AI Science Tracker

AI discoveries in 2023

122 discoveries announced in 2023

Discoveries

Universidad ECCI

Machine learning screens cinnamic-acid derivatives for Leishmania enzyme inhibition

Two random-forest classifiers and docking ranked an in-house set of cinnamic-acid derivatives; selected compounds were tested against recombinant L. major DHFR-TS. The paper reports IC50 values of 7.5, 6.1 and 10.1 µM for three named natural products versus 1.4 µM for methotrexate. These are enzyme assays, not whole-parasite treatment results.

Technology: Random-forest classifiers; structure-based 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: Primary article reports recombinant DHFR-TS inhibition for selected compounds and comparator.
  • Still unknown: Methods/results disagree on number of compounds; name/formula spellings and exact prior art need checking.
  • AI role: RF models and docking prioritize a small compound set for biochemical enzyme assays.

Evidence sources: https://doi.org/10.3390/molecules29010179

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: PMC-hosted primary full text; publisher primary article page. Material checked for the specific claims stated in the record..

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 closed-loop synthesis experimentally realizes unreported dye molecules

A closed-loop molecular discovery platform combined machine-learning property prediction, automated synthesis and characterization, and iterative model updates to explore dye-like molecules. In the published Science abstract, the team reports experimentally realizing 294 previously unreported molecules across three design–make–test–analyze cycles on four rarely reported scaffolds, while also identifying six top-performing molecules in a second case study using pre-existing chemical-space examples.

Technology: Autonomous molecular-discovery platform integrating generative and property-prediction machine-learning models, Bayesian optimization, and synthesis-planning software with automated chemical synthesis and characterization

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 published Science abstract, the team reports experimentally realizing 294 previously unreported molecules across three design–make–test–analyze cycles on four rarely reported scaffolds, while also identifying six top-performing molecules in a second case study using pre-existing chemical-space examples.
  • Still unknown: This is molecular discovery in dye-like chemical systems, not a validated drug or biological effect. The 294 number is from the published Science abstract; an earlier ChemRxiv preprint describes 312 unreported molecules, so version-specific counts differ and the published figure is used here. The measurements addressed absorption wavelength, lipophilicity and photo-oxidative stability; no environmental, toxicological or commercial benefit is claimed. Primary-source abstract and author-lab description checked; independent replication not assessed.
  • AI role: Generative ML proposed candidate dye-like structures; property-prediction models ranked molecules against absorption, lipophilicity, and photo-oxidative-stability targets; automated multi-step synthesis and measurements generated training data for model retraining and the next cycles. The closed-loop process converted model-guided suggestions into synthesized and characterized compounds.

Evidence sources: https://doi.org/10.1126/science.adi1407

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, 195 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 discovers D-amino-acid mixtures that boost anti-biofilm antibiotic cocktails

A machine-learning-guided high-throughput screen identified a five-component D-amino-acid mixture (“D-mix”) with about 78% anti-biofilm activity against Pseudomonas aeruginosa at 100 μM. Combining D-mix with half the lowest effective gentamicin dose retained about 91% bacterial-growth inhibition and about 93% anti-biofilm efficacy in vitro. In a small mouse lung-infection model, gentamicin/D-mix treatment reduced measured P. aeruginosa burden by about 95% in blood and 91% in lung tissue at 72 hours.

Technology: High-throughput screening coupled to machine-learning regression and iterative global optimization of D-amino-acid mixture ratios

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 small mouse lung-infection model, gentamicin/D-mix treatment reduced measured P. aeruginosa burden by about 95% in blood and 91% in lung tissue at 72 hours.
  • Still unknown: The in-vivo experiment used a small mouse P. aeruginosa infection model, not human treatment; study-specific bacterial counts and efficacy are reported at 72 hours. The composition is a selected mixture of five D-amino acids, with authors noting the full synergism mechanism remains unclear. D-mix alone was not bactericidal under the reported in-vitro conditions; effects are combination- and assay-dependent.
  • AI role: Experimental measurements of mixture biofilm effects were used to train and iteratively update ML regression models, which guided mixture selection over design rounds. The selected D-mix was then evaluated with different antibiotics in vitro; the lead gentamicin/D-mix combination was tested in a P. aeruginosa lung-infection mouse model.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/38126652/

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, 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.

MIT

A new class of antibiotic candidates against MRSA

A deep-learning model trained on about 39,000 compounds screened about 12 million commercial compounds; of about 280 purchased and tested, two from the same class reduced MRSA burden tenfold in two mouse models, with very low toxicity to human cells. Results from plates and mice; toxicity was partly predicted computationally.

Technology: Deep learning with chemical-structure explainability

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 283 compounds were bought and tested empirically; one structural class reduced bacterial burden in mouse skin and systemic thigh infection models, and cytotoxicity to human cells was measured.
  • Still unknown: The page does not list specific study limitations; next steps are chemical property analysis and potential clinical use.
  • AI role: Graph neural network ensembles trained on about 39,000 compounds with measured antibacterial and human-cell toxicity predicted both properties for about 12 million compounds and flagged candidates for purchase and testing.

Evidence sources: https://news.mit.edu/2023/using-ai-mit-researchers-identify-antibiotic-candidates-1220 · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-023-06887-8&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, 538 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.

Coscientist: autonomous chemical research with language models

A GPT-4-guided system designs, plans and runs experiments using web search, documentation, code and laboratory automation, performing six tasks including optimization of palladium cross-coupling reactions. The abstract does not make clear which tasks used real hardware; only the abstract was reviewed.

Technology: Coscientist (GPT-4)

Family GPT-4: 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 system was run on six tasks, including a palladium-catalysed cross-coupling reaction optimization that the abstract reports as successful.
  • Still unknown: The abstract does not state which tasks used physical hardware and which used simulation, and states no limitations.
  • AI role: GPT-4 planned and executed experiments by using web search, documentation, code execution and experimental automation tools.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-023-06792-0&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: 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, 1811 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.

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.

A machine-learning-selected P4 serum panel detects hepatocellular carcinoma and signals cirrhosis-to-cancer progression

A staged serum-proteomics study across 1,002 individuals used random-forest machine learning to select a four-marker panel, P4 (HABP2, CD163, AFP and PIVKA-II). In an independent validation set, the panel distinguished hepatocellular carcinoma from cirrhosis (AUC 0.979; sensitivity 0.925; specificity 0.915) and from healthy controls (AUC 0.992). In a prospective external cohort of cirrhosis patients followed for HCC, it predicted conversion with AUC 0.890 and median lead time of 11.4 months before imaging diagnosis.

Technology: Random-forest machine-learning model for serum-protein panel selection and HCC classification

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 an independent validation set, the panel distinguished hepatocellular carcinoma from cirrhosis (AUC 0.979; sensitivity 0.925; specificity 0.915) and from healthy controls (AUC 0.992). The prospectively collected cirrhosis cohort was split into 177 training and 76 validation participants for predictive-model construction; collection timing alone does not establish a prospective locked-model evaluation.
  • Still unknown: The prospective cohort included 253 cirrhosis patients, with 36 developing HCC; 8 of 65 participants were false positives for conversion by the reported endpoint. This is source-reported research performance, not proof of improved survival or a clinically implemented screening test. The lead time is measured relative to imaging diagnosis in the study cohorts.
  • AI role: The random-forest model evaluated serum protein panels and selected the four-marker P4 panel; researchers then tested classification in an independent validation set and assessed prediction of cirrhosis-to-HCC conversion in prospective follow-up data.

Evidence sources: https://www.nature.com/articles/s41467-023-44255-2.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 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, 104 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 denoising exposes strong size scaling in deep-earthquake rupture parameters

A multitask neural denoiser recovered teleseismic P-wave pulses from deep Mw 5.0+ earthquakes that had previously been too noisy for source analysis. The study found moment–duration scaling M₀ ∼ τ⁴, stress-drop scaling Δσ ∼ M₀⁰·²¹ and radiated-energy scaling Eᵣ ∼ M₀¹·²⁴; deep events had a median radiation efficiency near 5%, weak directivity and few subevents. After accounting for rigidity, no systematic depth trend in duration, stress drop or radiated energy remained across 100–700 km.

Technology: DenoTe, a transfer-learned multitask convolutional encoder–decoder neural network

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 multitask neural denoiser recovered teleseismic P-wave pulses from deep Mw 5.0+ earthquakes that had previously been too noisy for source analysis.
  • Still unknown: The study is a source-parameter reanalysis enabled by neural denoising, not prospective earthquake prediction. It reports scaling relations and interpretations from the available global event sample; stress and energy estimates remain subject to waveform, path-correction and source-model uncertainties.
  • AI role: Separated low-SNR teleseismic P-wave signals from overlapping noise, enabling source-parameter estimation from 920 deep-earthquake records, including moderate events ordinarily omitted from these analyses.

Evidence sources: https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2023JB027334

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, 12 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.

Google DeepMind

FunSearch: new mathematical results with a language model

A language model (PaLM 2) paired with an automatic evaluator found the largest cap sets on record in some dimensions, the biggest increase in 20 years, and bin-packing heuristics using fewer bins than established methods. The problem must first be expressed as code with an evaluation procedure.

Technology: FunSearch (PaLM 2 and automatic evaluator)

Family FunSearch: 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: Programs were scored automatically by an evaluator; on the cap set problem they found larger cap sets than the best known, and on online bin packing their heuristics beat established baselines.
  • Still unknown: The abstract does not explicitly state limitations; the blog says the method needs an automatic evaluation procedure and that the bin-packing heuristics were tailored to specific data.
  • AI role: A pretrained large language model (PaLM 2) generated candidate programs in an evolutionary search guided by an automatic evaluator.

Evidence sources: https://deepmind.google/blog/funsearch-making-new-discoveries-in-mathematical-sciences-using-large-language-models/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-023-06924-6&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: 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, 990 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 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.

Bayesian optimization identifies active non-noble oxygen-evolution oxides

Multiobjective Bayesian optimization guided robotic synthesis and electrochemical testing in a researcher-defined Co–Mn–Sb–Sn–Ti oxide space. The study identifies Mn-rich candidates and reports improved activity after Ti incorporation into MnOx following short accelerated-stress testing. Its 17-fold acceleration compares measurement time with a hypothetical exhaustive triplicated grid; it does not establish commercial electrolyzer lifetime or universal superiority.

Technology: Multiobjective Bayesian optimization with multitask Gaussian-process surrogate

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: Multiobjective Bayesian optimization with multitask Gaussian-process surrogate

Evidence sources: https://publikationen.bibliothek.kit.edu/1000167576

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.

MatterGen: stable materials generated from desired properties

A diffusion model generates stable, novel materials directly from desired properties: 2.9 times more stable, novel and unique structures than CDVAE, and over 250 materials with bulk modulus above 400 GPa versus 2 in the reference dataset. Results were checked with DFT calculations; experimental verification remains the ultimate test.

Technology: MatterGen (diffusion model)

Family MatterGen: 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: Generated structures were checked with DFT calculations and compared with CDVAE and screening baselines (2.9 times more stable, novel and unique structures than CDVAE, per the blog); experimental verification is stated as not yet done.
  • Still unknown: Limitations stated: DFT has known limitations and experimental verification remains the ultimate test; the arXiv abstract states no limitations.
  • AI role: A diffusion model generated new crystal structures conditioned on desired properties.

Evidence sources: https://arxiv.org/abs/2312.03687 · https://www.microsoft.com/en-us/research/blog/mattergen-property-guided-materials-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: 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.

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.

Google DeepMind

GNoME: 2.2 million new predicted crystals

A graph neural network predicted 2.2 million new crystals, including 380,000 of the most stable, raising the number of known stable materials from about 48,000 to 421,000. External laboratories synthesized 736 predicted structures. Stability is calculated and does not guarantee a material can be produced or will be useful.

Technology: GNoME (Graph Networks for Materials Exploration)

Family GNoME: profile · models page

Publication: Peer-reviewed paper · 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: Peer-reviewed paper
  • Test type: prospective or independent check
  • Tested: Predicted stability was checked computationally with DFT, and external laboratories independently synthesized 736 of the predicted structures.
  • Still unknown: The blog states that most of the 380,000 stable predictions had not been made in a lab and that stability is judged by current scientific standards; manufacturing and usefulness remain open.
  • AI role: A graph neural network trained on about 48,000 stable crystals predicted 2.2 million new crystals, including 380,000 stable ones.

Evidence sources: https://deepmind.google/blog/millions-of-new-materials-discovered-with-deep-learning/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-023-06735-9&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: 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, 1876 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 4/5 · 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.

A-Lab: an autonomous laboratory for synthesizing inorganic materials

In the Nature paper, the autonomous lab made compounds from computed targets over 17 days of continuous operation. The original press release reported 41 of 58 (71%), while the author’s January 19, 2026 correction reports 36 of 57: on review, 36 of 40 claimed successes were confirmed, 4 were inconclusive from diffraction alone, and ‘new’ meant new to the platform, not to science. One compound was removed because it was in the training data.

Technology: A-Lab (autonomous lab with language models and 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: The autonomous lab attempted 58 target materials in 17 days of continuous operation and produced 41 novel compounds; failed syntheses were analyzed to suggest improvements.
  • Still unknown: The abstract does not name the characterization methods and states no limitations beyond the failed-synthesis analysis; the record notes an erratum published in January 2026.
  • AI role: A robotic platform used computation, literature data, machine learning and active learning to plan and interpret solid-state powder syntheses.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-023-06734-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: 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, 1316 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.

