Bibliography
31 sources: the literature, benchmarks and guidance that this registry's methodology and its companion essay cite. Sources about a single molecule are on that molecule's record instead. Further reading, at the end, lists work the maintainer's research drew on that neither cites.
Download: CSV · BibTeX · CSL-JSON
Other counts of AI molecules in the clinic
- Jayatunga, M.K.P.; Ayers, M.; Bruens, L.; et al. (2024). How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons. Drug Discovery Today. doi:10.1016/j.drudis.2024.104009 Counted 75 molecules from AI-native companies that entered the clinic from 2015 to 2023 and estimated their Phase I and II success rates; its list also takes in molecules this registry codes as platform-only or physics-based, so its total sits above the strict count.
- Greenblatt, E.; Geoui, T. (2026). Dataset and analysis code for "28 Months Later: AI-Discovered, but Not AI-Developed. A citation-backed re-enumeration of the Jayatunga et al. clinical-trial cohort". Zenodo. doi:10.5281/zenodo.22814924 Re-checked the Jayatunga cohort with a citation for every field and added assets to reach 83, which the authors treat as a lower bound; 26 of its entries are strict in this registry.
- Haque, W.; Sai-Palm, K.; Rana, I.; et al. (2026). Artificial intelligence-enabled drug discovery: A growing but largely early-stage clinical pipeline. Cancer Treatment Reviews. doi:10.1016/j.ctrv.2026.103223 Counts 129 AI-enabled therapeutic assets from 69 companies that had entered an interventional trial by 1 March 2026, updating the authors' ASCO 2026 abstract (117 assets, 63 companies, to 1 July 2025). Reports observed success of 77.8% in Phase 1 and 19.5% in Phase 2, and 37% of assets on a novel target. It counts any meaningful AI contribution to discovery, design or optimisation, including repurposing, on company or third-party evidence, so it compares with this registry's strict and borderline entries together. Its supplement lists the assets; the crosswalk places each one here.
- Wellcome; Boston Consulting Group (2023). Unlocking the potential of AI in Drug Discovery: Current status, barriers and future opportunities. cms.wellcome.org BCG report for Wellcome whose list of 72 AI-derived clinical molecules is one of the four lists in this registry's crosswalk, and whose scenario model is the source of the 25–50% time and cost savings figure the essay traces.
Perspectives and reviews
- Bender, A.; Thomas, M.C.; Scannell, J.W.; et al. (2026). Artificial intelligence in drug discovery — what it is, where we stand and the path forward. Nature Reviews Drug Discovery. doi:10.1038/s41573-026-01496-2 Perspective by 16 authors, among them Jack Scannell, judging that AI has so far shown little clinically relevant impact on drug discovery and that effort has followed the available labeled data rather than the steps where drugs fail.
- Scannell, J.W.; Blanckley, A.; Boldon, H.; et al. (2012). Diagnosing the decline in pharmaceutical R&D efficiency. Nature Reviews Drug Discovery. doi:10.1038/nrd3681 Named Eroom's law, the halving of new drugs per billion dollars of R&D roughly every nine years since 1950, and listed a 'basic research–brute force' bias among the causes, a diagnosis the essay applies to AI throughput.
- Scannell, J.W.; Bosley, J. (2016). When Quality Beats Quantity: Decision Theory, Drug Discovery, and the Reproducibility Crisis. PLOS ONE. doi:10.1371/journal.pone.0147215 Decision-theory model showing that a small gain in how well a screening model predicts clinical outcome can outweigh a 10- to 100-fold gain in throughput, the core of the essay's validity argument.
- Cavalla, D. (2013). Predictive methods in drug repurposing: gold mine or just a bigger haystack?. Drug Discovery Today. doi:10.1016/j.drudis.2012.12.009 Early assessment of computational drug repurposing that asked whether prediction yields better candidates or merely more of them, the question the essay puts to AI discovery.
- Koller, D. (2026). Drug Discovery Has No Magic Wands. Deep Phenotype (Substack). deepphenotype.substack.com Essay by insitro's founder arguing that faster molecule design aimed at the wrong biological mechanisms only produces faster failures.
- Devereson, A.; Stanzl, E.; Yang, G. (2026). AI drug discovery: focusing on what matters most. McKinsey & Company. mckinsey.com Analysis of AI–pharma deal values and of nine AI programs that reached early milestones fast; all nine are strict in this registry, and the essay notes that three have since been stopped or shelved.
- Henry, A. (2026). We only understand about 20% of biology. So, what exactly are we feeding AI?. LinkedIn. linkedin.com UCB's chief scientific officer estimates that we understand no more than a fifth of basic biology and argues that more data about a poorly understood system does not by itself add understanding.
- Booth, B. (2026). The Fruits of AI Drug Abundance: Indigestion In The Near Term. LifeSciVC (blog). lifescivc.com Atlas Venture partner's prediction that AI, by speeding molecule design around known targets, will crowd hot targets with similar programs and lower industry R&D productivity for several years before raising it.
