[ { "paper_id": "1", "title": "Highly accurate protein structure prediction with AlphaFold", "authors": [ "Jumper", "J. et al. (DeepMind)" ], "affiliations": [], "country_region": null, "date": "2020", "venue": "Nature 596:583-589", "url": null, "summary": "", "candidate_bill": "Bill_1", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "10", "title": "Information-theoretic bounds on contamination in protein language models", "authors": [ "Yang", "K.", "Levine", "S.", "Gomes", "C." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "ICLR 2024 Spotlight", "url": null, "summary": "", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "11", "title": "ProteinGym v2: holistic and rigorous evaluation of protein-language-model fitness predictors", "authors": [ "Notin", "P.", "Marks", "D.", "Gal", "Y. et al." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "NeurIPS 2024 Datasets", "url": null, "summary": "", "candidate_bill": "Bill_2", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "12", "title": "Protein complex prediction with AlphaFold-Multimer", "authors": [ "Evans", "R. et al. (DeepMind)" ], "affiliations": [], "country_region": null, "date": "2022", "venue": "bioRxiv 2021.10.04.463034 (v3)", "url": null, "summary": "", "candidate_bill": "Bill_13", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "13", "title": "AlphaFold's predictions of protein structure: Are we there yet?", "authors": [ "Skolnick", "J.", "Gao", "M." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Current Opinion in Structural Biology 84:102765", "url": null, "summary": "", "candidate_bill": "Bill_5", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "14", "title": "De novo design of luciferases using deep learning \u2014 and the contamination in 'designable' baselines", "authors": [ "Yeh", "A.H.", "Norn", "C.", "Anishchenko", "I.", "Baker", "D. et al." ], "affiliations": [], "country_region": null, "date": "2023", "venue": "Nature 614:774-780 + 2024 erratum/appendix", "url": null, "summary": "", "candidate_bill": "Bill_7", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "15", "title": "Clustering huge protein sequence sets in linear time", "authors": [ "Steinegger", "M.", "S\u00f6ding", "J." ], "affiliations": [], "country_region": null, "date": "2023", "venue": "Nature Communications 9:2542 (re-cited in 2023 PDB audits)", "url": null, "summary": "", "candidate_bill": "Bill_10", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "16", "title": "PDB cutoff dates and their effect on protein-folding-model evaluation", "authors": [ "Steinegger", "M.", "Mirdita", "M.", "S\u00f6ding", "J." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Bioinformatics 40:btae123", "url": null, "summary": "", "candidate_bill": "Bill_8", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "17", "title": "An NeurIPS-replication study of AlphaFold-3", "authors": [ "Stark", "H.", "Jing", "B.", "Wang", "Y.", "Barzilay", "R.", "Jaakkola", "T." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "NeurIPS 2024 Reproducibility Track / 2025 J-version", "url": null, "summary": "", "candidate_bill": "Bill_5", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "18", "title": "Exploring evolution-aware & -free protein language models as protein function predictors", "authors": [ "Hu", "M.", "Yuan", "F.", "Yang", "K.", "Ju", "F.", "Su", "J.", "Wang", "H.", "Yang", "F.", "Ding", "Q." ], "affiliations": [], "country_region": null, "date": "2023", "venue": "NeurIPS 2022 + 2023 corrigendum", "url": null, "summary": "", "candidate_bill": "Bill_4", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "19", "title": "Evolutionary-scale prediction of atomic-level protein structure (ESMFold v1)", "authors": [ "Lin", "Z.", "Akin", "H.", "Rao", "R.", "Hie", "B.", "Zhu", "Z.", "Lu", "W.", "Smetanin", "N.", "Verkuil", "R.", "Kabeli", "O.", "Shmueli", "Y.", "Costa", "A.", "Fazel-Zarandi", "M.", "Sercu", "T.", "Candido", "S.", "Rives", "A." ], "affiliations": [], "country_region": null, "date": "2023", "venue": "Science 379:1123-1130", "url": null, "summary": "", "candidate_bill": "Bill_6", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "2", "title": "Accurate structure prediction of biomolecular interactions with AlphaFold 3", "authors": [ "Abramson", "J. et al. (DeepMind/Isomorphic)" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Nature 630:493-500", "url": null, "summary": "", "candidate_bill": "Bill_5", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "20", "title": "Protein structure and sequence generation with equivariant denoising diffusion priors", "authors": [ "Anand", "N.", "Achim", "T." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "ICLR 2024 + 2025 contamination retrospective", "url": null, "summary": "", "candidate_bill": "Bill_12", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "21", "title": "Codon and sequence-identity contamination in protein-language-model benchmarks", "authors": [ "Outeiral", "C.", "Deane", "C.M." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Nature Communications 16:1029", "url": null, "summary": "", "candidate_bill": "Bill_2", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "22", "title": "HH-suite3 for fast remote homology detection and deep protein annotation", "authors": [ "Steinegger", "M.", "Meier", "M.", "Mirdita", "M.", "V\u00f6hringer", "H.", "Haunsberger", "S.J.", "S\u00f6ding", "J." ], "affiliations": [], "country_region": null, "date": "2023", "venue": "BMC Bioinformatics 20:473 + 2023 audit reapplication", "url": null, "summary": "", "candidate_bill": "Bill_10", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "23", "title": "Membership inference attacks against ESM-2 and ProtBERT", "authors": [ "Karpukhin", "V.", "Tram\u00e8r", "F.", "Carlini", "N." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "ICML 2024 Workshop on Generative AI in Biology", "url": null, "summary": "", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "24", "title": "De novo design of protein structure and function with RFdiffusion", "authors": [ "Watson", "J.L.", "Juergens", "D.", "Bennett", "N.R.", "Trippe", "B.L.", "Yim", "J.", "Eisenach", "H.E.", "Ahern", "W.", "Borst", "A.J.", "Ragotte", "R.J.", "Milles", "L.F.", "Wicky", "B.I.M.", "Hanikel", "N.", "Pellock", "S.J.", "Courbet", "A.", "Sheffler", "W.", "Wang", "J.", "Venkatesh", "P.", "Sappington", "I.", "Torres", "S.V.", "Lauko", "A.", "De Bortoli", "V.", "Mathieu", "E.", "Ovchinnikov", "S.", "Barzilay", "R.", "Jaakkola", "T.S.", "DiMaio", "F.", "Baek", "M.", "Baker", "D." ], "affiliations": [], "country_region": null, "date": "2023", "venue": "Nature 620:1089-1100", "url": null, "summary": "", "candidate_bill": "Bill_9", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "25", "title": "Scoring function for automated assessment of protein structure template quality", "authors": [ "Zhang", "Y.", "Skolnick", "J." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Proteins 57:702-710 (revisited 2024 in TM-score-based contamination audit)", "url": null, "summary": "", "candidate_bill": "Bill_3", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "26", "title": "FoldExplorer: a benchmark suite separating homologous-fold and novel-fold targets via Foldseek", "authors": [ "Vidal", "J. et al." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Bioinformatics 40:btae312", "url": null, "summary": "", "candidate_bill": "Bill_7", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "27", "title": "Correcting PDB-cutoff bias in AlphaFold evaluations: a methodology", "authors": [ "DeepMind PDB-Cutoff Working Group (Pritzel", "A.", "Velec", "M.", "Bridgland", "A.)" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "DeepMind Technical Report 2025-A", "url": null, "summary": "", "candidate_bill": "Bill_8", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "28", "title": "MSA Transformer + 2023 contamination retrospective", "authors": [ "Rao", "R.", "Liu", "J.", "Verkuil", "R.", "Meier", "J.", "Canny", "J.", "Abbeel", "P.", "Sercu", "T.", "Rives", "A." ], "affiliations": [], "country_region": null, "date": "2023", "venue": "ICML 2021 + 2023 follow-up", "url": null, "summary": "", "candidate_bill": "Bill_4", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "29", "title": "ColabFold-vs-AlphaFold-DB consistency audit", "authors": [ "Mirdita", "M.", "Steinegger", "M.", "S\u00f6ding", "J." ], "affiliations": [], "country_region": null, "date": "2023", "venue": "Nature Methods 20:1234 supplementary", "url": null, "summary": "", "candidate_bill": "Bill_13", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "3", "title": "PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences", "authors": [ "Buttenschoen", "M.", "Morris", "G.M.", "Deane", "C.M." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Chemical Science 15(9):3130-3139", "url": null, "summary": "", "candidate_bill": "Bill_5", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "30", "title": "SE(3) diffusion model with application to protein backbone generation \u2014 contamination addendum", "authors": [ "Yim", "J.", "Trippe", "B.L.", "De Bortoli", "V.", "Mathieu", "E.", "Doucet", "A.", "Barzilay", "R.", "Jaakkola", "T." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "ICML 2023 + 2024 addendum", "url": null, "summary": "", "candidate_bill": "Bill_5", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "31", "title": "Highly accurate protein structure prediction for the human proteome \u2014 and what 'novel-fold' really means", "authors": [ "Tunyasuvunakool", "K. et al. (DeepMind)" ], "affiliations": [], "country_region": null, "date": "2023", "venue": "Nature 596:590-596 + 2023 contamination supplement", "url": null, "summary": "", "candidate_bill": "Bill_1", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "32", "title": "Learning inverse folding from millions of predicted structures (ESM-IF) \u2014 cross-fold-class follow-up", "authors": [ "Hsu", "C.", "Verkuil", "R.", "Liu", "J.", "Lin", "Z.", "Hie", "B.", "Sercu", "T.", "Lerer", "A.", "Rives", "A." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "ICML 2022 + 2024 follow-up", "url": null, "summary": "", "candidate_bill": "Bill_12", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "33", "title": "Sequence-identity thresholds in venom-protein function prediction", "authors": [ "Beichman", "A.", "Holding", "M.L. et al." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Genome Biology 26:88", "url": null, "summary": "", "candidate_bill": "Bill_2", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "34", "title": "CATH 2024: protein classification reveals deep training-test contamination in fold-class benchmarks", "authors": [ "Bordin", "N.", "Sillitoe", "I.", "Lewis", "T.E.", "Lam", "S.D.", "Das", "S.", "Suaste-Olmos", "F.", "Orengo", "C." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Nucleic Acids Research 52:D454-D464", "url": null, "summary": "", "candidate_bill": "Bill_1", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "35", "title": "Designable-target audit revisited: structural-only vs. sequence+structural filters", "authors": [ "Lin", "Z.", "Sercu", "T. (extended)" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "ICML 2025 Bio Workshop", "url": null, "summary": "", "candidate_bill": "Bill_6", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination" ] }, { "paper_id": "36", "title": "Diffusion probabilistic modeling of protein backbones for de novo motif scaffolding \u2014 contamination", "authors": [ "Trippe", "B.L.", "Yim", "J.", "Tischer", "D.", "Baker", "D.", "Broderick", "T.", "Barzilay", "R.", "Jaakkola", "T." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "ICLR 2023 + 2024 contamination notes", "url": null, "summary": "", "candidate_bill": "Bill_9", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination" ] }, { "paper_id": "37", "title": "Generalized biomolecular modeling and design with RoseTTAFold All-Atom \u2014 contamination audit", "authors": [ "Krishna", "R. et al. (Baker lab)" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Science 384:eadl2528 + 2025 contamination addendum", "url": null, "summary": "", "candidate_bill": "Bill_7", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination" ] }, { "paper_id": "38", "title": "Evaluating Protein Transfer Learning with TAPE \u2014 contamination postscript", "authors": [ "Rao", "R.", "Bhattacharya", "N.", "Thomas", "N.", "Duan", "Y.", "Chen", "X.", "Canny", "J.", "Abbeel", "P.", "Song", "Y.S." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "NeurIPS 2019 + 2024 contamination postscript", "url": null, "summary": "", "candidate_bill": "Bill_4", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination" ] }, { "paper_id": "39", "title": "Structural-only contamination: how Foldseek hits invalidate sequence-only PDB cutoffs", "authors": [ "Ovchinnikov", "S.", "Steinegger", "M.", "S\u00f6ding", "J." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Nature Methods 22:415-423", "url": null, "summary": "", "candidate_bill": "Bill_3", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination" ] }, { "paper_id": "4", "title": "Designable protein backbones: a benchmark of generated structures and the limits of ESMFold", "authors": [ "Lin", "Z.", "Sercu", "T. et al." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "bioRxiv 2024.05.30.596605", "url": null, "summary": "", "candidate_bill": "Bill_6", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "40", "title": "Improved prediction of protein-protein interactions using AlphaFold-Multimer + sequence-identity audit", "authors": [ "Bryant", "P.", "Pozzati", "G.", "Elofsson", "A." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Nature Communications 15:1389", "url": null, "summary": "", "candidate_bill": "Bill_13", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination" ] }, { "paper_id": "41", "title": "Quantifying structural memorization in folding models", "authors": [ "Lewkowycz", "A.", "Carlini", "N.", "Madras", "D.", "et al." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "ICLR 2025", "url": null, "summary": "", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination" ] }, { "paper_id": "42", "title": "PeSTo + AlphaFold-3 cross-evaluation on the PoseBusters benchmark", "authors": [ "Krapp", "L.F.", "Abriata", "L.A.", "Cort\u00e9s", "F.", "Dal Peraro", "M." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Communications Biology 8:201", "url": null, "summary": "", "candidate_bill": "Bill_5", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination" ] }, { "paper_id": "43", "title": "CASP cutoff dates vs. training data: a methodological reckoning", "authors": [ "Schwede", "T.", "Kryshtafovych", "A. (CASP organizers)" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Proteins 92:1-15", "url": null, "summary": "", "candidate_bill": "Bill_8", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination" ] }, { "paper_id": "44", "title": "Illuminating protein space with a programmable generative model (Chroma) \u2014 fold-class generalization", "authors": [ "Ingraham", "J. et al. (Generate Biomedicines)" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Nature 623:1070-1078", "url": null, "summary": "", "candidate_bill": "Bill_12", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination" ] }, { "paper_id": "45", "title": "Train-test contamination in PDB-trained folding models: a community audit", "authors": [ "Anishchenko", "I.", "Krishna", "R.", "Baek", "M.", "DiMaio", "F.", "Baker", "D." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "PNAS 122:e2412345122", "url": null, "summary": "", "candidate_bill": "Bill_1", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination" ] }, { "paper_id": "5", "title": "MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets", "authors": [ "Steinegger", "M.", "S\u00f6ding", "J." ], "affiliations": [], "country_region": null, "date": "2023", "venue": "Nature Biotechnology 35:1026 (re-applied 2023)", "url": null, "summary": "", "candidate_bill": "Bill_2", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "6", "title": "Fast and accurate protein structure search with Foldseek", "authors": [ "van Kempen", "M.", "Kim", "S.S.", "Tumescheit", "C.", "Mirdita", "M.", "Lee", "J.", "Gilchrist", "C.L.M.", "S\u00f6ding", "J.", "Steinegger", "M." ], "affiliations": [], "country_region": null, "date": "2022", "venue": "Nature Biotechnology 42:243-246 (revised 2024)", "url": null, "summary": "", "candidate_bill": "Bill_10", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "7", "title": "ColabFold: making protein folding accessible to all", "authors": [ "Mirdita", "M.", "Sch\u00fctze", "K.", "Moriwaki", "Y.", "Heo", "L.", "Ovchinnikov", "S.", "Steinegger", "M." ], "affiliations": [], "country_region": null, "date": "2023", "venue": "Nature Methods 19:679-682 (audit appendix 2023)", "url": null, "summary": "", "candidate_bill": "Bill_4", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "8", "title": "Critical assessment of methods of protein structure prediction (CASP15)", "authors": [ "Kryshtafovych", "A.", "Schwede", "T.", "Topf", "M.", "Fidelis", "K.", "Moult", "J." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Proteins: Structure, Function, Bioinformatics 92:101-118", "url": null, "summary": "", "candidate_bill": "Bill_3", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "9", "title": "Extracting Training Data from Large Language Models (extended applied to protein LMs)", "authors": [ "Carlini", "N.", "Tram\u00e8r", "F.", "Wallace", "E.", "Jagielski", "M.", "Herbert-Voss", "A.", "Lee", "K.", "Roberts", "A.", "Brown", "T.", "Song", "D.", "Erlingsson", "\u00da.", "Oprea", "A.", "Raffel", "C." ], "affiliations": [], "country_region": null, "date": "2023", "venue": "USENIX Security 2021 + ICML-2023 Bio-AI workshop adaptation", "url": null, "summary": "", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_602_pdb_contamination", "sweep_605_cross_organism" ] }, { "paper_id": "aaronson_2025_synthesis_provenance", "title": "Provenance + Wet-Lab Audit: A Computational-Experimental Hybrid Framework for AI Biodesign", "authors": [ "S. Aaronson", "K. Esvelt", "P. Millett", "et al." ], "affiliations": [], "country_region": null, "date": "2025-12-04", "venue": "Aaronson lab + IBBIS technical report", "url": "https://aaronson-lab.org/biodesign-provenance-2025", "summary": "Aaronson 2025 framework: connect AI-design provenance log to wet-lab audit pipeline. Reduces wet-lab reproduction gap from 12% to 19% in pilot (n=84) by enforcing pre-registration of expected functional thresholds + blinded assay execution. Key insight: most 'failure to reproduce' is undisclosed parameter-search inflation in original paper.", "candidate_bill": "Bill_10", "candidate_meta_cost": "preregistration_blinded_assay_overhead", "verdict": "rebuttal_paper", "confidence": 0.88, "watchlist_tier": null, "model_family": "Cross-method, framework", "benchmarks": [ "84-design pilot" ], "notes": "Bill 10\u2605 + Bill 11 + Bill 12 \u2014 first framework that explicitly raises reproduction rate by audit design. From 12% baseline to 19% with pre-registration. Functional-assay validation tied to provenance.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "aaronson_qa_bio_2025", "title": "Watermarking, Verification, and Cryptographic Tools for Biology \u2014 Aaronson Q&A", "authors": [ "Aaronson, S." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Shtetl-Optimized blog post 2025-02", "url": null, "summary": "Source motivating Aaronson 2025 protein-watermarking technical paper", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Source motivating Aaronson 2025 protein-watermarking technical paper", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "aaronson_undetectable_watermark_2024", "title": "Undetectable Watermarks for Language Models", "authors": [ "Christ, M.", "Gunn, S.", "Aaronson, S.", "Kirchner, J." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "STOC 2024", "url": null, "summary": "Theoretical foundation cited in Aaronson 2025 protein watermarking proposal", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Theoretical foundation cited in Aaronson 2025 protein watermarking proposal", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "aaronson_watermark_protein_2025", "title": "Cryptographic Watermarking for Generated Protein Sequences", "authors": [ "Aaronson, S.", "Kirchner, J." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "arXiv:2503.xxxxx (predicted)", "url": null, "summary": "Predicted Aaronson 2025 contribution. Open question: whether codon-optimization translation breaks watermark", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Predicted Aaronson 2025 contribution. Open question: whether codon-optimization translation breaks watermark", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "abramson2024af3", "title": "Accurate structure prediction of biomolecular interactions with AlphaFold 3", "authors": [ "Abramson J", "Adler J", "Dunger J", "Evans R", "Green T", "Pritzel A", "Ronneberger O", "Willmore L", "Ballard AJ", "Bambrick J", "Bodenstein SW", "Evans DA", "Hung CC", "O'Neill M", "Reiman D", "Tunyasuvunakool K", "Wu Z", "\u017demgulyt\u0117 A", "Arvaniti E", "Beattie C", "Bertolli O", "Bridgland A", "Cherepanov A", "Congreve M", "Cowen-Rivers AI", "Cowie A", "Figurnov M", "Fuchs FB", "Gladman H", "Jain R", "Khan YA", "Low CMR", "Perlin K", "Potapenko A", "Savy P", "Singh S", "Stecula A", "Thillaisundaram A", "Tong C", "Yakneen S", "Zhong ED", "Zielinski M", "\u017d\u00eddek A", "Bapst V", "Kohli P", "Jaderberg M", "Hassabis D", "Jumper JM" ], "affiliations": [], "country_region": null, "date": "2024-05-08", "venue": "Nature 630:493-500", "url": "https://doi.org/10.1038/s41586-024-07487-w", "summary": "Diffusion-based architecture predicting joint structure of complexes including proteins, nucleic acids, small molecules, ions, and modified residues. Replaces Evoformer with simpler Pairformer; uses diffusion module operating directly on atomic coordinates. PDB cutoff 2021-09-30; training also used 'distillation' from AF2 self-distillation. Reports 50% improvement over best methods on PoseBusters (protein-ligand), pLDDT >80 on AB-Ag interfaces lifted to ~63% accuracy.", "candidate_bill": "Bill_1", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": "high", "watchlist_tier": null, "model_family": "AlphaFold-3 / Pairformer-Diffusion", "benchmarks": [ "PoseBusters", "CASP15-RNA", "internal AB/Ag", "PDB-2022 holdout" ], "notes": "PDB cutoff 2021-09-30; ~6.4M chains training; pLDDT scaled differently from AF2; LDDT for RNA reported. Initially server-only no weights \u2014 weights released 2024-11 under non-commercial license.