Deep learning detects slow slip events in Cascadia GNSS data

A detector combining convolutional and attention-based neural networks, trained on synthetic geodetic time series, was applied to raw GNSS data in Cascadia from 2007 to 2022. The authors report 78 detections, including 35 of 40 events in the 2007–2017 comparison catalog and 43 additional events. Detection patterns align with tremor activity, but the paper says independent acquisition-data validation is needed, and shorter/low-confidence events remain uncertain. This method is specific to this station network and does not predict earthquakes.

Technology: Convolutional neural network and Transformer attention 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: Applied detector outputs were compared with a prior event catalog and tremor activity in real Cascadia GNSS data.
  • Still unknown: Some additional/short detections remain uncertain; article states validation with an independent acquisition dataset is needed. This is not earthquake prediction.
  • AI role: A CNN plus Transformer trained on synthetic GNSS series classified slow-slip windows in raw multistation GNSS data and produced event detections.

Evidence sources: https://www.nature.com/articles/s43247-023-01107-7

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 date: 16 Nov 2026. Review scope: Primary publisher full text, abstract, methods/results, affiliations and publication history..

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.

PANDA: pancreatic lesion detection on non-contrast CT with deep learning

PANDA was trained on non-contrast abdominal CT scans from 3,208 patients at a high-volume pancreatic cancer centre in Shanghai. In a retrospective evaluation of 20,530 consecutive patients across four clinical settings, the paper reports 92.9% sensitivity and 99.9% specificity for pancreatic lesion detection. The study also reports five cancers and 26 other pancreatic lesions missed by the initial standard of care. These results do not establish population-screening benefit; the authors call for prospective risk-benefit and cost-effectiveness studies and broader international validation.

Technology: PANDA (three-stage deep learning cascade on non-contrast CT)

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 retrospective evaluation of 20,530 consecutive patients across four clinical settings, PANDA's sensitivity and specificity for pancreatic lesion detection were reported, along with missed cancers relative to the initial standard of care.
  • Still unknown: The results do not establish population-screening benefit; the authors call for prospective risk-benefit and cost-effectiveness studies and broader international validation.
  • AI role: PANDA, a deep-learning model trained on non-contrast abdominal CT, detected pancreatic lesions.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41591-023-02640-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.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 403 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.

MAIP enriches high-throughput screening for antimalarial hits

The MAIP malaria-activity predictor was combined with a compound-selection workflow and high-throughput cascade to prioritize public-library compounds. The paper reports eight selected hits with potency and ADME profiles and 12-fold enrichment versus a random set. That enrichment is study-specific; private training-set overlap and chemical novelty remain unresolved.

Technology: MAIP machine-learning antimalarial activity 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: Primary article reports selected hits, experimental screening and 12-fold enrichment versus random selection.
  • Still unknown: Underlying partner data are partly private; selection and training overlap and exact compound prior art require review; not clinical evidence.
  • AI role: Predictor enriches candidate pool, with staged workflow and human selection preceding experimental screening.

Evidence sources: https://doi.org/10.1021/acsmedchemlett.3c00369

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 primary metadata/abstract; author-hosted primary full paper. Material checked for the specific claims stated in the record..

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.

Google DeepMind

GraphCast: global 10-day weather forecasts

It outperforms ECMWF’s HRES system on over 90% of 1,380 variables and test forecast lead times, and produces a 10-day forecast in under a minute on a single TPU. Comparison is limited to deterministic HRES; training uses ERA5 data.

Technology: GraphCast (graph neural network)

Family GraphCast: 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: 10-day forecasts of hundreds of variables at 0.25° in under a minute; outperforms the most accurate operational deterministic systems on 90% of 1,380 verification targets; supports better tropical-cyclone, atmospheric-river and extreme-temperature prediction.
  • Still unknown: Comparison with ensemble systems is not described in the abstract.
  • AI role: A machine-learning model trained directly on reanalysis data.

Evidence sources: https://doi.org/10.1126/science.adi2336

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, 2507 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.

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.

Machine-learning screening reveals enhanced formate selectivity of Cu–Ga for CO₂ electroreduction

A machine-learning high-throughput screen of 465 binary metallic catalysts predicted previously unreported Cu–Ga selectivity for formate production in CO₂ electroreduction. Electrochemical tests of electrodeposited Cu–Ga measured a maximum formate Faradaic efficiency of about 38.4% at −1.05 V_RHE, roughly four times that on bare Cu under the reported conditions. This is a catalyst selectivity result for the tested electrode and electrolyte, not evidence of industrial-scale performance.

Technology: DSTAR active-motif machine-learning models with a CO₂-reduction selectivity map

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 high-throughput screen of 465 binary metallic catalysts predicted previously unreported Cu–Ga selectivity for formate production in CO₂ electroreduction.
  • Still unknown: The reported ~38.4% Faradaic efficiency is for the tested electrodeposited Cu–Ga electrode at −1.05 V_RHE in the authors’ cell and electrolyte, and the comparison is to bare Cu under those conditions. The paper discusses other predicted alloy outcomes too; this entry counts the single Cu–Ga study/result once. No device-scale or long-duration performance is established.
  • AI role: The authors combined ML binding-energy predictions with a reaction-selectivity map to screen 465 binary metallic combinations and prioritize Cu–Ga and Cu–Pd behavior for follow-up. Experimental Cu–Ga measurements validated the previously unreported formate-selectivity prediction; the AI-enabled screening supplied the non-obvious target and selectivity hypothesis.

Evidence sources: https://www.nature.com/articles/s41467-023-43118-0

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 1% of works of the same type, year and subfield (OpenAlex work metrics, 195 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.

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.

AlphaZero finds improved finite extremal-graph bounds

A neural network-guided AlphaZero search with a curriculum generated triangle- and 4-cycle-free graphs that improve known lower bounds for the extremal function ex(n,{C3,C4}) over finite sizes; the proceedings report AlphaZero improvements for n=64–136 and release the graph constructions. The paper also finds improvements using a separate tabu-search method, which often matches or outperforms AlphaZero. This does not solve the asymptotic Erdős conjecture. The result was present in the arXiv v1 posted 2023-11-06; the 2024 IJCAI proceedings paper is a later version.

Technology: AlphaZero reinforcement learning with neural Pairformer representation and curriculum

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 paper reports released graphs with improved lower bounds for ex(n,{C3,C4}) for node counts 64–134. Incremental AlphaZero (neural network-guided Monte Carlo tree search) is explicitly reported to improve bounds for sizes 64–136 in the proceedings version; incremental tabu search also improves bounds and the authors report it can match or exceed AlphaZero. This is a finite-graph construction result, not a proof of the asymptotic Erdős conjecture.
  • Still unknown: The result consists of explicit finite constructions/lower bounds; the asymptotic conjecture remains open. The work does not establish that learned search outperforms strong local search.
  • AI role: AlphaZero with a neural Pairformer representation and a curriculum generated extremal graphs improving the established lower bounds for n=64–136; tabu search is a separate, non-neural method and its results are not attributed to AI.

Evidence sources: https://arxiv.org/html/2311.03583v1 · https://ijcai.org/proceedings/2024/0772 · https://ijcai.org/proceedings/2024/0772.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 arXiv v1 full text and metadata plus IJCAI 2024 proceedings page/PDF for AlphaZero versus tabu-search graph construction results, theorem limits 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.

Active learning identifies a high-performing four-metal perovskite oxygen-evolution catalyst

A graph-based machine-learning model and closed-loop active-learning experiments identified a four-metal perovskite oxide with an intrinsic oxygen-evolution overpotential of 391 mV at 10 mA cm⁻²oxide, among the lowest reported for four-metal perovskite oxides. The study also reports recovery of several experimentally observed composition trends. This is an electrocatalyst material result; the abstract does not establish commercial electrolyzer performance.

Technology: Graph-based machine-learning model with active learning and closed-loop 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: A graph-based machine-learning model and closed-loop active-learning experiments identified a four-metal perovskite oxide with an intrinsic oxygen-evolution overpotential of 391 mV at 10 mA cm⁻²oxide, among the lowest reported for four-metal perovskite oxides.
  • Still unknown: The 391 mV figure is the reported intrinsic overpotential at 10 mA cm⁻²oxide and is compared by the authors with four-metal perovskite oxides. This is a material-level electrocatalysis measurement, not a commercial device result. The reported finding is one selected composition from a model-guided campaign; the abstract does not provide a full uncertainty interval.
  • AI role: The model incorporated structural-characterization data and guided iterative material selection in a closed-loop experimental workflow. The authors state that it identified the composition with the reported 391 mV intrinsic overpotential; laboratory electrochemical measurements established the catalyst outcome.

Evidence sources: https://www.nature.com/articles/s41563-023-01707-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 1% of works of the same type, year and subfield (OpenAlex work metrics, 183 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.

MetNet-3: short-range neural forecasts in Google products

The model forecasts precipitation, temperature, wind and dew point up to 24 hours ahead at 1–4 km resolution, learning from about 1,000 stations, and outperforms HRRR and ENS in several regions. It operates in Google Search and the app in the United States and 27 European countries, with precipitation forecasts up to 12 hours. Preprint, no numerical values in the text and one case study.

Technology: MetNet-3

Family MetNet: 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: MetNet-3 was benchmarked against HRRR and ENS over the contiguous US for forecasts up to 24 hours, scored with CRPS; a single January 2019 case study is also shown.
  • Still unknown: The post has no limitations section; it notes operational coverage limited to the contiguous US and parts of Europe, hyperlocal wind results depend on test-station input, and the case study is a single example.
  • AI role: A neural model forecasts precipitation, temperature, wind and dew point up to 24 hours ahead, learning from about 1,000 stations.

Evidence sources: https://arxiv.org/abs/2306.06079 · https://research.google/blog/metnet-3-a-state-of-the-art-neural-weather-model-available-in-google-products/

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.

Machine learning exposed a structure in terminal-singularity classification

A neural-network classifier trained on eight-dimensional, Picard-rank-two toric varieties achieved 95% accuracy on a five-million-example test set and enabled rapid generation of a much larger predicted landscape of Q-Fano varieties. The authors report that the model’s classification pattern led them to formulate a new combinatorial criterion for terminal singularities, which they prove mathematically. The learned model suggested the route to the theorem; it did not supply the proof.

Technology: Feed-forward neural-network classifier trained on toric-variety weight matrices

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 report a 95% classification accuracy on a balanced test split of five million eight-dimensional toric-variety examples and use the model to explore a larger Q-Fano landscape. They formulate and prove a new combinatorial criterion for terminal singularities, motivated by the ML analysis.
  • Still unknown: The 95% is internal held-out classification performance, not independent external validation; the authors state the neural network does not itself perform the theorem’s divisibility-based test or proof.
  • AI role: The trained neural-network classifier revealed a tractable pattern in the classification landscape and directly motivated the authors’ new combinatorial criterion; the authors separately establish its correctness by proof.

Evidence sources: https://arxiv.org/abs/2310.20458 · https://arxiv.org/html/2310.20458v1

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 metadata and full text for posting date, author affiliations, learned-classifier results and connection to the stated mathematical criterion; linked NeurIPS proceedings identity 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.

Machine-learning models identify cerebrospinal-fluid protein signatures for multiple-sclerosis activity and disability progression

Using Olink proteomics of 1,463 proteins in CSF and plasma from two prospective Swedish cohorts, the study trained prediction models for multiple-sclerosis diagnosis, 2-year disease activity and long-term disability worsening. Lower CSF neurofilament light (NfL) predicted absence of disease activity at two years (replication AUC 0.77). A stepwise model using 11 CSF proteins (CXCL13, LTA, FCN2, ICAM3, LY9, SLAMF7, TYMP, CHI3L1, FYB1, TNFRSF1B and NfL) predicted normalized age-related MS severity; subgroup classification AUCs in the replication cohort ranged from 0.88 to 0.90.

Technology: Stepwise logistic regression for diagnosis and disease activity; stepwise linear regression for long-term disability severity

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 Olink proteomics of 1,463 proteins in CSF and plasma from two prospective Swedish cohorts, the study trained prediction models for multiple-sclerosis diagnosis, 2-year disease activity and long-term disability worsening.
  • Still unknown: The study included 143 people with early-stage MS and 43 healthy controls across discovery and replication cohorts; the replication MS cohort had 51 participants. The 2-year disease-activity comparison in replication was especially small (45 with activity versus 5 without). The 11-protein severity model predicted observational disability scores rather than treatment benefit. The plasma severity model had only 12 eligible replication samples and performed less convincingly. These are cohort-level results, not a clinically validated diagnostic assay or independent replication by another research group.
  • AI role: The study used statistical differential-expression testing to select disease-associated candidate proteins, then fitted logistic-regression models to predict MS diagnosis and 2-year disease activity and a linear-regression model to predict long-term normalized age-related MS severity. Stepwise feature selection produced the protein combinations; the frozen models were evaluated in a separate prospective cohort.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC10616092/

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, 112 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.

CAIMAN validates AI prescreening of typical and atypical large-bowel biopsies across UK and Portuguese cohorts

CAIMAN is a weakly supervised deep-learning model trained on pathologist-labeled whole-slide images to classify typical, atypical non-neoplastic and atypical neoplastic bowel biopsies. In three external cohorts comprising 1,536 slides from the UK and Portugal, it detected atypical biopsies with AUROC values of 0.9431, 0.9576 and 0.9636. At 99% sensitivity, it outperformed the prior IDaRS pipeline on two external cohorts, but not the third.

