n,tier,year,short,venue,demonstrated,stage_note,caveat,doi_or_src,link,method,target,title
1,1 · In silico only,2024,AlphaFold 3,Nature 630:493-500,"Predicts joint 3D structure of protein-ligand-nucleic acid complexes, substantially more accurate than prior docking and structure-prediction methods (incl. on PoseBusters). Enables the modelling layer AI drug discovery sits on.",Entirely computational. No molecule was synthesised or tested; benchmark accuracy only.,"Method/benchmark paper, not a drug candidate. Later addendum published (10.1038/s41586-024-08416-7).",10.1038/s41586-024-07487-w,https://doi.org/10.1038/s41586-024-07487-w,Diffusion-based co-folding model (DeepMind / Isomorphic Labs),Method paper — protein / nucleic acid / small-molecule / ion complexes,Accurate structure prediction of biomolecular interactions with AlphaFold 3
2,1 · In silico only,2024,PCMol (AlphaFold-conditioned generative design),J Chem Inf Model 64(21):8113-8122,Generated novel molecules conditioned on AlphaFold2 protein embeddings; embeddings cluster by target family and outperform raw sequence. Generated compounds scored favourably against known actives of held-out targets by docking.,Purely computational. Evaluation is docking scores + ligand similarity + embedding analysis.,No wet-lab validation whatsoever. Authors state generated molecules are only 'potentially active'.,10.1021/acs.jcim.4c00309,https://doi.org/10.1021/acs.jcim.4c00309,Multitarget transformer generative model conditioned on AlphaFold2 embeddings,"Multi-target, including proteins with sparse ligand data",AlphaFold Meets De Novo Drug Design: Leveraging Structural Protein Information in Multitarget Molecular Generative Models
3,1 · In silico only,2025,Boltz-2,Preprint (MIT CSAIL / Recursion),First co-folding model to predict structure and binding affinity together; ~1000x faster than free-energy perturbation with comparable rank correlation (Pearson ~0.62 on OpenFE FEP+). An independent prospective test on 557 Mac1-ligand complexes found its affinity predictions tracked measured potency better than AlphaFold3 or Chai-1 pose confidence.,"Purely computational. Affinity predictions benchmarked against already-measured assay data, not prospective wet-lab hits.",Preprint — not yet peer-reviewed. Prospective docking evaluations show degradation on non-'easy' targets with undefined pockets.,,https://doi.org/nan,Open-source co-folding model predicting structure + binding affinity jointly,General protein-small molecule affinity; benchmarked on OpenFE FEP+ and CASP16,Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction
4,2 · In vitro,2024,NGT — MC2R inhibitors,J Med Chem 67(21):19417-19427,Generated MC2R antagonists with no training data on the target. 12 compounds were synthesised and tested in a cAMP cell-based antagonist assay with full concentration-response curves and LC-MS; hits were selective across MCxR subtypes. Reported hit rate compared favourably to a conventional experimental HTS screen.,Synthesised and assayed in cells (EC50). No animal work.,Top potency was slightly lower than the conventional HTS comparator. Exact EC50 values sit in supplementary Table S4.,10.1021/acs.jmedchem.4c01763,https://doi.org/10.1021/acs.jmedchem.4c01763,Reinforcement-learning generative model over a 3-trillion-compound synthesis-on-demand library,Melanocortin-2 receptor (MC2R) — no prior antagonist structures in the literature,NGT: Generative AI with Synthesizability Guarantees Discovers MC2R Inhibitors from a Tera-Scale Virtual Screen