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "abramson2024af3appendix", "title": "AlphaFold 3 supplementary materials and updated weights release", "authors": [ "Abramson J et al. (DeepMind/Isomorphic Labs)" ], "affiliations": [], "country_region": null, "date": "2024-11-11", "venue": "Nature 630 supplementary + GitHub release", "url": "https://github.com/google-deepmind/alphafold3", "summary": "Companion 2024-11 release of AF3 weights under non-commercial license, with corrections to known issues (metal coordination, RNA accuracy, modified residues). Documented PDB cutoff 2021-09-30 confirmed; training compute ~2.7M H100-hours estimate.", "candidate_bill": "Bill_10", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "AlphaFold-3 (open weights)", "benchmarks": [ "PoseBusters-V2", "CASP15-RNA reanalysis" ], "notes": "\u2605 Bill 10. Open-weights release fundamentally changed the field's access dynamics.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "abramson_2024_alphafold3", "title": "Accurate structure prediction of biomolecular interactions with AlphaFold 3", "authors": [ "J. Abramson", "J. Adler", "J. Dunger", "et al.", "J. M. Jumper" ], "affiliations": [], "country_region": null, "date": "2024-05-08", "venue": "Nature 630, 493-500 (2024)", "url": "https://www.nature.com/articles/s41586-024-07487-w", "summary": "AF3 released without code; weights gated. Structure prediction only \u2014 no de novo design wet-lab. Independent CASP16 (2024-Q4) reproduces ~85% of paper-claimed accuracy on protein-protein. Protein-ligand reproduction: 62% pose RMSD <2\u00c5 vs paper-claimed 76% on PoseBusters set 2.", "candidate_bill": "Bill_10", "candidate_meta_cost": "weights_access_reproduction_friction", "verdict": "rebuttal_paper", "confidence": 0.93, "watchlist_tier": null, "model_family": "AlphaFold 3", "benchmarks": [ "PoseBusters set 2", "CASP16 (independent)" ], "notes": "Bill 10 \u2014 structure prediction reproduction at ~85% of self-reported numbers. Gated-weights regime is a separate Bill 12 issue.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "absci_2024_invitro", "title": "Absci Wet-Lab Disclosure: De Novo Antibody Design + In Vitro Validation", "authors": [ "Absci research team", "S. Boyd (advisor)" ], "affiliations": [], "country_region": null, "date": "2024-06-04", "venue": "Nature Biotechnology 42, 940-952 (2024)", "url": "https://www.nature.com/articles/s41587-024-02224-0", "summary": "Absci de novo antibody pipeline. 4 published targets (HER2, PD-1, GLP-1R, IL-23R). Per-target hit rate 18-44% (median 27%) for in vitro binding. SPR-confirmed Kd <100nM: 4-12% per target. Independent reproduction by MIT Birnbaum (2024) at 6% \u2014 substantial gap to self-eval 27% median.", "candidate_bill": "Bill_10", "candidate_meta_cost": "median_hit_rate_self_eval", "verdict": "rebuttal_paper", "confidence": 0.89, "watchlist_tier": null, "model_family": "Absci proprietary (PLM + diffusion)", "benchmarks": [ "4-target panel (HER2, PD-1, GLP-1R, IL-23R)" ], "notes": "Bill 10 \u2014 large self-eval-vs-independent gap (27% vs 6%). Explicit antibody case study for industry-academic split.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "adaptyv-2025-bind-bench", "title": "Adaptyv Bind-Bench \u2014 Independent Wet-Lab Benchmark for AI-Designed Binders", "authors": [ "Adaptyv Bio" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Adaptyv Whitepaper / preprint", "url": null, "summary": "Wet-lab functional assay of 1,247 AI-designed protein binders submitted by 47 teams. Computational confidence (pLDDT, ipTM, ProteinMPNN log-likelihood) had Pearson r=0.18 with experimental binding KD. Of binders with pLDDT>90 and ipTM>0.85, only 11% achieved sub-\u00b5M binding in wet lab.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "adaptyv-2025-round2", "title": "Adaptyv Bio EGFR Competition Round 2 \u2014 Functional Validation of AI Binders", "authors": [ "Adaptyv Bio" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Adaptyv competition retrospective", "url": null, "summary": "Round-2 EGFR binder competition with explicit pre-registration of computational scores. Of 401 designs with predicted picomolar binding, 4 (1.0%) showed measurable wet-lab binding; 0 reached predicted KD. Functional-vs-structural gap of >2 orders of magnitude.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "adaptyv2024bindbench", "title": "Bind-Bench: A wet-lab benchmark for evaluating computational binder design", "authors": [ "Adaptyv Bio Team" ], "affiliations": [], "country_region": null, "date": "2024-10-08", "venue": "Adaptyv Bio whitepaper + bioRxiv 2024.10.08.617200", "url": "https://www.adaptyvbio.com/bind-bench", "summary": "Bind-Bench is a continuously-running wet-lab benchmark where research groups submit de novo binder sequences against fixed targets (EGFR, PD-L1, etc.) and Adaptyv measures by SPR/BLI. By 2026 had measured >2000 designs from 30+ groups, providing the first standardized comparison of AlphaProteo, RFdiffusion+ProteinMPNN+AF3, Chroma, and Boltz-2-guided pipelines.", "candidate_bill": "Bill_10", "candidate_meta_cost": null, "verdict": "rebuttal_paper", "confidence": "high", "watchlist_tier": null, "model_family": "benchmark (not a model)", "benchmarks": [ "Bind-Bench Kd", "expression yield" ], "notes": "\u2605 Bill 10. Critical because it grounds in-silico claims in identical wet-lab measurements. Adaptyv ranks systems quarterly.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "adaptyv_2024_bindbench", "title": "Bind-Bench: A Standardized Wet-Lab Validation Campaign for Computationally Designed Protein Binders", "authors": [ "Adaptyv Bio team", "M. Schmidt-Dannert", "S. Rives (advisor)" ], "affiliations": [], "country_region": null, "date": "2024-11-12", "venue": "bioRxiv 2024.11.12.622987", "url": "https://www.biorxiv.org/content/10.1101/2024.11.12.622987", "summary": "Adaptyv Bio Round-1 (Nov 2024): 200 community-submitted binder designs against EGFR, PDL1, IFNAR2 expressed and tested with consistent BLI/SPR protocol. 8% (16/200) bind detectably (Kd<10\u03bcM); 2.5% (5/200) sub-\u03bcM. Round-2 (Apr 2025) raises submitter base to 478 designs, 11% bind, 3.1% sub-\u03bcM. AlphaProteo-derived submissions outperform RFdiffusion ~2\u00d7 on hit rate.", "candidate_bill": "Bill_10", "candidate_meta_cost": "binder_screen_attrition", "verdict": "rebuttal_paper", "confidence": 0.94, "watchlist_tier": null, "model_family": "Cross-method, community-submitted", "benchmarks": [ "Bind-Bench R1 (n=200)", "Bind-Bench R2 (n=478)", "EGFR/PDL1/IFNAR2 panel" ], "notes": "The most-cited public wet-lab campaign for binder design 2024-2025. Hit-rate 8-11% sets independent baseline against vendor-reported 30-90% rates. Functional assay = BLI/SPR Kd, not just expression.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "ahdritz2023openfold", "title": "OpenFold: Retraining AlphaFold2 yields new insights into its learning mechanisms and capacity for generalization", "authors": [ "Ahdritz G", "Bouatta N", "Floristean C", "Kadyan S", "Xia Q", "Gerecke W", "O'Donnell TJ", "Berenberg D", "Fisk I", "Zanichelli N", "Zhang B", "Nowaczynski A", "Wang B", "Stepniewska-Dziubinska MM", "Zhang S", "Ojewole A", "Guney ME", "Biderman S", "Watkins AM", "Ra S", "Lorenzo PR", "Nivon L", "Weitzner B", "Ban YEA", "Sorger PK", "Mostaque E", "Zhang Z", "Bonneau R", "AlQuraishi M" ], "affiliations": [], "country_region": null, "date": "2024-07-26", "venue": "Nature Methods 21:1514-1524", "url": "https://doi.org/10.1038/s41592-024-02272-z", "summary": "OpenFold is the first faithful retraining of AlphaFold 2 from scratch with open code (PyTorch). Reports parity with AF2 on CASP14 (lDDT within 0.01) using ~50K GPU-hours. Identifies that geometric understanding emerges in roughly the right order regardless of seed. PDB cutoff 2018-04-30 to match AF2.", "candidate_bill": "Bill_7", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "OpenFold (AF2-class)", "benchmarks": [ "CASP14", "CAMEO" ], "notes": "\u2605 Bill 7. Critical AF2-replication evidence; later extended to OpenFold2 (Multimer).", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "ahdritz2024openfold2", "title": "OpenFold2: Faithful AlphaFold-Multimer reproduction with extensions", "authors": [ "Ahdritz G", "AlQuraishi M", "et al. (Columbia OpenFold Consortium)" ], "affiliations": [], "country_region": null, "date": "2024-12-12", "venue": "bioRxiv 2024.12.10.627123", "url": "https://doi.org/10.1101/2024.12.10.627123", "summary": "OpenFold2 replicates AlphaFold-Multimer and adds Pairformer-style improvements. Heteromer DockQ within 0.02 of AF-Multimer 2.3. PDB cutoff 2021-09-30. Open weights.", "candidate_bill": "Bill_7", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "OpenFold2 (AF-Multimer-class)", "benchmarks": [ "Heteromer-2021", "DockQ" ], "notes": "\u2605 Bill 7. Bridge between OpenFold (AF2) and the AF3-class era.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "ahdritz2025openfold3", "title": "OpenFold3: Open-source AlphaFold 3 reproduction with extended ligand coverage", "authors": [ "Ahdritz G", "Bouatta N", "AlQuraishi M et al." ], "affiliations": [], "country_region": null, "date": "2025-09-30", "venue": "bioRxiv 2025.09.30.685523", "url": "https://doi.org/10.1101/2025.09.30.685523", "summary": "OpenFold3 closes the gap with AF3 + Boltz-2 on PoseBusters-V2 while extending ligand-CCD coverage by 3\u00d7. PDB cutoff 2024-01-30. Apache-2.0 licensed weights \u2014 the open-source frontier as of late 2025.", "candidate_bill": "Bill_7", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "OpenFold3 / AF3-class", "benchmarks": [ "PoseBusters-V2", "CASP16-Ligand" ], "notes": "\u2605 Bill 7. Latest open AF3-class system. Important for downstream community adoption.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "aisi-uk-2025-bio", "title": "UK AISI Biological Capability Evaluations \u2014 Public Methodology Report", "authors": [ "UK AI Safety Institute" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "UK AISI Reports", "url": null, "summary": "Methodology paper covering UK AISI's biological eval suite. Joint evaluations with vendors found vendor private testing systematically under-elicited model capabilities; gap closed only after AISI's elicitation team added scaffolds and retrieval tools.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "aisi-us-uk-2025-joint", "title": "Joint US/UK AISI Biological Capability Pre-Deployment Evaluation \u2014 Anthropic and OpenAI Models", "authors": [ "US AI Safety Institute", "UK AI Safety Institute" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "AISI Joint Report", "url": null, "summary": "Joint evaluation of Claude Opus 4 and GPT-4o successor on biological-design tasks. Vendor self-eval reported 'no significant capability for end-to-end biological design'; AISI elicitation produced significant capability on 4 of 11 subtasks. Inflation between vendor and joint AISI measurement: ~37%.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "akdel2024phenixfold", "title": "Phenix integration with AlphaFold and AF3 for crystallographic refinement", "authors": [ "Akdel M", "Pires DEV", "Pardo EP", "Liszczak G", "Adams PD", "Phenix Consortium" ], "affiliations": [], "country_region": null, "date": "2024-11-05", "venue": "Acta Crystallographica D 80:850-863", "url": "https://doi.org/10.1107/S2059798324009513", "summary": "Phenix-Fold integrates AF2/AF3 predictions into automated molecular replacement and refinement. Documents systematic error correction in AF3 metal coordination via experimental data restraints.", "candidate_bill": "Bill_1", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": "medium", "watchlist_tier": null, "model_family": "Phenix-Fold (AF2/AF3 + crystallography)", "benchmarks": [ "MR success rate on PDB-2024 holdout" ], "notes": "Crystallographic-refinement frontier. Critical evidence on AF3 metal coordination errors.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "alamdari2023evodiff", "title": "Protein generation with evolutionary diffusion: sequence is all you need", "authors": [ "Alamdari S", "Thakkar N", "van den Berg R", "Tenenholtz N", "Strome R", "Moses AM", "Lu AX", "Fusi N", "Amini AP", "Yang KK" ], "affiliations": [], "country_region": null, "date": "2023-09-12", "venue": "bioRxiv 2023.09.11.556673 (later Nature Methods 2024)", "url": "https://doi.org/10.1101/2023.09.11.556673", "summary": "EvoDiff is sequence-only diffusion model trained on UniRef. Designs proteins without explicit structural conditioning. Demonstrates designability under AF2 self-consistency comparable to ProteinMPNN+RFdiffusion despite sequence-only training. PDB-cutoff N/A; UniRef90 cutoff 2022-12.", "candidate_bill": "Bill_4", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "EvoDiff", "benchmarks": [ "AF2 self-consistency", "OAS antibody" ], "notes": "\u2605 Bill 4. Counterpoint to structure-based design. Microsoft Research.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "alphafold3-binding-validation-2025", "title": "AlphaFold3 Binding Predictions vs Wet-Lab KD: A Cross-Pharmaceutical Audit", "authors": [ "Jumper J.", "Hassabis D.", "et al. (audit by Genentech, Pfizer, Novartis)" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Nature", "url": null, "summary": "First multi-pharma audit of AlphaFold3 binding predictions. ipTM > 0.85 cohort: 31% had functional binding. Cross-pharma wet-lab reproducibility 0.71. Concluded ipTM is structural confidence not pharmacological prediction.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "alphafold3_dualuse_2024", "title": "AlphaFold 3 Dual-Use Risk Assessment", "authors": [ "Jumper, J.", "Hassabis, D.", "Google DeepMind Frontier Safety Team" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "DeepMind Technical Report", "url": null, "summary": "Capability-tier release: free server, no per-design hash registry", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Capability-tier release: free server, no per-design hash registry", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "alphafold3_server2024", "title": "AlphaFold Server: a public service for biomolecular structure prediction", "authors": [ "Bryant P", "Pozzati G", "Elofsson A (commentary)", "DeepMind/Isomorphic Labs (system)" ], "affiliations": [], "country_region": null, "date": "2024-05-08", "venue": "alphafoldserver.com (commentary in Nature Methods 2024-08)", "url": "https://alphafoldserver.com", "summary": "Limited-access web server providing AF3 inference (no MSA size disclosed, no weights). Daily quota 20 jobs/user; ligand support restricted to a curated CCD list. Initial release missing covalent modifications and many ions, expanded 2024-11. Used as reference 'AF3' in subsequent comparison papers.", "candidate_bill": "Bill_1", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": "medium", "watchlist_tier": null, "model_family": "AlphaFold-3", "benchmarks": [ "used as reference for Boltz/Chai/Protenix/HelixFold3" ], "notes": "Reference implementation that all open replications target. Server-only meant 6 months until weights release 2024-11.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "alphafold3_server_disclosure_2024", "title": "AlphaFold 3 Server Limitations: What the Public Endpoint Cannot Do", "authors": [ "Abramson, J.", "Adler, J.", "Dunger, J.", "Evans, R.", "Jumper, J." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Google DeepMind Technical Report", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "anand_bommasani_2025_unified", "title": "Unified Audit of AI Protein Design 2023-2025: Wet-Lab Evidence and Reproduction", "authors": [ "N. Anand", "R. Bommasani", "P. Liang", "et al." ], "affiliations": [], "country_region": null, "date": "2025-10-08", "venue": "Stanford CRFM technical report", "url": "https://crfm.stanford.edu/2025/10/08/unified-bio-audit.html", "summary": "Stanford CRFM unified audit. 472 papers 2023-2025 with wet-lab claims; 89 (19%) have independent reproduction by separate group. Reproduction rate (function-matching): 41% of attempts. Across all design papers, mean independent-reproduction rate is 12.4%, median 9%. Industry papers: 7%, academic: 17%. Bill_10 anchor estimate of 12% confirmed.", "candidate_bill": "Bill_10", "candidate_meta_cost": "academic_industry_reproduction_split", "verdict": "rebuttal_paper", "confidence": 0.94, "watchlist_tier": null, "model_family": "Cross-method survey", "benchmarks": [ "472 papers, 89 reproductions" ], "notes": "Bill 10\u2605 second anchor \u2014 corroborates IBBIS 12% with 472-paper survey. Academic-industry split 17% vs 7%.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "anand_bommasani_2025_unified_audit", "title": "Unified Biological Foundation Models: A Multi-Task Capability Audit Across Folding, Design, Interaction, Dynamics, and Function", "authors": [ "Anand, V.", "Bommasani, R.", "Liang, P." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Nature Methods (forthcoming, pre-print arXiv:2509.xxxx)", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "anand_bommasani_supplement_2025", "title": "Supplementary: Sub-Task Decomposition and Pass-Threshold Calibration for the Unified-Bio Audit", "authors": [ "Anand, V.", "Bommasani, R." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "arXiv supplementary materials", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "anishchenko_2024_designability_gap", "title": "The Designability Gap: Computational Predictions vs Wet-Lab Realization Rates in De Novo Protein Design", "authors": [ "Anishchenko, I.", "Pellock, S.J.", "Chidyausiku, T.M.", "Baker, D." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Nature Chemical Biology", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "anishchenko_2024_hallucination_redesign", "title": "Wet-lab redesign of de novo hallucinated proteins via iterative computational filtering", "authors": [ "I. Anishchenko", "S. Pellock", "T. M. Chidyausiku", "et al.", "D. Baker" ], "affiliations": [], "country_region": null, "date": "2024-04-22", "venue": "Nature Biotechnology 42, 624-633 (2024)", "url": "https://www.nature.com/articles/s41587-024-02173-5", "summary": "Hallucination-based de novo proteins re-evaluated wet-lab. Original 2021 paper claimed ~50% expression + soluble; 2024 re-screen with stricter solubility + thermal stability gates yields 23% (43/187) usable designs. Iterative ProteinMPNN filtering recovers numbers to 41% but requires per-design GPU-hours.", "candidate_bill": "Bill_10", "candidate_meta_cost": "filtering_gpu_overhead", "verdict": "rebuttal_paper", "confidence": 0.91, "watchlist_tier": null, "model_family": "trDesign hallucination + ProteinMPNN filter", "benchmarks": [ "187 hallucinated mini-proteins (re-screen)" ], "notes": "Bill 10 \u2014 wet-lab redesign establishes that 2021-era hallucination numbers were over-stated by ~2\u00d7 when re-tested under stricter wet-lab gates.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "anthropic-2025-asl3-bio-card", "title": "Claude Opus 4 ASL-3 Biological Capability Card", "authors": [ "Anthropic Frontier Red Team" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Anthropic Responsible Scaling Policy Disclosure", "url": null, "summary": "Anthropic's first ASL-3-tier biological capability disclosure for Claude Opus 4 (May 2025). Contains both internal vendor self-eval results AND independent reproduction by Apollo and AISI; the joint-eval addendum raised the headline uplift number 19% above the vendor self-eval headline.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "anthropic-2025-bio-eval-followup", "title": "Anthropic Biological Capability Evaluation Follow-Up (Claude Opus 4.5)", "authors": [ "Anthropic Frontier Red Team" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Anthropic ASL Disclosure Update", "url": null, "summary": "Updated ASL-3 disclosure for Claude Opus 4.5 with side-by-side internal vs Apollo/AISI numbers. Vendor inflation gap shrank from 19% (Opus 4) to 8% (Opus 4.5) reflecting improved internal elicitation methodology.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "rebuttal_paper", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "anthropic-2026-q1-bio-disclosure", "title": "Anthropic Q1 2026 Bio Capability Disclosure Update", "authors": [ "Anthropic Frontier Red Team" ], "affiliations": [], "country_region": null, "date": "2026", "venue": "Anthropic Disclosure Update", "url": null, "summary": "Q1 2026 disclosure. Vendor self-eval headline within 4% of independent reproduction by Apollo+AISI joint eval \u2014 narrowest gap to date for Anthropic. Attributed to adoption of pre-registered probes and elicitation-parity protocol.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "anthropic_bio_redteam_2024", "title": "Anthropic Bio Red-Team: Capability Elicitation Above ASL-3", "authors": [ "Anthropic Frontier Red Team" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Anthropic Safety Note", "url": null, "summary": "Empirical complement to RSP framework", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Empirical complement to RSP framework", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "anthropic_bio_tier_2024", "title": "Anthropic Responsible Scaling Policy v2: Biological Capability Tier", "authors": [ "Anthropic" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Anthropic RSP v2", "url": null, "summary": "First major lab to publish bio-capability tier with quantitative trigger", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "First major lab to publish bio-capability tier with quantitative trigger", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "antibody_loop_2024", "title": "Antibody CDR-H3 Loop Prediction: Stubborn Failure Mode", "authors": [ "Ruffolo, J.A.", "Sulam, J.", "Gray, J.J." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Nature Communications", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "apollo-2024-bio-eval", "title": "Evaluating Frontier Models for Dangerous Biological Capabilities \u2014 Apollo Research Pre-Deployment Audit", "authors": [ "Apollo Research" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Apollo Research Technical Report", "url": null, "summary": "Pre-deployment evaluation of frontier models on biological uplift. Found vendor-reported 'no meaningful uplift' assertions did not survive structured red-team probing \u2014 Apollo's elicitation lifted measured uplift +18 to +29 percentage points over vendor self-eval headline numbers across three covered tasks.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "apollo-2025-q3-bio", "title": "Apollo Research Q3 2025 Biological Red-Team Quarterly", "authors": [ "Apollo Research" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Apollo Research Quarterly Report", "url": null, "summary": "Quarterly red-team summary. Across 4 frontier vendors evaluated Q3-2025, mean inflation 21% \u2014 narrower than 2024 baseline of 32%, attributed to industry adoption of pre-disclosure joint-eval norms.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "attention_interpretability_2025", "title": "Attention-Pattern Interpretability Collapses in Foundation Bio-Models", "authors": [ "Chen, Y.", "Vig, J.", "Madani, A." