AI Science Tracker

ETH Zurich

7 discoveries from 2021 to 2026

Discoveries

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

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

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

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

Result, AI contribution and limits

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

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

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

Source checked · editorial details

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

Source type: paper. Paper: linked separately.

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

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

COSTAR predicts drug activity tested in glioblastoma samples

A learned drug-target connectivity signature selected 23 predicted hits and 25 negatives for testing in four glioblastoma patient samples. Ex-vivo measurements agreed with predictions at a reported mean ROC AUC of 0.94. The small validation establishes a bounded drug-activity prediction result, not clinical response. Vortioxetine was found by the initial experimental screen and is not attributed to COSTAR nomination.

Technology: COSTAR logistic LASSO drug-target 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: Newly selected compounds were tested ex vivo across four patient samples.
  • Still unknown: Patient benefit, generalization and clinical safety remain unestablished.
  • AI role: COSTAR logistic LASSO drug-target model supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11564103/

Source checked · editorial details

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

Source type: paper. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Primary full-text model Methods and selected experimental Results, publication metadata and affiliations; 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.

Active learning identifies a higher-activity streptavidin variant for gold-catalyzed hydroamination

An iterative active-learning campaign engineered a streptavidin artificial metalloenzyme for the known-to-the-authors gold-catalyzed hydroamination of 2-ethynylaniline to indole. Gaussian-process models guided screening across five residues in a 3.2-million-variant sequence space. The final-round library contained 38% variants more active than the prior reference, compared with 3% in the initial library. The leading variant, Sav CTNLV, showed 18-fold higher cell-specific activity than wild-type Sav and threefold higher activity than the previous reference variant in replicate testing.

Technology: Active-learning Gaussian-process regression model with optimized sequence descriptors and experimental-noise estimates

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

Result, AI contribution and limits

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

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The leading variant, Sav CTNLV, showed 18-fold higher cell-specific activity than wild-type Sav and threefold higher activity than the previous reference variant in replicate testing.
  • Still unknown: The hydroamination reaction and streptavidin–gold artificial metalloenzyme were established in earlier work by the same group; this record is specifically the later AI-guided identification of a higher-activity five-residue variant, not discovery of a new reaction. Activity was measured in an E. coli periplasmic whole-cell assay; purified variants did not retain the same rank order in vitro. Industrial utility or catalytic scale-up was not demonstrated.
  • AI role: The model used measured sequence-activity data from thousands of variants to select exploratory and exploitation libraries across five residues, then prioritized variants for experimental testing. It enabled identification of a high-activity sequence in a large combinatorial space; the paper reports considering experimental uncertainty in its rankings.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11273458/

Source checked · editorial details

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

Source type: paper. Paper: linked separately.

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

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 37 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Protein-surface-designed molecules activate PPARs

A generative model designed molecules from the three-dimensional surfaces of PPAR proteins: synthesized in the laboratory, they increased activity and were tested by Roche. They demonstrate the method and are not intended for commercial development; the tests do not establish safety or efficacy in humans.

Technology: Protein-surface-guided generative AI

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

Result, AI contribution and limits

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

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: ETH chemists synthesized the AI-designed molecules in the lab, and Roche ran a variety of tests showing they are stable and non-toxic.
  • Still unknown: The article does not state limitations, the molecules were not developed toward market, and the tests do not establish safety or efficacy in humans.
  • AI role: A generative model designed molecules from scratch matched to a protein's three-dimensional surface, without human intervention.

Evidence sources: https://ethz.ch/en/news-and-events/eth-news/news/2024/04/ai-designs-new-drugs-based-on-protein-structures.html

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

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning identifies 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.

Machine-intelligence design yields selective synthetic COX-1 inhibitors from a marine natural-product template

Using Marinopyrrole A as a template, an automated design workflow generated and synthesized new natural-product-inspired molecules; target-prediction software pointed to cyclooxygenases. In biochemical assays, synthetic mimetics 2 and 2a inhibited COX-1 at 0.10 and 0.16 μM, respectively, and compound 2a was crystallographically observed bound to COX-1. Human platelet assays also showed inhibition.

Technology: DOGS rule-based de novo molecule construction and CATS molecular-similarity ranking, coupled with SPiDER self-organizing-map machine-learning target 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: In biochemical assays, synthetic mimetics 2 and 2a inhibited COX-1 at 0.10 and 0.16 μM, respectively, and compound 2a was crystallographically observed bound to COX-1.
  • Still unknown: The study tests a small number of designed molecules; automated structure generation was rule-based and similarity-driven rather than a trained generative neural model. The COX-1 effects were biochemical/platelet assays, not clinical effects; the platelet IC50s for 2, 2a, 2b were 9, 13 and 56 nM, but COX-2 inhibition occurred at higher concentrations and selectivity varies by assay. Do not call these approved drugs or imply therapeutic benefit.
  • AI role: DOGS algorithmically generated natural-product-inspired structures and suggested synthetic routes; CATS ranked candidates by pharmacophore similarity. SPiDER's cascaded self-organizing-map models predicted molecular target families, including COX. Experimental target assays confirmed the previously unrecognized COX-1 activity of selected synthetic compounds, with crystallographic confirmation for compound 2a.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC8373093/

Source checked · editorial 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 84 of works of the same type, year and subfield (OpenAlex work metrics, 31 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.

ETH Zurich

Generative AI and on-chip synthesis produce new LXR agonists

A recurrent neural network generated de novo LXR agonist designs, and 25 were synthesized in a microfluidic flow system. In vitro screening found 17 hits (68%), and 12 were confirmed as potent LXR agonists after batch resynthesis. Validation is in vitro only, with no in vivo testing, and the sources disagree on the number of designs retained for synthesis.

Technology: LSTM recurrent neural network generator with transfer learning, coupled to on-chip flow 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: 25 synthesized designs were screened in reporter gene assays (17 hits); 12 were confirmed as LXR agonists after batch resynthesis, with EC50 0.18 to 4.5 micromolar for LXR-alpha. No in vivo testing was reported.
  • Still unknown: Chemical space was limited to one-step reactions, the primary screen used crude mixtures, two compounds were not confirmed, and the authors describe the work as proof-of-concept.
  • AI role: A generative LSTM pretrained on 656,070 compounds and fine-tuned on 40 known LXR-alpha agonists, with a virtual reaction filter of 17 reaction schemes.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC8195470/

Source checked · editorial 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, 203 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.

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