5,2 · In vitro,2025,RFdiffusion antibodies,Nature 649:183-193,"Designed VHHs, scFvs and full antibodies against specified epitopes. Cryo-EM structures (3.0-3.6 A) of designed binders against influenza HA and TcdB matched the design models closely, including all six CDR loop conformations for one scFv. Initial designs bound in the tens-to-hundreds of nM; affinity maturation produced single-digit nM binders.",Biophysical binding + high-resolution structural confirmation. No animal efficacy or PK.,"Independent third-party critiques (Zenodo 17797005, 17723030) question rigid-body complementarity assumptions, absence of mutational validation, and possible germline memorisation vs. true 'de novo' design.",10.1038/s41586-025-09721-5,https://doi.org/10.1038/s41586-025-09721-5,Fine-tuned RFdiffusion + ProteinMPNN; yeast display screening and OrthoRep affinity maturation,"User-specified epitopes: influenza HA, C. difficile toxin B, RSV site III, SARS-CoV-2 RBD, PHOX2B peptide-MHC",Atomically accurate de novo design of antibodies with RFdiffusion
6,3 · Animal (in vivo),2024,Wong et al. — new antibiotic class,Nature 626:177-185,"Experimentally profiled 39,312 compounds, then computationally screened 12,076,365; 283 were tested in the lab. Identified a structural class selectively active against MRSA and VRE that evaded substantial resistance, and reduced bacterial titres in mouse models of MRSA skin infection and systemic thigh infection. Mechanism: dissipating the bacterial membrane pH gradient.",In vitro screening plus in vivo mouse infection efficacy.,"Mouse infection models, not human. Mechanism is membrane-disrupting and mechanism-agnostic models take chemical structure as the only input.",10.1038/s41586-023-06887-8,https://doi.org/10.1038/s41586-023-06887-8,Ensembles of graph neural networks with substructure 'rationale' explanations,MRSA / vancomycin-resistant enterococci,Discovery of a structural class of antibiotics with explainable deep learning
7,3 · Animal (in vivo),2025,Krishnan et al. — de novo antibiotics,Cell,"Generated >36 million previously unenumerated compounds; 24 were synthesised and 7 empirically validated as antibacterial (~27% working true discovery rate). Leads NG1 (N. gonorrhoeae, novel LptA target) and DN1 (S. aureus/MRSA, membrane-acting) showed efficacy in mouse models of gonococcal vaginal infection and MRSA skin infection, with mechanisms distinct from clinical antibiotics.",In vitro MIC/MBC + in vivo mouse infection efficacy.,Mouse models only. Authors note the framework needs extension to pathogens such as M. tuberculosis and P. aeruginosa.,10.1016/j.cell.2025.07.033,https://doi.org/10.1016/j.cell.2025.07.033,"Generative models (CReM genetic algorithm, VAE) + graph neural network screening","N. gonorrhoeae, S. aureus incl. MRSA",A generative deep learning approach to de novo antibiotic design
8,3 · Animal (in vivo),2025,SK56 — GSDMD pore blocker,Nature Immunology 26:1660-1672,"From 12 AI-generated candidate peptides, SK56 emerged as most potent; binds GSDMD-NT directly (~0.25 uM by MST) and selectively blocks the pore without affecting IL-1beta or GSDMD cleavage. In LPS- and CLP-induced mouse sepsis, survival rose from 5% to 45%, with reduced multi-organ damage, and it remained effective when given 4-16 h after induction. Also validated in human whole-blood assays (37 cytokines reduced) and a human alveolar organoid-macrophage co-culture.",In vitro human cell/organoid assays + in vivo mouse sepsis efficacy.,"Survival still 45% at best. SK56 shows cross-affinity for GSDMC-NT, a specificity limitation.",10.1038/s41590-025-02280-x,https://doi.org/10.1038/s41590-025-02280-x,"Transformer-based atomic generative model (TransForPep) trained on ~40,000 protein-protein interfaces",GSDMD-NT pore (sepsis / inflammatory disease),Delaying pyroptosis with an AI-screened gasdermin D pore blocker mitigates inflammatory response
