examples included WRKY55, WRKY47, AT2G33710 and AT5G57150. The paper also reports that 75% of its ROS-predicted TFs had ROS or physiological stress functions supported by either this study or prior evidence.

Technology: Supervised-learning integrated gene-regulatory network (iGRN) built from motif, open-chromatin, transcription-factor-binding and expression-based networks

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Newly tested examples included WRKY55, WRKY47, AT2G33710 and AT5G57150.
  • Still unknown: The iGRN integrated existing experimental data and the ROS findings were tested in Arabidopsis stress assays, not crop-yield or field-resilience trials. The 75% figure combines prior literature with new phenotypic tests; it is not a rate of independently replicated new discoveries. The new TF functions are limited to tested oxidative-stress conditions and mutant lines.
  • AI role: The supervised integration model inferred TF–target regulatory links across multiple experimental input networks and predicted candidate TF functions, including ROS-response regulators. The investigators tested predicted TF candidates using Arabidopsis loss- or gain-of-function lines and ROS-specific phenotyping; these experiments supported 13 TFs without prior ROS/stress associations.

Evidence sources: https://www.nature.com/articles/s41477-021-00894-1.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 91 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

trRosetta releases over 6,300 predicted protein-family structures

Over 6,300 protein structures predicted with deep learning are available in Pfam to explore families without a known structure. These are computational models with estimated quality, not newly experimentally determined structures; this result is distinct from RoseTTAFold and the AlphaFold Database.

Technology: trRosetta

Family trRosetta: profile · models page

Publication: Announcement only · Test context: Test context unspecified in annotation. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: not reported
  • Tested: Pfam models are scored with lDDT (0.6 or higher considered reasonable); the article describes no independent validation.
  • Still unknown: The article does not state limitations and says more structures are to follow.
  • AI role: Deep-learning trRosetta predicted structures for protein families.

Evidence sources: https://www.ipd.uw.edu/2021/03/trrosetta-reveals-structures-for-every-protein-family/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 1/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AI-photoswitchable CRISPRi screening identifies new regulators of TFEB localization during starvation

An image-classification-guided pooled CRISPRi screen mapped factors that affect TFEB movement from nucleus to cytosol during prolonged starvation. The screen identified multiple candidate genes, and follow-up assays independently validated 21 of 64 neural-network-called hits; experiments on CREB5 further linked it to TFEB retention and starvation-associated autophagy/lysosomal protein expression.

Technology: AI-photoswitchable screening (AI-PS), using SVM classifiers and a convolutional neural network

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The screen identified multiple candidate genes, and follow-up assays independently validated 21 of 64 neural-network-called hits; experiments on CREB5 further linked it to TFEB retention and starvation-associated autophagy/lysosomal protein expression.
  • Still unknown: Only 21 of 64 neural-network-called TFEB hits were independently validated in secondary assays (about 32%); candidate effectors should not be treated as all confirmed. Results use cultured U2OS cells and prolonged starvation, and do not establish a physiological role in an organism. CREB5 received targeted follow-up but the overall screen has a modest validation rate.
  • AI role: SVM and CNN models classified single-cell images by subcellular protein localization; the system photoactivated cells with the target phenotype for FACS isolation and gRNA sequencing in pooled CRISPRi screening. AI enabled the genome-scale genotype-to-subcellular-phenotype screen.

Evidence sources: https://rupress.org/jcb/article/220/2/e202006180/211694/Image-based-pooled-whole-genome-CRISPRi-screening

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 86 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

MIT

Language models identify viral regions less prone to immune escape

Analysis of influenza, HIV, and coronavirus sequences identifies regions of surface proteins less likely to mutate to escape immune responses, including the S2 subunit of spike. These are computational leads for studying possible vaccine targets, not validated vaccines.

Technology: Protein-sequence language model

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Language models trained on about 60,000 HIV, 45,000 influenza and 4,000 coronavirus sequences were checked against earlier findings on influenza and HIV; the SARS-CoV-2 analysis is described as not yet peer-reviewed.
  • Still unknown: Flagged SARS-CoV-2 sequences still need investigation, and it is unknown how fast SARS-CoV-2 mutates, so how long current vaccines stay effective is unclear.
  • AI role: Adapted language-model methods to viral protein sequences to predict surface regions more or less likely to enable immune escape, as candidate vaccine targets.

Evidence sources: https://news.mit.edu/2021/model-viruses-escape-immune-0114 · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1126/science.abd7331&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 383 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Learned resistance models expose tested enzyme-substrate activities

Interrogating models trained on genome and resistance data nominated enzyme-substrate relationships for experimental testing. The authors report 22 previously unknown activities validated in their assay system. This is one mechanism study; it does not establish the prevalence or clinical contribution of each relationship in patients. Model performance varied substantially across organism and drug datasets.

Technology: Antibiotic-specific logistic-regression classifiers

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Predicted relationships received targeted experimental validation.
  • Still unknown: Clinical prevalence, generalization and independent replication remain unestablished.
  • AI role: Antibiotic-specific logistic-regression classifiers supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC8115898/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary article date, affiliations, selected logistic-regression Methods and experimental Results; supplements not reviewed..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

AlphaFold wins CASP14 for protein folding

AlphaFold achieved a median GDT score of 92.4 across the target set (about 1.6 Å mean error) and 87.0 on the hardest targets, which organizers judged a solution to the problem of single-protein structures. DeepMind notes that protein complexes and interactions with other molecules remain open challenges.

Technology: AlphaFold 2

Family AlphaFold: profile · models page

Publication: Announcement only · Test context: prospective or independent check. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: prospective or independent check
  • Tested: Blind assessment at CASP14 on recently determined protein structures: median GDT of 92.4 across targets and 87.0 on the hardest free-modelling targets, judged by the organizers as a solution to single-protein structure prediction.
  • Still unknown: Protein complexes, interactions with DNA, RNA or small molecules, and full side-chain placement remain open; not every predicted structure will be perfect.
  • AI role: Attention-based neural network using evolutionary sequence alignments and residue-pair representations to predict 3D protein structure from sequence.

Evidence sources: https://deepmind.google/blog/alphafold-a-solution-to-a-50-year-old-grand-challenge-in-biology/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 5/5 · Validation 4/5 · Novelty 5/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Akita: predicting 3D genome folding from DNA sequence

Akita, a convolutional network, predicts chromatin contact maps for about 1 Mb of DNA sequence. On held-out data it reaches a Pearson R of 0.61 and Spearman R of 0.56, and it learns an orientation-specific grammar for CTCF binding sites. Caveat: the predictions are computational; the paper applies them to in silico mutagenesis, eQTL interpretation and structural-variant prediction.

Technology: Akita (convolutional neural network, Basenji-based trunk)

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Akita's contact-map predictions were evaluated on held-out data and applied in silico to mutagenesis, eQTL interpretation and structural-variant prediction.
  • Still unknown: Not reported in the abstract read; the predictions are computational and no experimental validation is described there.
  • AI role: A convolutional network predicting chromatin contact maps from about 1 Mb of DNA sequence, which learned an orientation-specific CTCF binding grammar.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41592-020-0958-x&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 309 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

OPEX-guided experiments reveal bacterial cross-stress responses

Learned models selected bacterial stress-combination experiments and incorporated their measurements. The study reported cross-stress protection and vulnerability patterns, with follow-up supporting roles for stress-response machinery. This is a laboratory microbiology result, not a clinical resistance prediction or evidence of treatment efficacy.

Technology: OPEX machine-learning experimental selection

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Selected experiments and subsequent biological validation tested stress responses.
  • Still unknown: Transfer beyond the studied organism and conditions, clinical outcomes and independent replication were not assessed.
  • AI role: OPEX machine-learning experimental selection supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/33024104/ · https://escholarship.org/content/qt5cc185t3/qt5cc185t3.pdf?t=qlp51p · https://tagkopouloslab.ucdavis.edu/news/2020/10/new-article-in-nature-communication/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary PubMed abstract/date/affiliations, indexed deposited publisher-paper abstract and author laboratory announcement..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Random-forest screening identifies experimentally active anti-CRISPR families

A learned model ranked candidate protein families for experimental follow-up. Testing supported two previously unknown anti-CRISPR proteins in the studied system. The much larger predicted family catalog remains computational and is not counted as experimentally validated discoveries. Assay activity does not establish native-host function or improved genome-editing performance.

Technology: Extremely randomized forest of genomic-context features

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Selected predicted candidates received functional laboratory tests.
  • Still unknown: Broader family validation, native-host activity and practical utility remain unestablished.
  • AI role: Extremely randomized forest of genomic-context features supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://www.nature.com/articles/s41467-020-17652-0.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary published PDF abstract, affiliations and model-training Methods; experimental scope checked against prior primary review..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

An attention-based neural network generates functional secretion signal peptides

An attention-based Transformer trained on signal-peptide data from Swiss-Prot generated candidate secretion sequences. Experimental testing in an industrial Bacillus subtilis strain found that model-generated signal peptides directed secreted enzyme activity competitive with industrially used signal peptides. The result establishes functional candidates in the reported host and assay; it does not establish broad performance across proteins or production hosts. The primary institutional record gives the article publication date as 2020-07-10.

Technology: Attention-based neural network (Transformer) trained on Swiss-Prot signal peptide sequences

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The primary abstract reports experimental testing in an industrial Bacillus subtilis strain and secreted enzyme activity competitive with industrially used signal peptides.
  • Still unknown: The bounded abstract-level claim does not quantify all peptide/host combinations or demonstrate generalization across unrelated proteins and production conditions.
  • AI role: An attention-based Transformer trained on Swiss-Prot generated signal-peptide sequences, which were then tested experimentally for protein secretion.