Technology: CAIMAN (Colorectal AI Model for Abnormality Detection), a weakly supervised deep-learning whole-slide image classifier

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 three external cohorts comprising 1,536 slides from the UK and Portugal, it detected atypical biopsies with AUROC values of 0.9431, 0.9576 and 0.9636.
  • Still unknown: Retrospective validation on archived biopsies does not establish clinical deployment or improved patient outcomes. CAIMAN did not outperform IDaRS at 99% sensitivity in one of the three external validation cohorts; the authors call for larger multicentre validation and a clinical trial. The reported 99% sensitivity is paired with cohort-specific specificity, not a universal performance guarantee.
  • AI role: AI learned biopsy-level histological patterns from slide-level diagnostic labels, then classified whole-slide images into typical, atypical non-neoplastic or atypical neoplastic classes. The trained frozen model was evaluated on three external cohorts; the reported finding is a specific retrospective clinical diagnostic-validation result.

Evidence sources: https://wrap.warwick.ac.uk/id/eprint/180883/

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, 50 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 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 image search helps identify Pavo, an isolated star-forming dwarf galaxy

A search combining low-surface-brightness galaxy detection and machine-learning candidate classification identified Pavo in Dark Energy Camera Legacy Survey imaging. Follow-up Magellan imaging resolved young and old stellar populations and measured a tip-of-the-red-giant-branch distance of about 1.99 Mpc, establishing Pavo as an extremely isolated, low-mass star-forming dwarf galaxy just beyond the Local Group.

Technology: Machine-learning image candidate classifier used with low-surface-brightness galaxy search 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: A search combining low-surface-brightness galaxy detection and machine-learning candidate classification identified Pavo in Dark Energy Camera Legacy Survey imaging.
  • Still unknown: The AI contributed candidate classification within a broader image-search workflow. The discovery was verified by follow-up observations and astrophysical analysis. AAS conference abstract from 2024 also describes this same Pavo discovery and is not a separate result.
  • AI role: Classified image-search candidates in DECaLS imaging; the search team selected Pavo for follow-up, and telescope imaging and color-magnitude analysis established its distance and galaxy properties.

Evidence sources: https://arxiv.org/abs/2310.01478

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, 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.

Integrated radiogenomic machine learning predicts neoadjuvant chemotherapy response in metastatic ovarian cancer

The IRON ensemble combined baseline clinical data, CA-125, ctDNA and CT radiomics from primary and metastatic lesions to estimate tumor-volume change during neoadjuvant chemotherapy for high-grade serous ovarian carcinoma. In an independent 42-patient Barts cohort, the integrated model achieved AUC 0.78 for RECIST 1.1 response classification versus 0.47 for a clinical-only model; radiomics-containing scores significantly correlated with volume change in internal holdout. The study also measured spatial heterogeneity, with omental lesions showing better response than pelvic/ovarian disease.

Technology: IRON ensemble combining elastic net, support-vector regression and random forest

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 an independent 42-patient Barts cohort, the integrated model achieved AUC 0.78 for RECIST 1.1 response classification versus 0.47 for a clinical-only model; radiomics-containing scores significantly correlated with volume change in internal holdout. The model used historical Cambridge and Barts cohorts with baseline and post-treatment imaging, including an external retrospective test set.
  • Still unknown: Model analysis of two prospective observational cohorts (training n=72, internal holdout n=20, external n=42), not a decision trial. External ctDNA unavailable and imputed to training means. Sample size limits generalizability; treatment benefit from model-guided decisions was not tested.
  • AI role: Pretreatment clinical, blood and multi-lesion radiomic features predicted volumetric response to neoadjuvant chemotherapy. Held-out and external evaluation showed radiomics added predictive information beyond clinical variables.

Evidence sources: https://www.nature.com/articles/s41467-023-41820-7.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.

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 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.

Neural active learning guides stronger bone-scaffold architectures

Learned architecture models and finite-element simulations guided fabrication of scaffold subunits. A larger scaffold assembled through human adaptation of the learned structures showed about 20% greater load-bearing capacity than the uniform comparator. This is a mechanical bench-test result, not evidence of implantation, bone healing or clinical benefit.

Technology: GAD-MALL autoencoder and convolutional 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: Printed structures received mechanical tests.
  • Still unknown: Independent replication and in-vivo outcomes were not established.
  • AI role: GAD-MALL autoencoder and convolutional active learning supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://www.nature.com/articles/s41467-023-42415-y · https://pure.mpg.de/pubman/item/item_3643037_1/component/file_3643048/s41467-023-42415-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: linked separately.

Review date: 16 Nov 2026. Review scope: Indexed primary publisher affiliations, date and selected design Results; indexed institutional publisher PDF mechanical comparison. 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.

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.

Vesuvius Challenge: first letters read from an unopened Herculaneum scroll

For the first time, a complete word (‘porphyras’, purple) was read from inside a never-opened carbonized scroll; a model trained on crack patterns yielded 13 letters independently and unanimously annotated by papyrologists. The letters are faint and lack context; other readings are not ruled out.

Technology: Machine-learning models for ink detection

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: A model trained on crack patterns produced 13 letters, which papyrologists independently annotated with varying confidence; the page links only to code and no paper.
  • Still unknown: The letters are faint and lack context, word boundaries are unclear, and the page states other readings are not ruled out and the papyrological team was still investigating.
  • AI role: Machine-learning models trained on ink crackle patterns (and a second model with pseudo-labeling) read letters from the scroll.

Evidence sources: https://scrollprize.org/firstletters

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 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.

Bayesian optimization identifies a methane-combustion catalyst formula

Bayesian optimization guided the experimental search over Pt/Pd and Ce/Zr ratios in PtPd@CeZrOx core–shell nanospheres. The authors report that two optimization rounds comprising 18 experiments found a formula with methane-conversion T50 approaching 330 °C, lower than prior reports cited by the study. The full assay conditions and comparison basis were not reviewed here; this is a bounded laboratory result, not a claim of general superiority, durability, or deployment.

Technology: Bayesian optimization for experimental catalyst-formula 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 publisher abstract reports that BO guided two experimental optimization rounds (18 experiments total) and that the selected PtPd@CeZrOx formula reached a methane-conversion T50 approaching 330 °C.
  • Still unknown: Full test conditions, exact comparator basis, replicate variability and durability were not assessed in this abstract-limited review. The reported T50 does not establish field emissions performance or general superiority.
  • AI role: Bayesian optimization selected multi-component Pt/Pd and Ce/Zr catalyst-formula ratios across iterative experiments; the reported T50 is the measured outcome associated with the selected formula.

Evidence sources: https://onlinelibrary.wiley.com/doi/10.1002/anie.202313068 · https://pubmed.ncbi.nlm.nih.gov/37823845/

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 abstract/indexed article text and PubMed primary metadata/abstract read. Full article Methods/Results and supporting information were not accessible..

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.

BTSbot workflow automatically confirms and classifies SN 2023tyk

BTSbot selected SN 2023tyk from ZTF alerts and requested robotic spectroscopy; SNIascore supplied its Type Ia classification. The integrated workflow publicly reported the result without human intervention in those steps. This is one transient result, not a new supernova class or a claim that BTSbot alone performed the observation. It preceded the expanded 2024 study.

Technology: BTSbot convolutional classifier and SNIascore

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: Robotic spectroscopy supported the Type Ia classification.
  • Still unknown: Universal discovery priority and broader workflow completeness were not assessed.
  • AI role: BTSbot convolutional classifier and SNIascore supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://www.wis-tns.org/astronotes/astronote/2023-265 · https://news.northwestern.edu/stories/2023/10/first-supernova-detected-confirmed-classified-and-shared-by-ai?fj=1

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 date: 16 Nov 2026. Review scope: Indexed primary TNS AstroNote and institutional announcement, plus indexed published-paper abstract. Direct TNS retrieval 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.

CNN-assisted search confirms quadruply lensed quasar PS J2107−1611

A convolutional neural network trained on simulated lens images screened 27,760 preselected quasar candidates with multiple nearby Pan-STARRS detections. Researchers visually inspected the top 2,000 scores and selected follow-up targets; spectroscopy and multi-survey imaging established PS J2107−1611 as a quadruply imaged quasar at redshift 2.673, with a 4.3-arcsecond image separation and a lensed host-galaxy arc. Its fold-image optical fluxes are roughly 15 times fainter than a smooth lens-mass model predicts, while radio flux ratios are consistent with that model. The date is the original preprint posting; the journal version appeared in November 2023.

Technology: Convolutional neural network trained on simulated quadruply lensed quasars, applied to Pan-STARRS gri image cutouts

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 convolutional neural network trained on simulated lens images screened 27,760 preselected quasar candidates with multiple nearby Pan-STARRS detections.
  • Still unknown: This is a single confirmed system from a prefiltered candidate set; the authors visually reviewed high-scoring candidates and performed follow-up. Spectroscopy confirms the source is a quasar at z=2.673, while the four-image lens interpretation is supported by high-resolution imaging, lens modeling, a host-galaxy arc, and wavelength-dependent flux ratios. The ∼15-fold optical fold-pair anomaly is relative to the smooth macromodel; the cause is not resolved in this letter. Detailed JWST/VLT/EVN follow-up was left for future work.
  • AI role: The CNN ranked image cutouts of 27,760 quasar catalog entries that already had multiple nearby Pan-STARRS detections. The team visually examined the top 2,000, selected 131 promising systems, and followed targets with imaging/spectroscopy; the model therefore enabled prioritization but did not itself confirm lensing. Human follow-up established the quasar redshift and lens configuration.

Evidence sources: https://arxiv.org/abs/2310.04494

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 54 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 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.

Extrapolative machine learning identifies a multielement RWGS catalyst

In a closed loop combining machine-learning prediction with experiments, researchers tested 300 reverse water-gas shift catalyst formulations over 44 cycles, starting from 45 data points. They report Pt(3)/Rb(1)-Ba(1)-Mo(0.6)-Nb(0.2)/TiO₂ as the optimal tested composition; Nb was absent from the initial dataset. The result is specific to the study’s catalyst tests and conditions, not commercial CO₂ conversion. The earlier ChemRxiv version is not used to claim priority here.

Technology: Extrapolative machine learning; closed-loop prediction and experiment

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 300 experimentally tested catalysts over 44 ML-plus-experiment cycles and identifies the best tested formulation.
  • Still unknown: Commercial operation, independent replication and performance beyond reported experimental conditions are not established.
  • AI role: Machine-learning predictions guide sequential catalyst selection; experimental results update the closed-loop search.

Evidence sources: https://www.nature.com/articles/s41467-023-41341-3.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 date: 16 Nov 2026. Review scope: Publisher PDF indexed primary full text and publication metadata; direct HTML endpoint blocked by cookie redirect..

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

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.

RETFound: a foundation model for disease detection from retinal images

RETFound is self-pretrained on unlabelled retinal images, reportedly about 1.6 million from Moorfields Eye Hospital, and then adapted to disease tasks. The authors report better diagnosis of sight-threatening eye diseases than comparison models, better prediction of systemic conditions such as heart failure and myocardial infarction, and strong performance with fewer labels. The evaluation is retrospective, with no clinical deployment study, and the model was released openly.

Technology: RETFound (masked-autoencoder foundation model for retinal images)

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: RETFound, self-pretrained on unlabelled retinal images, was adapted to disease tasks and compared with other models on diagnosis of sight-threatening eye diseases and on prediction of heart failure and myocardial infarction.
  • Still unknown: The abstract states no limitations of this work; the evaluation is retrospective with no clinical deployment study.
  • AI role: A self-supervised foundation model pretrained on retinal images and fine-tuned for disease tasks.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-023-06555-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 1% of works of the same type, year and subfield (OpenAlex work metrics, 1177 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.

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.

PhotoCat guides experimental discovery of photocatalytic reactions

The PhotoCat Transformer was pretrained on general reaction data and fine-tuned on the authors’ photocatalysis database to predict products and reaction conditions. In the 2023 ChemRxiv version, the authors report using it to guide experimental validation of previously unreported photocatalytic reactions. The later 2026 journal version describes four novel validated reactions, whereas the preprint reports five; this entry makes no exact-count claim and treats the journal article as a revised version of the same study.

Technology: Transformer-based deep-learning reaction predictor pretrained on USPTO and fine-tuned on PhotoCatDB, with reaction-condition inputs

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 PhotoCat Transformer was pretrained on general reaction data and fine-tuned on a photocatalysis database to predict products and conditions. The 2023 ChemRxiv version reports wet-lab validation of five predicted reactions; the 2026 journal abstract reports four novel validated reactions.
  • Still unknown: The preprint and journal version differ in their reported count of validated novel reactions. This record states the shared model-guided experimental discovery claim and makes no exact-count claim.
  • AI role: A Transformer reaction predictor was pretrained on USPTO and fine-tuned on PhotoCatDB to rank photocatalytic reaction outcomes and conditions, guiding experimental validation.

