AI Science Tracker

ESM / ESMFold

18 discoveries name this family, 2022 to 2026

Where ESM / ESMFold appears in the catalog

Why linked?

These associations match family names in discovery titles and technology text. They are not verified model-version assignments or causal attribution. Missing links do not establish absence of use. A record may link to multiple families.

Represented fields

biology · 16 linked records
medicine · 2 linked records

Test contexts

direct test · 16 linked records
retrospective or benchmark · 2 linked records

Annotation coverage

Evidence annotation present · 18 linked records

Examples

Up to three linked records, newest publication date first, then stable ID. This selection is not a ranking.

Discoveries

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

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

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

Family ESM / ESMFold: profile · models page

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

Result, AI contribution and limits

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

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

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

Source checked · editorial details

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

Source type: paper. Paper: linked separately.

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

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

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

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

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

Family ESM / ESMFold: profile · models page

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

Result, AI contribution and limits

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

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

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

Source checked · editorial details

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

Source type: paper. Paper: linked separately.

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

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

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

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

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

Family ESM / ESMFold: profile · models page

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

Result, AI contribution and limits

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

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

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

Source checked · editorial details

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

Source type: paper. Paper: linked separately.

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

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Diffusion-guided peptide design yields antimicrobial candidate Arcinin

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

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

Family ESM / ESMFold: profile · models page

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

Result, AI contribution and limits

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

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

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

Source checked · editorial details

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

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

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

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

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

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

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

Family ESM / ESMFold: profile · models page

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

Result, AI contribution and limits

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

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

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

Source checked · editorial details

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

Source type: paper. Paper: linked separately.

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

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

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

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

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

Family ESM / ESMFold: profile · models page

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

Result, AI contribution and limits

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

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

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

Source checked · editorial details

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

Source type: other. Paper: linked separately.

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

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine-learning-guided mining identifies 91 previously unknown PET hydrolases, including acid-active enzymes

Researchers iteratively combined sequence mining, machine-learning activity/property predictors and high-throughput assays to screen natural enzyme homologs. They experimentally identified 91 previously unknown PET hydrolases among the 212 candidates assayed; 35 had activity on crystalline PET at pH 4.5, and four matched or exceeded a benchmark enzyme under that screening condition.

Technology: Three-round iterative machine-learning pipeline (including VAE/rank model, supervised activity predictors, ESM embeddings and property predictors)

Family ESM / ESMFold: profile · models page

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

Result, AI contribution and limits

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

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Researchers iteratively combined sequence mining, machine-learning activity/property predictors and high-throughput assays to screen natural enzyme homologs.
  • Still unknown: The article defines PET hydrolase by demonstrated PET-depolymerizing activity and notes that activity may be promiscuous relative to the enzyme’s native role. Assays are small-scale and 48 hours; the authors say further enzyme engineering is needed for meaningful bioreactor studies. This is distinct from the catalogued AI-engineered FAST-PETase: it discovers natural homologs and diverse acid/temperature activity profiles rather than optimizing one known PETase scaffold.
  • AI role: Machine-learning activity, thermal-stability and acid-tolerance predictions prioritized natural homologs across successive mining rounds; new assay results updated later predictors. Researchers expressed/purified over 200 candidates and measured PET hydrolysis to verify activity.

Evidence sources: https://doi.org/10.1021/acscatal.5c03460

Source checked · editorial details

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

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

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

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 39 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ESM-guided metagenome mining identifies seven new Cas12a subtypes with distinct activities

AIL-Scan fine-tuned an evolutionary-scale protein language model (ESM) on Cas proteins to search metagenomic sequences and identify previously undocumented Cas12a lineages. The study classified seven previously unreported Cas12a subtypes with distinct CRISPR-locus organization, then biochemically characterized 16 Cas12a proteins from the candidate clades, documenting different DNA-cleavage and PAM-recognition preferences. The authors also demonstrated a specific assay strategy for distinguishing the KRAS G12C variant using a Cas12a variant with broader PAM recognition.