Evidence sources: https://authors.library.caltech.edu/records/5dqa2-44d08

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary Caltech institutional record with the article abstract, author contribution statement, publication date, and affiliation; the linked PDF download was unavailable in this session..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AcRanker-guided discovery of AcrIIA20 and AcrIIA21 anti-CRISPR proteins

AcRanker, an XGBoost ranking model using amino-acid sequence features, ranked proteins from prophage regions in self-targeting Streptococcus genomes. Experimental testing identified two previously unknown type II-A anti-CRISPRs: AcrIIA20 (ML1), which strongly inhibited Streptococcus iniae Cas9 and weakly inhibited S. pyogenes Cas9, and AcrIIA21 (ML8), which inhibited S. pyogenes, S. aureus and S. iniae Cas9 with lower potency.

Technology: AcRanker, an XGBoost pairwise ranking model trained on amino-acid composition and grouped di- and tri-mer features

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Experimental testing identified two previously unknown type II-A anti-CRISPRs: AcrIIA20 (ML1), which strongly inhibited Streptococcus iniae Cas9 and weakly inhibited S. pyogenes Cas9, and AcrIIA21 (ML8), which inhibited S. pyogenes, S. aureus and S. iniae Cas9 with lower potency.
  • Still unknown: The authors selected 10 candidates for follow-up, successfully purified nine and reported three validated candidates across their experiments, including ML3/AcrIIA12, which had also been identified separately as an LmoCas9 inhibitor during the study. AcrIIA20 and AcrIIA21 were roughly tenfold less potent against SpyCas9 than the established AcrIIA4 control; tested potency and host range are assay-specific. The 2025 Evo paper reports distinct generated proteins EvoAcr1–5 and separate assays; shared anti-CRISPR function does not make these identical discoveries.
  • AI role: The model ranked candidate proteins from predicted prophage regions for anti-CRISPR likelihood; the researchers selected candidates from those rankings and biochemically tested them. It directly prioritized the proteins later named AcrIIA20 and AcrIIA21, whose Cas9 inhibition was measured in vitro.

Evidence sources: https://academic.oup.com/nar/article/48/9/4698/5819938

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 125 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Learned genomic models predict Baltic microbial niche placement

Models trained on functional-gene profiles predicted microbial distribution along salinity, depth and sampling-fraction gradients in Baltic Sea observations. The study links genomic repertoire to ecological niche placement within this ecosystem. It does not demonstrate that associated genes causally determine habitat or that the models transfer elsewhere. The genome clusters are not separately counted discoveries.

Technology: Genome-feature regression and tree-based learning

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Predictions were compared with observed distributions in the sampled ecosystem.
  • Still unknown: Causal effects, field manipulation and cross-ecosystem transfer were not established.
  • AI role: Genome-feature regression and tree-based learning supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://www.nature.com/articles/s42003-020-0856-x · https://epic.awi.de/52845/1/Alneberg_et_al_2020.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Indexed primary publisher abstract, date and affiliations, plus institutional publisher-PDF metadata; Methods review retained from frozen research..

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

Predicted structures of SARS-CoV-2 proteins

DeepMind openly shared AlphaFold predictions, without waiting for review, for six poorly studied viral proteins (membrane, ORF3a, Nsp2, Nsp4, Nsp6, and the C-terminal domain of papain-like protease). They were not experimentally verified; an ORF3a experimental structure deposited in the PDB in June 2020 matched the second predicted model well.

Technology: AlphaFold

Family AlphaFold: profile · models page

Publication: Announcement only · Test context: prospective or independent check. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: prospective or independent check
  • Tested: Predictions compared with experimental structures: the spike prediction matched the PDB structure, and the second-ranked ORF3a model agreed well with later experimental work.
  • Still unknown: Predictions were not experimentally verified overall; the system was still in development and accuracy is not certain; released before peer review; the top ORF3a model misplaced some transmembrane helices.
  • AI role: AlphaFold free-modelling predicted structures for six understudied SARS-CoV-2 proteins, with per-residue confidence scores.

Evidence sources: https://deepmind.google/blog/computational-predictions-of-protein-structures-associated-with-covid-19/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 4/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Interpretable multi-omics machine learning reveals unexpected drug mechanisms in Huntington disease models

An interpretable machine-learning model integrated transcriptomic, epigenomic, metabolomic and proteomic data from Huntington-disease cell models to identify previously unknown mechanisms for protective compounds. The analysis predicted autophagy activation for one compound group and mitochondrial-respiration inhibition for another; follow-up imaging, biochemical and energetics assays experimentally confirmed these effects in the cell model.

Technology: Interpretable network-optimization model integrating multi-omics over a molecular-interaction network

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The analysis predicted autophagy activation for one compound group and mitochondrial-respiration inhibition for another; follow-up imaging, biochemical and energetics assays experimentally confirmed these effects in the cell model.
  • Still unknown: Findings were experimentally established in STHdhQ111 murine striatal cells, with autophagy effects also checked in human SH-SY5Y and HEK293 cell lines. The paper notes that the model prioritizes pathways but does not resolve every causal step or demonstrate therapeutic efficacy in patients.
  • AI role: The model integrated multiple omics and prioritized cellular pathways as candidate mechanisms of action; researchers tested those predictions with cellular imaging, biochemical assays and energetics measurements. It revealed unexpected mechanisms for compounds with unrelated known targets.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC6976599/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 76 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Open the filtered catalog for charts and filters, or browse all organizations and model families.