AI Science Tracker

California Institute of Technology

8 discoveries from 2020 to 2026

Discoveries

Machine learning identifies the first enzyme for phenazine–thiol conjugation

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

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

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

Result, AI contribution and limits

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

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

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

Source checked · editorial details

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

Source type: paper. Paper: linked separately.

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

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Active learning identifies a protoglobin variant with 99% yield and cis-selective cyclopropanation

An iterative machine-learning campaign engineered the protoglobin from Pyrobaculum arsenaticum (ParPgb) for non-native cyclopropanation of 4-vinylanisole. After two model-guided rounds following an initial library screen, variant MPFDY gave 99% total cyclopropane yield and a 14:1 cis:trans ratio in the reported assay, compared with a parent variant that produced about 40% yield and favored trans product. The result demonstrates a specific sequence combination with improved yield and stereoselectivity in this enzyme scaffold.

Technology: ALDE (Active Learning-assisted Directed Evolution), using a bootstrapped deep-neural-network ensemble and Thompson sampling

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

Result, AI contribution and limits

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

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: After two model-guided rounds following an initial library screen, variant MPFDY gave 99% total cyclopropane yield and a 14:1 cis:trans ratio in the reported assay, compared with a parent variant that produced about 40% yield and favored trans product.
  • Still unknown: Cyclopropanation is an established non-native enzyme reaction class; the study's distinct result is the ALDE-guided ParPgb sequence and its measured product yield/selectivity. The 99% yield and 14:1 ratio are study assay results, not evidence of industrial process readiness. The study is a single enzyme-engineering campaign.
  • AI role: ALDE trained on measured sequence–yield data, quantified model uncertainty, and selected variants across two iterative rounds in a five-residue, 3.2-million-sequence design space. Experimental screening measured product yield and stereoselectivity; the model-guided variants included MPFDY, whose combined mutations improved the targeted cis-product objective where tested single mutations and their naive recombinations did not.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11739421/

Source checked · editorial details

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

Source type: paper. Paper: linked separately.

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

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 186 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Interpretable machine learning discovers ultrapermeable polymer membranes for gas separation

Multitask machine-learning models trained on experimental gas permeability data screened over 9 million hypothetical polymers. They identified hundreds of never-before-seen ultrapermeable membranes, with molecular dynamics simulations confirming the predicted permeabilities of the top candidates. The validation is computational rather than fabricated membranes, so the candidates remain unproven experimentally.

Technology: Multitask machine learning with model interpretation, plus molecular dynamics

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

Result, AI contribution and limits

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

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Molecular dynamics simulations checked the predicted permeabilities of the top candidates among over 9 million screened hypothetical polymers; no membranes were fabricated.
  • Still unknown: The authors suggest many candidates could be translated to reality, which is not tested; no limitations section is reported.
  • AI role: Interpretable multitask machine-learning models trained on experimental gas permeability data screened hypothetical polymers.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1126/sciadv.abn9545&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, 257 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.

Semi-supervised learning selects a lithium thioborate for vacancy engineering

Semi-supervised learning screened lithium-containing structures and selected Li3BS3 for experimental characterization. The parent material was a poor room-temperature conductor. Researchers then introduced vacancies by Si substitution and amorphized the substituted material; the paper reports approximately four orders of magnitude higher ionic conductivity for a-Li2.95B0.95Si0.05S3. The model selected the framework, while human-led defect engineering produced the high-conductivity sample. This does not establish that pristine Li3BS3 is superionic.

Technology: Semi-supervised learning; agglomerative clustering

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

Result, AI contribution and 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 selected Li3BS3 for characterization and reports increased ionic conductivity after Si substitution and amorphization, with the combined vacancy-engineering treatment improving conductivity by roughly four orders of magnitude.
  • Still unknown: The ML model selected a framework candidate; vacancy engineering and synthesis were subsequent researcher-led steps. The primary page was checked through indexed full-text results because direct article opening returned 403.
  • AI role: Semi-supervised learning used literature-derived conductivity labels and structural descriptors to rank lithium-containing structures and select Li3BS3 for experimental characterization.

Evidence sources: https://authors.library.caltech.edu/records/evkz6-5jd40 · https://pubs.rsc.org/en/content/articlehtml/2023/ee/d2ee03499a

Source checked · editorial 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: CaltechAUTHORS primary manuscript record and deposited manuscript metadata; RSC indexed primary full-text results and methods..