Evidence sources: https://chemrxiv.org/engage/api-gateway/chemrxiv/assets/orp/resource/item/65004fd999918fe537d200ab/original/photo-cat-an-artificial-intelligence-driven-synthesis-planning-platform-for-photocatalysis.pdf · https://pmc.ncbi.nlm.nih.gov/articles/PMC12913628/

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 v2 full-text PDF and version-history metadata; primary PMC full-text journal version and publication metadata. The preprint and journal versions report different counts (five versus four); the summary preserves this 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.

Machine learning helps recover dimension from quantum periods for a class of Fano varieties

A feed-forward neural network classified the dimension of Fano varieties from quantum-period coefficients with 98% accuracy on the authors’ dataset. The authors report that machine learning helped formulate rigorous asymptotic results, which prove dimension recovery for specified low-Picard-rank toric Fano classes. The broader conjecture that the quantum period uniquely determines every Fano variety remains open; the empirical classification accuracy is not itself a proof. The journal article appeared 2023-09-08, before the same work’s 2023-09-11 arXiv posting.

Technology: Feed-forward 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 article combines a dataset classification result with rigorous asymptotic proofs for specified classes.
  • Still unknown: 98% is empirical accuracy on the paper’s dataset; the broad uniqueness conjecture remains open. Journal publication preceded arXiv submission.
  • AI role: A neural network analyzed quantum-period coefficients and helped identify dimension-related patterns used to formulate asymptotic results, which the authors prove for restricted classes.

Evidence sources: https://kasprzyk.work/research/pdf/ml_dimension.pdf · https://www.nature.com/articles/s41467-023-41157-1

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: 16 Nov 2026. Review scope: Primary author-hosted full manuscript plus primary publisher article metadata; checked authors, method, theorem scope 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.

AI-assisted seismic mapping reveals segmented fault activity at the Discovery transform fault

A machine-learning earthquake-detection and relocation workflow applied to 2008 ocean-bottom seismometer records found that the western Discovery transform-fault strand may comprise seven patches with distinct seismicity. The authors also identify a seismicity-gap zone that they interpret as possibly linked to an active small off-ridge rift.

Technology: PhaseNet phase picker, GaMMA phase association, and machine-learning-enhanced seismic event workflow

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 machine-learning earthquake-detection and relocation workflow applied to 2008 ocean-bottom seismometer records found that the western Discovery transform-fault strand may comprise seven patches with distinct seismicity.
  • Still unknown: The seven-patch segmentation and possible off-fault rift relationship are interpretations of one year of local 2008 seismic data, not a causal or universal transform-fault model. The location and tectonic interpretations are supported by the new catalog and follow-up cross-correlation relocation; the source describes the rift explanation as possible.
  • AI role: The AI workflow detected P/S arrivals from noisy ocean-bottom recordings and associated event phases; cross-correlation relocation and geological interpretation by the research team produced the fault-pattern interpretation.

Evidence sources: https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2023GC011043

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, 22 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.

Deep learning detects an emerging anthropogenic warming fingerprint in daily precipitation

A convolutional neural network trained on 80 CESM2 large-ensemble members detected an emerging climate-change signal in observed daily precipitation fields. Across four satellite, merged-observation and reanalysis products, the model’s annual estimated global-mean temperature exceeded the historical internal-variability range from the mid-2010s onward. The signal was driven mainly by precipitation variability on timescales shorter than 10 days, with hotspots including the tropical eastern Pacific and mid-latitude storm tracks; rain-gauge checks over the eastern United States were consistent with the increase in high-frequency variability.

Technology: Deep detection (DD) convolutional neural network

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 convolutional neural network trained on 80 CESM2 large-ensemble members detected an emerging climate-change signal in observed daily precipitation fields.
  • Still unknown: This is an observational detection and attribution result using a CNN trained on climate-model simulations; it is not a direct causal experiment or a forecast of individual storms. The long-term mean precipitation change remained indistinguishable from natural variability in the analysis. Precipitation products disagree in some estimates, especially where direct ocean measurements are sparse, and the reported hotspots depend on the selected datasets and attribution methods. The paper reports an emerging signal beginning in the mid-2010s, not a uniform increase in all regions or at every timescale.
  • AI role: The CNN was trained on paired daily precipitation anomaly maps and annual global-mean surface-air-temperature values from 80 CESM2 Large Ensemble members. Applied to observed precipitation products, it learned spatial and temporal patterns that predicted the warming metric and exposed an emergent signal that conventional linear detection did not resolve. Explainability analyses (occlusion sensitivity, with SHAP and integrated-gradients checks) localized the strongest contributions to high-frequency precipitation variability.

Evidence sources: https://www.nature.com/articles/s41586-023-06474-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: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 115 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.

A neuroprosthesis translates attempted speech at 62 words per minute

For a participant with ALS, intracortical electrodes and an AI algorithm converted attempted speech into text at 62 words per minute. With a 125,000-word vocabulary, the error rate was 23.8%, versus 9.1% for only 50 words. Proof of concept in one person, with an invasive implant and experimental device.

Technology: Neural network for phoneme decoding and 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: One participant with ALS attempted speech in real time; word error rates were measured at 50-word and 125,000-word vocabularies, with decoding at 62 words per minute.
  • Still unknown: The abstract states no limitations; the results come from one participant, the large-vocabulary error rate (23.8%) is substantial and decoding is well below conversational speed.
  • AI role: An AI algorithm decoded spiking activity from intracortical microelectrode arrays into text.

Evidence sources: https://med.stanford.edu/news/all-news/2023/08/brain-implant-speech-als.html · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-023-06377-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: 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, 675 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.

Woman with paralysis communicates through a brain-controlled avatar

For a participant with stroke-related paralysis, an electrode array on the cortex and AI algorithms translated attempted speech into text, synthetic voice and avatar movements. Text reached nearly 80 words per minute after individual training. Experimental study in one person; the system requires a physical connection to computers.

Technology: Neural speech decoding and AI facial animation

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: One clinical-trial participant with paralysis attempted silent speech; decoding was measured in real time at a median of 78 words per minute with a 25% median word error rate.
  • Still unknown: The abstract does not explicitly state limitations; the single-participant study and 25% word error rate are the main caveats it implies.
  • AI role: Deep-learning models decoded text from surface speech-cortex signals, synthesized personalized speech and animated a facial avatar.

Evidence sources: https://engineering.berkeley.edu/news/2023/08/novel-brain-implant-helps-paralyzed-woman-speak-using-a-digital-avatar/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-023-06443-4&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: 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, 600 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 assists a proved dimension bound for affine Deligne–Lusztig varieties

The authors used computational data and machine-learning pattern exploration to study affine Deligne–Lusztig varieties. They report rediscovering the previously known virtual-dimension formula, then using the exploration to formulate and prove a new lower bound on actual dimension. The mathematical theorem, not model accuracy or rediscovery of the known formula, is the specific result. The 2024 journal article and 2023 arXiv v1 are versions of one study.

Technology: Linear models, support-vector machines and neural networks

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 paper reports mathematical proof of a lower bound prompted by machine-learning exploration.
  • Still unknown: The virtual-dimension formula was previously known and is not the new theorem; the work is not a formal automated proof.
  • AI role: Models explored generated mathematical data and highlighted patterns that led mathematicians to formulate and prove a new lower bound.

Evidence sources: https://arxiv.org/html/2308.11355v1 · https://link.springer.com/article/10.1007/s42543-024-00086-8

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 date: 16 Nov 2026. Review scope: Primary arXiv v1 full text and primary journal indexed full text, including methods, theorem, affiliations and version dates..

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.

Active-learning search experimentally realizes a high-zT Cu₁₋ₓAgₓGaTe₂ thermoelectric

Compressed-sensing symbolic regression and active learning screened the Cu₁₋ₓAgₓGaTe₂ thermoelectric series and guided synthesis of selected compositions. The published study reports a p-type Cu₀.₄₅Ag₀.₅₅GaTe₂ sample with a measured thermoelectric figure of merit zT of about 1.90 at 770 K, using measured heat capacity in the analysis. This is one paper-level materials result from a targeted search, not a claim that every composition in the series is new.

Technology: SISSO compressed-sensing symbolic regression within an active-learning 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: Compressed-sensing symbolic regression and active learning screened the Cu₁₋ₓAgₓGaTe₂ thermoelectric series and guided synthesis of selected compositions.
  • Still unknown: The zT value is the final peer-reviewed paper’s reported result using measured heat capacity; an earlier preprint version of the same study reported a different higher estimate (about 2.8 at 827 K), so that preprint value is not substituted here or counted as a separate finding. The reported result is a laboratory thermoelectric sample at elevated temperature, not a device or commercial-performance demonstration. Active learning searched a defined composition series and does not establish that every searched composition was novel.
  • AI role: The authors used SISSO to derive interpretable descriptors from their experimental thermoelectric dataset, then used active learning to rank prospective compositions for synthesis and measurement. The model guided candidate selection; the researchers synthesized and characterized the samples and measured the properties.

Evidence sources: https://pubs.rsc.org/en/content/articlelanding/2023/ta/d3ta03990k

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 57 of works of the same type, year and subfield (OpenAlex work metrics, 10 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.

Neanderthal and Denisovan peptides reconstructed as antibiotic candidates

AI analysis of proteins from extinct hominins identified sequences that were then synthesized and tested against pathogens in culture and in mice. Some peptides showed antimicrobial activity; these are preclinical results, with no demonstration of safety or efficacy in humans.

Technology: AI for molecular de-extinction

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: Fragments were tested for antimicrobial activity in vitro, and representative fragments reduced A. baumannii infection in mouse skin abscess and thigh infection models.
  • Still unknown: The abstract does not explicitly state limitations; the findings rest on in vitro work and two mouse models rather than human trials.
  • AI role: A random forest model (panCleave) predicted protease cleavage sites across proteomes to find antimicrobial fragments in modern human and extinct hominin proteins.

Evidence sources: https://www.engineering.upenn.edu/stories/for-a-new-generation-of-antibiotics-scientists-are-bringing-extinct-molecules-back-to-life-and-discovering-the-hidden-genetics-of-immunity-along-the-way/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1016/j.chom.2023.07.001&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, 165 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.

Avoiding aircraft contrails with AI forecasts

In a trial with American Airlines and Breakthrough Energy, pilots flew 70 flights over six months using AI forecasts to avoid forming contrails, reducing them by 54% compared with flights without the forecasts. Flights that avoided contrails used 2% more fuel; across an airline’s flights, the impact could fall to 0.3% if only a small share need rerouting.

Technology: AI contrail-formation forecasts and satellite imagery

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: Google, American Airlines and Breakthrough Energy flew 70 pilot flights over six months using AI forecasts to avoid contrail-forming altitudes; satellite imagery reported a 54% contrail reduction.
  • Still unknown: The article does not state a control group, the baseline for the 54% figure, statistical methods or peer review; fuel impact (2% more fuel, 0.3% airline-wide) is an estimate.
  • AI role: AI contrail forecasts guided pilots to avoid altitudes likely to form contrails.

Evidence sources: https://blog.google/technology/ai/ai-airlines-contrails-climate-change/

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 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.

MASAI: AI-assisted mammography screening in a randomized trial

In the initial analysis of 80,033 women, AI-assisted screening detected 244 cancers versus 203 with standard double reading and reduced the reading workload by 44%. A Lund update on January 30, 2026, after more than two years and about 106,000 women, reports higher sensitivity without more false positives. The comparison is in a Swedish setting; it does not demonstrate reduced mortality.

Technology: AI for mammogram triage and marking

Publication: Publication status unspecified in annotation · 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: Not reported
  • Test type: prospective or independent check
  • Tested: A prospective randomized single-site trial compared AI-supported screening (about 40,003 women) with standard double reading (about 40,030 women), measuring cancers detected, false positives and radiologist workload.
  • Still unknown: The Lund article states limits: single site, no mortality outcomes, and the main analysis awaits two-year follow-up; the January 2026 update could not be verified on the page read.
  • AI role: AI flagged higher-risk mammography exams for two physicians or marked suspicious findings for a single reader.

Evidence sources: https://www.lu.se/artikel/ai-stodd-modell-saker-mammografiscreening

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 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.

Physics-informed machine learning guides experimental Curie-temperature validation in Ce–Zr–Fe alloys

A machine-learning model trained on measured Curie temperatures selected two compositions in the (Zr1−xCex)Fe2 system, (Zr0.16Ce0.84)Fe2 and (Zr0.94Ce0.06)Fe2, for experimental validation. The study also reports a relationship between electronic states near the Fermi level and Curie temperature, including a breakdown of de Gennes scaling below a rare-earth-content threshold. This is a materials discovery study, not evidence of a finished permanent magnet or deployed technology.

Technology: Physics-informed machine-learning model trained on experimentally measured Curie temperatures

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 (Zr0.16Ce0.84)Fe2 and (Zr0.94Ce0.06)Fe2 for experimental Curie-temperature validation of the model-guided predictions.
  • Still unknown: The exact Curie-temperature measurements and uncertainty are not included here. This study does not establish a finished permanent magnet, device-level performance, or zero-carbon deployment.
  • AI role: A learned model trained on experimentally measured Curie temperatures was used to predict multicomponent magnetic-material properties and select the two Ce–Zr–Fe compositions for experimental validation. The paper also reports that the model captured electronic-structure trends associated with Curie temperature.

Evidence sources: https://pubs.acs.org/doi/10.1021/acs.chemmater.3c00892

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 abstract, article metadata and supporting-information description. Full article numerical Curie-temperature results and uncertainty were 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.