Technology: AIL-Scan, using fine-tuned ESM-2 protein language-model embeddings and a machine-learning classifier to identify Cas proteins and predict Cas12a trans-cleavage activity

Family ESM / ESMFold: profile · models page

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

Result, AI contribution and limits

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

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The study classified seven previously unreported Cas12a subtypes with distinct CRISPR-locus organization, then biochemically characterized 16 Cas12a proteins from the candidate clades, documenting different DNA-cleavage and PAM-recognition preferences.
  • Still unknown: The paper reports metagenomic classification and biochemical assays, not a clinical genome-editing or diagnostic application. The PAM/SNP detection demonstration is an experimental proof of concept. The trans-cleavage prediction model was trained on a small labeled set (69 proteins), with two unreported candidates used as an explicit prediction check; CRISPR loci and small-gene annotation in metagenomes remain incomplete. Distinct from other Cas-protein discovery papers by target family, candidate set and measured result.
  • AI role: The fine-tuned ESM model screened metagenomic proteins for Cas identity and generated embeddings; supervised models on small experimental labels predicted trans-cleavage activity to help select uncharacterized candidates. Researchers then manually inspected candidate loci and experimentally measured binding, cleavage and PAM preferences. AI supported sequence-family identification and candidate prioritization; laboratory assays established the reported biochemical properties.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC12375029/

Source checked · editorial details

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

Source type: paper. Paper: linked separately.

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

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 27 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

PepMLM designs experimentally binding target-conditioned peptide ligands, including peptides used for targeted protein degradation

PepMLM fine-tunes ESM-2 using masked-language modeling conditioned on a target protein sequence to generate de novo linear peptide binders without structural input. In the reported experiments, selected peptides showed sequence-specific binding to targets including NCAM1 and AMHR2; incorporation into ubiquitin-based antibody constructs enabled targeted degradation of disease-relevant proteins, including Huntington’s disease drivers and viral phosphoproteins.

Technology: PepMLM, ESM-2 fine-tuned with target-sequence-conditioned masked language modeling

Family ESM / ESMFold: profile · models page

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

Result, AI contribution and limits

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

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: In the reported experiments, selected peptides showed sequence-specific binding to targets including NCAM1 and AMHR2; incorporation into ubiquitin-based antibody constructs enabled targeted degradation of disease-relevant proteins, including Huntington’s disease drivers and viral phosphoproteins.
  • Still unknown: The reported evidence is preclinical binding/degradation assays, not therapeutic efficacy, pharmacokinetics or safety. The article notes that experimental optimization/stabilization of its selected peptides was not performed and further validation is needed for viral-target applications. Primary publisher full text/abstract checked; no independent replication assessed.
  • AI role: Given a target protein sequence, PepMLM generates candidate linear binder peptide sequences by reconstructing masked peptide regions. The authors tested selected designs in binding and degradation assays and reported target-specific activity in the stated systems.

Evidence sources: https://www.nature.com/articles/s41587-025-02761-2

Source checked · editorial details

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

Source type: paper. Paper: linked separately.

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

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 103 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

BEAUT uncovers new gut-microbial bile-acid enzymes and the novel bile acid 3-acetoDCA

The BEAUT workflow combines ESM-2 protein-language-model embeddings with a neural classifier and active-site-pocket-similarity data augmentation to search gut microbial proteins for bile-acid enzymes. The authors predicted over 600,000 candidate enzymes and experimentally tested selected predictions; assays identified new bile-acid activities including MABH and ADS. ADS produces 3-acetoDCA, a previously undescribed bile-acid carbon skeleton, reported in human feces and shown to alter microbial composition and promote Lactobacillus-associated effects in experimental studies.

Technology: BEAUT, ESM-2 protein-language-model embeddings plus dense neural-network classifier, augmented using active-site-pocket similarity

Family ESM / ESMFold: profile · models page

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

Result, AI contribution and limits

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

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The authors predicted over 600,000 candidate enzymes and experimentally tested selected predictions; assays identified new bile-acid activities including MABH and ADS.
  • Still unknown: The 600,000 figure is a prediction set, not experimentally characterized enzymes; authors tested selected subsets. Physiological claims are based on study assays and animal/microbial experiments, not demonstrated human clinical effects. The journal full text was inaccessible in this reader; the Peking University research team's detailed release summarizes its Cell paper, methods and experiments.
  • AI role: Screened microbial protein sequences for likely bile-acid-metabolizing function and prioritized candidate enzymes; experimental characterization of selected AI predictions identified enzyme activities and the ADS/3-acetoDCA pathway.

