AI Science Tracker

University of Washington Institute for Protein Design

9 discoveries from 2021 to 2025

Discoveries

Computer-designed proteins neutralize snake toxins

Small, stable proteins designed with AI bind and neutralize three-finger toxins from cobras and mambas in the lab; in mice, they yield 80–100% survival against lethal neurotoxin doses, depending on the toxin. Laboratory and animal studies only; traditional antivenoms remain central.

Technology: AI protein design (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: Lab tests showed the designed proteins neutralized multiple subtypes of three-finger toxins; in mice they gave complete protection from lethal neurotoxin exposure, with 80-100% survival depending on the toxin and antitoxin.
  • Still unknown: The page does not mention human trials, toxin types other than elapid three-finger toxins, and traditional antivenoms remain central to treatment.
  • AI role: Designed small, stable proteins on the computer that bind key regions of the toxins.

Evidence sources: https://www.ipd.uw.edu/2025/01/neutralizing-deadly-snake-toxins/

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, 196 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.

Deep learning identifies previously unreported protein interactions in human bacterial pathogens

A RoseTTAFold2-Lite pipeline combined residue coevolution, deep-learning complex prediction and AlphaFold modelling to screen protein pairs across 19 human bacterial pathogens. In a validation set selected using the deep-learning pipeline, co-immunoprecipitation detected four of six predicted interactions. Three pairs lacked prior direct interaction evidence: E. coli UbiE–YcaR, P. aeruginosa PA4105–PA4106, and P. aeruginosa WbpG–HisF2. The PA4105–PA4106 structural similarity to proteins in 3-thiaglutamate biosynthesis and the WbpG–HisF2 result provide functional hypotheses, not established native-pathogen mechanisms.

Technology: RoseTTAFold2-Lite integrated with direct coupling analysis and AlphaFold complex modelling

Family AlphaFold: profile · models page

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

Result, AI contribution and limits

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

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: In a validation set selected using the deep-learning pipeline, co-immunoprecipitation detected four of six predicted interactions.
  • Still unknown: The positive interaction evidence comes from heterologous co-immunoprecipitation/pull-down assays, with two independent experiments per positive pair, rather than native-pathogen systems. The validation subset was small (four positives among six AI-selected candidates); most of the thousands of predicted interactions remain untested, and false positives and false negatives are expected. Proposed roles for PA4105–PA4106 and UbiE–YcaR remain hypotheses; this paper does not establish their physiological function or a therapeutic target.
  • AI role: RoseTTAFold2-Lite used residue–residue coevolution and predicted complex structures to rank candidate protein pairs in a proteome-wide screen. The authors selected six pairs from the final deep-learning pipeline for co-immunoprecipitation tests; four yielded positive interaction signals. Three were described as lacking previously reported direct interaction evidence, including UbiE–YcaR, PA4105–PA4106 and WbpG–HisF2. The models enabled the experimental target selection; they did not themselves establish physical interactions.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11445079/

Source checked · editorial details

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

Source type: paper. Paper: linked separately.

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

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 52 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AI-accelerated virtual screening identifies small-molecule binders to KLHDC2 and Nav1.7

The OpenVS platform combined target-specific neural-network active learning with RosettaVS physics-based docking to search billion-scale compound libraries. The study reported seven low-micromolar KLHDC2 degron-pocket binders and four Naᵥ1.7 VSD4 binders. KLHDC2 hit C29 competed with a degron peptide at IC50 about 2.9 μM and its predicted pose was supported by a 2.0 Å crystal structure; the strongest Naᵥ1.7 hit, Z8739902234, had IC50 1.3 μM in whole-cell electrophysiology.

Technology: OpenVS active-learning platform with target-specific feed-forward neural-network triage, coupled to RosettaVS physics-based docking and RosettaGenFF-VS scoring

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

Result, AI contribution and limits

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

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The study reported seven low-micromolar KLHDC2 degron-pocket binders and four Naᵥ1.7 VSD4 binders.
  • Still unknown: Early ligand-hit discovery only: binding/inhibition was measured in biochemical competition, biolayer interferometry and cell electrophysiology assays, not in animals or patients. The paper states further optimization is needed to improve KLHDC2 compounds toward nanomolar potency. KLHDC2 C29’s crystal structure is experimental support for one predicted binding pose; target modulation, therapeutic efficacy and safety were not established. The reported target hit rates apply to manually filtered, synthesized subsets, not the full billion-molecule library.
  • AI role: The target-specific neural network used docking-labeled data to guide which molecules were prioritized for additional expensive docking over iterative multi-billion-compound screens; RosettaVS/RosettaGenFF-VS performed physics-based docking and pose/affinity scoring. The AI-guided screen nominated molecules that were synthesized and experimentally tested. Experiments confirmed KLHDC2 degron displacement and a crystal structure matching the predicted pose, plus Naᵥ1.7 binding/inhibition in patch clamp. The paper attributes discovery to the combined AI-accelerated and physics-based workflow, not to a generative model alone.

Evidence sources: https://www.nature.com/articles/s41467-024-52061-7.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, 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 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Antibodies designed from scratch with RFdiffusion

Baker’s group showed RFdiffusion can be adapted to generate antibodies: in laboratory tests, designs bound influenza virus and other targets as intended. Proof-of-concept study, first circulated as a preprint in March 2024; experts quoted by Nature call it promising but an early step; no affinity values reported.

Technology: RFdiffusion

Family RFdiffusion: 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: Laboratory tests confirmed that the designed antibodies bind influenza virus and other targets as intended; no affinity values are reported.
  • Still unknown: The work is described as a proof-of-concept study, and no specific limitations are listed.
  • AI role: RFdiffusion was tuned to generate antibodies from scratch.

Evidence sources: https://www.ipd.uw.edu/2024/04/nature-a-landmark-moment-scientists-use-ai-to-design-antibodies-from-scratch/

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 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.

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.

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.

ProteinMPNN: protein sequences designed in about one second

A neural network generates the amino-acid sequence for a given protein shape in about one second, over 200 times faster than the best previous software, without specialist tuning. A related study shows that proteins designed with ProteinMPNN fold correctly much more often, including nanoscale rings and complex assemblies. Researchers describe it as the beginning of machine learning in protein design.

Technology: ProteinMPNN

Family ProteinMPNN: profile · models page

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

Result, AI contribution and limits

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

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Speed is compared with prior software; a related Science study tested lab-made designs, which folded as intended more often, including nanoscale rings imaged by electron microscopy.
  • Still unknown: The article gives no quantitative success rates and notes protein data is scarce, requiring trial and error.
  • AI role: Neural network generated amino-acid sequences for given protein shapes.

Evidence sources: https://www.ipd.uw.edu/2022/09/proteinmpnn-excels-at-creating-new-proteins/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1126/science.add2187&resultType=core&format=json

Source checked · editorial details

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

Source type: press release. Paper: linked separately.

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

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 2354 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

A neural network “hallucinates” new protein structures

The network predicted about 2,000 new sequences as foldable; over 100 were produced in the lab, and for three, nuclear magnetic resonance and X-ray crystallography confirmed the predicted shapes. Only three proteins were characterized in detail; using the method to create useful functions (drugs, enzymes) is a future goal.

Technology: Neural network for structure prediction (hallucination)

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

Result, AI contribution and limits

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

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: 129 hallucinated sequences were synthesised, 27 gave monodisperse proteins, and three structures determined by X-ray crystallography or NMR matched the hallucinated models.
  • Still unknown: Only three proteins were characterised in detail, and using the method for useful functions is a future goal.
  • AI role: Neural network hallucinated new protein sequences predicted to fold.

Evidence sources: https://newsroom.uw.edu/news-releases/deep-learning-dreams-new-protein-structures · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1038/s41586-021-04184-w&resultType=core&format=json

Source checked · editorial details

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

Source type: press release. Paper: linked separately.

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

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 653 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

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

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

RoseTTAFold: accessible prediction of protein structures

A three-track neural network (sequence, residue interactions, and 3D structure) computes a structure in about ten minutes on a single gaming computer. Baker’s group predicted hundreds of new structures, including little-known proteins in the human genome. The web server received over 4,500 requests in the previous month. The page does not state limitations.

Technology: RoseTTAFold

Family RoseTTAFold: profile · models page

Publication: Peer-reviewed paper · Test context: Test context unspecified in annotation. Catalog categories, not an evidence grade.

Result, AI contribution and limits

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

  • Publication status: Peer-reviewed paper
  • Test type: not reported
  • Tested: The page describes the tool's speed and hundreds of new structures but does not describe how predictions were checked.
  • Still unknown: Not reported.
  • AI role: Three-track neural network predicted protein structures.

Evidence sources: https://www.ipd.uw.edu/2021/07/rosettafold-accurate-protein-structure-prediction-accessible-to-all/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1126/science.abj8754&resultType=core&format=json

The legacy label “Not reported” does not distinguish information absent from reviewed material from information not assessed. It is not a scientific finding of absence.

Source checked · editorial details

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

Source type: press release. Paper: linked separately.

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

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 5653 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 1/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

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

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

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