CMS makes the first hadron-collider observation of boosted Z→cc̄ decay

In 138 fb⁻¹ of 13 TeV proton-proton data, CMS observed boosted Z→cc̄ production in association with jets for the first time at a hadron collider. The measured signal strength relative to the Standard Model prediction is 1.00 with combined uncertainty +0.19/−0.17. A deep neural network (DeepDoubleX) identified large-radius jets from charm-quark pairs; the paper also sets a 95% confidence-level upper limit of 47 times the Standard Model prediction on boosted H→cc̄ production.

Technology: DeepDoubleX deep neural network for double-charm jet tagging

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 138 fb⁻¹ of 13 TeV proton-proton data, CMS observed boosted Z→cc̄ production in association with jets for the first time at a hadron collider.
  • Still unknown: The first-observation claim is specifically for Z→cc̄ in association with jets at high transverse momentum at a hadron collider; it is not the first observation of Z decay generally. The neural tagger is an important analysis tool, while the observation derives from the full CMS event selection and likelihood fit. In the same paper, boosted H→cc̄ was not observed; only an upper limit is reported.
  • AI role: The neural network processed secondary-vertex and particle-flow candidate features to classify boosted two-prong jets as charm versus light flavor and charm versus bottom. Its charm tagging separated the Z→cc̄ signal from dominant multijet backgrounds and supported the first observation in the boosted high-transverse-momentum regime.

Evidence sources: https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.131.041801

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.

Machine learning guides proofs of braid and flat-braid invariants

A one-hidden-layer multilayer perceptron classified examples of three-strand braids and suggested algebraic criteria. Authors converted the observed patterns into proofs. The paper proves a complete criterion for triviality of flat braids on three strands; its result does not establish a complete invariant for arbitrary braids or the neural network itself as a proof system.

Technology: Supervised multilayer perceptron

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 specific mathematical or algorithmic result summarized here.
  • Still unknown: The result is attributed to the paper; no independent proof review beyond the authors’ source text is claimed.
  • AI role: The MLP classified examples and exposed patterns used by human authors to formulate and prove exact invariants.

Evidence sources: https://arxiv.org/html/2307.12185

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 preprint full text or abstract reviewed; evidence 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.

Machine-learning analysis of plasma metabolomic fingerprints yields a 21-metabolite gastric-cancer panel validated across multicentre cohorts

In a multicentre study, machine-learning analysis of plasma metabolic fingerprints from 1,944 retrospective participants selected a 21-metabolite panel that differentiated gastric-cancer from non-cancer cohorts; five model types achieved AUCs of 0.921–0.971 in discovery and 0.907–0.940 in external verification. Prospective testing in 264 participants using both NPELDI-MS and UPLC-MS reported diagnostic AUCs of 0.855–0.918 and 0.856–0.916, respectively. The paper also reports a four-metabolite prognostic score based on retrospective survival data.

Technology: Neural network, lasso regression, ridge regression, support-vector machine and random-forest models applied to plasma metabolic fingerprints

Publication: Peer-reviewed paper · 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: Peer-reviewed paper
  • Test type: prospective or independent check
  • Tested: In a multicentre study, machine-learning analysis of plasma metabolic fingerprints from 1,944 retrospective participants selected a 21-metabolite panel that differentiated gastric-cancer from non-cancer cohorts; five model types achieved AUCs of 0.921–0.971 in discovery and 0.907–0.940 in external verification.
  • Still unknown: The panel was evaluated in selected clinical cohorts and is not established for routine screening or clinical use. Prospective validation was at one lead center, and only six panel metabolites were robustly detected by confirmatory UPLC-MS under the reported testing. Model performance does not show that the panel is causal or improves outcomes. Primary full text/abstract checked; no independent replication assessed.
  • AI role: The study used machine-learning feature selection/modeling (including logistic-regression coefficient filtering) to identify and evaluate a 21-metabolite diagnostic panel from high-throughput mass-spectrometry fingerprints; classifiers were then validated on independent retrospective and prospective cohorts.

Evidence sources: https://gut.bmj.com/content/72/11/2051

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, 95 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

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.

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.

Huawei Cloud

Pangu-Weather: medium-range forecasts outperform the operational ECMWF model

According to the authors, it is the first AI model to outperform ECMWF’s IFS at medium range: five-day Z500 RMSE of 296.7 versus 333.7, a global 24-hour forecast in 1.4 seconds on a GPU, and mean cyclone-track error of 120 km at three days versus 162 km. Source reviewed: Huawei Cloud blog, not the Nature paper (July 6, 2023); it underestimates extremes.

Technology: Pangu-Weather

Family Pangu-Weather: 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: Trained on 39 years of global data; stronger deterministic results on reanalysis data than ECMWF's operational IFS in all tested variables; also evaluated on extreme weather and ensemble forecasts.
  • Still unknown: Comparisons are initialized with reanalysis data, as the abstract states; performance from operational analyses is not given there.
  • AI role: Three-dimensional deep networks with Earth-specific priors and hierarchical temporal aggregation.

Evidence sources: https://doi.org/10.1038/s41586-023-06185-3

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, 2232 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.

NowcastNet: extreme-precipitation forecasts up to three hours ahead

A model combining neural networks and physical constraints forecasts precipitation up to three hours ahead; 62 meteorologists preferred it in 71% of 2,400 evaluated events. Tsinghua reports it was already deployed on the National Meteorological Center’s operational SWAN 3.0 platform. Expert preference does not measure reduced damage or guarantee accuracy for every event.

Technology: NowcastNet

Publication: Peer-reviewed paper · 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: Peer-reviewed paper
  • Test type: prospective or independent check
  • Tested: 62 forecasters from 23 meteorological stations ran retrospective and prospective checks on about 2,400 extreme-precipitation events and ranked NowcastNet first in 71% of cases against leading methods.
  • Still unknown: No limitations of NowcastNet are stated; the abstract reports no verification scores beyond expert preference, and the page says the team plans to extend the approach to other physics problems.
  • AI role: A neural network combined with physics-based precipitation evolution forecasts extreme precipitation up to three hours ahead from radar data.

Evidence sources: https://www.thss.tsinghua.edu.cn/info/1023/2133.htm · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-023-06184-4&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: 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, 576 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · 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.

Deep learning detects elevated tissue-derived cfDNA across cancer and cirrhosis cohorts

A study built a fragment-level methylation atlas from 521 noncancer tissue samples spanning 29 human tissue types and used it with cfSort, a supervised deep neural network, to estimate tissue composition in plasma cell-free DNA. Applying cfSort to patient data, the investigators found higher cfDNA fractions from the affected tissue in liver, lung, colorectal and stomach cancer and cirrhosis cohorts than in healthy controls; the estimated fractions also increased with cancer stage. The tissue-origin signal was reproduced on an independent whole-genome-bisulfite-sequencing dataset for liver cancer. In serial samples from four non-small-cell lung cancer patients receiving anti-PD-1 therapy, estimated liver or kidney tissue fractions tracked measured biochemical markers of organ injury.

Technology: cfSort, an ensemble of two dense deep neural networks trained on simulated mixtures using a fragment-level tissue methylation atlas

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: Applying cfSort to patient data, the investigators found higher cfDNA fractions from the affected tissue in liver, lung, colorectal and stomach cancer and cirrhosis cohorts than in healthy controls; the estimated fractions also increased with cancer stage.
  • Still unknown: The disease cohorts were retrospective and the clinical analyses demonstrate association and measurement feasibility, not prospective diagnostic benefit or improved outcomes. The clinical groups were 21 cirrhosis patients and 201 cancer patients, with 100 healthy controls; cross-platform liver-cancer validation used 24 cases and 32 controls. Anti-PD-1 injury monitoring involved only four NSCLC patients at three timepoints. The authors note that disease-specific methylation markers could improve cancer detection beyond tissue-of-origin fractions.
  • AI role: The team used the supervised cfSort model to deconvolve methylation profiles in plasma cfDNA and estimate contributions from 29 tissue types. This enabled the reported comparison of affected-tissue cfDNA fractions in cancer and cirrhosis patient cohorts versus controls and longitudinal tracking of tissue-injury signals during immunotherapy. The atlas and model were also evaluated against simulated mixtures and independent methylation data.

Evidence sources: https://par.nsf.gov/servlets/purl/10428542

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, 77 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 Fe–Mn–Ni catalysts for NOx reduction

An artificial neural network trained on literature data and a genetic algorithm helped identify Fe–Mn–Ni catalyst compositions for ammonia selective catalytic reduction. Researchers synthesized and tested model-guided candidates and report good NOx conversion for Fe₀.₃₁₉Mn₀.₂₀₆Ni₀.₄₇₅ over the tested temperature range. The work combines model selection, human screening and controlled experiments; the ratio is not a complete oxide formula and does not demonstrate industrial exhaust performance. The article text differs on whether the workflow involved two or four iterations, so this summary avoids an iteration count.

Technology: Artificial neural network and 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: The paper reports synthesis and catalytic conversion testing of selected Fe–Mn–Ni formulations.
  • Still unknown: Article reports inconsistent iteration counts; test range and lab results do not establish durability or practical flue-gas performance.
  • AI role: An ANN and genetic search ranked candidate Fe–Mn–Ni compositions; experimental results were used in the iterative optimization workflow.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC10666265/ · https://pubmed.ncbi.nlm.nih.gov/37393584/

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, PubMed metadata/affiliation and publisher date 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.

Machine-learning-guided experiments identify a high-power-factor PbSe:SnSb thermoelectric family

A closed-loop error-correction learning workflow guided synthesis of PbSe alloyed with SnSb and documented a previously unexplored low-temperature thermoelectric family. At 300 K, the best tested composition, 2 wt% SnSb-doped PbSe, had a measured power factor above 850 µW mK⁻² and more than twice that of PbSe. Density-functional calculations attribute the modeled transport improvement to Sn/Sb interactions that reduce hole effective mass.

Technology: Closed-loop error-correction learning with a neural-network update to a prior materials-property model; composition representations compared included Magpie, Roost and CrabNet

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 closed-loop error-correction learning workflow guided synthesis of PbSe alloyed with SnSb and documented a previously unexplored low-temperature thermoelectric family.
  • Still unknown: The result is specific to the tested powder-processing, composition and measurement protocol, including synthesis below 300 °C; it is not evidence of device-level or commercial performance. The reported mechanism for the increased power factor is supported by DFT calculations, not a direct measurement of carrier effective mass. The paper compares the 2 wt% composition with PbSe and reports more than 2× power factor at room temperature.
  • AI role: A model trained on published experimental thermoelectric data ranked candidate compositions, and a dense neural-network error-correction model incorporated local experimental comparisons to adapt predictions to the authors’ synthesis and measurement conditions. Iterative model-guided experiments selected the PbSe:SnSb family and 2 wt% SnSb-doped PbSe for characterization. AI supplied composition down-selection and adaptive experiment guidance; the authors synthesized and measured the samples.

Evidence sources: https://advanced.onlinelibrary.wiley.com/doi/10.1002/adma.202302575

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 70 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 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.

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.

CogSim helps predict fusion ignition at the NIF

Neural networks, Bayesian inference and physics simulations helped assess, before the experiment, the NIF design that produced 3.15 MJ of fusion energy from 2.05 MJ of laser energy delivered to the target on December 5, 2022. Experts designed the experiment; AI estimated its outcome and uncertainties. This is not the energy balance of the entire facility; the study was later published in Science in 2025.

Technology: CogSim (neural networks, transfer learning and Bayesian 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: The CogSim team predicted the December 5, 2022 shot outcome before the experiment (about 50.2% chance of exceeding break-even), and the result of 3.15 MJ from 2.05 MJ matched the prediction after peer review and independent diagnostics.
  • Still unknown: The page says the approach does not guarantee success and that techniques still have much to prove; one good prediction does not validate a model, and testing on more experiments was planned for 2023.
  • AI role: A generative model combining simulations, deep learning, experimental data and Bayesian statistics estimated the outcome and uncertainties of the NIF design before the shot.

Evidence sources: https://lmf.llnl.gov/news/high-performance-computing-ai-cognitive-simulation-helped-llnl-conquer-fusion-ignition · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1126/science.adm8201&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 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 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.

Sequence-conditioned CogMol designs SARS-CoV-2 spike and main-protease inhibitors

IBM Research and academic collaborators used the CogMol deep generative foundation model to design compounds from target protein sequence information alone for SARS-CoV-2 spike receptor-binding domain and main protease. In vitro inhibition was reported for two of four synthesized candidates for each target; the most potent spike-RBD inhibitor also showed activity against multiple variants in live-virus neutralization assays.

Technology: CogMol, a sequence-conditioned deep generative foundation model using a molecular variational autoencoder and protein–ligand binding 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: In vitro inhibition was reported for two of four synthesized candidates for each target; the most potent spike-RBD inhibitor also showed activity against multiple variants in live-virus neutralization assays.
  • Still unknown: The source describes a proof of concept with a small selected synthesis set from roughly 100 ranked designs. Only in-vitro target assays were performed; the leading RBD inhibitor was also tested in live-virus neutralization against variants. The paper says ADME attributes were not experimentally validated and further work is needed to confirm inhibitor binding mechanisms. These are preclinical leads, not medicines or demonstrated clinical benefit.
  • AI role: The team sampled CogMol conditioned on each target's amino-acid sequence to generate candidate small molecules; candidates were filtered and ranked in silico, synthesized and assayed. The model generated compounds without target-specific fine-tuning or target-structure input at inference; assays found micromolar inhibition for 2/4 synthesized candidates for each target.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC10284550/

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, 66 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.