Evidence sources: https://cqb.pku.edu.cn/info/1101/3119.htm

Source checked · editorial details

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

Source type: press release. Paper: linked separately.

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

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 58 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

VenusMine identifies PET hydrolase KbPETase

VenusMine combines structure-based search with protein-language-model representations and learned property filters to prioritize candidate PET hydrolases from large sequence collections. The authors tested 34 selected proteins and report PET-hydrolyzing activity for 14; KbPETase had an optimum activity temperature 32 °C above the IsPETase comparator. This is laboratory enzyme evidence, not demonstrated waste-processing performance. A 2024 preprint reportedly contains an earlier version; its exact version-to-result chronology was not freshly resolved, so this draft uses the journal version date and does not claim first discovery priority.

Technology: VenusMine; ProstT5; ESM2

Family ESM / ESMFold: profile · models page

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

Result, AI contribution and limits

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

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: 34 selected proteins were tested; the source reports PET-hydrolyzing activity for 14 and characterizes KbPETase experimentally.
  • Still unknown: Scale-up, real-world plastic-waste processing, independent replication and exact relation of the earlier preprint’s experiments remain outside this bounded review.
  • AI role: VenusMine uses structure search, protein-language-model embeddings and learned property predictions to select enzyme candidates for experimental testing.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC12228687/

Source checked · editorial details

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

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Indexed primary article text including title, methods, results and publication metadata; direct PMC open triggered CAPTCHA..

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Experimental feedback guides ESM-2 to brighter CreiLOV variants

Reinforcement Learning from eXperimental Feedback (RLXF) used a learned fluorescence reward model to align ESM-2 with measured CreiLOV sequence-function data. In the peer-reviewed study, the authors expressed 60 generated variants in E. coli and measured cellular fluorescence; six of ten aligned ESM-2 five-mutant designs exceeded CreiLOV fluorescence, and three aligned designs exceeded CreiLOV-T7S. The first bioRxiv version (posted 2025-05-08) already described experimentally enhanced fluorescence from RLXF-aligned designs; detailed measurements here are from the 2026 journal version. The earlier ICLR workshop comparison was computational and did not report new wet-lab fluorescence assays.

Technology: 650-million-parameter ESM-2 protein language model aligned by reinforcement learning from experimental feedback using a learned sequence-fluorescence reward model

Family ESM / ESMFold: profile · models page

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

Result, AI contribution and limits

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

  • Publication status: Preprint
  • Test type: direct test
  • Tested: The journal article reports 60 CreiLOV designs expressed in E. coli BL21(DE3) and measured for total cellular fluorescence. Six of ten aligned ESM-2 five-mutant designs exceeded CreiLOV fluorescence, and three aligned designs exceeded the prior CreiLOV-T7S control; the brightest reported design was 1.2-fold above T7S.
  • Still unknown: The measurements use total cellular fluorescence and are specific to tested designs/conditions; broader cellular contexts, protein function beyond fluorescence, and independent reproduction are not established. The qualitative enhanced-fluorescence result is present in the 2025 preprint abstract; detailed design-set metrics are verified from the later journal version.
  • AI role: A learned sequence-fluorescence reward model trained on measured CreiLOV variant data guides supervised fine-tuning and PPO alignment of ESM-2. The aligned model generates the candidate sequences that were expressed and assayed.

Evidence sources: https://www.biorxiv.org/content/10.1101/2025.05.02.651993v1 · https://www.nature.com/articles/s41467-026-77557-2 · https://openreview.net/pdf?id=yuiQMh52Bm

Source checked · editorial details

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

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary bioRxiv v1 abstract, posting-date and affiliation metadata; primary GEM@ICLR workshop paper full text for methods and experimental outcomes; primary Nature Communications indexed text for later journal-version comparison..