9,3 · Animal (in vivo),2025,ISM5939 — ENPP1 inhibitor,Nature Communications 16:4793,"Identified ENPP1 as a safer STING-modulating target than direct STING agonism; designed orally bioavailable ISM5939, reportedly reaching a hit series in ~3 months. In murine syngeneic models (MC38, CT26, LLC1, 4T1, EMT6) it synergised with PD-1/PD-L1 blockade and chemotherapy, with a wider therapeutic index and less cytokine release than direct STING agonists (ADU-S100, diABZI, MSA-2).",In vivo mouse tumour models; no human dosing in this paper.,Syngeneic mouse models. The parent company's related clinical-stage assets are separate studies.,10.1038/s41467-025-59874-0,https://doi.org/10.1038/s41467-025-59874-0,PandaOmics target discovery + Chemistry42 generative chemistry (structure-based drug design),ENPP1 (innate immune checkpoint) in solid tumours,Oral ENPP1 inhibitor designed using generative AI as next generation STING modulator for solid tumors
10,3 · Animal (in vivo),2024,BGM0504 — dual GLP-1/GIP agonist,Scientific Reports 14:16680,"MD revealed a K20 salt bridge with the receptor ECDs invisible in cryo-EM, explaining why acylation at K20 kills activity; relocating the acyl chain to the C-terminus produced BGM0504. cAMP EC50 0.031 nM (GLP-1R) and 0.182 nM (GIPR), 2-3x more potent than tirzepatide. db/db mice showed dose-dependent glucose and weight reduction superior to tirzepatide; cynomolgus monkeys showed prolonged half-life and elevated exposure.",In vitro potency + mouse efficacy + non-human primate PK.,Physics-based MD rather than machine learning — included as an 'AI-adjacent' computational-design comparator. Its human trial data are separate and not part of this paper.,10.1038/s41598-024-66998-8,https://doi.org/10.1038/s41598-024-66998-8,500 ns molecular dynamics simulations (three parallel runs) guiding structure optimisation,"GLP-1R / GIPR (type 2 diabetes, obesity)",Molecular dynamics-guided optimization of BGM0504 enhances dual-target agonism for combating diabetes and obesity
11,4 · Human clinical,2024,INS018_055 / rentosertib — preclinical + Phase 1,Nature Biotechnology,"First-in-class TNIK inhibitor derived from AI target ID plus generative chemistry in ~18 months from target to preclinical candidate. Showed anti-fibrotic activity by oral, inhaled and topical routes across bleomycin mouse lung fibrosis and skin/kidney fibrosis models, plus favourable PK/safety across species. A randomised, double-blind, placebo-controlled Phase 1 (NCT05154240, n=78 healthy volunteers) showed safety, tolerability and PK.",Animal efficacy + human Phase 1 safety/PK. Phase 1 was not designed to show efficacy.,The human component is safety and pharmacokinetics in healthy volunteers only — no efficacy signal at this stage.,10.1038/s41587-024-02143-0,https://doi.org/10.1038/s41587-024-02143-0,PandaOmics (target discovery) + Chemistry42 (generative chemistry),"TNIK, for idiopathic pulmonary fibrosis (IPF)",A small-molecule TNIK inhibitor targets fibrosis in preclinical and clinical models
12,4 · Human clinical,2025,Rentosertib (ISM001-055) — Phase 2a,Nature Medicine 31(8):2602-2610,"The first randomized, placebo-controlled, peer-reviewed clinical proof-of-concept for an AI-discovered and AI-designed drug. 71 IPF patients across 22 centres in China, 12 weeks. Primary endpoint (treatment-emergent adverse events) was similar across arms. On a secondary endpoint, 60 mg QD gave a mean FVC change of +98.4 mL (95% CI 10.9-185.9) versus -20.3 mL (95% CI -116.1 to 75.6) on placebo, with a dose-dependent trend. By mid-2026 a Phase III was initiated.","Human Phase 2a, randomized and placebo-controlled — the furthest any AI-designed molecule had reached at publication.","Small arms (n=17-18 each), 12-week duration, and the FVC result is a SECONDARY endpoint — the trial was not powered to demonstrate efficacy, and the confidence intervals are wide.",10.1038/s41591-025-03743-2,https://doi.org/10.1038/s41591-025-03743-2,Generative AI design (TNIK inhibitor); GENESIS-IPF trial NCT05938920,"TNIK, idiopathic pulmonary fibrosis",A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial