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.

ATLAS graph neural network enables first observation of four-top-quark production

Using 140 fb⁻¹ of 13 TeV proton–proton collisions, ATLAS observed simultaneous production of four top quarks in same-sign dilepton and multilepton events, with 6.1σ observed (4.3σ expected) significance. The measured cross-section is 22.5⁺⁶·⁶₋₅·₅ fb, consistent with the Standard Model prediction within 1.8σ. A graph neural network separated the four-top signal from dominant backgrounds in the signal extraction.

Technology: Event-level graph neural network built with graph_nets/TensorFlow

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 140 fb⁻¹ of 13 TeV proton–proton collisions, ATLAS observed simultaneous production of four top quarks in same-sign dilepton and multilepton events, with 6.1σ observed (4.3σ expected) significance.
  • Still unknown: This is an ATLAS observation from its own 140 fb⁻¹ analysis after earlier ATLAS/CMS evidence; it is one paper-level observation, not a separate count for CMS or later combinations. The measured cross-section uncertainty is large and its central value is 1.8σ above the SM prediction, so this is not evidence of a significant deviation from the Standard Model. Background control regions and systematic uncertainties are part of the analysis.
  • AI role: The GNN discriminant classified events using a fully connected graph of reconstructed jets, leptons and missing transverse momentum, with node, edge and global features. It separated four-top signal from background and supplied the discriminant used in the signal extraction and cross-section fit.

Evidence sources: https://link.springer.com/article/10.1140/epjc/s10052-023-11573-0

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 1% of works of the same type, year and subfield (OpenAlex work metrics, 148 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.

Google DeepMind

AlphaDev: reinforcement learning discovers faster sorting routines

Working at the assembly-instruction level, the system found sorting routines up to 70% faster for sequences of 3–5 elements and about 1.7% faster for sequences over 250,000 elements; they were adopted in the LLVM libc++ standard library. A hashing algorithm 30% faster was added to Abseil. The largest gains are for short sequences.

Technology: AlphaDev (reinforcement learning)

Family AlphaDev: 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: AlphaDev found sort 3 and sort 5 routines with fewer instructions than human benchmarks, matched sort 4, and had lower latency on variable sorts. Libc++ versions were up to 70% faster for five elements and about 1.7% faster above 250,000 elements.
  • Still unknown: Hashing result (30% faster) not seen in the text read. Search is at assembly level; the paper reports optimality proof only for sort 3 (17 instructions). Results cover tested CPUs: ARMv8, Intel Skylake, AMD Zen 2.
  • AI role: AlphaDev, an AlphaZero-based reinforcement learning agent with a transformer encoder, built assembly programs instruction by instruction, rewarded for correctness and latency; the team translated results to C++.

Evidence sources: https://doi.org/10.1038/s41586-023-06004-9

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, 238 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 guides direct synthesis of higher-silica faujasite

Machine-learning models learned relationships between crystallization composition/conditions and zeolite characteristics from prior synthesis records, guiding human selection of conditions for faujasite synthesis. The team reports directly synthesized, unseeded, organic-structure-directing-agent-free FAU with Si/Al 3.5, compared with the prior reported maximum 2.8. Following ion exchange, laboratory tests report improved propane cracking and dehydrogenation rates per proton site relative to the lower-silica comparator. The FAU framework was already known; this result concerns composition and route, not a new framework or industrial performance.

Technology: Geometric harmonics, neural networks and 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: Researchers synthesized and characterized FAU and tested acid catalysts in propane cracking/dehydrogenation under laboratory conditions.
  • Still unknown: Known framework; activity comparison is assay-specific and does not establish industrial gains.
  • AI role: Models trained on prior synthesis records identified composition/condition settings that guided direct high-silica FAU synthesis.

Evidence sources: https://authors.library.caltech.edu/records/makq9-r3046

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 Caltech institutional record and deposited published article/supplement; date, affiliations, methods and abstract/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.

Abaucin: a targeted antibiotic against Acinetobacter baumannii

A model trained on about 7,500 compounds screened 6,680 new compounds in under two hours; of 240 tested in the laboratory, nine were antibiotics, including abaucin, which was effective against wound infections in mice. Activity is highly narrow (no effect on the other bacteria tested) and still preclinical.

Technology: Machine learning for compound 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: About 7,500 molecules were screened in vitro, a neural network predicted new compounds that were tested in the laboratory (about 240 tested, nine antibiotics), and abaucin controlled an A. baumannii infection in a mouse wound model.
  • Still unknown: The abstract states no explicit limitations; the press release says activity is highly narrow (no effect on the other bacteria tested) and the compound remains preclinical.
  • AI role: A neural network trained on screening results predicted structurally new compounds active against A. baumannii.

Evidence sources: https://news.mit.edu/2023/using-ai-scientists-combat-drug-resistant-infections-0525 · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41589-023-01349-8&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 date: 16 Nov 2026. Review scope: Original institutional announcement checked for the specific results, AI contribution and limitations cited in the antibiotic dossier or mathematical follow-up pilot. Full-paper reanalysis, external replication and an exhaustive follow-up search were not performed..

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 420 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
  • same-team extension · 14 Aug 2025: MIT identifies halicin and abaucin as earlier screening results motivating generative design of new NG1/DN1 candidates. The later study does not establish clinical efficacy of these earlier candidates. Source. Recorded 16 Nov 2026.

Follow-up checked 16 Nov 2026 within scope: Bounded check of the original announcements cited by the events recorded here, including explicit links to later program work. This is not an exhaustive literature search or ongoing monitoring commitment..

Directed relationships from this entry; they do not merge catalog counts.

  • 2020-02-halicin · extends: The source explicitly traces this A. baumannii screening study to the earlier halicin program. Distinct candidate and experiment; not independent confirmation of halicin. Source

Active learning maps wind-tunnel terrain equivalence

A learned surrogate selected automated wind-tunnel experiments that mapped terrain configurations with similar second-order wind statistics. Different terrain shapes could meet the study's equivalence criterion. This does not imply identical full turbulence fields or equal structural safety; higher-order statistics can differ. Estimated experiment savings are not a fully executed randomized comparison with alternative sampling.

Technology: Gaussian-process sequential 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: Automated terrain changes and measured wind profiles defined experimental equivalence regions.
  • Still unknown: Generality beyond the facility, full-field equivalence and safety outcomes remain unestablished.
  • AI role: Gaussian-process sequential active learning supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC10209069/

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 retrieved via browser request: publication metadata, Gaussian-process Methods and terrain-study 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.

ExoMiner and multiplicity validate 69 new exoplanets

ExoMiner V1.2 combines deep learning with information about other signals in the same star system to validate 69 planets in Kepler data. This is statistical validation of transit signals, not new direct observation; the result is distinct from the 301 planets validated in 2021.

Technology: ExoMiner V1.2 / multiplicity boost

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: ExoMiner V1.2 with multiplicity information was applied to Kepler data and validated 69 planets, compared against baseline classifiers and prior validation work.
  • Still unknown: The arXiv page states no limitations, and the statistical validation and false-positive tests are not described on the page read.
  • AI role: A deep-learning classifier (ExoMiner V1.2) combined with information about other signals in the same star system to validate transit signals.

Evidence sources: https://arxiv.org/abs/2305.02470 · https://newsroom.usra.edu/discovery-of-69-new-exoplanets-using-machine-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: 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.

Nonlocal quantum interference without entanglement of intrinsic properties

An AI- and graph-theory-based prediction was confirmed in two experiments: four photons exhibit nonlocal interference without requiring entanglement of their intrinsic properties. The superposition concerns their generation locations; applications to metrology remain prospective.

Technology: AI and graph theory for quantum experiments

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-photon nonlocal interference without entanglement of intrinsic properties was demonstrated experimentally, and the abstract reports the interference controlled by an undetected photon.
  • Still unknown: The abstract states no limitations; the press release says applications to metrology remain prospective.
  • AI role: The press release says an AI- and graph-theory-based prediction was confirmed in two experiments; the Nature Communications abstract does not describe the AI prediction.

Evidence sources: https://mpl.mpg.de/news/article/where-quantum-particles-are-born-ai-reveals-new-insights-beyond-entanglement · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41467-023-37228-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: 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, 44 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.

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.

New zirconium-indium-nickel superconductor identified with machine learning

Predictive models and experts guided fabrication of candidates and identified a zirconium, indium and nickel alloy with a superconducting transition around 9 K. The press release reports a three-month discovery; this is not a room-temperature superconductor or a demonstration of industrial use.

Technology: MITHRIL / machine learning

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: Closed-loop ML predictions were tested experimentally: candidates were fabricated and measured, and results fed back to refine the model; a Zr-In-Ni superconductor with transition around 9 K was found in four cycles.
  • Still unknown: The arXiv page states no limitations and omits synthesis, transition temperature and model accuracy details; the press release notes training-data human bias and chemically impossible generated compositions, and cites an npj Computational Materials paper that was not read.
  • AI role: Machine-learning models trained on public superconductor datasets predicted and ranked candidate compositions for fabrication.

Evidence sources: https://www.jhuapl.edu/news/news-releases/230503-ai-discovers-novel-superconductor · https://arxiv.org/abs/2212.11855

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.

Baidu Research

LinearDesign: more stable mRNA sequences designed by an algorithm

The algorithm, based on a parsing technique from natural-language processing, finds the most stable mRNA sequence for a given protein in about eleven minutes. According to Baidu’s press release, whose study appeared in Nature, sequences produced this way showed greater stability, protein expression and antibody response than comparators in animal experiments. Figures are company-reported; the Baidu release copy on EurekAlert was reviewed, not a baidu.com page.

Technology: LinearDesign

Family LinearDesign: 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: LinearDesign-designed mRNA for COVID-19 and VZV was compared with wildtype and codon-optimized sequences in animal experiments, measuring half-life, expression and antibody response; the figures are company-reported and the release describes no independent replication.
  • Still unknown: The Nature abstract states no specific study limitations and the release states none; the release says benefits such as lower doses and fewer side effects are possibilities not tested.
  • AI role: A parsing-based algorithm (LinearDesign) optimized mRNA structural stability and codon usage for a given protein.

Evidence sources: https://www.eurekalert.org/news-releases/987999 · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-023-06127-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, 560 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.

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.

Machine-learning classification shows rigidity-corrected moment release is similar across earthquake depths

Random-forest classification of source-time-function characteristics for 3,675 global Mw > 5.5 earthquakes showed that depth-dependent rigidity explains much of the differences in rupture duration, length and stress drop between shallow, intermediate and deep earthquakes. Once medium rigidity is corrected, the analyzed events exhibit similar moment-release processes, supporting a constant-strain-drop interpretation across depths and host-rock types.

Technology: Random-forest machine-learning classification and 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: Random-forest classification of source-time-function characteristics for 3,675 global Mw > 5.5 earthquakes showed that depth-dependent rigidity explains much of the differences in rupture duration, length and stress drop between shallow, intermediate and deep earthquakes.
  • Still unknown: This is a global observational analysis of 3,675 moderate-to-large earthquakes. Its constant-strain-drop conclusion is a supported scaling interpretation, not proof that every earthquake rupture mechanism is identical; depth class and source-time-function estimates have sampling limitations.
  • AI role: Compared source-time-function features across depth classes and identified surrounding-medium rigidity as the principal explanatory feature, enabling the rigidity-corrected cross-depth comparison.

Evidence sources: https://www.nature.com/articles/s41561-023-01176-5

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, 20 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.

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.

Longitudinal EHR machine learning identifies patients who may need autoimmune rheumatic-disease evaluation years earlier

A random-forest model trained on longitudinal electronic health records from 161,584 people across two institutions identified records associated with high-specificity autoantibody testing for systemic autoimmune rheumatic diseases (SARDs). The model achieved AUROC 0.93 in internal validation and 0.87 on the external All of Us cohort. In a separate BioMe cohort, model probabilities tracked SARD and other autoimmune diagnoses; records could predict later autoantibody testing and rheumatology encounters up to five years before those events. The study also found higher autoantibody yield among model-prioritized individuals than under the comparison clinical-testing rate.

Technology: Random-forest machine-learning model using longitudinal electronic-health-record diagnoses, medications, laboratory results and vital signs

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 random-forest model trained on longitudinal electronic health records from 161,584 people across two institutions identified records associated with high-specificity autoantibody testing for systemic autoimmune rheumatic diseases (SARDs).
  • Still unknown: The model predicts clinical suspicion and need for rheumatological testing, not a SARD diagnosis. The study was retrospective and opportunistic; autoantibody testing and rheumatology encounters are imperfect proxies for clinical need, and some diseases were not included in the training target. The authors state prospective studies are needed to determine whether model-guided testing improves care or outcomes. The five-year lead is retrospective prediction of later evaluation events, not proven earlier diagnosis or patient benefit.
  • AI role: The model ranked patients by probability of receiving SARD-specific autoantibody testing, using clinical data preceding the test or rheumatology encounter. Investigators evaluated performance in internal validation, the external All of Us cohort and an independent BioMe cohort, and tested whether higher scores tracked diagnosed autoimmune disease and identified later testing or encounters earlier in retrospective records.