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ESM-Ezy mining identifies multicopper oxidases and asparaginases with experimentally improved catalytic properties

The ESM-1b semantic-embedding mining strategy nominated low-sequence-similarity enzyme candidates. In the reported experimental assessment, 44% of tested multicopper oxidases outperformed their query enzymes in at least one measured property (including catalytic efficiency, heat/solvent tolerance or pH stability); 51% were reported to excel in environmental-remediation applications. Among mined L-asparaginases, 40% had higher specific activity and catalytic efficiency than query enzymes.

Technology: ESM-Ezy, using ESM-1b protein-language-model embeddings and semantic-space similarity

Family ESM / ESMFold: profile · models page

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

Result, AI contribution and limits

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

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: In the reported experimental assessment, 44% of tested multicopper oxidases outperformed their query enzymes in at least one measured property (including catalytic efficiency, heat/solvent tolerance or pH stability); 51% were reported to excel in environmental-remediation applications.
  • Still unknown: Percentages describe the selected candidates and tested property comparisons in this study, not the fraction of all database enzymes that will be superior. The reported remediation suitability is assay-specific; environmental or industrial deployment is not established. Primary abstract/full-text search result checked; no independent replication assessed.
  • AI role: ESM-1b sequence embeddings and semantic-space similarity ranked low-similarity candidate enzymes from sequence databases for experimental follow-up; assays assessed catalytic properties of selected multicopper oxidases and L-asparaginases.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/40188191/

Source checked · editorial details

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

Source type: paper. Paper: linked separately.

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

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 38 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Mal-ID maps disease-associated immune-receptor patterns across infections and autoimmune conditions

Mal-ID analyzed B-cell and T-cell receptor repertoires from 593 people, including patients with COVID-19, HIV, systemic lupus erythematosus and type-1 diabetes, influenza-vaccination recipients and healthy controls. Its combined models classified held-out immune states and highlighted disease-associated receptor patterns; interpretation identified distinct autoreactivity features in lupus and type-1 diabetes and antigen-associated receptor sequences in COVID-19. The authors also evaluated classification on independent COVID-19 cohorts.

Technology: Mal-ID (MAchine Learning for Immunological Diagnosis), an ensemble using repertoire feature models and ESM-2 protein-language-model embeddings

Family ESM / ESMFold: 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 authors also evaluated classification on independent COVID-19 cohorts.
  • Still unknown: This is a patient-sample diagnostic-validation and immune-biology study, not evidence of clinical deployment or benefit. Validation cohorts and disease-specific sample sizes are limited; external COVID-19 classification involved small cohorts and thresholds were tuned in some analyses. Additional disease-specific validation, threshold calibration, sequencing-depth assessment and interpretation alongside standard clinical assessment remain necessary. Its 2024 bioRxiv preprint is the same work, not a separate finding.
  • AI role: Three machine-learning representations of B-cell and T-cell receptor data were combined to classify immune states and prioritize disease-associated receptor sequences, gene segments and isotypes for biological interpretation.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/39977494/

Source checked · editorial 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, 132 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.

CHOOSER uses protein foundation models to discover functional Casλ homologs

The CHOOSER framework fine-tuned ESM-2 protein-language-model representations for alignment-free CRISPR-Cas discovery and pre-crRNA self-processing predictions. It nominated 11 Casλ homologs, nearly doubling the known Casλ set. Wet-lab work on selected homologs confirmed pre-crRNA processing for five of six tested proteins; EphcCasλ also cleaved target DNA and showed trans-cleavage activity, with a TTR PAM reported.

Technology: CHOOSER, a protein foundation-model framework fine-tuning ESM-2 representations for Cas homolog identification and Cas12 pre-crRNA-processing classification

Family ESM / ESMFold: profile · models page

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

Result, AI contribution and limits

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

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Wet-lab work on selected homologs confirmed pre-crRNA processing for five of six tested proteins; EphcCasλ also cleaved target DNA and showed trans-cleavage activity, with a TTR PAM reported.
  • Still unknown: This is a specific functional genomics result, not evidence of a ready-to-use diagnostic or gene-editing tool. Only selected candidates were purified and tested; the paper reports pre-crRNA activity in five of six tested Casλ proteins and DNA-cleavage validation for a subset. EphcCasλ activity and PAM characterization are in-vitro findings. The article has a publisher-listed author correction dated 2025-01-09; this record relies on the version-of-record text and retains limited scope.
  • AI role: CHOOSER mined candidate proteins adjacent to CRISPR arrays using learned sequence representations rather than alignment-only detection and prioritized Cas12 homologs by predicted pre-crRNA processing. The study reported 11 Casλ homologs and experimentally tested selected predictions; purified EphcCasλ was validated for self-processing pre-crRNA, DNA cleavage and trans-cleavage.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11576732/

Source checked · editorial details

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

Source type: paper. Paper: linked separately.