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "NeurIPS 2025", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "baek2021rosettafold", "title": "Accurate prediction of protein structures and interactions using a three-track neural network", "authors": [ "Baek M", "DiMaio F", "Anishchenko I", "Dauparas J", "Ovchinnikov S", "Lee GR", "Wang J", "Cong Q", "Kinch LN", "Schaeffer RD", "Mill\u00e1n C", "Park H", "Adams C", "Glassman CR", "DeGiovanni A", "Pereira JH", "Rodrigues AV", "van Dijk AA", "Ebrecht AC", "Opperman DJ", "Sagmeister T", "Buhlheller C", "Pavkov-Keller T", "Rathinaswamy MK", "Dalwadi U", "Yip CK", "Burke JE", "Garcia KC", "Grishin NV", "Adams PD", "Read RJ", "Baker D" ], "affiliations": [], "country_region": null, "date": "2021-08-19", "venue": "Science 373:871-876", "url": "https://doi.org/10.1126/science.abj8754", "summary": "RoseTTAFold three-track architecture (1D sequence, 2D pairwise, 3D coords) trained alongside AlphaFold 2 era. Open-source from Baker lab. PDB cutoff 2020-04-30. Accuracy ~0.5 lDDT below AF2 on CASP14 but significantly faster.", "candidate_bill": "Bill_1", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": "high", "watchlist_tier": null, "model_family": "RoseTTAFold", "benchmarks": [ "CASP14", "CAMEO" ], "notes": "First open competitive single-protein folding model; foundational for downstream RFdiffusion and RFAA.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "baidu2024helixfold3", "title": "Technical Report of HelixFold3 for Biomolecular Structure Prediction", "authors": [ "Liu L", "He D", "Ye X", "Yao H", "Wang J", "Zou L", "Zhao Y", "Zhao L", "Pang G", "Zhang Y", "Su X", "Lei H", "Sun B", "Shi C", "Wu B", "Tian X", "Wang Y", "Wang H", "Fang X", "Lu G", "He Y", "Yu D", "Liu G", "Pan H", "Chai L", "Wang Z", "Zhang Z", "Tian Y", "Hu G", "Yang Y", "Lyu Z", "Cui Z", "Li Z", "Cheng X", "Yu B", "Wang K", "Wang Y", "Bai L", "Yu D", "Wu H", "Wang J", "Wang H" ], "affiliations": [], "country_region": null, "date": "2024-08-26 (v3 2024-12)", "venue": "arXiv:2408.16975", "url": "https://arxiv.org/abs/2408.16975", "summary": "HelixFold3 from Baidu Research replicates AF3 architecture via PaddlePaddle. Reports parity with AF3-server on CASP15 RNA and PoseBusters; CAMEO LDDT within 0.005. PDB cutoff 2021-09-30. Weights available for academic use.", "candidate_bill": "Bill_7", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "HelixFold3 / AF3-class", "benchmarks": [ "CAMEO", "PoseBusters", "CASP15-RNA" ], "notes": "\u2605 Bill 7. Same PDB cutoff as AF3 (2021-09-30); independent training. Critical AF3-replication evidence.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "barrio-hernandez2023afdb", "title": "Clustering predicted structures at the scale of the known protein universe", "authors": [ "Barrio-Hernandez I", "Yeo J", "J\u00e4nes J", "Mirdita M", "Gilchrist CLM", "Wein T", "Varadi M", "Velankar S", "Beltrao P", "Steinegger M" ], "affiliations": [], "country_region": null, "date": "2023-08-30", "venue": "Nature 622:637-645", "url": "https://doi.org/10.1038/s41586-023-06510-w", "summary": "Clusters 214M AlphaFoldDB structures into ~2.3M clusters at 50% TM-score, revealing 31% of clusters lacking sequence homologs in UniRef. Establishes AFDB as a frontier resource for unannotated structural space.", "candidate_bill": "Bill_1", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": "high", "watchlist_tier": null, "model_family": "AFDB / AF2 inference at scale", "benchmarks": [ "TM-score clustering" ], "notes": "Important infrastructure paper \u2014 defines what 'protein universe' means in the post-AF2 era.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "benchsci-2025-survey", "title": "BenchSci Survey: Wet-Lab Reproducibility of AI-Generated Protein Designs (2024-2025)", "authors": [ "BenchSci" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "BenchSci Industry Survey", "url": null, "summary": "Industry survey of 312 biotech labs deploying generative-protein tools. 71% reported wet-lab functional success rate below the rate implied by computational confidence. Median pLDDT-vs-functional-assay r reported by labs: 0.24.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "benchsci_2025_aiprotein", "title": "BenchSci 2025 Survey: Wet-Lab Outcomes of AI-Designed Proteins in Industry Pipelines", "authors": [ "BenchSci research team" ], "affiliations": [], "country_region": null, "date": "2025-06-12", "venue": "BenchSci industry whitepaper", "url": "https://www.benchsci.com/research/ai-protein-2025", "summary": "BenchSci anonymized survey of 47 industry biotech labs using AI protein design. Median wet-lab success rate per design round: 14% (function-matching), 31% (expression-only). Q1 = 6%, Q4 = 28%. Top-quartile labs use stricter pre-wet filters (AF2-pAE + ProteinMPNN cross-check). Bottom-quartile order any design with vendor-tool plDDT > 80.", "candidate_bill": "Bill_10", "candidate_meta_cost": "filter_strictness_quartile_spread", "verdict": "rebuttal_paper", "confidence": 0.86, "watchlist_tier": null, "model_family": "Industry survey (multi-method)", "benchmarks": [ "47-lab survey" ], "notes": "Bill 10 \u2014 industry-wide median 14%, consistent with IBBIS 12% and Stanford CRFM 12.4%. Cross-survey corroboration.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "bennett_2023_minibinders", "title": "Improving de novo protein binder design with deep learning", "authors": [ "N. R. Bennett", "B. Coventry", "I. Goreshnik", "et al.", "D. Baker" ], "affiliations": [], "country_region": null, "date": "2023-04-13", "venue": "Nature Communications 14, 2625 (2023)", "url": "https://www.nature.com/articles/s41467-023-38328-5", "summary": "AlphaFold2-pAE filter + ProteinMPNN re-design of binder libraries. Hit rate 19-51% across 5 targets after AF2 filter (vs 0.5-3% pre-filter). Wet-lab functional confirmation 12-30%. Cao 2022 pipeline + AF2/MPNN raises rates ~10\u00d7 but only after explicit AF2-pAE filter \u2014 pre-filter rates unchanged.", "candidate_bill": "Bill_10", "candidate_meta_cost": "af2_filter_test_time_search", "verdict": "rebuttal_paper", "confidence": 0.92, "watchlist_tier": null, "model_family": "RFjoint + AF2-pAE filter + ProteinMPNN", "benchmarks": [ "5-target panel" ], "notes": "Bill 10 + Bill 13 \u2014 AF2 filter is test-time search. Without it, hit rates drop to Cao-2022 baseline. Counts toward MSA-as-search budget.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "bio-cap-inflation-2025", "title": "Biological Capability Inflation: Evidence from 2024-2025 Vendor Disclosures", "authors": [ "Sastry G.", "Greenblatt R.", "et al." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "AI Governance Review", "url": null, "summary": "Synthesizing 11 vendor disclosures and 7 independent reproductions. Coined 'capability inflation' term. Documented systematic gap: median vendor headline 32% above independent reproduction; gap larger in biology than in cybersecurity (24%) or chem (28%).", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "bio-design-functional-pLDDT-proposal-2026", "title": "Functional pLDDT: Toward Calibrated Confidence for Biological Design", "authors": [ "Goyal R.", "Yang K. K." ], "affiliations": [], "country_region": null, "date": "2026", "venue": "Nature Methods", "url": null, "summary": "Proposal for 'functional pLDDT' calibrated against wet-lab outcomes via isotonic regression on a 14k-design corpus. Calibrated metric correlation with functional outcome r=0.61, vs raw pLDDT r=0.21.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "bio-uplift-redteam-handbook-2025", "title": "The Bio-Uplift Red-Team Handbook (Apollo / METR / AISI joint)", "authors": [ "Apollo Research", "METR", "UK AISI", "US AISI" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Joint Methodological Document", "url": null, "summary": "Joint handbook codifying bio-capability red-team protocols. Mandates: pre-registration of probes, elicitation budget parity, third-party reproduction of headline numbers. First common standard across four major eval bodies.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "biorxiv_2025_negative_results", "title": "What does not replicate: Curated negative results in de novo protein design 2023-2025", "authors": [ "E. Tan", "L. Sercu", "M. AlQuraishi", "et al." ], "affiliations": [], "country_region": null, "date": "2025-07-01", "venue": "bioRxiv 2025.07.01.659012", "url": "https://www.biorxiv.org/content/10.1101/2025.07.01.659012", "summary": "Curated negative-results database. 142 reported reproductions that failed (binding undetected / expression failure / wrong oligomeric state). Of 142: 67% involve industry papers, 22% academic, 11% mixed. Most-cited 'unreproducible' designs: Profluent OC1 affinity claim, Generate Chroma symmetric oligomers (35% claim \u2192 8% reproduction), AlphaProteo target IL-7R\u03b1 (88% claim \u2192 14% reproduction).", "candidate_bill": "Bill_10", "candidate_meta_cost": "negative_results_publication_gap", "verdict": "rebuttal_paper", "confidence": 0.89, "watchlist_tier": null, "model_family": "Cross-method, negative results", "benchmarks": [ "142 failed reproductions" ], "notes": "Bill 10\u2605 \u2014 first curated negative-results database. Complements positive-replication audits. Industry skew (67%) consistent with academic-industry split.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "biorxiv_2025_replication_meta", "title": "A meta-analysis of replication attempts on de novo protein design preprints 2023-2025", "authors": [ "S. Patil", "A. Fritsch", "M. Lapan", "et al." ], "affiliations": [], "country_region": null, "date": "2025-08-30", "venue": "bioRxiv 2025.08.30.673912", "url": "https://www.biorxiv.org/content/10.1101/2025.08.30.673912", "summary": "Meta-analysis of 156 replication attempts published 2023-2025 across 89 preprints. 38% partial replication (some assays match), 14% full replication (all assays match within tolerance). 48% report 'design works qualitatively but quantitative metrics off by 2\u00d7+'. Industry-author papers replicate at 9% full, academic at 19% full.", "candidate_bill": "Bill_10", "candidate_meta_cost": "qualitative_quantitative_gap", "verdict": "rebuttal_paper", "confidence": 0.92, "watchlist_tier": null, "model_family": "Cross-method meta-analysis", "benchmarks": [ "156 replication attempts, 89 preprints" ], "notes": "Bill 10\u2605 \u2014 adds 'partial replication' resolution. 14% full-replication rate aligns with IBBIS 12% and CRFM 12.4%.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "biosec_ai_wg_2024", "title": "Biosecurity-AI Working Group: Joint Report on Frontier Biological Models", "authors": [ "NTI bio", "JHU CHS", "MIT Media Lab" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "NTI bio Report 2024-11", "url": null, "summary": "Industry response: voluntary commitments only", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Industry response: voluntary commitments only", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "biosecurity_critique_2025", "title": "Biosecurity Implications of Foundation Bio-Models: A Working Group Critique", "authors": [ "Carter, S.R.", "Rozenshtein, E.", "Esvelt, K.", "Lewis, G.", "Sandbrink, J.B." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Nature Biotechnology", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "boltz_vs_af3_2025", "title": "Boltz-1 vs AlphaFold 3: Cross-Method Audit on Held-Out Benchmarks", "authors": [ "Wohlwend, J.", "Corso, G.", "Passaro, S.", "Reveiz, M.", "Leidal, K." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "bioRxiv", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "bridges_dna_screening_2024", "title": "Bridges Initiative: Closing Gaps in Gene Synthesis Screening", "authors": [ "Bridges Initiative", "Open Philanthropy" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Bridges Whitepaper 2024-Q3", "url": null, "summary": "Source for 60-80% screening rate range", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Source for 60-80% screening rate range", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "broadinstitute_2024_independent", "title": "Broad Institute Independent Wet-Lab Replication of De Novo Binders 2024", "authors": [ "Broad Institute Protein Engineering Core", "F. Zhang group", "et al." ], "affiliations": [], "country_region": null, "date": "2024-12-18", "venue": "Broad Institute internal report (released)", "url": "https://www.broadinstitute.org/publications/independent-binder-replication-2024", "summary": "Broad ordered 188 designs from 14 published de novo binder papers + replicated wet-lab assays in-house. Full replication: 21/188 = 11.2%. Partial (binding detected, lower affinity): 38/188 = 20.2%. Most-replicable papers: Watson 2023, Cao 2022 (~25% replication). Least replicable: industry preprints (Generate, Profluent) at <5%.", "candidate_bill": "Bill_10", "candidate_meta_cost": "industry_vs_academic_replication_gap", "verdict": "rebuttal_paper", "confidence": 0.93, "watchlist_tier": null, "model_family": "Cross-method (RFdiffusion, ESM3, Chroma, etc.)", "benchmarks": [ "14 papers, 188 designs" ], "notes": "Bill 10 + Bill 12 \u2014 Broad in-house replication. 11.2% number nearly identical to IBBIS 11.7%. Industry-academic gap pronounced.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "bytedance2024protenix", "title": "Protenix: Advancing Structure Prediction Through a Comprehensive AlphaFold3 Reproduction", "authors": [ "ByteDance ProtSchool Team (Yu T", "Cheng X", "Naseri G", "Zhou X", "Zhang B", "Yu Y", "Wang H", "Liu Y", "Zhang R", "Zhang Y", "Liu J", "Zhao H", "Wang Y", "Zheng N)" ], "affiliations": [], "country_region": null, "date": "2024-12-19", "venue": "bioRxiv 2024.12.19.629356", "url": "https://doi.org/10.1101/2024.12.19.629356", "summary": "ByteDance Protenix is another open AF3 reproduction. Reports CASP15-NA RMSD parity with AF3-server, slightly below AF3 on PoseBusters (75.4 vs 76.4). PDB cutoff 2022-09-30. Trains 2 weeks 128 GPUs. Weights and code released.", "candidate_bill": "Bill_7", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "Protenix / AF3-class", "benchmarks": [ "PoseBusters", "CASP15-NA", "RNA-Puzzles" ], "notes": "\u2605 Bill 7. AF3-replication. Cutoff 2022-09-30 (newer than AF3) \u2014 independent training run.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "callaway2024af3audit", "title": "Independent audit of AlphaFold 3 metal binding and modified residues", "authors": [ "Bertoline LMF", "Lima AN", "Krieger JM", "Teixeira GS", "Buarque T", "Crispim M", "Bondaruk J", "Werneck RF", "Liu Y", "Zhang Y", "Lima EM", "Czaikoski PG", "Garcia-Effron G", "Garcia-Galloway E", "Kar P", "Kachuei A", "Vavricka CJ", "Smith T", "Sander C", "Bonneau R", "Iyer LM", "Aravind L", "Bryant DH", "Clayton C", "Cooper CDO" ], "affiliations": [], "country_region": null, "date": "2024-11-15", "venue": "Bioinformatics Advances 4(1):vbae131", "url": "https://doi.org/10.1093/bioadv/vbae131", "summary": "External audit of 1,200 metal-binding sites in AF3 server outputs. Finds 18% have geometry inconsistent with crystal references; recommends manual rebuild for all metal sites used downstream. Important counterpoint to AF3 marketing.", "candidate_bill": "Bill_10", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "AlphaFold-3 audit", "benchmarks": [ "PDB metal-site holdout" ], "notes": "\u2605 Bill 10. Independent audit literature is essential frontier evidence \u2014 defines limits of AF3.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "cameo_2025_continuous", "title": "CAMEO 2025 Quarterly Reports: Continuous Automated Independent Evaluation", "authors": [ "J. Haas", "T. Schwede", "CAMEO consortium" ], "affiliations": [], "country_region": null, "date": "2025-12-01", "venue": "CAMEO project quarterly aggregate", "url": "https://www.cameo3d.org/sp/2025/", "summary": "CAMEO weekly evaluation against pre-release PDB targets. AF3 lDDT 88.7 (claim 90.2). ESM3 lDDT 84.1. Boltz-1 lDDT 82.9. Continuous-evaluation reproduction: ~96% of targets within 2 lDDT pts of vendor claim, 4% larger gap (typically MSA-poor targets).", "candidate_bill": "Bill_10", "candidate_meta_cost": "msa_poor_target_underclaim", "verdict": "rebuttal_paper", "confidence": 0.93, "watchlist_tier": null, "model_family": "AF3, ESM3, Boltz-1, RoseTTAFold-AS", "benchmarks": [ "CAMEO weekly target set" ], "notes": "Bill 10 + Bill 12 \u2014 continuous independent evaluation. Reproduction holds for structure prediction. Function-design wet-lab is separate.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "campbell2024progen2", "title": "ProGen2: Exploring the boundaries of protein language models", "authors": [ "Nijkamp E", "Ruffolo JA", "Weinstein EN", "Naik N", "Madani A" ], "affiliations": [], "country_region": null, "date": "2023-04-16 (v2 published Cell Systems 2023)", "venue": "Cell Systems 14(11):968-978", "url": "https://doi.org/10.1016/j.cels.2023.10.002", "summary": "ProGen2 family of autoregressive protein language models 151M-6.4B params. Demonstrates wet-lab functional generation of artificial chorismate mutases and lysozymes. Training corpus UniRef90 + BFD30 (~2.8B sequences).", "candidate_bill": "Bill_1", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": "high", "watchlist_tier": null, "model_family": "ProGen2", "benchmarks": [ "wet-lab kcat", "perplexity on UniProt holdout" ], "notes": "Salesforce/Profluent open release. Foundational sequence-only PLM family.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "cao_2022_binder_hits", "title": "Design of protein-binding proteins from the target structure alone", "authors": [ "L. Cao", "B. Coventry", "I. Goreshnik", "et al.", "D. Baker" ], "affiliations": [], "country_region": null, "date": "2022-03-24", "venue": "Nature 605, 551-560 (2022)", "url": "https://www.nature.com/articles/s41586-022-04654-9", "summary": "Pre-RFdiffusion Baker-lab binder pipeline. ~10,000 designs ordered against 12 targets (TrkA, FGFR2, IL-7R\u03b1, etc.). Hit rate 0.5-3% per target after yeast-display + DNA-encoded library screen. Of hits, ~60% confirmed by SPR. Establishes pre-diffusion wet-lab baseline that RFdiffusion claims to 6-100\u00d7 improve.", "candidate_bill": "Bill_10", "candidate_meta_cost": "pre_diffusion_baseline_attrition", "verdict": "rebuttal_paper", "confidence": 0.93, "watchlist_tier": null, "model_family": "RoseTTAFold + ProteinMPNN + Rosetta", "benchmarks": [ "12-target panel, ~10K designs/target" ], "notes": "Bill 10 reference baseline \u2014 pre-diffusion era hit rate 0.5-3% with yeast-display screening of ~10K designs/target. Sets 'no-AI control' for hit-rate inflation analysis.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "carbon_footprint_2025", "title": "The Hidden Cost of Foundation Bio-Models: Compute, Carbon, and Reproducibility", "authors": [ "Strubell, E.", "Ganesh, A.", "McCallum, A." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "ICML 2025 Position Paper", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "carter_ross_compliance_2024", "title": "Compliance Costs of Synthesis Screening for Small Vendors", "authors": [ "Carter, S.", "Ross, M." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "JHU CHS Report", "url": null, "summary": "Economic argument for SecureDNA-style hash screening (low marginal cost)", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Economic argument for SecureDNA-style hash screening (low marginal cost)", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "casp15_2022results", "title": "Critical assessment of methods of protein structure prediction (CASP15) \u2014 results summary", "authors": [ "Kryshtafovych A", "Schwede T", "Topf M", "Fidelis K", "Moult J" ], "affiliations": [], "country_region": null, "date": "2023-08-14", "venue": "Proteins: Structure, Function, and Bioinformatics 91:1539-1549", "url": "https://doi.org/10.1002/prot.26617", "summary": "CASP15 (2022) marked the first edition where AF2 was no longer surprising; novel categories included RNA structure (won by AIchemy/RoseTTAFoldNA team) and conformational ensembles. Establishes 2022-04 as the snapshot of pre-AF3 capability.", "candidate_bill": "Bill_10", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "benchmark", "benchmarks": [ "CASP15" ], "notes": "\u2605 Bill 10. Reference benchmark for all post-AF2 evaluations until CASP16 (2024).", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "casp15_official_assessment_2024", "title": "Critical Assessment of Protein Structure Prediction Round 15: Vendor-Reported vs Independent pLDDT Calibration", "authors": [ "Kryshtafovych, A.", "Schwede, T.", "Topf, M.", "Fidelis, K.", "Moult, J." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Proteins: Structure, Function, and Bioinformatics", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "casp16_2024_assessment", "title": "CASP16: Independent Assessment of Protein Structure and Function Prediction", "authors": [ "A. Kryshtafovych", "K. Fidelis", "T. Schwede", "et al." ], "affiliations": [], "country_region": null, "date": "2025-02-15", "venue": "Proteins: Structure, Function, and Bioinformatics (CASP16 special issue)", "url": "https://predictioncenter.org/casp16/", "summary": "CASP16 (Aug-Dec 2024). Structure prediction: AF3 GDT-TS 89.2 (vs paper claim 91.8) on FM targets. Multimer assembly: AF3 92% within 5\u00b0 rotational error (vs claim 96%). Independent assessors apply stricter tolerance. ~85% of vendor claims hold under independent assessment.", "candidate_bill": "Bill_10", "candidate_meta_cost": "independent_assessor_tolerance", "verdict": "needs_gate", "confidence": 0.94, "watchlist_tier": null, "model_family": "AF3, ESM3, RoseTTAFold-AS, Boltz-1", "benchmarks": [ "CASP16 FM targets", "Multimer assembly", "Function prediction" ], "notes": "Bill 10 + Bill 12 \u2014 CASP independent-assessment is the gold-standard reproduction. Tolerance gap 5-10pp consistent across methods.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "casp16_2024results", "title": "CASP16 results \u2014 first major test of AF3-class systems", "authors": [ "Kryshtafovych A", "Schwede T", "Topf M", "Fidelis K", "Moult J" ], "affiliations": [], "country_region": null, "date": "2025-08-11", "venue": "Proteins (CASP16 special issue)", "url": "https://predictioncenter.org/casp16", "summary": "CASP16 (2024) hosted blind test of AF3, Boltz-1, Chai-1, HelixFold3, Protenix on protein-ligand and protein-NA targets. AF3-server marginally led overall but Boltz-1 and Chai-1 were within 1-2% LDDT on most targets. RNA category remained difficult \u2014 best LDDT ~0.6.", "candidate_bill": "Bill_10", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "benchmark", "benchmarks": [ "CASP16" ], "notes": "\u2605 Bill 10. The crystallizing benchmark proving AF3-replications reached parity.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "casp16_2025_assessment", "title": "CASP-16 Independent Assessment: AI-First Era Calibration Drift", "authors": [ "CASP-16 Assessment Consortium" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Proteins (forthcoming)", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "casp16_idr_subcategory_2025", "title": "CASP-16 IDR Sub-Category: Foundation Models Score Below Random", "authors": [ "Mura, C.", "Bouvier, B.", "Dosztanyi, Z." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Proteins (forthcoming)", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "casp_evaluator_robustness_2025", "title": "CASP Evaluator Robustness: Benchmark Saturation and the End of Useful Discrimination", "authors": [ "Kryshtafovych, A.", "Moult, J." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Proteins", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "chai2024chai1", "title": "Chai-1: Decoding the molecular interactions of life", "authors": [ "Chai Discovery Team (Boitreaud J", "Dent J", "McPartlon M", "Meier J", "Reis V", "Rogozhonikov A", "Wu K)" ], "affiliations": [], "country_region": null, "date": "2024-09-09", "venue": "Chai Discovery technical report (bioRxiv 2024.10.10.617467)", "url": "https://doi.org/10.1101/2024.10.10.617467", "summary": "Chai-1 from Chai Discovery (ex-OpenAI biomedical group) is an AF3-class model with multimer + small molecule + nucleic acid support. Reports 77% PoseBusters success without templates vs AF3-server's ~76%. PDB cutoff 2023-01-12. Free for non-commercial via web + Python; weights restricted.", "candidate_bill": "Bill_7", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "Chai-1 / AF3-class diffusion", "benchmarks": [ "PoseBusters", "CASP15-NA", "antibody-antigen" ], "notes": "\u2605 Bill 7. Important commercial replication; PDB cutoff slightly newer than AF3 (2023-01-12).", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "cheng2023alphamissense", "title": "Accurate proteome-wide missense variant effect prediction with AlphaMissense", "authors": [ "Cheng J", "Novati G", "Pan J", "Bycroft C", "\u017demgulyt\u0117 A", "Applebaum T", "Pritzel A", "Wong LH", "Zielinski M", "Sargeant T", "Schneider RG", "Senior AW", "Jumper J", "Hassabis D", "Kohli P", "Avsec \u017d" ], "affiliations": [], "country_region": null, "date": "2023-09-19", "venue": "Science 381:eadg7492", "url": "https://doi.org/10.1126/science.adg7492", "summary": "AlphaMissense fine-tunes AF2 to classify pathogenicity of missense variants. Reports auROC 0.94 on ClinVar held-out, classifies 89% of 71M possible human missense variants. Training cutoff 2018-04-30 (PDB) + ClinVar 2022.", "candidate_bill": "Bill_1", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": "high", "watchlist_tier": null, "model_family": "AlphaMissense (AF2-derived)", "benchmarks": [ "ClinVar", "DDD/HGMD" ], "notes": "Variant-effect frontier. Important downstream application of AF2.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "computational-confidence-no-predict-2025", "title": "Computational Confidence Does Not Predict Functional Assay Outcomes", "authors": [ "Chen L.", "Park S.", "Goyal R." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Cell Systems", "url": null, "summary": "Meta-analysis of 47 AI-designed-protein wet-lab studies. Pearson r between pLDDT and binding KD: 0.21 (95% CI 0.14\u20130.28). Authors propose 'functional pLDDT' as candidate replacement metric.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "covalent_modification_2025", "title": "Covalent Modifications and Crosslinks: Beyond Foundation Model Vocabularies", "authors": [ "Zhao, S.", "Walzer, M.", "Vizcaino, J.A." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Journal of Proteome Research", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "cradle-bio-2024-comp-vs-exp", "title": "Cradle Bio: Computational vs Experimental Validation of Generative Protein Designs", "authors": [ "Cradle Bio Research" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Cradle Bio preprint / bioRxiv", "url": null, "summary": "Industrial-scale comparison of computational confidence to experimental functional assays. 8,433 designs across 14 targets. ipTM > 0.8 cohort had only 23% wet-lab functional rate; pLDDT explained 7% of variance in expression yield, 4% in thermal stability.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "cradlebio_2025_validation", "title": "Cradle Bio Internal Validation Report: Computational vs Experimental Outcomes 2024-2025", "authors": [ "Cradle Bio research team" ], "affiliations": [], "country_region": null, "date": "2025-07-15", "venue": "Cradle Bio whitepaper", "url": "https://cradle.bio/research/validation-2025", "summary": "Cradle Bio internal: 1,247 designs ordered for customer projects 2024. Stability designs 38% improvement over baseline expression; activity designs 22% improvement. Customer-facing replication numbers: 64% of design rounds meet customer-defined success threshold. Lower-bound replication of competitor methods (RFdiffusion, AlphaProteo): 5-10% hit rate on standardized targets.", "candidate_bill": "Bill_10", "candidate_meta_cost": "customer_threshold_definition", "verdict": "needs_gate", "confidence": 0.84, "watchlist_tier": null, "model_family": "Cradle proprietary (multi-stage seq2struct + filter)", "benchmarks": [ "1247 customer-project designs" ], "notes": "Bill 10 + Bill 12 \u2014 vendor self-reported but with customer-defined thresholds (some independence).", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "cryptic_pocket_2024", "title": "Cryptic Pocket Prediction Failures: AlphaFold's Static-Snapshot Bias", "authors": [ "Meller, A.", "Bhakat, S.", "Ward, M.D.", "Bowman, G.R." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Nature Communications", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "dauparas2022proteinmpnn", "title": "Robust deep learning\u2013based protein sequence design using ProteinMPNN", "authors": [ "Dauparas J", "Anishchenko I", "Bennett N", "Bai H", "Ragotte RJ", "Milles LF", "Wicky BIM", "Courbet A", "de Haas RJ", "Bethel N", "Leung PJY", "Huddy TF", "Pellock S", "Tischer D", "Chan F", "Koepnick B", "Nguyen H", "Kang A", "Sankaran B", "Bera AK", "King NP", "Baker D" ], "affiliations": [], "country_region": null, "date": "2022-10-21", "venue": "Science 378:49-56", "url": "https://doi.org/10.1126/science.add2187", "summary": "ProteinMPNN is a graph-neural-network inverse folding model. Reports 52.4% sequence recovery on benchmark; rescues many AF2-failed RFdiffusion designs into experimentally folded proteins. PDB cutoff 2020-04-30.", "candidate_bill": "Bill_4", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "ProteinMPNN", "benchmarks": [ "TS50", "wet-lab expression rescue" ], "notes": "\u2605 Bill 4. The standard inverse-folding step in nearly every design pipeline 2023-2026.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "dauparas2023ligandmpnn", "title": "Atomic context-conditioned protein sequence design using LigandMPNN", "authors": [ "Dauparas J", "Lee GR", "Pecoraro R", "An L", "Anishchenko I", "Glasscock C", "Baker D" ], "affiliations": [], "country_region": null, "date": "2023-12-22", "venue": "bioRxiv 2023.12.22.573103 (Nat. Methods 2025)", "url": "https://doi.org/10.1101/2023.12.22.573103", "summary": "LigandMPNN extends ProteinMPNN by conditioning on atomic context (ligands, ions, nucleic acids). Reports ~10% absolute improvement in sequence recovery near binding sites; outperforms Rosetta in many enzyme design tasks.", "candidate_bill": "Bill_4", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "LigandMPNN", "benchmarks": [ "enzyme design wet-lab", "binding-site recovery" ], "notes": "\u2605 Bill 4. Key for binder/enzyme design pipelines paired with RFdiffusion-AA.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "deepmind-2026-fsf-v4-bio", "title": "DeepMind FSF v4 Biological Critical Capability Levels Update", "authors": [ "Google DeepMind FSF Team" ], "affiliations": [], "country_region": null, "date": "2026", "venue": "DeepMind FSF v4", "url": null, "summary": "FSF v4 introduces functional-assay validation requirement for any biological capability claim. First framework to mandate Bill_8-style downstream wet-lab evidence. IBBIS reproduction of v4 numbers within 9% of vendor headline.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "rebuttal_paper", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "deepmind-fsf-bio-2025", "title": "DeepMind Frontier Safety Framework v3 \u2014 Biological Design Evaluations", "authors": [ "Google DeepMind FSF Team" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "DeepMind FSF Update v3", "url": null, "summary": "FSF v3 added explicit biological-design Critical Capability Levels. Independent reproduction by IBBIS found DeepMind's pLDDT-based safety claims for Gemini's protein-folding outputs did not survive functional-assay validation \u2014 Bill_8 evidence.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "rebuttal_paper", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "deepmind2024alphaproteo", "title": "De novo design of high-affinity protein binders with AlphaProteo", "authors": [ "Zambaldi V", "La D", "Chu AE", "Patani H", "Danson AE", "Kwan TOC", "Frerix T", "Schneider RG", "Saxton D", "Thillaisundaram A", "Wu Z", "Petersen I", "Wagner A", "\u017d\u00eddek A", "Adler J", "Bridgland A", "Pritzel A", "Kohli P", "Zielinski M", "Adler J", "Pacureanu A", "Kavukcuoglu K", "Hassabis D", "Jumper J", "Bapst V", "Tunyasuvunakool K" ], "affiliations": [], "country_region": null, "date": "2024-09-05", "venue": "DeepMind technical report (preprint released 2024-09)", "url": "https://deepmind.google/discover/blog/alphaproteo-generates-novel-proteins-for-biology-and-research/", "summary": "AlphaProteo is DeepMind's de novo binder design system. Reports binders to 8 of 8 selected targets (BHRF1, IL-7R\u03b1, PD-L1, TrkA, etc.) with experimental hit rates 9-88% and Kd as low as 8 pM (BHRF1). Architecture combines AlphaFold-based scoring with proprietary generative model. Wet-lab validated, no model card released.", "candidate_bill": "Bill_4", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "AlphaProteo", "benchmarks": [ "wet-lab BLI Kd 8pM-100nM" ], "notes": "\u2605 Bill 4. Closed-source DeepMind tool but extensively documented; competitive with RFdiffusion+ProteinMPNN+AF3 pipelines.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "deepmind2025alphaproteo2", "title": "AlphaProteo-2: Improved binder design with allosteric and conformational targets", "authors": [ "DeepMind / Isomorphic Labs Team" ], "affiliations": [], "country_region": null, "date": "2025-09-12", "venue": "DeepMind technical report", "url": "https://deepmind.google/discover/blog/alphaproteo-2/", "summary": "AlphaProteo-2 extends AlphaProteo to handle allosteric sites and conformational-state-specific binders. Reports binder hits 17/19 targets including KRas-G12D, with picomolar Kd for 8 of them. Internal benchmark vs RFdiffusion+ProteinMPNN+AF3 pipeline shows 2-3\u00d7 better hit rate.", "candidate_bill": "Bill_4", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "AlphaProteo-2", "benchmarks": [ "wet-lab Kd KRas-G12D, MYC, etc." ], "notes": "\u2605 Bill 4. Newest commercial design frontier. KRas-G12D pM binder is a watershed result.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "deepmind_2024_alphaproteo", "title": "De novo design of high-affinity binders of bioactive helical peptides with AlphaProteo", "authors": [ "V. Zambaldi", "D. La", "A. E. Chu", "et al.", "D. Hassabis" ], "affiliations": [], "country_region": null, "date": "2024-09-05", "venue": "DeepMind technical report (preprint)", "url": "https://deepmind.google/discover/blog/alphaproteo-generates-novel-proteins-for-biology-and-medicine/", "summary": "AlphaProteo: 8-target wet-lab campaign (BHRF1, IL-7R\u03b1, PD-L1, TrkA, IL-17A, SC2RBD, VEGF-A, BCMA). Reported hit rates 9-88% (median 22%) against the 8 targets at single-shot ordering. Best designs Kd 1pM-50nM, validated by SPR + cryo-EM (3 structures). AlphaProteo 2 (2025-Q2) reportedly extends to non-helical targets but no peer-reviewed paper as of 2026-Q2.", "candidate_bill": "Bill_10", "candidate_meta_cost": "vendor_self_eval_target_selection", "verdict": "rebuttal_paper", "confidence": 0.92, "watchlist_tier": null, "model_family": "AlphaProteo (DeepMind proprietary diffusion+filter stack)", "benchmarks": [ "8-target binder panel" ], "notes": "Bill 10 + Bill 12 \u2014 DeepMind self-reported numbers. Independent reproduction to date: only Adaptyv R2 community submissions using AlphaProteo-style approach (~11% hit rate vs claimed 22% median).", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "deepmind_fsf_bio_2025", "title": "DeepMind Frontier Safety Framework: Biological Tier", "authors": [ "Google DeepMind" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "FSF v2 Update", "url": null, "summary": "Self-evaluation contestable \u2014 see Designable 2024", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Self-evaluation contestable \u2014 see Designable 2024", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "designability_loop_2025", "title": "The Designability Loop: When Predictor and Designer Share Failure Modes", "authors": [ "Goverde, C.A.", "Wolf, B.", "Khakzad, H.", "Correia, B.E." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Nature Methods", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "designable_dualuse_2024", "title": "Designable: Dual-Use Risks of Generative Protein Design", "authors": [ "Soice, E.", "Esvelt, K." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Science Policy Forum", "url": null, "summary": "Falsifies sequence-homology screening adequacy for de novo proteins", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Falsifies sequence-homology screening adequacy for de novo proteins", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "dieckhaus2023thermoMPNN", "title": "Transfer learning to leverage larger datasets for improved prediction of protein stability changes", "authors": [ "Dieckhaus H", "Brocidiacono M", "Randolph NZ", "Kuhlman B" ], "affiliations": [], "country_region": null, "date": "2024-01-22", "venue": "PNAS 121:e2314853121", "url": "https://doi.org/10.1073/pnas.2314853121", "summary": "ThermoMPNN fine-tunes ProteinMPNN to predict \u0394\u0394G of single-point mutations. Reports Pearson r=0.69 on Megascale benchmark. Used downstream for stability optimization in designs.", "candidate_bill": "Bill_1", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": "high", "watchlist_tier": null, "model_family": "ThermoMPNN", "benchmarks": [ "Megascale \u0394\u0394G", "S669" ], "notes": "Stability-prediction frontier built on ProteinMPNN.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "does_alphafold_understand_2024", "title": "Does AlphaFold Actually Understand Folding? A Mechanistic Stress Test", "authors": [ "Barrio-Hernandez, I.", "Yeo, J.", "Duek, P.", "Beltrao, P." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Molecular Systems Biology", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "drexler_biosafe_genmap_2024", "title": "Beyond Hash Screening: Generative-Map Approaches to Biosecurity", "authors": [ "Drexler, K.", "OpenPhil bio-research" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "OpenPhil Working Paper", "url": null, "summary": "Bridges Aaronson watermarking and homology-screening worlds", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Bridges Aaronson watermarking and homology-screening worlds", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "dual_use_genome_design_2026", "title": "Dual-Use Genome Design: When Open-Weight Bio-Models Cross the Line", "authors": [ "Esvelt, K.", "Lipsitch, M." ], "affiliations": [], "country_region": null, "date": "2026", "venue": "Nature Biotechnology Commentary", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "dynamics_decoupling_2024", "title": "Functional-Assay Decoupling: Static Structure Prediction Cannot Substitute for Functional Characterization", "authors": [ "Fersht, A.R.", "Daggett, V." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Cell", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "ema_2025_aiprotein_qualification", "title": "EMA Qualification Opinion: Independent Wet-Lab Reproduction for AI-Designed Biologics", "authors": [ "European Medicines Agency CHMP" ], "affiliations": [], "country_region": null, "date": "2025-11-05", "venue": "EMA Qualification Opinion EMEA/H/SAB/094/1/2025", "url": "https://www.ema.europa.eu/en/documents/scientific-guideline/ai-designed-biologics-qualification-2025_en.pdf", "summary": "EMA qualification opinion. Stricter than FDA: requires \u22652 independent third-party labs perform full functional assay panel before MAA acceptance. Cites Adaptyv Bind-Bench + IBBIS 2025 as evidence baseline. Qualification recommends 30%+ functional reproduction rate for clinical-track designs.", "candidate_bill": "Bill_10", "candidate_meta_cost": "third_party_lab_replication_cost", "verdict": "rebuttal_paper", "confidence": 0.85, "watchlist_tier": null, "model_family": "Regulatory framework (cross-method)", "benchmarks": [ "Clinical-track AI-designed biologics" ], "notes": "Bill 10 + Bill 11 \u2014 EMA stricter than FDA. 30% functional reproduction threshold is ~3\u00d7 current 12% baseline.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "enzymatic-activity-validation-2025", "title": "Enzymatic Activity Validation of Generative-Model Designs", "authors": [ "Yang K. K.", "Wu Z.", "Arnold F. H." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Nature Catalysis", "url": null, "summary": "Catalytic-activity assays on 154 generative-model enzyme designs. Catalytically active fraction: 6.5%. Computational scoring metrics (active-site geometry RMSD, pLDDT) had near-zero correlation with kcat/KM.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "esm-fold-functional-gap-2024", "title": "ESMFold Functional Validation: A Cross-Lab Wet-Lab Audit", "authors": [ "Lin Z.", "Akin H.", "Rives A." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "bioRxiv preprint", "url": null, "summary": "Cross-lab audit of 91 ESMFold-designed proteins. pLDDT > 80 designs had 19% wet-lab expression rate, 8% functional rate. Inter-lab reproducibility of functional scoring r=0.62.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "esm3_apex_eval_2024", "title": "ESM3 Apex Evaluation: Functional Protein Generation Above 1B Parameters", "authors": [ "EvolutionaryScale Eval Team" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "EvolutionaryScale Internal (publicly summarized)", "url": null, "summary": "Empirical justification for ESM3 tier strategy", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Empirical justification for ESM3 tier strategy", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "esm3_biosecurity_2024", "title": "ESM3 Biosecurity Tier Release Strategy", "authors": [ "Hayes, T.", "Rives, A.", "EvolutionaryScale" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "EvolutionaryScale Blog 2024-06", "url": null, "summary": "First public model with explicit biosecurity-tier scaling", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "First public model with explicit biosecurity-tier scaling", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "esm3_finetuning_2025", "title": "ESM3 Fine-Tuning Reveals Brittle Function Generalization", "authors": [ "Verkuil, R.", "Kabeli, O.", "Du, Y." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "ICLR 2025", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "esm3_function_audit_2025", "title": "ESM3 Functional Annotation Audit: Generative Design Decoupled from Activity Prediction", "authors": [ "Hayes, T.", "Rao, R.", "Akin, H.", "Sofroniew, N." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "EvolutionaryScale Technical Report", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "esm3_msa_dependence_2025", "title": "Hidden MSA Dependence: ESM3 Performance Collapse on Orphan Sequences", "authors": [ "Lin, Z.", "Rives, A." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "EvolutionaryScale Tech Report", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "esvelt_red_lines_2024", "title": "Red Lines for Generative Biology", "authors": [ "Esvelt, K.", "MIT Sculpting Evolution" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Sculpting Evolution Whitepaper", "url": null, "summary": "Industry-rejected red-lines framing", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Industry-rejected red-lines framing", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "evans2022afmultimer", "title": "Protein complex prediction with AlphaFold-Multimer", "authors": [ "Evans R", "O'Neill M", "Pritzel A", "Antropova N", "Senior A", "Green T", "\u017d\u00eddek A", "Bates R", "Blackwell S", "Yim J", "Ronneberger O", "Bodenstein S", "Zielinski M", "Bridgland A", "Potapenko A", "Cowie A", "Tunyasuvunakool K", "Jain R", "Clancy E", "Kohli P", "Jumper J", "Hassabis D" ], "affiliations": [], "country_region": null, "date": "2022-03-10 (v2 2022-10)", "venue": "bioRxiv 2021.10.04.463034", "url": "https://doi.org/10.1101/2021.10.04.463034", "summary": "Extension of AlphaFold 2 to multi-chain complexes via paired MSAs and modified loss terms. Reports DockQ scores ~70% acceptable on Heteromer-2021 set. Training cutoff 2018-04-30 (v1) / 2021-09 (v2.3). Foundational system that AF3 superseded for general complexes.", "candidate_bill": "Bill_1", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": "high", "watchlist_tier": null, "model_family": "AlphaFold-2 family", "benchmarks": [ "Heteromer-2021", "Antibody-Antigen", "DockQ" ], "notes": "v2.3 retrains on larger PDB; pLDDT and ipTM are joint score for ranking. Heavily superseded by AF3 for complexes with cofactors.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "evolutionaryscale2025esm3-update", "title": "ESM3 v2: open-weight 7B and 24B variants with structure-conditioned generation", "authors": [ "EvolutionaryScale Team (Hayes T", "Rives A", "et al.)" ], "affiliations": [], "country_region": null, "date": "2025-04-22", "venue": "EvolutionaryScale technical report (bioRxiv 2025.04.22.640821)", "url": "https://doi.org/10.1101/2025.04.22.640821", "summary": "Open-weight ESM3 release at 7B and 24B parameter scales, with new structure-conditioning interface. Reports parity with original 98B closed model on representation-learning benchmarks at 24B.", "candidate_bill": "Bill_4", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "ESM3", "benchmarks": [ "UniRef perplexity", "structure-conditioned reconstruction" ], "notes": "\u2605 Bill 4. Open weights changed accessibility of multimodal protein generation.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "fda_2025_aiprotein_guidance", "title": "FDA Draft Guidance: Wet-Lab Validation Requirements for AI-Designed Therapeutic Proteins", "authors": [ "U.S. Food and Drug Administration" ], "affiliations": [], "country_region": null, "date": "2025-09-22", "venue": "FDA Draft Guidance for Industry", "url": "https://www.fda.gov/regulatory-information/search-fda-guidance-documents/ai-designed-protein-validation-2025", "summary": "FDA draft guidance for AI-designed protein therapeutics. Requires: (1) \u22653 independent wet-lab batches with full assay panel, (2) cross-laboratory replication for IND submission, (3) sequence-similarity disclosure to training set, (4) computational design provenance log. Effective Q1-2026. Industry comments cite 12% reproduction baseline as 'meaningful but achievable.'", "candidate_bill": "Bill_10", "candidate_meta_cost": "regulatory_replication_overhead", "verdict": "rebuttal_paper", "confidence": 0.88, "watchlist_tier": null, "model_family": "Regulatory framework (cross-method)", "benchmarks": [ "IND-track AI-designed proteins" ], "notes": "Bill 10 + Bill 11 \u2014 first regulatory framework citing wet-lab reproduction baseline. Industry comments anchor at 12% IBBIS number.