Benchmarks
- Wong, C.H.; Siah, K.W.; Lo, A.W. (2019). Estimation of clinical trial success rates and related parameters. Biostatistics. doi:10.1093/biostatistics/kxx069 Estimated phase-transition success rates from trial records for 2000–2015 (Phase I 66.4%, Phase II 58.3%) with a path-by-path method that imputes missing phases; the essay's main historical benchmark for Phase I.
- Biotechnology Innovation Organization; Informa Pharma Intelligence; QLS Advisors (2021). Clinical Development Success Rates and Contributing Factors 2011–2020. bio.org Industry benchmark from 12,728 phase transitions in 2011–2020 that puts Phase II success at 28.9%, the low end of the historical range the essay sets AI-derived molecules against.
- Deloitte (2026). Measuring the return from pharmaceutical innovation (16th edition). deloitte.com Annual analysis of R&D returns at the 20 largest drugmakers; source of the essay's average cost of about $2.7bn per new drug in 2025, counting the cost of failures.
- Minikel, E.V.; Painter, J.L.; Dong, C.C.; et al. (2024). Refining the impact of genetic evidence on clinical success. Nature. doi:10.1038/s41586-024-07316-0 Found that drug mechanisms with human genetic support are about 2.6 times as likely to succeed from Phase I to launch, the essay's strongest evidence that better target validity changes outcomes.
- U.S. National Library of Medicine (2026). ClinicalTrials.gov API v2. ClinicalTrials.gov. clinicaltrials.gov Source of every record in trials.csv and of the industry comparator (18,194 industry-sponsored Phase 1–2 drug trials started 2018–2024, 18.6% of closed trials terminated) against which the registry's termination rate is set.
Methods and regulatory evidence
- International Council for Harmonisation (2023). Assessment and control of DNA reactive (mutagenic) impurities in pharmaceuticals to limit potential carcinogenic risk M7(R2). ICH. database.ich.org Harmonized guideline under which two complementary computational predictions, one expert rule-based and one statistical, can stand in for Ames testing of some drug impurities; the essay's example of computation a regulator accepts.
- Hasselgren, C.; Bercu, J.; Cayley, A.; et al. (2020). Management of pharmaceutical ICH M7 (Q)SAR predictions – The impact of model updates. Regulatory Toxicology and Pharmacology. doi:10.1016/j.yrtph.2020.104807 Industry analysis of how updates to the expert-system and statistical (Q)SAR models used under ICH M7 change their predictions, which supports the essay's point that M7 relies on a trained statistical model.
- Li, Y.; Sun, H.; Zhang, Z. (2025). The Evolution and Future Directions of PBPK Modeling in FDA Regulatory Review. Pharmaceutics. doi:10.3390/pharmaceutics17111413 Analysis of FDA review documents finding that 65 of the 245 novel drugs approved in 2020–24 submitted PBPK models as pivotal evidence, the essay's example of modeling accepted in place of some clinical studies.
- Furuhama, A.; Kitazawa, A.; Yao, J.; et al. (2023). Evaluation of QSAR models for predicting mutagenicity: outcome of the Second Ames/QSAR international challenge project. SAR and QSAR in Environmental Research. doi:10.1080/1062936X.2023.2284902 Second Ames/QSAR challenge, in which 21 models predicted the mutagenicity of 1,589 chemicals blind and averaged 46% sensitivity and 88% specificity; the essay's evidence that computational filters err in both directions.
- Monticello, T.M.; Jones, T.W.; Dambach, D.M.; et al. (2017). Current nonclinical testing paradigm enables safe entry to First-In-Human clinical trials: The IQ consortium nonclinical to clinical translational database. Toxicology and Applied Pharmacology. doi:10.1016/j.taap.2017.09.006 IQ Consortium database comparing nonclinical findings with first-in-human outcomes, in which animal studies flagged 48% of the adverse effects later seen in people; the essay's measure of the incumbent filter.
- Lyu, J.; Wang, S.; Balius, T.E.; et al. (2019). Ultra-large library docking for discovering new chemotypes. Nature. doi:10.1038/s41586-019-0917-9 Ultra-large docking study that made and tested molecules across the score range, low scorers included, and saw the hit rate fall to zero; the essay's example of a model's rejections being tested prospectively.
- Neves, P.; McClure, K.; Verhoeven, J.; et al. (2022). Global Reactivity Models are Impactful in Industrial Synthesis Applications. ChemRxiv. doi:10.26434/chemrxiv-2022-5775s Janssen study of reaction-yield models that ran reactions the model predicted would fail as a negative control, the closest published case the essay's fact check found to a company testing its model's rejections.