Evidence sources: https://www.nature.com/articles/s41467-023-37996-7.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.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 40 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 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.

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.

Machine-learning screening yields novel potent acetylcholinesterase and butyrylcholinesterase inhibitors

NCATS researchers trained machine-learning QSAR models and applied them to a roughly 360,000-compound library. Experimental colorimetric and orthogonal fluorescence assays confirmed new AChE/BChE inhibitors; the study reports 22 AChE and 67 BChE inhibitors with IC50 below 5 μM, including novel structural series, with some submicromolar compounds.

Technology: Ensemble comparison of machine-learning QSAR classifiers (including XGBoost, random forest, SVM, neural network and naïve Bayes)

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 colorimetric and orthogonal fluorescence assays confirmed new AChE/BChE inhibitors; the study reports 22 AChE and 67 BChE inhibitors with IC50 below 5 μM, including novel structural series, with some submicromolar compounds.
  • Still unknown: These are biochemical inhibitor assays and early medicinal-chemistry leads; they do not establish brain exposure, in-vivo efficacy, safety or clinical benefit. Several known compounds were newly identified as inhibitors, so the count of all assay hits should not be described as entirely novel molecules.
  • AI role: Models trained on quantitative high-throughput screening data prioritized compounds in a ~360,000-compound collection for AChE/BChE inhibition and selectivity assays. Researchers tested 2,762 high-scoring compounds in colorimetric assays and used an orthogonal fluorescence assay to confirm predicted inhibitors; subsequent structure–activity analysis identified novel scaffolds.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC10688023/

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.

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.

Machine learning finds ribbon-knot certificates and rules out candidate pairs

Bayesian optimization and reinforcement learning search for sequences of knot moves that certify a knot is ribbon. The primary v1 reports that both knots in 843 specified pairs were shown ribbon, eliminating those pairs as potential counterexamples in the studied strategy. It also recovers a separately independently established set of 1,705 ribbon knots; these are not claimed as novel. This does not resolve the smooth four-dimensional Poincaré conjecture.

Technology: Bayesian optimization; reinforcement learning

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 specific mathematical or algorithmic result summarized here.
  • Still unknown: The result is attributed to the paper; no independent proof review beyond the authors’ source text is claimed.
  • AI role: The algorithms search move sequences; a successful path is a verifiable certificate that authors use to prove ribbon status.

Evidence sources: https://arxiv.org/html/2304.09304v1

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 preprint full text or abstract reviewed; evidence 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.

PRIMO: a sharper image of the M87 black hole

A machine-learning method trained on more than 30,000 simulated black-hole images reconstructed M87* from 2017 Event Horizon Telescope data, using the array’s full resolution: the ring is about half as wide as in the 2019 image. It fills gaps with patterns learned from simulations, without quantified uncertainty.

Technology: PRIMO (dictionary learning)

Publication: Publication status unspecified in annotation · 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: Not reported
  • Test type: retrospective or benchmark
  • Tested: The PRIMO reconstruction, built from over 30,000 simulated black-hole images, was checked for consistency with the 2017 EHT data and the 2019 image; no held-out tests are described.
  • Still unknown: The article states that generating the image requires assuming an appropriate form of the missing information and describes no quantified uncertainties; implications for the black hole are framed as future possibilities.
  • AI role: A machine-learning method trained on simulated images filled gaps in the 2017 EHT data to reconstruct the M87* image.

Evidence sources: https://news.research.gatech.edu/2023/04/13/sharper-look-m87-black-hole

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 10% of works of the same type, year and subfield (OpenAlex work metrics, 56 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.

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.

Machine-learning analysis identifies kinematic evidence for a protoplanet in the HD 142666 disk

A machine-learning model identified a strong, localized non-Keplerian gas-motion feature in the protoplanetary disk around HD 142666. Hydrodynamic simulations of a system containing a 5-Jupiter-mass planet at 75 au reproduced the observed kinematic structure; the authors conclude that HD 142666 hosts an embedded planet by current field criteria. The planet is inferred from disk kinematics rather than directly imaged.

Technology: EfficientNetV2 and RegNet image classifiers (six channel-count variants), trained on synthetic planet/no-planet disk observations

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 machine-learning model identified a strong, localized non-Keplerian gas-motion feature in the protoplanetary disk around HD 142666.
  • Still unknown: The paper’s planet conclusion is an inference from gas kinematics supported by simulations, not a direct image or independent dynamical mass measurement. The models were classifiers, not trained to localize planets, and the 5-Jupiter-mass, 75-au configuration is a simulation that reproduces the observed feature, not a separately measured planet parameter.
  • AI role: The classifiers applied to ALMA data assigned high planet-presence scores and their activation maps highlighted the previously unreported non-Keplerian velocity kink; the authors then used standard hydrodynamic simulations to test whether an embedded planet reproduced that feature.

Evidence sources: https://iopscience.iop.org/article/10.3847/1538-4357/acc737

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 40 of works of the same type, year and subfield (OpenAlex work metrics, 3 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.

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.

Machine-learning-guided calculations predict four rare-earth-free magnetic Fe–Co–Si compounds

A crystal-graph neural network screened 357,480 hypothetical Fe–Co–Si structures; adaptive genetic search and density-functional calculations then identified four predicted ternary structures with calculated saturation polarization ≥1.0 T, easy-axis anisotropy ≥1.0 MJ/m³, and Curie temperature >840 K. Their calculated energies lie within 70 meV/atom of the ternary convex hull, supporting but not proving experimental accessibility. The paper proposes the structures for synthesis; it does not experimentally demonstrate them.

Technology: Crystal graph convolutional neural network (CGCNN) with an adaptive genetic algorithm

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 crystal-graph neural network screened 357,480 hypothetical Fe–Co–Si structures; adaptive genetic search and density-functional calculations then identified four predicted ternary structures with calculated saturation polarization ≥1.0 T, easy-axis anisotropy ≥1.0 MJ/m³, and Curie temperature >840 K.
  • Still unknown: This is a theoretical, ML-guided first-principles prediction; the paper proposes the structures for synthesis and does not report experimental synthesis. The abstract describes four compounds; the results section mentions five structures meeting initial thresholds, of which four are dynamically stable. Report the four final candidates and their calculated properties only. The work establishes neither actual permanent-magnet performance nor synthesizability.
  • AI role: CGCNN models rapidly screened hypothetical structures and predicted formation energy and magnetic polarization; first-principles results were fed back to train improved models across successive generations, while an adaptive genetic algorithm searched lower-energy structures. ML materially narrowed the space for the final density-functional magnetic-property calculations.

Evidence sources: https://journals.aps.org/prmaterials/abstract/10.1103/PhysRevMaterials.7.034410

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 75 of works of the same type, year and subfield (OpenAlex work metrics, 18 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.

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.

Reinforcement-learning-guided CdS shell-growth route using AlphaFlow

AlphaFlow used reinforcement learning with an automated droplet reactor to select and optimize reagent sequences for CdS-shell growth on supplied CdSe quantum dots. Under the authors’ aged, diluted starting-material comparison, the selected route achieved a 26 nm larger absorption-peak shift and 450% higher photoluminescence intensity after six cycles than the conventional sequence; freshly diluted starting material showed a different comparison. This is a lab route to a known nanomaterial, not discovery of a new compound or proof of universal superiority.

Technology: Reinforcement-learning agent with neural reward regression and boosted-tree viability classification in an automated fluidic 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: Primary Nature Communications article reports a six-cycle experimental comparison on aged diluted starting material; conditions and starting materials constrain result.
  • Still unknown: Earlier route disclosures and independent replication are not established.
  • AI role: Reinforcement-learning-guided reaction sequence selection and optimization linked to measured shell-growth signals.

Evidence sources: https://www.nature.com/articles/s41467-023-37139-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: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary publisher full text checked for identity, dates, organization affiliations, AI method and bounded experimental 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.

Adaptive learning finds higher-activity Fe–N–C oxygen-reduction catalysts

An adaptive-learning design-of-experiments loop guided robotic high-throughput synthesis of Fe–N–C oxygen-reduction electrocatalysts across a six-dimensional process/composition space. Starting from 36 characterized samples, the team synthesized 14 additional catalyst formulations from a nominal search space of roughly 10 billion combinations. Four outperformed the best sample in the initial dataset; the best showed about 33% higher ORR activity, reported as 16.3 mA mg⁻¹ at 0.8 V in rotating-disk measurements.

Technology: Adaptive learning framework with machine-learning surrogates, statistical inference, uncertainty quantification, and global optimization, paired with robotic high-throughput synthesis

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 adaptive-learning design-of-experiments loop guided robotic high-throughput synthesis of Fe–N–C oxygen-reduction electrocatalysts across a six-dimensional process/composition space.
  • Still unknown: The finding concerns optimized Fe–N–C catalyst formulations and synthesis conditions within the study’s defined six-dimensional search; it is not discovery of a new catalyst family or proof that all approximately 10 billion combinations were tested. The 33% improvement is relative to the best sample in the initial 36-sample dataset. Activity was measured using rotating-ring disk electrode testing, not a fuel-cell device or long-duration operating demonstration. The study reports 14 new synthesized catalysts and four that beat the initial samples; these are one paper-level result, not separate entries.
  • AI role: The model-based design-of-experiments tool used current measurements and uncertainty to select subsequent synthesis conditions in a six-dimensional search space. It guided the selection of 14 samples for synthesis and measurement, producing four catalysts that exceeded the initial best. Robotic high-throughput synthesis and RRDE testing generated the experimental feedback; researchers interpreted the results.

Evidence sources: https://www.sciencedirect.com/science/article/pii/S0378775322015609

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, 27 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.

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.

Virtual screening identifies antiplasmodial beta-haematin inhibitor chemotypes

Molecular docking and Bayesian/SVM classifiers screened a ChemDiv 300k library. Researchers tested prioritized commercial compounds and reported coumarin and iminodipyridinopyrimidine chemotypes active against beta-haematin and whole-cell P. falciparum, including chloroquine-resistant parasites. This is one bounded in-vitro campaign, not clinical evidence.

Technology: Bayesian and support-vector classifiers; molecular 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: Primary abstract reports prioritized compounds with beta-haematin and whole-cell parasite activity.
  • Still unknown: A 2023 conference abstract may disclose an early coumarin result; the article’s specific two-chemotype campaign is retained without priority claim; no clinical evidence.
  • AI role: Sequential ligand/structure virtual screening prioritizes commercial compounds for biochemical and parasite assays.

Evidence sources: https://doi.org/10.1038/s41598-023-29273-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: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: PMC-hosted primary full article; PubMed primary record/abstract. Material checked for the specific claims stated in the record..

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-generated catalyst design led to experimentally tested polymerization catalysts

A regression-transformer model fine-tuned on historical ring-opening-polymerization data generated candidate catalysts. Researchers selected synthetically accessible, modified surrogates of one model-generated catalyst for laboratory testing. One surrogate, 5c, reached 60% conversion in 32 minutes for monomer 2a and 100% conversion in 45 minutes for monomer 2d, with reported dispersities of 1.06 and 1.34, respectively. The first ChemRxiv v1 (2022-06-27) described computational generation and recovery of a literature-known catalyst. The experimental results appeared in ChemRxiv v2, posted 2023-02-13, and the 2023 peer-reviewed journal article followed on 2023-06-21. The measured tests concern modified surrogates, not the unmodified generated structure.

Technology: Regression Transformer generative model fine-tuned on historical polymerization data represented with Chemical Markdown Language

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 reports model-derived catalyst 3g and synthetically accessible modified surrogates 5c/5d. Surrogate 5c produced 60% conversion of monomer 2a in 32 minutes (dispersity 1.06) and 100% conversion of monomer 2d in 45 minutes (dispersity 1.34).
  • Still unknown: The tested catalysts were modified surrogates rather than the unmodified model-generated 3g structure. These are laboratory ring-opening-polymerization outcomes, not process-scale or commercial performance. The earlier 2022 v1 reported model generation and literature-known catalyst recovery; the experimentally validated result is reported in v2 on 2023-02-13 and in the later journal article.
  • AI role: A regression transformer generated catalyst 3g from historical polymerization data; researchers tested synthetically accessible modified surrogates 5c and 5d that preserved its endocyclic guanidine motif.

Evidence sources: https://chemrxiv.org/engage/chemrxiv/article-details/62b60865e84dd185e60214af · https://chemrxiv.org/engage/chemrxiv/article-details/63e668409da0bc6b33b1870c · https://pmc.ncbi.nlm.nih.gov/articles/PMC10284867/ · https://pubmed.ncbi.nlm.nih.gov/37344485/

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 v1/v2 records and linked full v1 PDF for version chronology; primary PubMed Central journal full text for model-to-catalyst relationship and measured outcomes; PubMed metadata for journal publication date and 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.

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.

Unsupervised machine learning reveals previously missed hydraulic-fracturing seismicity in Live Oak County

An unsupervised repeating-signal detector, combined with template matching, exposed small earthquake bursts in a region of the Eagle Ford shale where conventional catalogs had detected little seismicity. The authors report that the new events temporally coincided with hydraulic-fracturing activity, and their regional analysis found detectable seismicity began after injection flux exceeded about two million barrels per month despite similar cumulative injection volume; this supports, but does not prove, a causal role for injection rate and is limited by sparse station coverage.