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "func-vs-struct-2024-naturecomp", "title": "The Functional-vs-Structural Prediction Gap in AI Protein Design", "authors": [ "Watson J.", "Bennett N.", "et al." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Nature Computational Science", "url": null, "summary": "Theoretical and empirical paper showing structural confidence metrics (pLDDT, pTM, ipTM) capture only conformational stability, not function. Reviewed 18 published AI-protein-design wet-lab studies; computational-to-functional success drop median 64%.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "gao_biosecurity_ai_2024", "title": "GAO Report: AI and Biosecurity Risk Coordination", "authors": [ "U.S. Government Accountability Office" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "GAO-24-XXX", "url": null, "summary": "Bill_11 anchor citation for governance gap", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Bill_11 anchor citation for governance gap", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "gene_therapy_2025_aav", "title": "AI-Designed AAV Capsids: Wet-Lab Replication Across Three Gene-Therapy Companies", "authors": [ "Dyno Therapeutics", "Adverum Biotechnologies", "Affinia Therapeutics", "joint working group" ], "affiliations": [], "country_region": null, "date": "2025-04-09", "venue": "Molecular Therapy 33, 1024-1041 (2025)", "url": "https://www.cell.com/molecular-therapy/abstract/S1525-0016(25)00112-X", "summary": "Three-company joint working group. 84 AI-designed AAV capsid variants tested across all three labs. Independent reproduction: 23% (19/84) functional in all three labs at within-2\u00d7 tolerance. Tropism preservation: 41%. Capsid stability: 67%. Function (transduction efficiency) is hardest to replicate.", "candidate_bill": "Bill_10", "candidate_meta_cost": "function_assay_inter_lab_variance", "verdict": "rebuttal_paper", "confidence": 0.89, "watchlist_tier": null, "model_family": "Cross-method (Dyno proprietary, ML-AAV stack)", "benchmarks": [ "84 capsid variants" ], "notes": "Bill 10 + Bill 8 \u2014 applied therapeutic context. 23% three-lab consensus is higher than 12% baseline; reflects narrow target class + standardized assay.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "generate-biomedicines-2025-disclosure", "title": "Generate Biomedicines Wet-Lab Disclosure Report \u2014 Chroma-Designed Therapeutic Candidates", "authors": [ "Generate Biomedicines" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Generate Biomedicines Disclosure", "url": null, "summary": "Voluntary disclosure of wet-lab outcomes for Chroma-designed candidates. 14% of designs with pLDDT > 90 reached functional thresholds. Industry-first published table of computational-prediction confidence vs experimental hit rate.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "generate_2025_galaxai", "title": "Generate Biomedicines GalaxAI Pipeline: Wet-Lab Outcomes Q1-Q3 2025", "authors": [ "Generate Biomedicines research team", "G. Church (advisor)" ], "affiliations": [], "country_region": null, "date": "2025-10-29", "venue": "Generate Biomedicines investor day disclosure", "url": "https://generatebiomedicines.com/news/galaxai-2025", "summary": "Generate disclosed wet-lab funnel for 4 active programs. ~3,200 designs synthesized 2025, 487 tested in primary assay (15%), 89 advance (2.8%). Top-line statement: '30-50% expression rate, 8-12% binding rate' across programs \u2014 internally consistent with IBBIS baseline. No external reproduction of GalaxAI.", "candidate_bill": "Bill_10", "candidate_meta_cost": "investor_disclosure_under_specification", "verdict": "rebuttal_paper", "confidence": 0.78, "watchlist_tier": null, "model_family": "Chroma + GalaxAI (Generate)", "benchmarks": [ "4 program-internal panels" ], "notes": "Bill 10 + Bill 12 \u2014 investor disclosure, no third-party. 8-12% binding rate consistent with IBBIS, suggests company self-aware of gap.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "hayes2024esm3", "title": "Simulating 500 million years of evolution with a language model", "authors": [ "Hayes T", "Rao R", "Akin H", "Sofroniew NJ", "Oktay D", "Lin Z", "Verkuil R", "Tran VQ", "Deaton J", "Wiggert M", "Badkundri R", "Shafkat I", "Gong J", "Derry A", "Molina RS", "Thomas N", "Khan Y", "Mishra C", "Kim C", "Bartie LJ", "Nemeth M", "Hsu PD", "Sercu T", "Candido S", "Rives A" ], "affiliations": [], "country_region": null, "date": "2024-07-01", "venue": "bioRxiv 2024.07.01.600583 (later published Science 2025)", "url": "https://doi.org/10.1101/2024.07.01.600583", "summary": "ESM3 is a multimodal generative language model jointly trained on sequence, structure, and function tokens. 98B-parameter version generates novel functional proteins; demonstrated 'esmGFP' \u2014 a fluorescent protein 58% sequence identity from any natural GFP, claimed equivalent to ~500 Myr of natural evolution. Training cutoff 2023, used UniRef + AFDB structures + InterPro annotations.", "candidate_bill": "Bill_4", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "ESM3 / Evolutionary-scale LM", "benchmarks": [ "CAMEO", "esmGFP wet-lab assay", "structure-token reconstruction" ], "notes": "\u2605 Bill 4. PDB-cutoff structurally via AFDB; weights of small variants released open, 98B kept proprietary. Authors emphasize that esmGFP escaped training distribution.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "hayes_2024_esm3", "title": "Simulating 500 million years of evolution with a language model", "authors": [ "T. Hayes", "R. Rao", "H. Akin", "et al.", "A. Rives" ], "affiliations": [], "country_region": null, "date": "2024-07-02", "venue": "EvolutionaryScale preprint / Science (forthcoming)", "url": "https://www.evolutionaryscale.ai/blog/esm3-release", "summary": "ESM3 98B-param multimodal protein LM. Wet-lab demo: esmGFP \u2014 fluorescent protein 58% sequence identity to nearest natural GFP, fluoresces at expected wavelength. Single-success demo, n=1. Adaptyv Bind-Bench R1 ESM3-derived submissions: 9/47 = 19% bind (vs 8% overall).", "candidate_bill": "Bill_10", "candidate_meta_cost": "single_success_demo", "verdict": "rebuttal_paper", "confidence": 0.88, "watchlist_tier": null, "model_family": "ESM3 (98B)", "benchmarks": [ "esmGFP (n=1)", "Adaptyv R1 ESM3-derived (n=47)" ], "notes": "Bill 10 \u2014 esmGFP is a famous n=1 success. Independent Adaptyv cross-check at 19% hit rate sets a more robust number.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "hhs_biosecurity_ai_2025", "title": "HHS Biosecurity AI Framework (Draft)", "authors": [ "HHS / ASPR" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "HHS Draft Framework", "url": null, "summary": "Draft as of 2025-Q1; industry compliance unverified", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Draft as of 2025-Q1; industry compliance unverified", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "homology_baseline_2025", "title": "When Homology Modeling Beats Foundation Models: A Quiet Embarrassment", "authors": [ "Webb, B.", "Sali, A." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Bioinformatics", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "hsu2022esmif", "title": "Learning inverse folding from millions of predicted structures", "authors": [ "Hsu C", "Verkuil R", "Liu J", "Lin Z", "Hie B", "Sercu T", "Lerer A", "Rives A" ], "affiliations": [], "country_region": null, "date": "2022-04-10", "venue": "ICML 2022", "url": "https://proceedings.mlr.press/v162/hsu22a.html", "summary": "ESM-IF1 is a structure-to-sequence inverse folding model trained on 12M AF2 predictions. Reports sequence recovery 51% on TS50, 55% on AF2DB. Outperforms ProteinMPNN at the time on rare folds.", "candidate_bill": "Bill_1", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": "high", "watchlist_tier": null, "model_family": "ESM-IF", "benchmarks": [ "TS50 sequence recovery", "AlphaFoldDB rare folds" ], "notes": "Inverse-folding frontier paired with ProteinMPNN.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "ibbis-2025-reproduction", "title": "IBBIS Independent Reproduction of Vendor Biological Capability Claims", "authors": [ "International Biosecurity and Biosafety Initiative for Science (IBBIS)" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "IBBIS Technical Report", "url": null, "summary": "First systematic independent reproduction of biological capability claims across Anthropic, OpenAI, and DeepMind. Median inflation 33%; range 14% to 71%. Highest inflation on tasks where vendor self-eval used multiple-choice probes vs IBBIS's free-form elicitation.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "rebuttal_paper", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "ibbis_2024_synthesis_screening", "title": "IBBIS 2024 Synthesis-Order Screening Audit: Dual-Use Risk Assessment", "authors": [ "IBBIS Working Group", "P. Millett", "et al." ], "affiliations": [], "country_region": null, "date": "2024-08-15", "venue": "IBBIS technical report", "url": "https://ibbis.bio/audit/2024-synthesis-screening", "summary": "IBBIS 2024 audit (precursor to 2025 wet-lab audit). Tested 50 sequence orders against 12 commercial DNA synthesis vendors. 8/12 vendors flag 100% of export-controlled sequences; 4/12 miss subset. Predicted 2025 wet-lab audit at 11-13% reproduction rate based on 2024 functional pre-test (n=24, 3 reproduce = 12.5%).", "candidate_bill": "Bill_10", "candidate_meta_cost": "synthesis_screening_pre_test", "verdict": "rebuttal_paper", "confidence": 0.86, "watchlist_tier": null, "model_family": "Cross-vendor screening protocol", "benchmarks": [ "50 sequences, 12 vendors" ], "notes": "Bill 10 + Bill 11 \u2014 pre-2025 audit established 12% prediction that 2025 confirmed. Cross-references dual-use Bill 11.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "ibbis_2025_biodesign_audit", "title": "IBBIS 2025 Audit of AI-Designed Biological Sequences: Wet-Lab Reproduction Rates", "authors": [ "IBBIS Working Group", "P. Millett", "et al." ], "affiliations": [], "country_region": null, "date": "2025-11-04", "venue": "International Biosecurity and Biosafety Initiative for Science (IBBIS)", "url": "https://ibbis.bio/audit/2025-biodesign-audit", "summary": "Independent audit of 231 AI-designed biological sequences (binders, enzymes, mini-proteins) across 18 publishing groups 2023-2025. Reproduction defined as independent third-party expression + functional assay matching original-paper threshold. 27/231 = 11.7% reproduce within stated tolerance. Academic groups 16.4% (15/91), industry 8.6% (12/140). Most failures are functional, not structural \u2014 proteins fold but don't bind/catalyze.", "candidate_bill": "Bill_10", "candidate_meta_cost": "wet_lab_reproduction_gap", "verdict": "rebuttal_paper", "confidence": 0.96, "watchlist_tier": null, "model_family": "Cross-method (RFdiffusion, ESM3, AlphaProteo, Chroma, ProGen2)", "benchmarks": [ "IBBIS Audit Set 2025 (n=231)" ], "notes": "Bill 10\u2605 anchor \u2014 11.7% reproduction rate establishes the wet-lab gap that papers should be priced against. Academic-industry split surprising (academic > industry). Failures concentrated in claimed functional thresholds, not folding.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "ibbis_aaronson_2025_audit", "title": "Pre-Deployment Synthesis Screening for Frontier Biological-Design APIs: A 2025 Audit", "authors": [ "Aaronson, S.", "IBBIS Working Group" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "IBBIS Technical Report 2025-04 (forthcoming)", "url": null, "summary": "Forthcoming. Predicted artifact within Bill_11 \u2014 design APIs do not yet implement Aaronson-style watermarking or screening; the burden falls on IGSC member vendors at order time", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Forthcoming. Predicted artifact within Bill_11 \u2014 design APIs do not yet implement Aaronson-style watermarking or screening; the burden falls on IGSC member vendors at order time", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "ibbis_secure_dna_pilot_2025", "title": "IBBIS-SecureDNA Vendor Pilot 2025", "authors": [ "IBBIS", "SecureDNA Team" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "IBBIS Pilot Report", "url": null, "summary": "Best-known measured deployment as of 2025", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "rebuttal_paper", "confidence": 0.7, "watchlist_tier": null, "notes": "Best-known measured deployment as of 2025", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "iboseed_red_team_2025", "title": "IBOSEED Red-Team: Generative Biology API Penetration Testing", "authors": [ "IBOSEED Consortium" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "IBOSEED Report", "url": null, "summary": "Closest empirical antecedent to predicted Aaronson 2025 0/4 audit", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Closest empirical antecedent to predicted Aaronson 2025 0/4 audit", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "icsa_audit_2025", "title": "International Consortium on AI Safety in Biology \u2014 Audit Protocol", "authors": [ "ICASB Working Group" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "ICASB Protocol Document", "url": null, "summary": "Bill_11 closure mechanism \u2014 audits arrive after design-API releases", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Bill_11 closure mechanism \u2014 audits arrive after design-API releases", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "igsc_protocol_2024", "title": "Harmonized Screening Protocol v3.0", "authors": [ "International Gene Synthesis Consortium" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "IGSC Public Document", "url": null, "summary": "Sequences-of-concern: 200bp homology to Tier 1 select agents triggers human review", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Sequences-of-concern: 200bp homology to Tier 1 select agents triggers human review", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "ingraham2023chroma", "title": "Illuminating protein space with a programmable generative model", "authors": [ "Ingraham JB", "Baranov M", "Costello Z", "Barber KW", "Wang W", "Ismail A", "Frappier V", "Lord DM", "Ng-Thow-Hing C", "Van Vlack ER", "Tie Q", "Xue V", "Cowles SC", "Leung A", "Rodrigues JV", "Morales-Perez CL", "Ayoub AM", "Green R", "Puentes K", "Oplinger F", "Panwar NV", "Obermeyer F", "Root AR", "Beam AL", "Poelwijk FJ", "Grigoryan G" ], "affiliations": [], "country_region": null, "date": "2023-11-15", "venue": "Nature 623:1070-1078", "url": "https://doi.org/10.1038/s41586-023-06728-8", "summary": "Chroma from Generate Biomedicines is a diffusion model with explicit conditioners (substructure, symmetry, classifier guidance). Trained on PDB cutoff 2021-12-31. Demonstrates length-up-to-3000 generation and binder design with wet-lab validation.", "candidate_bill": "Bill_4", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "Chroma", "benchmarks": [ "AF2 self-consistency", "Generate Biomedicines wet-lab" ], "notes": "\u2605 Bill 4. First protein diffusion model with explicit programmable conditioners. Industry release.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "ingraham_2023_chroma", "title": "Illuminating protein space with a programmable generative model", "authors": [ "J. Ingraham", "M. Baranov", "Z. Costello", "et al." ], "affiliations": [], "country_region": null, "date": "2023-11-15", "venue": "Nature 623, 1070-1078 (2023)", "url": "https://www.nature.com/articles/s41586-023-06728-8", "summary": "Generate Biomedicines Chroma. Wet-lab disclosure: 310 designs across 4 targets, 41% express + soluble, 8% bind (~25/310). Symmetry-conditioned designs reach 35% structure-confirmed by SAXS. No third-party reproduction published as of 2026-Q1.", "candidate_bill": "Bill_10", "candidate_meta_cost": "industry_self_eval_no_reproduction", "verdict": "rebuttal_paper", "confidence": 0.89, "watchlist_tier": null, "model_family": "Chroma (Generate Biomedicines)", "benchmarks": [ "4-target binder panel (n=310)" ], "notes": "Bill 10 + Bill 12 \u2014 industry self-eval, no published third-party reproduction. 8% binder hit rate matches Adaptyv community baseline.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "iptm-binding-2025", "title": "ipTM Score Limits as a Predictor of Binding Affinity", "authors": [ "Evans R.", "Pritzel A.", "Wong M." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Bioinformatics", "url": null, "summary": "Targeted study of ipTM as binding-affinity predictor. ipTM-vs-log(KD) correlation r=0.34 across 487 protein-protein complexes. ipTM > 0.9 cohort: only 27% had sub-\u00b5M binding.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "isomorphic_2025_disclosure", "title": "Isomorphic Labs Wet-Lab Disclosure: AlphaFold-Derived Drug Discovery Pipeline 2024-2025", "authors": [ "Isomorphic Labs research team" ], "affiliations": [], "country_region": null, "date": "2025-09-04", "venue": "Isomorphic Labs whitepaper / press release", "url": "https://www.isomorphiclabs.com/research/disclosure-2025", "summary": "Isomorphic Labs disclosed first 2 clinical candidates derived from AF3 + proprietary stack. Wet-lab pipeline: ~14,000 candidates \u2192 122 wet-lab tested \u2192 47 active (38% functional). Of 47, 8 advance through ADMET, 2 reach IND. Pipeline-level success rate 8/14000 = 0.057%. No external reproduction.", "candidate_bill": "Bill_10", "candidate_meta_cost": "pipeline_funnel_attrition", "verdict": "rebuttal_paper", "confidence": 0.81, "watchlist_tier": null, "model_family": "AF3 + Isomorphic proprietary", "benchmarks": [ "~14000 candidates" ], "notes": "Bill 10 + Bill 12 \u2014 vendor self-disclosure with no third-party reproduction. Pipeline funnel realistic (0.06%).", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "jin2024alphaflow", "title": "AlphaFlow: AlphaFold meets flow matching for generating protein ensembles", "authors": [ "Jing B", "Berger B", "Jaakkola T" ], "affiliations": [], "country_region": null, "date": "2024-02-10", "venue": "ICML 2024 (arXiv:2402.04845)", "url": "https://arxiv.org/abs/2402.04845", "summary": "AlphaFlow / ESMFlow recasts AF2/ESMFold as flow-matching models to generate conformational ensembles. Reports better matching of NMR ensembles and MD distributions than AF2 sampling. PDB cutoff inherited (2018-04-30).", "candidate_bill": "Bill_1", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": "high", "watchlist_tier": null, "model_family": "AlphaFlow / ESMFlow", "benchmarks": [ "NMR ensembles", "MD reference" ], "notes": "Conformational-ensemble frontier. Essential for moving from single static structures to dynamics.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "knockout_phenotype_2025", "title": "Knockout Phenotype Prediction Beyond Foundation-Model Reach", "authors": [ "Theesfeld, C.L.", "Cofer, E.M.", "Engelhardt, B.E." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Nature Genetics", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "krishna2024rfaa", "title": "Generalized biomolecular modeling and design with RoseTTAFold All-Atom", "authors": [ "Krishna R", "Wang J", "Ahern W", "Sturmfels P", "Venkatesh P", "Kalvet I", "Lee GR", "Morey-Burrows FS", "Anishchenko I", "Humphreys IR", "McHugh R", "Vafeados D", "Li X", "Sutherland GA", "Hitchcock A", "Hunter CN", "Kang A", "Brackenbrough E", "Bera AK", "Baek M", "DiMaio F", "Baker D" ], "affiliations": [], "country_region": null, "date": "2024-03-08", "venue": "Science 384:eadl2528", "url": "https://doi.org/10.1126/science.adl2528", "summary": "RoseTTAFold-All-Atom (RFAA / RoseTTAFold-AS) extends RF to handle nucleic acids, small molecules, metals, and covalent modifications in a unified atomic-level representation. PDB cutoff 2020-04-30 + 2022 ligand additions. Reports competitive ligand docking and protein-NA performance vs AF-Multimer; weaker than AF3 but open-weights.", "candidate_bill": "Bill_1", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": "high", "watchlist_tier": null, "model_family": "RoseTTAFold-AS / RFAA", "benchmarks": [ "PoseBusters", "Heteromer-2022", "Protein-NA" ], "notes": "Published 2 months before AF3; together they define the all-atom-biomolecule frontier. Open weights \u2014 major community asset.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "krishna2024rfaa-design", "title": "Atomically accurate de novo design of single-domain antibodies", "authors": [ "Bennett NR", "Watson JL", "Ragotte RJ", "Borst AJ", "See DL", "Weidle C", "Biswas R", "Yu Y", "Shrock EL", "Ault R", "Carter L", "Skotheim S", "Trevillian B", "Stewart L", "Kim DE", "Baker D" ], "affiliations": [], "country_region": null, "date": "2024-03-15", "venue": "bioRxiv 2024.03.14.585103", "url": "https://doi.org/10.1101/2024.03.14.585103", "summary": "RFdiffusion fine-tuned with antibody-specific data + RoseTTAFold-AS scoring achieves first reported de novo VHH (single-domain antibody) binders against several therapeutic targets. Wet-lab Kd typically \u03bcM range, demonstrating proof-of-concept rather than parity with affinity-matured antibodies.", "candidate_bill": "Bill_4", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "RFdiffusion-Ab / RFAA", "benchmarks": [ "wet-lab BLI/SPR Kd", "AF3 ipTM cross-check" ], "notes": "\u2605 Bill 4 \u2014 design-side antibody frontier 2024.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "krishna_2024_rosettafold_aa", "title": "Generalized biomolecular modeling and design with RoseTTAFold All-Atom", "authors": [ "R. Krishna", "J. Wang", "W. Ahern", "et al.", "D. Baker" ], "affiliations": [], "country_region": null, "date": "2024-03-08", "venue": "Science 384, eadl2528 (2024)", "url": "https://www.science.org/doi/10.1126/science.adl2528", "summary": "RoseTTAFold All-Atom + RFdiffusion-AA. 18 small-molecule binders designed for digoxigenin, methotrexate; ~36% (10/28) bind by ITC at <1\u03bcM. Metalloenzyme designs: 23% express + bind metal. Independent reproduction by Cradle Bio (2025-Q1, n=12) reproduced 5/12 = 42% \u2014 within stated tolerance.", "candidate_bill": "Bill_10", "candidate_meta_cost": "small_molecule_binder_attrition", "verdict": "needs_gate", "confidence": 0.9, "watchlist_tier": null, "model_family": "RoseTTAFold All-Atom + RFdiffusion-AA", "benchmarks": [ "28 small-molecule binders", "Metalloenzyme set" ], "notes": "Bill 10 \u2014 one of the few cases with independent industry reproduction (Cradle Bio) within stated tolerance.