- Mas, P.; Filoche-Rommé, B.; Bianciotto, M.; et al. (2026). Back to the future of lead optimization: benchmarking compound prioritization strategies. Digital Discovery. doi:10.1039/d5dd00387c Benchmark of compound prioritization strategies noting that compounds designed but never made cannot be scored without an oracle, the measurement gap behind the essay's call for a reject-pile holdout.
- Jin, Q.; Wang, Z.; Floudas, C.S.; et al. (2024). Matching patients to clinical trials with large language models. Nature Communications. doi:10.1038/s41467-024-53081-z TrialGPT, a large-language-model tool for matching patients to trials that reports how well it excludes as well as matches, the reason the essay says such tools are rarely, not never, audited for wrongful screen-outs.
Other sources cited in the essay
- Ishino, Y.; Shinagawa, H.; Makino, K.; et al. (1987). Nucleotide sequence of the iap gene, responsible for alkaline phosphatase isozyme conversion in Escherichia coli, and identification of the gene product. Journal of Bacteriology. doi:10.1128/jb.169.12.5429-5433.1987 First report, in E. coli, of the repeated sequences later named CRISPR, the start of the essay's 36-year span from curiosity to an approved CRISPR medicine.
- Jinek, M.; Chylinski, K.; Fonfara, I.; et al. (2012). A Programmable Dual-RNA–Guided DNA Endonuclease in Adaptive Bacterial Immunity. Science. doi:10.1126/science.1225829 Showed in vitro that Cas9 can be programmed with a guide RNA to cut chosen DNA, the start of the essay's 11-year span from tool to approved medicine.
- Yoon, P.H.; Athukoralage, J.S.; Ameisen, E.; et al. (2026). Autonomous AI agents discover reverse transcriptases with tandem repeat arrays. bioRxiv. doi:10.64898/2026.10.05.755809 Preprint behind Anthropic's announcement that Claude agents found reverse transcriptases with tandem repeat arrays, including the reruns that missed the array; the essay's case study of AI throughput versus validated discovery.
- Korn, A.M.; Hillhouse, A.E.; Sun, L.; et al. (2021). Comparative Genomics of Three Novel Jumbo Bacteriophages Infecting Staphylococcus aureus. Journal of Virology. doi:10.1128/jvi.02391-20 Genomic description of three jumbo Staphylococcus aureus phages, cited by the essay to show that the enzyme at the center of Anthropic's finding had already been described in 2021.
Further reading
Not cited in the essay or the methodology; part of the research behind them.
- Bender, A.; Cortés-Ciriano, I. (2021). Artificial intelligence in drug discovery: what is realistic, what are illusions? Part 1: Ways to make an impact, and why we are not there yet. Drug Discovery Today. doi:10.1016/j.drudis.2020.12.009 Estimates which stages of drug discovery matter most and finds that clinical success rates dominate: deciding which compound to take forward, and how to run its trials, matters more than speed or cost. AI has so far focused on how to make a compound rather than which one to make, judged on clinical efficacy and safety.
- Bender, A.; Cortés-Ciriano, I. (2021). Artificial intelligence in drug discovery: what is realistic, what are illusions? Part 2: a discussion of chemical and biological data. Drug Discovery Today. doi:10.1016/j.drudis.2020.11.037 Argues that AI has advanced less in drug discovery than in image or speech recognition partly because of its data: chemical and biological data are scarcer and less relevant to outcomes in patients, and better understanding of biology and more practically relevant data are needed.
- Handa, K.; Thomas, M.C.; Kageyama, M.; Iijima, T.; Bender, A. (2023). On the difficulty of validating molecular generative models realistically: a case study on public and proprietary data. Journal of Cheminformatics. doi:10.1186/s13321-023-00781-1 Trains a generative model on a project's early compounds and asks whether it produces the project's later ones. Among its top 100 designs it rediscovered 1.60% of later-stage compounds on public projects but 0.00% on in-house ones (0.04% across all 5,000), so retrospective tests on public data flatter generative models.
- Lindus Therapeutics (2026). Announcing a $1M Phase 2a Trial in Neurology. Lindus Therapeutics (company blog). lindustx.com A clinical-trial company's account of running a 25-participant Phase 2a trial in ME/CFS for under $1 million, with AI drafting the trial documents, searching health records for recruits and flagging data problems. A case for AI lowering trial costs rather than designing drugs; costs are self-reported against modeled benchmarks, and no results have been released.
- Walker, A.; Hepler, C. (2026). Deterministic Intelligence in a Probabilistic World. Alacrita (white paper). alacrita.com A drug-development consultancy's review of the clinical record of AI-derived molecules for investors: a late-2026 scorecard, the published Phase 1 and Phase 2 success rates set against industry benchmarks, five technical lanes with their failure modes, worked likelihood-of-approval arithmetic and a table of the regulatory instruments. It has no cohort of its own, and its Phase 1 chart compares a rate of successful Phase 1 completions with an industry rate of moving on to Phase 2. It describes zasocitinib's design as mainly physics-based.