Technology: Repeating Signal Detector (RSD), an unsupervised machine-learning waveform-family 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: An unsupervised repeating-signal detector, combined with template matching, exposed small earthquake bursts in a region of the Eagle Ford shale where conventional catalogs had detected little seismicity.
  • Still unknown: The RSD detections were manually inspected and some event locations were poorly constrained; the authors note station density limits detection and do not establish causation from temporal correlation alone. This paper extends an earlier Eagle Ford study but reports a distinct Live Oak County seismicity region and subsequent time-period catalog.
  • AI role: RSD searched continuous seismograms without relying on preselected earthquake templates, clustered similar signals, and revealed small-magnitude event bursts missed by the previous catalog. Researchers manually inspected signals and analyzed their timing relative to hydraulic-fracturing records.

Evidence sources: https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2022JB025436

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 61 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 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.

Google Research

Real-time wildfire perimeters from satellite imagery

A model using geostationary imagery tracks fire perimeters with updates every 10–15 minutes and, according to the page, is more accurate than similar products and the previous model based on NOAA GOES-16/17 data; shown in Search and Maps for large fires in the United States, Mexico, much of Canada and Australia. Resolution is 2 km; clouds can hide fires; figures appear only in an image.

Technology: Model on geostationary satellite imagery

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: Fire perimeters were compared with official fire scars using precision, recall and distance metrics; a comparison with an earlier GOES-16/17 model and a 2022 US test set unavailable during training are also reported.
  • Still unknown: The post states limitations: smoke and clouds hide fires, resolution is about 2 km and coarser away from nadir, and coverage is limited to large fires in the listed regions; no paper is cited.
  • AI role: A model using geostationary satellite imagery tracks fire perimeters with updates every 10-15 minutes.

Evidence sources: https://research.google/blog/real-time-tracking-of-wildfire-boundaries-using-satellite-imagery/

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.

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.

Deep learning searches for technosignatures around 820 nearby stars

A deep-learning method analyzed 150 TB of Green Bank Telescope data from 820 stars, previously searched in 2017 without interesting signals, and found eight narrowband signals of interest with nonzero drift, present only in observations where the target was in view. Follow-up observations did not find them again; no claim of extraterrestrial origin.

Technology: Deep learning on radio data

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 deep-learning method analyzed Green Bank Telescope data from 820 stars and flagged eight narrowband candidates; the page reports no follow-up results and no injection tests, and the abstract gives no false-positive figures.
  • Still unknown: The page states no limitations; rejecting human radio interference is named as the core challenge, and the full paper's methods, which the listing says were later published in Nature Astronomy, were not read.
  • AI role: A beta-Convolutional Variational Autoencoder flagged technosignature candidates in a semi-unsupervised manner.

Evidence sources: https://arxiv.org/abs/2301.12670 · https://www.eurekalert.org/news-releases/978138

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 2/5 · 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.

Elastic-net proteomic risk modeling and multi-cohort validation identify proteins associated with incident diabetes

An untargeted analysis of 4,955 plasma proteins in the ARIC cohort identified 140 associations with incident diabetes; 64 were significant in an internal sample and 47 (including 22 reported as novel) were associated in the external Singapore Multi-Ethnic Cohort. Elastic-net Cox modeling selected a 27-protein panel for 20-year risk prediction. In internal validation, adding those proteins to conventional risk factors including fasting glucose and HbA1c increased the C statistic from 0.720 to 0.731. Genetic-instrument analyses supported possible causal roles for SHBG, ATP1B2 and GSTA1, although ATP1B2’s genetic and observational directions differed.

Technology: Elastic-net Cox regression for protein feature selection and incident-diabetes risk prediction

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: An untargeted analysis of 4,955 plasma proteins in the ARIC cohort identified 140 associations with incident diabetes; 64 were significant in an internal sample and 47 (including 22 reported as novel) were associated in the external Singapore Multi-Ethnic Cohort.
  • Still unknown: The 47-protein associations were externally tested in Singapore MEC; the elastic-net predictive model itself was internally validated in ARIC, not independently externally validated. Its incremental C statistic over covariates including fasting glucose and HbA1c was modest (0.011; P=0.04), and the protein model overestimated observed diabetes risk across internal-validation deciles. The causal claims are genetic-instrument inferences, not experimental proof; ATP1B2 had opposite directions in genetic-instrument versus observational analyses. Cohort associations do not establish clinical utility.
  • AI role: Elastic net selected a 27-protein combination from candidate proteins to improve 20-year incident-diabetes risk prediction beyond clinical and cardiometabolic covariates. The model was trained in the ARIC discovery subset and internally evaluated in a held-out ARIC subset. Separate Cox association and genetic-instrument analyses reported the externally validated protein associations; the record does not attribute those analyses to machine learning.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC10090896/

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 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.

Support-vector regression selects more selective tetrahydropyran catalysts

Models of catalyst fragments and reaction settings selected catalysts for asymmetric tetrahydropyran synthesis. A tested catalyst reached an enantiomeric ratio of 91.5:8.5, above the training examples' maximum of 82:18. The result improves a known catalyst class for a focused reaction; it does not establish invention of the reaction or broad generality.

Technology: Support-vector regression with tunable fragment descriptors

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: Previously untested catalyst substituents were prepared and experimentally evaluated.
  • Still unknown: Exhaustive catalyst prior art, independent replication and broader reaction scope were not assessed.
  • AI role: Support-vector regression with tunable fragment descriptors supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://onlinelibrary.wiley.com/doi/full/10.1002/anie.202218659

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 article dates, affiliations and selected model/experimental-validation Results; supplementary data 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.

CNN-augmented focal mechanisms reveal stress-linked faulting changes around major southern California earthquakes

CNN phase and first-motion pickers supplied additional waveform polarities and arrivals for a 297,478-event southern California focal-mechanism catalog, about 40% larger than the prior catalog. The authors used the expanded catalog to observe systematic post-mainshock faulting-style changes spatially correlated with estimated static stress perturbations from the 2010 El Mayor–Cucapah and 2019 Ridgecrest earthquakes, including changes more than 100 km from the epicentres.

Technology: Convolutional neural networks for seismic phase arrival and first-motion polarity picking

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: CNN phase and first-motion pickers supplied additional waveform polarities and arrivals for a 297,478-event southern California focal-mechanism catalog, about 40% larger than the prior catalog.
  • Still unknown: Catalog quality varies with location and station coverage; the authors report uncertainty in some solutions and frame stress processes as associations/inferences. The catalog provides improved observations, not a claim of earthquake predictability.
  • AI role: Extracted additional phase arrivals and P-wave first-motion polarities from noisy seismograms; combined with existing observations in HASH focal-mechanism solutions to increase catalog coverage and resolution for small earthquakes.

Evidence sources: https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2022JB025975

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, 47 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.

Random-forest metabolite profiling reduces false-positive newborn screens for four inherited disorders

A 121-analyte LC–MS/MS panel combined established markers, untargeted-metabolomics features and chromatographically resolved isobars with random-forest classification of California newborn-screen dried blood spots. In 883 archived infant samples, the study reports that the second-tier approach maintained 100% sensitivity while reducing false positives by 83% for glutaric acidemia type I, 84% for methylmalonic acidemia, 100% for ornithine transcarbamylase deficiency and 51% for very-long-chain acyl-CoA dehydrogenase deficiency. Several long-chain acylcarnitine and hydroxy-acylcarnitine isobar features contributed to classification, although their chemical identities were unresolved.

Technology: Random Forest classifier trained on targeted LC–MS/MS metabolite profiles

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 883 archived infant samples, the study reports that the second-tier approach maintained 100% sensitivity while reducing false positives by 83% for glutaric acidemia type I, 84% for methylmalonic acidemia, 100% for ornithine transcarbamylase deficiency and 51% for very-long-chain acyl-CoA dehydrogenase deficiency.
  • Still unknown: This was a proof-of-principle second-tier simulation on residual archived California newborn-screen specimens, not a prospective implementation study. The validation sample was not a fully independent external set, and archived-sample degradation and processing batch effects limited biomarker discovery; some discriminating isobar identities and biological significance remain unknown. The reported sensitivity and false-positive reductions are cohort-specific and do not establish improved health outcomes or routine clinical utility.
  • AI role: The classifier modeled multianalyte profiles from true-positive and false-positive screen-positive samples to distinguish the groups for four inherited metabolic disorders. Feature-importance analysis showed that complex profiles, including unresolved long-chain acylcarnitine/hydroxy-acylcarnitine isobars, contributed to classification; the model's second-tier performance was assessed against known case status.

Evidence sources: https://onlinelibrary.wiley.com/doi/full/10.1002/jimd.12591

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 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.

Sybil: lung-cancer risk prediction up to six years from a low-dose CT

The model predicts risk from a single CT with a six-year C-index of 0.75 (NLST), 0.81 (MGH) and 0.80 (CGMH), and a one-year ROC area between 0.86 and 0.94. Retrospective validation on three independent datasets; trained only on smokers, with prospective testing and testing in nonsmokers as next steps.

Technology: Sybil (neural network on CT scans)

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 trained on NLST low-dose CT scans and validated on a held-out NLST set and two independent cohorts (MGH and CGMH), with six-year C-index 0.75, 0.81 and 0.80; the abstract describes no prospective testing.
  • Still unknown: The abstract states that future study is required to understand clinical applications; the press release says prospective testing and testing in nonsmokers are next steps.
  • AI role: A deep learning model predicts lung-cancer risk from a single low-dose chest CT.

Evidence sources: https://news.mit.edu/2023/ai-model-can-detect-future-lung-cancer-0120 · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1200/JCO.22.01345&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, 332 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.

A liver-cancer target and molecule identified with Pharma.AI

Pharma.AI combined PandaOmics, Chemistry42 and an AlphaFold-predicted structure to identify a hepatocellular-carcinoma target and produce a hit molecule within 30 days of target selection, synthesizing seven compounds. A second round produced a more potent hit; an early Chemical Science study, with no clinical efficacy demonstrated.

Technology: Pharma.AI, PandaOmics, Chemistry42 and 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: Selected compounds were synthesized and tested in biological assays: the first hit bound CDK20 (Kd about 9.2 µM) and a second AI-driven round gave a more potent compound with greater antiproliferative activity in liver cancer cells; the abstract describes no in vivo testing.
  • Still unknown: The abstract notes AlphaFold predictions have varying confidence, the first hit came from only 7 synthesized compounds, and potency values vary widely; no clinical efficacy is demonstrated.
  • AI role: Pharma.AI platforms (PandaOmics, Chemistry42) used AlphaFold-predicted structures to select a liver-cancer target and generate candidate molecules.

Evidence sources: https://www.utoronto.ca/news/researchers-use-ai-powered-database-design-potential-cancer-drug-30-days · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1039/D2SC05709C&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, 403 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.

Autonomous X-ray mapping reveals previously unseen block-copolymer morphologies, including a nanoscale ladder

An autonomous experiment planner selected synchrotron X-ray measurements across a combinatorial block-copolymer blend sample. The resulting map and targeted electron microscopy revealed previously unreported alternating, skew/bilayer, and ladder-like self-assembled morphologies, and the authors infer a general mechanism of local subdomain partitioning.

Technology: gpCAM autonomous experimentation framework for selecting high-value measurements

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 autonomous experiment planner selected synchrotron X-ray measurements across a combinatorial block-copolymer blend sample.
  • Still unknown: The newly discovered results are nanoscale morphologies in one specific polymer-blend/template system; the paper does not show that each morphology is already useful in a device. AI selected measurements, while domain researchers performed electron-microscopy confirmation and interpretation.
  • AI role: The algorithm autonomously chose where to collect synchrotron X-ray scattering data and updated its model online; researchers then used electron microscopy to validate and interpret the novel morphologies.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC9839324/

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, 47 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 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.

Machine learning guides a proposed UCd3 intermetallic phase assignment

A support-vector classifier using composition-derived elemental features predicted a Cu₃Au-type structure for UCd₃ and guided synthesis. Powder X-ray diffraction from the second of two samples showed peaks the authors fit to a UCd₃ phase with a = 4.648(2) Å; the sample also contained uranium, uranium oxide and cadmium, while the first sample differed. This is the authors’ proposed phase assignment from a multiphase powder, not phase-pure material or independent structural confirmation.

Technology: Support-vector machine discriminant analysis; Gaussian-process regression for feature completion

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 report synthesis and powder X-ray diffraction; the proposed phase was present in the second of two samples.
  • Still unknown: Only one synthesis batch/sample (the second of two samples) showed the additional peaks assigned by the authors to UCd₃; the powder was multiphase. The checked 2022 primary review says UCd₁₁ was the only previously reported U-Cd compound, but this is not an independent confirmation of the proposed UCd₃ structure.
  • AI role: SVM classification of composition-derived descriptors predicted a Cu₃Au-type structure and motivated experimental synthesis/testing.

Evidence sources: https://www.cambridge.org/engage/api-gateway/coe/assets/orp/resource/item/63b23bbd16e9a864303c5bf0/original/machine-learning-descriptors-in-materials-chemistry-prediction-and-experimental-validation-synthesis-of-novel-intermetallic-u-cd3.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 ChemRxiv v1 full text, affiliation/date metadata, methods and diffraction discussion, plus primary UCd11 literature review for prior-compound context..

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.

Open the filtered catalog for charts and filters, or browse all organizations and model families.