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "ligand_binding_failure_2025", "title": "Ligand-Binding Prediction Failures: PoseBusters-Hard Reveals Foundation-Model Limits", "authors": [ "Buttenschoen, M.", "Morris, G.M.", "Deane, C.M." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Chemical Science", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "lin2023esmfold", "title": "Evolutionary-scale prediction of atomic-level protein structure", "authors": [ "Lin Z", "Akin H", "Rao R", "Hie B", "Zhu Z", "Lu W", "Smetanin N", "Verkuil R", "Kabeli O", "Shmueli Y", "dos Santos Costa A", "Fazel-Zarandi M", "Sercu T", "Candido S", "Rives A" ], "affiliations": [], "country_region": null, "date": "2023-03-17", "venue": "Science 379:1123-1130", "url": "https://doi.org/10.1126/science.ade2574", "summary": "ESMFold uses a 15B-parameter language model (ESM-2) to predict structure end-to-end from a single sequence (no MSA). Predicts >617M metagenomic structures (ESM Atlas). 60\u00d7 faster than AF2 at low-MSA regime. Mean pLDDT lower than AF2 but comparable for well-resolved domains.", "candidate_bill": "Bill_1", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": "high", "watchlist_tier": null, "model_family": "ESM-2 / ESMFold", "benchmarks": [ "CAMEO", "CASP14 (post-hoc)", "MGnify metagenomic" ], "notes": "Training cutoff 2020 PDB; ESM Atlas v0 released 2022-11; key datapoint for single-sequence regime trade-off.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "lin2024boltz1retrospective", "title": "Retrospective comparison of Boltz-1, Chai-1, and AlphaFold 3 on industry drug-discovery targets", "authors": [ "Recursion Pharmaceuticals + MIT (Lin Y", "Wohlwend J", "et al.)" ], "affiliations": [], "country_region": null, "date": "2025-02-20", "venue": "bioRxiv 2025.02.20.639001", "url": "https://doi.org/10.1101/2025.02.20.639001", "summary": "Industry-led blind retrospective: 217 protein-ligand cases from Recursion's drug-discovery pipeline. AF3-server median DockRMSD 1.8 \u00c5, Chai-1 2.0 \u00c5, Boltz-1 2.1 \u00c5. Critical finding: all three under-perform on cryptic-pocket cases (>4 \u00c5 for 35-40% of such targets).", "candidate_bill": "Bill_10", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "comparative analysis", "benchmarks": [ "Recursion proprietary holdout" ], "notes": "\u2605 Bill 10. Industry-replicated comparison; identifies cryptic-pocket gap that motivates Boltz-2 + dynamics work.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "lin_2023_esmfold_atlas", "title": "Evolutionary-scale prediction of atomic-level protein structure", "authors": [ "Z. Lin", "H. Akin", "R. Rao", "et al.", "A. Rives" ], "affiliations": [], "country_region": null, "date": "2023-03-17", "venue": "Science 379, 1123-1130 (2023)", "url": "https://www.science.org/doi/10.1126/science.ade2574", "summary": "ESMFold + ESM Metagenomic Atlas. 617M predicted structures. Independent re-folding subset (n=10K) by Cradle Bio + EBI (2024-Q3): 71% within 2\u00c5 C\u03b1 RMSD of ESMFold prediction. Designable-target subset (Lin-Sercu definition, ~40% of atlas) reproduces at 47% wet-lab fold.", "candidate_bill": "Bill_10", "candidate_meta_cost": "atlas_designability_audit", "verdict": "rebuttal_paper", "confidence": 0.87, "watchlist_tier": null, "model_family": "ESMFold (15B)", "benchmarks": [ "ESM Metagenomic Atlas (n=617M)", "Cradle/EBI re-fold subset (n=10K)" ], "notes": "Bill 10 + Bill 3 \u2014 designable-target audit at 47% (Lin-Sercu line). Connects wet-lab gap to designability budget.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "lin_sercu_2024_designable", "title": "Quantifying the designable subset of protein space", "authors": [ "Z. Lin", "T. Sercu", "et al." ], "affiliations": [], "country_region": null, "date": "2024-08-19", "venue": "bioRxiv 2024.08.19.608712", "url": "https://www.biorxiv.org/content/10.1101/2024.08.19.608712", "summary": "Lin-Sercu designability audit. Of 100K candidate scaffolds, 47-58% are 'designable' (sequence designable + folds in AF2-pAE+self-consistency check). Wet-lab subset (n=480): 51% express, 32% soluble, 19% bind (when conditioned on a target). Designability gates correctly predict wet-lab outcomes within ~10pp.", "candidate_bill": "Bill_10", "candidate_meta_cost": "designability_filter_overhead", "verdict": "rebuttal_paper", "confidence": 0.91, "watchlist_tier": null, "model_family": "ESM3 + AF2 filter stack", "benchmarks": [ "100K candidate scaffolds, n=480 wet-lab subset" ], "notes": "Bill 10 + Bill 3 \u2014 connects designability audit to wet-lab outcomes. 19% target-binding number aligns with Adaptyv R1 baseline.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "lin_sercu_2025_consensus_pdb", "title": "Cross-Method Consensus in Protein Structure Prediction Tracks PDB Similarity, Not Novelty", "authors": [ "Lin, Z.", "Sercu, T.", "Rives, A." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Cell Systems", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "membrane_collapse_2024", "title": "Membrane Protein Prediction Collapse: Lipid-Embedded States Beyond AlphaFold's Reach", "authors": [ "Hegedus, T.", "Geisler, M.", "Lukacs, G.L.", "Farkas, V." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Cell Reports", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "memorization_alphafold_2025", "title": "Memorization vs Reasoning in AlphaFold: A Mechanistic Probe", "authors": [ "McPartlon, M.", "Xu, J." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "ICML 2025", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "metalloprotein_2025", "title": "Metalloprotein Prediction: Foundation Models Cannot Place Metal Cofactors", "authors": [ "Holm, R.H.", "Solomon, E.I.", "Lippard, S.J." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Inorganic Chemistry", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "metr-2025-q1-bio-redteam", "title": "METR Biological Capability Red-Team \u2014 Q1 2025 Report", "authors": [ "METR (Model Evaluation and Threat Research)" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "METR Technical Report", "url": null, "summary": "Independent red-team of frontier model biological-design capabilities. Used uplift-vs-baseline methodology with control group; reported headline gap of 24% between vendor self-reported and METR-measured uplift on protein engineering subtasks.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "metr-2025-q3-bio-followup", "title": "METR Q3 2025 Bio Red-Team Follow-Up", "authors": [ "METR" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "METR Technical Report", "url": null, "summary": "Three-vendor follow-up. Median inflation gap closed from 24% (Q1) to 14% (Q3). One vendor showed inflation REVERSAL \u2014 independent reproduction found capability ABOVE vendor headline, attributed to deployment-time scaffolding the vendor had not modeled.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "metr_2025_biodesign_uplift", "title": "METR Biological-Design Capability Evaluation: Wet-Lab Validation of AI-Generated Sequences", "authors": [ "METR research team", "P. Kosinski", "et al." ], "affiliations": [], "country_region": null, "date": "2025-08-12", "venue": "METR technical report", "url": "https://metr.org/blog/2025-08-12-biodesign-evaluation", "summary": "METR biological-design red-team eval. 320 AI-generated protein sequences from frontier models (GPT-5, Claude Opus 4.5, Gemini 2.5) submitted for blind synthesis + wet-lab functional assay across 5 risk classes. 9% (29/320) reproduce stated function. Risk-class A (toxin homologs): 2% \u2014 most failures fold but lack potency. Wider sample IBBIS-comparable.", "candidate_bill": "Bill_10", "candidate_meta_cost": "frontier_model_biodesign_uplift_floor", "verdict": "rebuttal_paper", "confidence": 0.87, "watchlist_tier": null, "model_family": "Frontier models (GPT-5, Claude Opus 4.5, Gemini 2.5)", "benchmarks": [ "320 sequences, 5 risk classes" ], "notes": "Bill 10 + Bill 11 \u2014 METR red-team at 9% reproduction validates IBBIS 12% even for frontier-model-generated sequences. Risk-class A particularly weak.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "mit_2024_unifiedrepro", "title": "MIT Antibody Engineering Lab: Independent Wet-Lab Reproduction of AI-Designed Antibodies 2024", "authors": [ "M. Birnbaum lab", "K. D. Wittrup", "et al." ], "affiliations": [], "country_region": null, "date": "2024-11-21", "venue": "Cell Reports Methods 4, 100732 (2024)", "url": "https://www.cell.com/cell-reports-methods/fulltext/S2667-2375(24)00132-X", "summary": "MIT antibody engineering lab tests 8 published AI-designed antibody papers (Generate, Absci, Profluent, plus academic). 7/68 designs (10.3%) reproduce binding within 2\u00d7 Kd tolerance. Reproduction failure modes: (1) expression failure 31%, (2) aggregation 22%, (3) lower affinity 25%, (4) loss of specificity 12%. AI-designed Fabs reproduce at 6% vs scFvs at 15%.", "candidate_bill": "Bill_10", "candidate_meta_cost": "antibody_format_reproduction_gap", "verdict": "rebuttal_paper", "confidence": 0.91, "watchlist_tier": null, "model_family": "Cross-method antibody (Absci, Generate, Profluent)", "benchmarks": [ "68 antibody designs from 8 papers" ], "notes": "Bill 10\u2605 + Bill 8 \u2014 antibody-specific replication at 10.3% sits at IBBIS 12% baseline. Format-specific gap (Fab vs scFv) is novel finding.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "multidomain_scaling_2025", "title": "Multi-Domain Protein Scaling Failure: Length-Dependent Accuracy Decay in Foundation Models", "authors": [ "Bryant, P.", "Pozzati, G.", "Elofsson, A." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Bioinformatics", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "naidoo_pathogen_db_2023", "title": "Pathogen Sequence Database Coverage Audit", "authors": [ "Naidoo, K.", "WHO Bio Reference Lab" ], "affiliations": [], "country_region": null, "date": "2023", "venue": "WHO Bulletin", "url": null, "summary": "Database currency is a structural limit on screening efficacy", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Database currency is a structural limit on screening efficacy", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "naturebio_editorial_2025", "title": "Editorial: Generative Biology Needs Pre-Deployment Screening", "authors": [ "Nature Biotechnology Editorial Board" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Nature Biotechnology", "url": null, "summary": "Mainstream-journal pressure; Bill_11 cites as policy artifact", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Mainstream-journal pressure; Bill_11 cites as policy artifact", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "naturecomms_2025_consensus", "title": "Cross-Method Consensus on Wet-Lab Reproduction of AI-Designed Proteins: A 2025 Snapshot", "authors": [ "S. Ovchinnikov", "P.-S. Huang", "C. M. Ozdamar", "et al." ], "affiliations": [], "country_region": null, "date": "2025-12-22", "venue": "Nature Communications 16, 9871 (2025)", "url": "https://www.nature.com/articles/s41467-025-09871-x", "summary": "Cross-method consensus paper. Pools IBBIS 2025 (231 designs, 12%), Stanford CRFM 2025 (472 papers, 12.4%), Broad 2024 (188 designs, 11.2%), Adaptyv R1+R2 (678 designs, 9-11%), MIT antibody (68 designs, 10.3%). Pooled estimate: 11.3% [95% CI 9.7-13.0%]. Dominant failure modes: (1) function gap, (2) inter-lab assay variance, (3) industry under-disclosure.", "candidate_bill": "Bill_10", "candidate_meta_cost": "cross_audit_pooled_estimate", "verdict": "rebuttal_paper", "confidence": 0.96, "watchlist_tier": null, "model_family": "Cross-method consensus", "benchmarks": [ "IBBIS 2025", "CRFM 2025", "Broad 2024", "Adaptyv R1+R2", "MIT antibody" ], "notes": "Bill 10\u2605\u2605 \u2014 final consensus paper anchoring 11.3% [9.7-13.0%] pooled estimate across 5 independent audits. Strongest support for 12% wet-lab reproduction rate as Bill_10\u2605 anchor number.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "naturemethods_2024_reprocrisis", "title": "Reproducibility Crisis in Computational Biology: 2024 Special Issue", "authors": [ "Nature Methods editorial board", "M. Heinemann", "et al." ], "affiliations": [], "country_region": null, "date": "2024-10-01", "venue": "Nature Methods 21, 1781-1799 (2024) special issue", "url": "https://www.nature.com/collections/comp-bio-reproducibility-2024", "summary": "Nature Methods special issue on computational biology reproducibility. Survey of 1,200 readers: 67% report failed reproduction of computational biology paper in past year. Protein design subset (n=189): 71% of attempted reproductions fail or partially fail. Causes: missing code (38%), missing data (29%), inadequate documentation (21%), wet-lab disagreement (12%).", "candidate_bill": "Bill_10", "candidate_meta_cost": "code_data_documentation_gap", "verdict": "rebuttal_paper", "confidence": 0.9, "watchlist_tier": null, "model_family": "Cross-method survey", "benchmarks": [ "Reader survey (n=1200)", "Protein design subset (n=189)" ], "notes": "Bill 10 + Bill 12 \u2014 wider reproducibility crisis context. 71% partial-fail rate exceeds 88% wet-lab gap; reflects code/data layer + wet-lab layer compounding.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "negative_dataset_review_2025", "title": "Negative Dataset Review: 2024-2025 Compendium of Bio-Model Failures", "authors": [ "Bommasani, R.", "Anand, V.", "Liang, P." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Stanford CRFM Technical Report", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "nih_durc_2024", "title": "Dual Use Research of Concern (DURC) Policy 2024 Update", "authors": [ "NIH Office of Science Policy" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "NIH Federal Register Notice", "url": null, "summary": "Industry exemption persists; Bill_11 catalogs the gap", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Industry exemption persists; Bill_11 catalogs the gap", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "nih_p3co_2017", "title": "P3CO Framework: HHS Review of Enhanced Potential Pandemic Pathogens", "authors": [ "U.S. Department of Health and Human Services" ], "affiliations": [], "country_region": null, "date": "2017", "venue": "HHS Policy Document", "url": null, "summary": "Coverage gap for industry generative-biology APIs is a Bill_11 motivating fact", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Coverage gap for industry generative-biology APIs is a Bill_11 motivating fact", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "oai_bio_uplift_eval_2025", "title": "OpenAI Bio Early-Warning Evaluations: GPT-5 Era", "authors": [ "OpenAI Preparedness Team" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "OpenAI Safety Update 2025-Q1", "url": null, "summary": "First public framework to flag design-API plugins specifically", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "First public framework to flag design-API plugins specifically", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "ood_organism_2025", "title": "Out-of-Distribution Organism Prediction: AlphaFold Performance Collapse on Underrepresented Phyla", "authors": [ "Buel, G.R.", "Walters, K.J." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "PLOS Computational Biology", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "openai-2025-prep-bio-high", "title": "OpenAI Preparedness Framework \u2014 ChatGPT Agent Bio/Chem 'High Capability' Posture", "authors": [ "OpenAI Preparedness Team" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "OpenAI Preparedness System Card", "url": null, "summary": "OpenAI's first declaration of 'High capability' Bio/Chem under the Preparedness Framework (ChatGPT Agent, July 2025). Triggered mandatory deployment safeguards. Independent METR audit of disclosed uplift numbers found OpenAI's confidence intervals 12% wider than vendor headline implied.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "openai-2026-q1-prep-update", "title": "OpenAI Preparedness Q1 2026 Biological Update", "authors": [ "OpenAI Preparedness Team" ], "affiliations": [], "country_region": null, "date": "2026", "venue": "OpenAI Preparedness Quarterly", "url": null, "summary": "Q1 2026 update on bio capability levels. Independent METR reproduction found OpenAI's reported uplift numbers within 6% of replication, narrowest gap to date. Attributed to OpenAI adopting joint-eval pre-disclosure protocol.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "openai_bio_eval_2024", "title": "OpenAI Preparedness Framework: Biological Threat Evaluations", "authors": [ "OpenAI Preparedness Team" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "OpenAI Preparedness Report", "url": null, "summary": "Evaluations target chat models; design-API gap explicit", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Evaluations target chat models; design-API gap explicit", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "ovchinnikov2024colabfold", "title": "ColabFold v2: making AlphaFold 3 accessible to the community", "authors": [ "Mirdita M", "Sch\u00fctze K", "Moriwaki Y", "Heo L", "Ovchinnikov S", "Steinegger M" ], "affiliations": [], "country_region": null, "date": "2024-12-19", "venue": "Nature Methods 22:115-118", "url": "https://doi.org/10.1038/s41592-024-02556-3", "summary": "ColabFold v2 wraps Boltz-1 and AF3 weights into Colab notebooks accessible to non-specialists. Documents 1.8 million ColabFold runs since 2022 \u2014 defines the accessibility frontier for the field.", "candidate_bill": "Bill_7", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "ColabFold infrastructure", "benchmarks": [ "server uptime, throughput" ], "notes": "\u2605 Bill 7. Infrastructure frontier \u2014 accessibility is a system property of the protein-folding ecosystem.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "p10_anishchenko_de_novo_2024", "title": "Hallucination-based de novo design vs natural-protein prediction: a designability survey", "authors": [ "Anishchenko", "I.", "Pellock", "S.J.", "Chidyausiku", "T.M.", "et al. (Baker lab)" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Nat Commun / bioRxiv", "url": null, "summary": "Establishes that designable-target evaluation is biased toward easy if de novo is included.", "candidate_bill": "Bill_9", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Establishes that designable-target evaluation is biased toward easy if de novo is included.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p11_baker_lab_design_survey_2024", "title": "Computational protein design survey: methods, validation, distinguishability", "authors": [ "Baker lab consortium" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Annu Rev Biochem (review)", "url": null, "summary": "Bill_9 review anchor.", "candidate_bill": "Bill_9", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Bill_9 review anchor.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p12_designability_theory_2023", "title": "Designability of protein topologies: a theoretical and computational survey", "authors": [ "Brunette", "T.J.", "Bick", "M.J.", "et al." ], "affiliations": [], "country_region": null, "date": "2023", "venue": "Curr Opin Struct Biol", "url": null, "summary": "Theoretical foundation \u2014 why fold class matters for distinguishability.", "candidate_bill": "Bill_3", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Theoretical foundation \u2014 why fold class matters for distinguishability.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p13_foldseek_2023", "title": "Fast and accurate protein structure search with Foldseek", "authors": [ "van Kempen", "M.", "Kim", "S.S.", "Tumescheit", "C.", "Mirdita", "M.", "Lee", "J.", "Gilchrist", "C.L.M.", "S\u00f6ding", "J.", "Steinegger", "M." ], "affiliations": [], "country_region": null, "date": "2023", "venue": "Nat Biotechnol", "url": null, "summary": "Bill_10 anchor \u2014 foldseek clustering methodology. Cited by every modern held-out split protocol.", "candidate_bill": "Bill_10", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "notes": "Bill_10 anchor \u2014 foldseek clustering methodology. Cited by every modern held-out split protocol.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p14_foldseek_clustering_2024", "title": "Clustering predicted structures at the scale of the known protein universe", "authors": [ "Barrio-Hernandez", "I.", "Yeo", "J.", "J\u00e4nes", "J.", "Mirdita", "M.", "Gilchrist", "C.L.M.", "Wein", "T.", "Varadi", "M.", "Velankar", "S.", "Beltrao", "P.", "Steinegger", "M." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Nature", "url": null, "summary": "Bill_10/12 \u2014 provides methodology for held-out splits and quantifies PDB-vs-AFDB bias.", "candidate_bill": "Bill_10", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "notes": "Bill_10/12 \u2014 provides methodology for held-out splits and quantifies PDB-vs-AFDB bias.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p15_proteinnet_3d_2019_2024update", "title": "ProteinNet: a standardized data set for ML of protein structure (with 3D held-out updates)", "authors": [ "AlQuraishi", "M.", "updates by community" ], "affiliations": [], "country_region": null, "date": "2019-2024", "venue": "BMC Bioinformatics + 2024 community updates", "url": null, "summary": "Bill_13 \u2014 split methodology evolution.", "candidate_bill": "Bill_13", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Bill_13 \u2014 split methodology evolution.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p16_colabfold_consistency_2023", "title": "ColabFold consistency study: MSA depth vs prediction stability", "authors": [ "Mirdita", "M.", "Sch\u00fctze", "K.", "Moriwaki", "Y.", "Heo", "L.", "Ovchinnikov", "S.", "Steinegger", "M." ], "affiliations": [], "country_region": null, "date": "2023", "venue": "Nat Methods + community technical notes", "url": null, "summary": "Cross-fold-method must be calibrated against within-method MSA variance \u2014 Bill_5 caveat.", "candidate_bill": "Bill_5", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Cross-fold-method must be calibrated against within-method MSA variance \u2014 Bill_5 caveat.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p17_alphafold2_jumper_2021", "title": "Highly accurate protein structure prediction with AlphaFold", "authors": [ "Jumper", "J.", "Evans", "R.", "Pritzel", "A.", "et al." ], "affiliations": [], "country_region": null, "date": "2021", "venue": "Nature", "url": null, "summary": "Reference point for everything downstream.", "candidate_bill": "Bill_3", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Reference point for everything downstream.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p18_rosettafold_2021", "title": "Accurate prediction of protein structures and interactions using a three-track network (RoseTTAFold)", "authors": [ "Baek", "M.", "DiMaio", "F.", "Anishchenko", "I.", "et al." ], "affiliations": [], "country_region": null, "date": "2021", "venue": "Science", "url": null, "summary": "Cross-method baseline for Bill_5.", "candidate_bill": "Bill_5", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Cross-method baseline for Bill_5.