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

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 26 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Structure-informed model selects antibody variants active against SARS-CoV-2 escape variants

ESM-IF1 scored substitutions in antibody–antigen complexes, and researchers experimentally screened variants of two clinical antibodies. The journal article reports up to 25-fold neutralization improvement against BQ.1.1 and up to 37-fold affinity improvement against XBB.1.5; these are distinct assays/variants. The preprint abstract reports a different neutralization figure, so this record uses journal values and does not backdate them to preprint v1.

Technology: ESM-IF1 inverse-folding protein model

Family ESM / ESMFold: profile · models page

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

Result, AI contribution and limits

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

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: PMC-hosted primary article reports experimental screening and improved neutralization/affinity results for antibody variants.
  • Still unknown: Exact v1 result content and sequence-level overlap with earlier group studies not completely established; no clinical efficacy.
  • AI role: Structure-informed inverse-folding scores guided substitutions followed by antibody binding and neutralization tests.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11616794/

Source checked · editorial details

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

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

Review date: 16 Nov 2026. Review scope: Primary source checked for identity, date, model role and bounded reported result..

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ESM3 and esmGFP: a fluorescent protein designed by a language model

ESM3 designed esmGFP, a fluorescent protein of 229 amino acids with 96 mutations relative to the closest known protein (about 58% similarity). In a second round of 96 designs, some proteins glowed as brightly as natural GFPs. The comparison with ‘more than 500 million years of evolution’ is an analogy-based estimate.

Technology: ESM3

Family ESM / ESMFold: profile · models page

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

Result, AI contribution and limits

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

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: In two rounds of 96 designs, lab measurements of fluorescence found some proteins as bright as natural GFPs; esmGFP differs by 96 mutations and is about 58% similar to the closest natural protein.
  • Still unknown: The page lists no formal limitations; the designed B8 was about 50 times dimmer and slower to mature than natural GFPs.
  • AI role: ESM3 generated new GFP candidates prompted with structural residues from natural GFP.

Evidence sources: https://www.evolutionaryscale.ai/blog/esm3-release

Source checked · editorial details

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

Source type: blog. Paper: linked separately.

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

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Meta AI

ESM Metagenomic Atlas: over 600 million protein structures

Meta AI published predicted structures for almost the entire MGnify90 database (over 600 million), generated in about two weeks on roughly 2,000 GPUs with a 15-billion-parameter language model, up to 60 times faster than previous methods. These are predictions, and high-confidence structures are a much smaller subset.

Technology: ESMFold (15-billion-parameter language model)

Family ESM / ESMFold: profile · models page

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

Result, AI contribution and limits

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

  • Publication status: Preprint
  • Test type: retrospective or benchmark
  • Tested: The preprint reports structure prediction up to 60 times faster while maintaining accuracy versus state-of-the-art methods and an atlas of 617 million structures; the blog describes no experimental checks.
  • Still unknown: The blog lists no limitations and notes some sequences fold into disordered forms; high-confidence predictions are a subset.
  • AI role: Language model (15B parameters) predicted protein structures from sequence.

Evidence sources: https://ai.meta.com/blog/protein-folding-esmfold-metagenomics/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1101/2022.07.20.500902&resultType=core&format=json

Source checked · editorial details

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

Source type: blog. Paper: linked separately.

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

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Open the interactive profile for the timeline, or browse all organizations and model families.

3 model releases

Epoch AI data (CC BY 4.0), newest first. A discovery is attributed to a model family, not to a specific release.

DateModelAccess
ESM3 (98B) Unreleased
ESM2-15B Open weights (unrestricted)
ESM1b Open weights (unrestricted)