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 phase models identify synergistic photoanode mixtures

Learning from measured diffraction and photocurrent data identified Bi-Cu-V oxide phase mixtures whose photoactivity exceeded an additive non-interacting-phase model. BiVO4-like phases showed beneficial associations with copper-vanadate phases. This is an experimentally grounded composition-property result; the underlying electronic, catalytic and morphological mechanisms remain unresolved, and device-level solar-fuel efficiency was not established.

Technology: Deep Reasoning Networks and matrix factorization

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: Measured photocurrents and phase distributions were compared with a learned additive baseline.
  • Still unknown: Independent replication, detailed mechanisms and practical device performance were not assessed.
  • AI role: Deep Reasoning Networks and matrix factorization supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://www.nature.com/articles/s41524-022-00747-1

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: Indexed primary publisher abstract, selected phase-model and measured-synergy Results, date and affiliations; supplementary material not reviewed..

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning-assisted directed evolution with in silico combinatorial screening

A path-independent machine-learning protocol screens full combinatorial protein libraries in silico, with training-set design that avoids near-zero-fitness variants. On a four-site epistatic fitness landscape the optimized protocol reached the global fitness maximum up to 81-fold more often than single-step greedy optimization. Caveat: validation is on a combinatorial fitness landscape dataset, not a new wet-lab evolution campaign.

Technology: Machine-learning-assisted directed evolution (in silico combinatorial screening)

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

Result, AI contribution and limits

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

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: The ML-assisted directed evolution protocol was compared with single-step greedy optimization on a four-site epistatic fitness landscape.
  • Still unknown: Not reported in the abstract read.
  • AI role: A path-independent machine-learning protocol that screens full combinatorial libraries in silico, with training-set design avoiding near-zero-fitness variants.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1016/j.cels.2021.07.008&resultType=core&format=json

Source checked · editorial details

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

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

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

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 220 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Deep learning finds comet C/2020 T2 in a single ZTF image

The Tails EfficientDet model identified the extended object C/2020 T2 in one Zwicky Transient Facility Twilight Survey image on 2020 October 7. Follow-up orbital work confirmed it as a long-period comet. The study reports this as the first comet discovery assisted by deep learning; it also recovered a previously known comet, which is not counted as a new discovery here.

Technology: Tails, an EfficientDet-based deep-learning comet detector

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

Result, AI contribution and 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 Tails EfficientDet model identified the extended object C/2020 T2 in one Zwicky Transient Facility Twilight Survey image on 2020 October 7.
  • Still unknown: The model surfaced the object, while human/standard follow-up established its orbit and comet classification. The entry counts only the newly discovered C/2020 T2, not recovery of P/2016 J3/P/2021 A3 or the detector’s performance metrics.
  • AI role: Detected the comet’s extended coma/tail morphology in a single survey image and surfaced C/2020 T2 for discovery/follow-up; this enabled detection without requiring multi-epoch orbit linking at the initial search stage.

Evidence sources: https://arxiv.org/abs/2102.13352

Source checked · editorial details

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

Source type: paper. Paper: linked separately.

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

Citations: at percentile 77 of works of the same type, year and subfield (OpenAlex work metrics, 16 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.

An attention-based neural network generates functional secretion signal peptides

An attention-based Transformer trained on signal-peptide data from Swiss-Prot generated candidate secretion sequences. Experimental testing in an industrial Bacillus subtilis strain found that model-generated signal peptides directed secreted enzyme activity competitive with industrially used signal peptides. The result establishes functional candidates in the reported host and assay; it does not establish broad performance across proteins or production hosts. The primary institutional record gives the article publication date as 2020-07-10.

Technology: Attention-based neural network (Transformer) trained on Swiss-Prot signal peptide sequences

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

Result, AI contribution and limits

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

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The primary abstract reports experimental testing in an industrial Bacillus subtilis strain and secreted enzyme activity competitive with industrially used signal peptides.
  • Still unknown: The bounded abstract-level claim does not quantify all peptide/host combinations or demonstrate generalization across unrelated proteins and production conditions.
  • AI role: An attention-based Transformer trained on Swiss-Prot generated signal-peptide sequences, which were then tested experimentally for protein secretion.

Evidence sources: https://authors.library.caltech.edu/records/5dqa2-44d08

Source checked · editorial details

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

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary Caltech institutional record with the article abstract, author contribution statement, publication date, and affiliation; the linked PDF download was unavailable in this session..

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

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