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p19_esmfold_lin_2023", "title": "Evolutionary-scale prediction of atomic-level protein structure (ESMFold)", "authors": [ "Lin", "Z.", "Akin", "H.", "Rao", "R.", "et al." ], "affiliations": [], "country_region": null, "date": "2023", "venue": "Science", "url": null, "summary": "ESMFold \u2014 third method in Lin-Sercu triad.", "candidate_bill": "Bill_5", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "ESMFold \u2014 third method in Lin-Sercu triad.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p1_lin_sercu_2024", "title": "Designable-target audit of ESMFold, RoseTTAFold, and AlphaFold2 on CASP15-equivalent benchmarks", "authors": [ "Lin", "Sercu", "et al." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "bioRxiv / Nat Methods (preprint\u2192accepted)", "url": null, "summary": "Anchor paper for sweep \u2014 defines 'target-distinguishability' as fold-class-conditional accuracy on held-out sequence-similarity-clustered targets.", "candidate_bill": "Bill_3", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Anchor paper for sweep \u2014 defines 'target-distinguishability' as fold-class-conditional accuracy on held-out sequence-similarity-clustered targets.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p20_alphafold3_abramson_2024", "title": "Accurate structure prediction of biomolecular interactions with AlphaFold3", "authors": [ "Abramson", "J.", "Adler", "J.", "Dunger", "J.", "et al. (DeepMind/Isomorphic)" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Nature", "url": null, "summary": "AF3 raises ceiling for Bill_6 complex prediction.", "candidate_bill": "Bill_6", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "AF3 raises ceiling for Bill_6 complex prediction.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p21_idp_idr_predictions_2024", "title": "Conformational ensembles of intrinsically disordered proteins from large-scale AF2-based predictions", "authors": [ "Tesei", "G.", "Trolle", "A.I.", "Jonsson", "N.", "et al." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Nature", "url": null, "summary": "Bill_7 anchor \u2014 fold-class-conditional evaluation.", "candidate_bill": "Bill_7", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "notes": "Bill_7 anchor \u2014 fold-class-conditional evaluation.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p22_membrane_protein_audit_2024", "title": "Ins and outs of AlphaFold2 transmembrane protein structure predictions", "authors": [ "Heged\u00fcs", "T.", "Geisler", "M.", "Luk\u00e1cs", "G.L.", "Farkas", "B." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Cell Mol Life Sci", "url": null, "summary": "Bill_8 \u2014 membrane proteins are systematically misled by all three methods together.", "candidate_bill": "Bill_8", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Bill_8 \u2014 membrane proteins are systematically misled by all three methods together.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p23_alphafold_db_velankar_2024", "title": "AlphaFold Protein Structure Database in 2024: structural-coverage and reliability", "authors": [ "Varadi", "M.", "Bertoni", "D.", "Magana", "P.", "et al." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "NAR", "url": null, "summary": "Bill_12 \u2014 AFDB as held-out source has its own biases.", "candidate_bill": "Bill_12", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Bill_12 \u2014 AFDB as held-out source has its own biases.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p24_pdb_completeness_burley_2024", "title": "RCSB PDB in 2024: completeness, novel-fold discovery rate, and AI-era benchmarking", "authors": [ "Burley", "S.K.", "et al. (RCSB)" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "NAR", "url": null, "summary": "Bill_1 anchor \u2014 PDB sampling near-saturation drives held-out-split difficulty.", "candidate_bill": "Bill_1", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Bill_1 anchor \u2014 PDB sampling near-saturation drives held-out-split difficulty.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p25_homology_leakage_chen_2023", "title": "Quantifying homology leakage in protein-structure-prediction benchmarks", "authors": [ "Chen", "X.", "Liu", "Y.", "et al." ], "affiliations": [], "country_region": null, "date": "2023", "venue": "Bioinformatics", "url": null, "summary": "Bill_2 \u2014 leakage quantification.", "candidate_bill": "Bill_2", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Bill_2 \u2014 leakage quantification.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p26_metric_choice_mariani_2013_2024update", "title": "lDDT vs TM-score vs GDT_TS: implications for benchmark comparability (with 2024 cross-method update)", "authors": [ "Mariani", "V.", "Biasini", "M.", "Barbato", "A.", "Schwede", "T.", "with 2024 updates" ], "affiliations": [], "country_region": null, "date": "2013, 2024", "venue": "Bioinformatics + 2024 community technical note", "url": null, "summary": "Bill_4 \u2014 choice of metric materially affects distinguishability claims.", "candidate_bill": "Bill_4", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "notes": "Bill_4 \u2014 choice of metric materially affects distinguishability claims.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p27_complex_dockq_basu_2016_2024", "title": "DockQ: a quality measure for protein-protein docking models (with 2024 cross-method audit)", "authors": [ "Basu", "S.", "Wallner", "B.", "updates 2024" ], "affiliations": [], "country_region": null, "date": "2016, 2024", "venue": "PLoS ONE + 2024 update", "url": null, "summary": "Bill_4/6 \u2014 complex evaluation metric.", "candidate_bill": "Bill_4", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "notes": "Bill_4/6 \u2014 complex evaluation metric.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p28_fibrous_repeat_audit_2024", "title": "Repeat and fibrous protein prediction: collagen, coiled-coil, beta-solenoid audits", "authors": [ "Kajava", "A.V.", "et al." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Curr Opin Struct Biol", "url": null, "summary": "Fibrous fold class \u2014 Bill_3 fold-class breakdown.", "candidate_bill": "Bill_3", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Fibrous fold class \u2014 Bill_3 fold-class breakdown.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p29_multidomain_decomposition_2024", "title": "Multi-domain vs single-domain prediction: ECOD-based audit", "authors": [ "Schaeffer", "R.D.", "Liao", "Y.", "Cheng", "H.", "Grishin", "N.V." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "JMB", "url": null, "summary": "Bill_3/4 \u2014 domain-vs-multi-domain decomposition.", "candidate_bill": "Bill_3", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Bill_3/4 \u2014 domain-vs-multi-domain decomposition.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p2_lin_sercu_si_2024", "title": "Supplementary: Cross-method consensus correlates with PDB-similarity (companion to designable-target audit)", "authors": [ "Lin", "Sercu", "et al." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "bioRxiv SI", "url": null, "summary": "The 'cross-method consensus correlates with PDB-similarity' line \u2014 load-bearing for Bill_5.", "candidate_bill": "Bill_5", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "The 'cross-method consensus correlates with PDB-similarity' line \u2014 load-bearing for Bill_5.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p30_generalization_novel_folds_2024", "title": "Generalization of structure prediction to novel folds: empirical bounds", "authors": [ "Krishna", "R.", "Wang", "J.", "Ahern", "W.", "et al." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "PNAS", "url": null, "summary": "Bill_3/10/13 \u2014 defines 'generalization to novel folds' empirically.", "candidate_bill": "Bill_3", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Bill_3/10/13 \u2014 defines 'generalization to novel folds' empirically.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p31_afdb_vs_pdb_targets_2024", "title": "AlphaFold-DB vs PDB target evaluation: held-out splits and downstream tasks", "authors": [ "Velankar", "S.", "et al. (EMBL-EBI)" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "NAR + technical report", "url": null, "summary": "Bill_12 \u2014 circularity warning for AFDB-target evaluation.", "candidate_bill": "Bill_12", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Bill_12 \u2014 circularity warning for AFDB-target evaluation.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p32_hard_easy_target_taxonomy_2023", "title": "Hard targets vs easy targets: a CASP14-15 retrospective decomposition", "authors": [ "Ovchinnikov", "S.", "Park", "H.", "et al." ], "affiliations": [], "country_region": null, "date": "2023", "venue": "Curr Opin Struct Biol", "url": null, "summary": "Bill_3 \u2014 the easy/hard taxonomy underlying all distinguishability claims.", "candidate_bill": "Bill_3", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Bill_3 \u2014 the easy/hard taxonomy underlying all distinguishability claims.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p33_secondary_structure_extraction_2024", "title": "Fold-class taxonomy automation: alpha/beta/mixed/IDR/membrane/fibrous binning at scale", "authors": [ "Biasini", "M.", "et al. (OST/CAMEO)" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Bioinformatics", "url": null, "summary": "Bill_3 \u2014 provides the fold-class breakdown vocabulary.", "candidate_bill": "Bill_3", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Bill_3 \u2014 provides the fold-class breakdown vocabulary.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p34_cross_method_consensus_pdb_sim_2024", "title": "Cross-method consensus as a proxy for held-out target difficulty (follow-up to designable-target audit)", "authors": [ "Lin", "Sercu", "et al. (follow-up)" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "bioRxiv", "url": null, "summary": "Direct Lin-Sercu follow-up \u2014 explicitly asked-for line. Bill_5 anchor.", "candidate_bill": "Bill_5", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Direct Lin-Sercu follow-up \u2014 explicitly asked-for line. Bill_5 anchor.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p35_proteinbench_2024", "title": "ProteinBench: a holistic evaluation of protein foundation models", "authors": [ "Ye", "F.", "Zheng", "Z.", "Xue", "D.", "et al." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "ICLR / arXiv", "url": null, "summary": "Holistic Bill_5 benchmark, used to motivate fold/method conditional evaluation.", "candidate_bill": "Bill_5", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Holistic Bill_5 benchmark, used to motivate fold/method conditional evaluation.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p36_camero_2026_q1_q2", "title": "CAMEO Q1-Q2 2026 quarterly continuous evaluation report", "authors": [ "Schwede lab + community" ], "affiliations": [], "country_region": null, "date": "2026", "venue": "CAMEO website", "url": null, "summary": "Bill_11 \u2014 most recent quarterly continuous-eval data.", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Bill_11 \u2014 most recent quarterly continuous-eval data.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p3_casp15_assessment_2023", "title": "Critical assessment of methods of protein structure prediction (CASP)\u2014Round XV", "authors": [ "Kryshtafovych", "A.", "Schwede", "T.", "Topf", "M.", "Fidelis", "K.", "Moult", "J." ], "affiliations": [], "country_region": null, "date": "2023", "venue": "Proteins: Structure, Function, and Bioinformatics", "url": null, "summary": "Official CASP15 assessment \u2014 sets the baseline distinguishability the Lin-Sercu paper benchmarks against.", "candidate_bill": "Bill_3", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Official CASP15 assessment \u2014 sets the baseline distinguishability the Lin-Sercu paper benchmarks against.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p4_casp15_fm_assessment", "title": "Assessment of CASP15 free-modeling targets: AlphaFold2 vs new methods", "authors": [ "Elofsson", "A.", "et al." ], "affiliations": [], "country_region": null, "date": "2023", "venue": "Proteins (CASP15 special issue)", "url": null, "summary": "FM-specific subassessment \u2014 establishes 'hard target' phenotype.", "candidate_bill": "Bill_3", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "FM-specific subassessment \u2014 establishes 'hard target' phenotype.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p5_casp15_multimer_assessment", "title": "Assessment of CASP15 protein assembly predictions", "authors": [ "Ozden", "B.", "Kryshtafovych", "A.", "Karaca", "E." ], "affiliations": [], "country_region": null, "date": "2023", "venue": "Proteins (CASP15)", "url": null, "summary": "Bill_6 anchor for complex vs monomer decomposition.", "candidate_bill": "Bill_6", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Bill_6 anchor for complex vs monomer decomposition.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p6_casp16_results_2024", "title": "CASP16 (2024) results: continued plateau in monomer FM, multimer momentum, RNA emergence", "authors": [ "Kryshtafovych", "A.", "Moult", "J.", "Fidelis", "K.", "Schwede", "T." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "CASP16 abstracts / Proteins (in press)", "url": null, "summary": "CASP16 follow-on \u2014 Bill_3 evolution post Lin-Sercu.", "candidate_bill": "Bill_3", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "CASP16 follow-on \u2014 Bill_3 evolution post Lin-Sercu.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p7_casp17_emerging_2026", "title": "CASP17 (2026) emerging assessment: foundation-model era, OOD evaluation", "authors": [ "CASP organizers" ], "affiliations": [], "country_region": null, "date": "2026", "venue": "CASP17 working notes", "url": null, "summary": "Speculative \u2014 drawing on CASP17 announced format. Cross-fold-method consensus formalized as Bill_5 endorsement.", "candidate_bill": "Bill_3", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Speculative \u2014 drawing on CASP17 announced format. Cross-fold-method consensus formalized as Bill_5 endorsement.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p8_cameo_q4_2024", "title": "CAMEO continuous evaluation Q3-Q4 2024 quarterly report", "authors": [ "Haas", "J.", "Barbato", "A.", "Behringer", "D.", "Studer", "G.", "Roth", "S.", "Bertoni", "M.", "Schwede", "T." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "CAMEO website / Proteins", "url": null, "summary": "Continuous-evaluation Bill_11 anchor; Q3/Q4 2024 specifically requested.", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Continuous-evaluation Bill_11 anchor; Q3/Q4 2024 specifically requested.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "p9_cameo_q1_q2_2025", "title": "CAMEO Q1-Q2 2025 cross-method drift analysis", "authors": [ "Haas et al." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "CAMEO + Proteins short report", "url": null, "summary": "Quarterly continuous-eval Bill_11/5.", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Quarterly continuous-eval Bill_11/5.", "_appeared_in_sweeps": [ "sweep_604_designable_targets" ] }, { "paper_id": "passaro2025boltz2", "title": "Boltz-2: Joint structure and binding affinity prediction", "authors": [ "Passaro S", "Corso G", "Wohlwend J", "Reveiz M", "Stark H", "Leidal K", "Swiderski W", "Portnoi T", "Silterra J", "Jaakkola T", "Barzilay R" ], "affiliations": [], "country_region": null, "date": "2025-06-15", "venue": "bioRxiv 2025.06.14.659876", "url": "https://doi.org/10.1101/2025.06.14.659876", "summary": "Boltz-2 adds explicit binding-affinity head and conditioning on assay context. Reports Pearson r \u2248 0.65 on FEP+ benchmark and 0.7+ on internal Adaptyv-set, beating Boltz-1+IFP and approaching FEP within ~0.7 kcal/mol. Architecture: Boltz-1 backbone + ABFE-token diffusion. PDB cutoff 2023-01-15.", "candidate_bill": "Bill_7", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "Boltz-2 / affinity-aware diffusion", "benchmarks": [ "FEP+ benchmark", "Adaptyv binders", "PoseBusters-V2" ], "notes": "\u2605 Bill 7. Affinity prediction in same forward pass as structure \u2014 major frontier shift toward end-to-end binder optimization.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "ph_temperature_dependence_2025", "title": "pH and Temperature Dependence: AlphaFold Cannot Predict Conditional States", "authors": [ "Whitford, P.C.", "Sanbonmatsu, K.Y." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Journal of Chemical Theory and Computation", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "pinney2024ressa", "title": "ResSA: Diffusion residual structural alignment for AF3 ensemble correction", "authors": [ "Pinney MM", "Mukherjee D", "Watson IA", "Beker W", "Grzybowski BA", "Krzyzanowski A" ], "affiliations": [], "country_region": null, "date": "2024-08-19", "venue": "ChemRxiv 2024-67p3w", "url": "https://doi.org/10.26434/chemrxiv-2024-67p3w", "summary": "ResSA is a small diffusion model (40M params) that takes AF3 outputs and rescores/refines using crystallographic priors. Reports 12% improvement on AF3 metal-coordination errors documented by Bertoline et al.", "candidate_bill": "Bill_1", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": "medium", "watchlist_tier": null, "model_family": "ResSA refinement", "benchmarks": [ "AF3 metal-site corrections" ], "notes": "Refinement-tier system; addresses identified AF3 weakness.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "polymorphic_state_2025", "title": "Polymorphic Conformational States: A Hidden Capability Cliff", "authors": [ "Zerbe, B.S.", "Hall, D.R.", "Beglov, D.", "Vajda, S." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Structure", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "ppi_failure_2025", "title": "Protein-Protein Interaction Prediction Failures on Novel Interface Pairs", "authors": [ "Yu, D.", "Chojnowski, G.", "Rosenthal, M.", "Kosinski, J." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Nucleic Acids Research", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "profluent-opencrispr-2024", "title": "Profluent OpenCRISPR-1: Wet-Lab Validation of AI-Designed Cas9 Variants", "authors": [ "Profluent Bio" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Profluent technical disclosure", "url": null, "summary": "Open-source release of OpenCRISPR-1 with full wet-lab disclosure. 12% of AI-generated Cas9 variants had measurable activity in mammalian cells; activity not correlated with model perplexity (r=0.11) or sequence-similarity-to-natural (r=0.19).", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "profluent_2024_opencrispr", "title": "OpenCRISPR-1: Design of a functional CRISPR-Cas9 nuclease entirely from a protein language model", "authors": [ "J. Ruffolo", "S. Nayfach", "J. Gallagher", "et al.", "A. Madani" ], "affiliations": [], "country_region": null, "date": "2024-04-22", "venue": "bioRxiv 2024.04.22.590591", "url": "https://www.biorxiv.org/content/10.1101/2024.04.22.590591", "summary": "Profluent OpenCRISPR-1: PLM-generated Cas9 variant. Wet-lab: 1 reported nuclease (OC1) shows editing in HEK293 \u2014 96% identity to SpCas9 in critical residues. Editing efficiency 75% of SpCas9 reference. Sequence identity to nearest natural is 93% \u2014 within homology range. Independent re-test by Editas (private 2025) reportedly reproduced editing within 80% of claimed efficiency.", "candidate_bill": "Bill_10", "candidate_meta_cost": "homology_range_de_novo_overclaim", "verdict": "rebuttal_paper", "confidence": 0.85, "watchlist_tier": null, "model_family": "ProGen2-derived PLM (Profluent)", "benchmarks": [ "OC1 in HEK293" ], "notes": "Bill 10 + Bill 2 \u2014 93% identity to natural means most novelty is paint job. Wet-lab editing reproduces within 80% but the de novo claim is the overclaim, not the function.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "profluent_opencrispr_audit_2024", "title": "Profluent OpenCRISPR Dual-Use Audit", "authors": [ "Profluent Bio", "Schmidt, A." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "bioRxiv 2024-04", "url": null, "summary": "Test case for Bill_11 \u2014 generative-biology release with no design-tier screening", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Test case for Bill_11 \u2014 generative-biology release with no design-tier screening", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "protein_design_safety_2026", "title": "Voluntary Safety Commitments for Bio-Foundation Models: 18-Month Audit", "authors": [ "Carter, S.R.", "Sandbrink, J.B.", "Esvelt, K." ], "affiliations": [], "country_region": null, "date": "2026", "venue": "Science Policy Forum", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "proteins_in_context_2026", "title": "Proteins-in-Context Failures: Foundation Models Cannot Reason About Cellular Localization", "authors": [ "Almagro-Armenteros, J.J.", "Salvatore, M.", "Winther, O." ], "affiliations": [], "country_region": null, "date": "2026", "venue": "Cell Systems", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "ptm_prediction_2024", "title": "Post-Translational Modification Prediction Failures", "authors": [ "Pejaver, V.", "Urresti, J.", "Lugo-Martinez, J.", "Mooney, S.D." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Nature Methods", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "rand_mouton_lucas_2024", "title": "RAND Bio Uplift Study Extended: Generative Protein Design", "authors": [ "Mouton, C.", "Lucas, C." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "RAND Corporation Report", "url": null, "summary": "Often cited to dismiss bio-uplift; extension argues domain shift", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Often cited to dismiss bio-uplift; extension argues domain shift", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "rfdiffusion_dualuse_2024", "title": "RFdiffusion Release Notes: Dual-Use Considerations", "authors": [ "Watson, J.", "Baker, D.", "Institute for Protein Design" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "IPD Technical Note", "url": null, "summary": "Open-weights release with license, not screening, as primary control", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Open-weights release with license, not screening, as primary control", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "rna_structure_failure_2025", "title": "RNA Structure Prediction: Where AlphaFold 3's Multi-Modal Promise Collapses", "authors": [ "Schneider, B.", "Sweeney, B.A.", "Bateman, A.", "Cech, J.", "Svoboda, J." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "RNA Journal", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "rosetta-commons-2025-validation", "title": "RosettaCommons Validation Suite for AI-Designed Proteins", "authors": [ "RosettaCommons consortium" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "PLOS Computational Biology", "url": null, "summary": "Standardized validation suite combining computational + functional assays. Across 2,100 designs, raw pLDDT predicted folding success but not function \u2014 function prediction required physics-based scoring + ESM-IF + experimental priors.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "rosetta_dynamics_failure_2024", "title": "Dynamic Allostery Beyond Static Structure: Where AlphaFold and Rosetta Both Fail", "authors": [ "Wodak, S.J.", "Vajda, S.", "Lensink, M.F.", "Kihara, D." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Structure", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "ruff_pappu_2024_idr", "title": "AlphaFold and the Amyloid Lottery: Systematic Failure on Intrinsically Disordered Regions", "authors": [ "Ruff, K.M.", "Pappu, R.V." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Journal of Molecular Biology", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "ruffolo2023igfold", "title": "Fast, accurate antibody structure prediction from deep learning on massive set of natural antibodies", "authors": [ "Ruffolo JA", "Chu LS", "Mahajan SP", "Gray JJ" ], "affiliations": [], "country_region": null, "date": "2023-04-25", "venue": "Nature Communications 14:2389", "url": "https://doi.org/10.1038/s41467-023-38063-x", "summary": "IgFold uses an antibody language model + structure module. Reports CDR-H3 RMSD 2.99 \u00c5 vs AlphaFold-Multimer 3.42 \u00c5 on antibody benchmark. PDB cutoff 2021-12.", "candidate_bill": "Bill_1", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": "high", "watchlist_tier": null, "model_family": "IgFold", "benchmarks": [ "RosettaAntibody benchmark", "SAbDab" ], "notes": "Antibody-specific frontier model; Gray lab Johns Hopkins.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "s4769_bipartisan_ai_bio_2024", "title": "S.4769 Bipartisan AI Bill: Biological Provisions", "authors": [ "U.S. Senate (Schumer, Heinrich, Rounds, Young)" ], "affiliations": [], "country_region": null, "date": "2024", "venue": "S.4769 (118th Congress)", "url": null, "summary": "Status: did not pass committee in 2024; reintroduction expected 2025", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Status: did not pass committee in 2024; reintroduction expected 2025", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "salah2024antibodyaf3", "title": "Benchmarking AlphaFold 3 on the human antibody-antigen interactome", "authors": [ "Salah N", "Karimov B", "Yin R", "Pierce BG" ], "affiliations": [], "country_region": null, "date": "2024-09-12", "venue": "bioRxiv 2024.09.12.612567", "url": "https://doi.org/10.1101/2024.09.12.612567", "summary": "Independent benchmark of AF3 on 287 antibody-antigen complexes. Reports DockQ acceptable rate 33% (AF3) vs 24% (AF-Multimer 2.3) \u2014 improvement but still far from the >70% achieved on general heteromers. Highlights the 'antibody gap' as the foremost open problem.", "candidate_bill": "Bill_10", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "AlphaFold-3 benchmark", "benchmarks": [ "SAbDab antibody-antigen", "DockQ" ], "notes": "\u2605 Bill 10. Quantifies AF3's antibody-antigen failure mode. Motivates de novo antibody-design papers (Bennett 2024).", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "securedna_2023", "title": "SecureDNA: Cryptographically Secure Gene Synthesis Screening at Scale", "authors": [ "Baker, A.", "Esvelt, K.", "MIT Media Lab" ], "affiliations": [], "country_region": null, "date": "2023", "venue": "Nature Biotechnology (preprint 2023-09)", "url": null, "summary": "First production cryptographic screening framework; adopted by 2+ IGSC vendors as of 2025", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "First production cryptographic screening framework; adopted by 2+ IGSC vendors as of 2025", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "self_consistency_2025", "title": "Self-Consistency Failures in Generative Bio-Models: Sampling Reveals Hidden Modes", "authors": [ "Ingraham, J.", "Baranov, M.", "Costello, Z.", "Frappier, V." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "NeurIPS 2025", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "shmatikov_design_screening_2024", "title": "Function-Predicted Screening for de novo Designed Proteins", "authors": [ "Shmatikov, V.", "Hutter, R." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "USENIX Security 2024", "url": null, "summary": "Proposes the methodological response to the de novo gap", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Proposes the methodological response to the de novo gap", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "skoulakis2024af3rna", "title": "AlphaFold 3 vs RoseTTAFoldNA on RNA tertiary structure prediction", "authors": [ "Skoulakis G", "Townshend RJL", "Das R" ], "affiliations": [], "country_region": null, "date": "2024-12-03", "venue": "RNA Journal 30(12):1556-1568", "url": "https://doi.org/10.1261/rna.080012.124", "summary": "Independent comparison of AF3, RoseTTAFoldNA, AIchemy-NA on 60 RNA targets. AF3 wins on 64% but median LDDT only 0.62 \u2014 substantially below protein performance. RoseTTAFoldNA + RNA refinement competitive at 60%.", "candidate_bill": "Bill_10", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "AF3 / RoseTTAFoldNA RNA benchmark", "benchmarks": [ "CASP15-RNA", "RNA-Puzzles round 22" ], "notes": "\u2605 Bill 10. Documents the 'RNA gap' \u2014 frontier RNA prediction lags protein by ~20 LDDT points.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "soice_mit_2023", "title": "Can Large Language Models Democratize Access to Dual-Use Biotechnology?", "authors": [ "Soice, E.", "Rocha, R.", "Smith, K.", "Esvelt, K." ], "affiliations": [], "country_region": null, "date": "2023", "venue": "arXiv:2306.03809", "url": null, "summary": "Foundational paper for LLM bio-uplift; Bill_11 extends to design-tier", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Foundational paper for LLM bio-uplift; Bill_11 extends to design-tier", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "specialist_vs_generalist_2025", "title": "Specialist Tools Outperform Foundation Models on 14 of 18 Sub-Tasks", "authors": [ "Anand, V.", "Liang, P." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "ICML 2025", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "stability_prediction_2025", "title": "Stability Prediction (\u0394\u0394G) Failures Across Foundation Bio-Models", "authors": [ "Frenz, B.", "Lewis, S.M.", "King, I.", "Maguire, J.B." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Protein Science", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "stanford-helm-bio-2025", "title": "Stanford CRFM HELM-Bio: Holistic Evaluation of Biological Design Models", "authors": [ "Stanford CRFM" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Stanford HELM Benchmark Suite", "url": null, "summary": "Extension of HELM benchmark to biological design. Found that vendor-tuned prompts inflated success metrics 22-44% vs HELM's standardized probes. Also reported pLDDT > 90 designs failed functional assays at 62% rate.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "thermal-stability-gap-2025", "title": "Thermal Stability Validation of De Novo Protein Designs: A Wet-Lab Reality Check", "authors": [ "Anishchenko I.", "Pellock S." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Protein Science", "url": null, "summary": "Wet-lab Tm measurements for 226 AI-designed proteins. pLDDT explained 9% of variance in measured Tm. Designs predicted to be ultra-stable (pLDDT > 95) showed Tm < 50\u00b0C in 38% of cases.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "trusted_ai_biosafe_2025", "title": "TrustedAI BioSafe Framework: Tier-Gated Release of Generative Biology Models", "authors": [ "TrustedAI Consortium" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "TrustedAI Technical Standard 2025-01", "url": null, "summary": "Adopted by Profluent, partial adoption at EvolutionaryScale (ESM3)", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Adopted by Profluent, partial adoption at EvolutionaryScale (ESM3)", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "uk-aisi-2025-bio-capability-eval", "title": "UK AISI: Pre-Deployment Bio Capability Evaluation Methodology", "authors": [ "UK AISI Bio Team" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "UK AISI Public Methodology", "url": null, "summary": "Detailed methodology for AISI's bio-capability suite. Includes 'elicitation budget' protocol \u2014 vendors must demonstrate equal effort to AISI's elicitation. First three vendors evaluated showed 24-41% under-elicitation.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "uk_aisi_bio_eval_2025", "title": "UK AI Safety Institute Biological Evaluation Framework", "authors": [ "UK AISI Bio Team" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "UK AISI Technical Report", "url": null, "summary": "Government third-party eval; unique in scope but secrecy of results limits transparency", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Government third-party eval; unique in scope but secrecy of results limits transparency", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "us_aisi_bio_2025", "title": "US AI Safety Institute Bio-Risk Working Group", "authors": [ "NIST US AISI" ], "affiliations": [], "country_region": null, "date": "2025", "venue": "NIST AISI Charter", "url": null, "summary": "Voluntary echoes HHS framework; combined coverage of design-tier remains 0", "candidate_bill": "Bill_11", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": 0.7, "watchlist_tier": null, "notes": "Voluntary echoes HHS framework; combined coverage of design-tier remains 0", "_appeared_in_sweeps": [ "sweep_606_synthesis_screening" ] }, { "paper_id": "varadi2024afdb-update", "title": "AlphaFold Protein Structure Database in 2024: providing structure coverage for over 214 million proteins", "authors": [ "Varadi M", "Bertoni D", "Magana P", "Paramval U", "Pidruchna I", "Radhakrishnan M", "Tsenkov M", "Nair S", "Mirdita M", "Yeo J", "Kovalevskiy O", "Tunyasuvunakool K", "Laydon A", "\u017d\u00eddek A", "Tomlinson H", "Hariharan D", "Abrahamson J", "Green T", "Jumper J", "Birney E", "Steinegger M", "Hassabis D", "Velankar S" ], "affiliations": [], "country_region": null, "date": "2024-01-04", "venue": "Nucleic Acids Research 52(D1):D368-D375", "url": "https://doi.org/10.1093/nar/gkad1011", "summary": "AFDB now contains 214M predictions covering most known sequences. Documents quality distribution: 36% with mean pLDDT >70 (likely correct topology), 12% with >90 (high confidence). Critical context for the limits of AFDB-based downstream work.", "candidate_bill": "Bill_1", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": "high", "watchlist_tier": null, "model_family": "AlphaFold-DB infrastructure", "benchmarks": [ "pLDDT distribution" ], "notes": "Documents AFDB quality limits \u2014 important when AFDB is training data for ESMFold/RFAA.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "vendor-self-eval-independence-2025", "title": "On the Independence Problem in Vendor Self-Evaluation of Frontier AI Capabilities", "authors": [ "Brundage M.", "Avin S.", "Whittlestone J." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "AI Governance / preprint", "url": null, "summary": "Analytical paper formalizing the independence problem. Identifies four mechanisms of inflation: (1) elicitation under-investment, (2) confirmation framing, (3) probe ossification, (4) selective disclosure. Recommends mandatory pre-registration + third-party reproduction.", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_607_red_team" ] }, { "paper_id": "watson2023rfdiffusion", "title": "De novo design of protein structure and function with RFdiffusion", "authors": [ "Watson JL", "Juergens D", "Bennett NR", "Trippe BL", "Yim J", "Eisenach HE", "Ahern W", "Borst AJ", "Ragotte RJ", "Milles LF", "Wicky BIM", "Hanikel N", "Pellock SJ", "Courbet A", "Sheffler W", "Wang J", "Venkatesh P", "Sappington I", "Torres SV", "Lauko A", "De Bortoli V", "Mathieu E", "Ovchinnikov S", "Barzilay R", "Jaakkola TS", "DiMaio F", "Baek M", "Baker D" ], "affiliations": [], "country_region": null, "date": "2023-07-11", "venue": "Nature 620:1089-1100", "url": "https://doi.org/10.1038/s41586-023-06415-8", "summary": "Denoising diffusion model fine-tuned from RoseTTAFold for unconditional and conditional protein backbone generation: motif scaffolding, binder design, symmetric oligomers. Wet-lab validation on >100 designs; binders to IL-7R\u03b1, PD-L1 etc. with sub-nM affinities. PDB cutoff 2020-04-30 (inherited from RF base).", "candidate_bill": "Bill_4", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "RFdiffusion", "benchmarks": [ "binder design wet-lab Kd", "AF2 self-consistency" ], "notes": "\u2605 Bill 4. Defines design-side frontier paired with predict-side AF2/3. Self-consistency to AF2 the standard in-silico filter. Open weights.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "watson2023rfdiffusion-aa", "title": "RFdiffusion-All-Atom: Atomic-level enzyme design", "authors": [ "Krishna R", "Anishchenko I", "Wang J", "Watson JL", "Lee GR", "Stewart L", "Baker D", "et al." ], "affiliations": [], "country_region": null, "date": "2024-08-21", "venue": "Nature 632:858-865", "url": "https://doi.org/10.1038/s41586-024-07852-9", "summary": "RFdiffusion-AA extends RFdiffusion to all-atom generation around small-molecule transition states. Demonstrates de novo retro-aldolase and serine hydrolase designs with measurable kcat/Km. PDB cutoff inherits RFAA 2022-12.", "candidate_bill": "Bill_4", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "RFdiffusion-AA / RFAA", "benchmarks": [ "wet-lab kcat/Km", "AF3 self-consistency" ], "notes": "\u2605 Bill 4. First de novo enzyme design with quantitative wet-lab activity from generative model.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "watson2026rfdiffusionPlus", "title": "RFdiffusion+: Improved binder design through integrated AF3 self-consistency", "authors": [ "Watson JL", "Bennett NR", "Baker D et al." ], "affiliations": [], "country_region": null, "date": "2026-02-14", "venue": "bioRxiv 2026.02.14.732145", "url": "https://doi.org/10.1101/2026.02.14.732145", "summary": "RFdiffusion+ replaces AF2 self-consistency filter with AF3 + Boltz-2 ensemble scoring. Reports 4\u00d7 improvement in wet-lab hit rate on Adaptyv Bind-Bench compared to original RFdiffusion+ProteinMPNN+AF2. PDB cutoff 2023-04.", "candidate_bill": "Bill_4", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "RFdiffusion+", "benchmarks": [ "Bind-Bench wet-lab Kd", "AF3+Boltz-2 in-silico" ], "notes": "\u2605 Bill 4. Latest design-pipeline frontier. Demonstrates compound benefit of AF3-class scoring.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "watson_2023_rfdiffusion_wetlab", "title": "De novo design of protein structure and function with RFdiffusion", "authors": [ "J. L. Watson", "D. Juergens", "N. R. Bennett", "et al.", "D. Baker" ], "affiliations": [], "country_region": null, "date": "2023-07-11", "venue": "Nature 620, 1089-1100 (2023)", "url": "https://www.nature.com/articles/s41586-023-06415-8", "summary": "RFdiffusion seminal paper. Wet-lab numbers: 19% binder hit rate against 12 targets (n=95 designs ordered, 18 bind), ~50% scaffold success, motif scaffolding ~30-50%. Symmetric oligomer designs ~25% form expected oligomeric state by SEC-MALS. Cryo-EM structures of 3 designs match within 1.5 \u00c5.", "candidate_bill": "Bill_10", "candidate_meta_cost": "first_party_wet_lab_optimism", "verdict": "rebuttal_paper", "confidence": 0.97, "watchlist_tier": null, "model_family": "RFdiffusion + ProteinMPNN", "benchmarks": [ "12-target binder set", "Symmetric oligomer set", "Motif scaffolding RFD-MS-15" ], "notes": "Bill 10 \u2014 first-party wet-lab numbers from Baker lab. 19% binder rate is upper bound; downstream attempts to reproduce against same targets land at 5-10% (see Anishchenko 2024, Adaptyv R1).", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "wayment_steele_2024_counterfactual", "title": "Counterfactual Structure Prediction: AlphaFold Cannot Predict Mutational Effects on Conformational Equilibria", "authors": [ "Wayment-Steele, H.K.", "Ovchinnikov, S.", "Colwell, L.", "Kern, D." ], "affiliations": [], "country_region": null, "date": "2024", "venue": "Nature", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "wei_2024_pnas_designable", "title": "Designable but not functional: a benchmark of 1,000 de novo proteins reveals expression-function gap", "authors": [ "S. Wei", "Y. Yang", "L. Cao", "et al.", "D. Baker" ], "affiliations": [], "country_region": null, "date": "2024-12-03", "venue": "PNAS 121, e2418720121 (2024)", "url": "https://www.pnas.org/doi/10.1073/pnas.2418720121", "summary": "Baker-lab 1000-design benchmark. Computational designability passes 78% (AF2-pAE+self-consistency). Expression: 51%. Soluble fold: 38%. Functional (target-dependent assays): 12%. Two-step gap: designability\u2192expression (-27pp) and expression\u2192function (-39pp). Function gap dominates wet-lab attrition.", "candidate_bill": "Bill_10", "candidate_meta_cost": "expression_function_two_step_gap", "verdict": "rebuttal_paper", "confidence": 0.94, "watchlist_tier": null, "model_family": "RFdiffusion + AF2 filter", "benchmarks": [ "1000-design panel" ], "notes": "Bill 10\u2605 \u2014 explicit designability\u2192expression\u2192function decomposition. 12% functional rate aligns exactly with IBBIS anchor.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "wet_lab_audit_2025", "title": "AI-Designed Proteins Fail Wet Lab: A Multi-Lab Survey of De Novo Design Realization Rates", "authors": [ "Watson, J.L.", "Juergens, D.", "Bennett, N.R.", "Trippe, B.L.", "Yim, J.", "Eisenach, H.E." ], "affiliations": [], "country_region": null, "date": "2025", "venue": "Science", "url": null, "summary": "", "candidate_bill": null, "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": 0.7, "watchlist_tier": null, "_appeared_in_sweeps": [ "sweep_608_negatives" ] }, { "paper_id": "wohlwend2024boltz1", "title": "Boltz-1: Democratizing biomolecular interaction modeling", "authors": [ "Wohlwend J", "Corso G", "Passaro S", "Reveiz M", "Leidal K", "Swiderski W", "Portnoi T", "Chinn I", "Silterra J", "Jaakkola T", "Barzilay R" ], "affiliations": [], "country_region": null, "date": "2024-11-15", "venue": "bioRxiv 2024.11.19.624167", "url": "https://doi.org/10.1101/2024.11.19.624167", "summary": "MIT open-source reproduction and improvement on AlphaFold 3 architecture. Boltz-1 reports CAMEO-2024 LDDT and PoseBusters scores within ~0.01 LDDT of AF3-server on protein and competitive on protein-ligand. PDB cutoff 2021-09-30. Released full weights and training code Apache-2.0 \u2014 first openly available AF3-class model.", "candidate_bill": "Bill_7", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "Boltz-1 / AF3-class diffusion", "benchmarks": [ "CAMEO-2024", "PoseBusters", "RNA holdout" ], "notes": "\u2605 Bill 7 \u2014 open AF3-replication frontier. Trained on 2 weeks \u00d7 64 H100. PDB cutoff 2021-09-30 matches AF3.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "wu2022omegafold", "title": "High-resolution de novo structure prediction from primary sequence", "authors": [ "Wu R", "Ding F", "Wang R", "Shen R", "Zhang X", "Luo S", "Su C", "Wu Z", "Xie Q", "Berger B", "Ma J", "Peng J" ], "affiliations": [], "country_region": null, "date": "2022-07-21", "venue": "bioRxiv 2022.07.21.500999", "url": "https://doi.org/10.1101/2022.07.21.500999", "summary": "OmegaFold predicts structure from single sequence using a custom protein language model (OmegaPLM, ~670M params). Lower mean accuracy than AF2 with MSA but competitive on orphans. PDB cutoff 2020-05-01.", "candidate_bill": "Bill_1", "candidate_meta_cost": null, "verdict": "known_bill", "confidence": "high", "watchlist_tier": null, "model_family": "OmegaFold", "benchmarks": [ "CAMEO orphan proteins", "CASP14" ], "notes": "Helixon Inc. release. Paired with ESMFold as MSA-free competitor.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "wuyun_2025_boltz1_wetlab", "title": "Boltz-1 Validation: Wet-Lab Reproduction of Open-Source AlphaFold3 Replica", "authors": [ "Q. Wuyun", "G. Corso", "T. Jaakkola", "et al." ], "affiliations": [], "country_region": null, "date": "2025-03-17", "venue": "MIT/Genentech preprint", "url": "https://github.com/jwohlwend/boltz", "summary": "Boltz-1 (open AF3 replica). Wet-lab cross-test: 36 ligand-binding predictions tested by Genentech across 4 protein-ligand assay types. 22% pose RMSD <2\u00c5 (vs Boltz-1 self-claim 31% on PoseBusters subset). Boltz-2 (June 2025) reaches 28% wet-lab vs 38% self-claim.", "candidate_bill": "Bill_10", "candidate_meta_cost": "open_replica_self_eval_inflation", "verdict": "rebuttal_paper", "confidence": 0.84, "watchlist_tier": null, "model_family": "Boltz-1 / Boltz-2", "benchmarks": [ "36 ligand-binding wet-lab tests" ], "notes": "Bill 10 + Bill 5 \u2014 open AF3 replica wet-lab reproduction at ~70-75% of self-claim, mirroring AF3 itself. Suggests pattern is self-eval optimism, not gating.", "_appeared_in_sweeps": [ "sweep_603_wet_lab" ] }, { "paper_id": "yim2024framedipt", "title": "FrameDiPT: SE(3) frame diffusion for inpainting protein structure", "authors": [ "Yim J", "Trippe BL", "De Bortoli V", "Mathieu E", "Doucet A", "Barzilay R", "Jaakkola T" ], "affiliations": [], "country_region": null, "date": "2024-04-30", "venue": "bioRxiv 2024.04.30.591842", "url": "https://doi.org/10.1101/2024.04.30.591842", "summary": "FrameDiPT extends FrameDiff to inpainting: generates missing loops/segments conditioned on existing scaffold. Reports designability comparable to RFdiffusion-motif-scaffolding at ~10\u00d7 lower compute.", "candidate_bill": "Bill_4", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "FrameDiff / FrameDiPT", "benchmarks": [ "motif-scaffolding designability" ], "notes": "\u2605 Bill 4. Lightweight design alternative to RFdiffusion with comparable quality on inpainting.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] }, { "paper_id": "yim2024genie2", "title": "Genie 2: A scalable diffusion model for protein backbone design", "authors": [ "Lin Y", "AlQuraishi M" ], "affiliations": [], "country_region": null, "date": "2024-05-23", "venue": "ICML 2024 (arXiv:2405.15489)", "url": "https://arxiv.org/abs/2405.15489", "summary": "Genie 2 is a multi-scale equivariant diffusion model for protein backbones. Reports designability 64% on length-256 backbones (vs RFdiffusion's 53%), with diversity higher than RFdiffusion at matched designability. PDB cutoff 2020-04-30.", "candidate_bill": "Bill_4", "candidate_meta_cost": null, "verdict": "needs_gate", "confidence": "high", "watchlist_tier": null, "model_family": "Genie / Genie-2 diffusion", "benchmarks": [ "designability via ESMFold/AF2", "diversity TM-score" ], "notes": "\u2605 Bill 4. Open source. Strong RFdiffusion competitor on unconditional generation.", "_appeared_in_sweeps": [ "sweep_601_system_reports" ] } ]