The first OpenBind release: An open experimental structure–affinity dataset and benchmark for structure-based AI 1 OpenBind is introduced as an open-science effort to generate experimentally coherent structure–affinity datasets for structure-based AI. This first public release targets enteroviral 2A protease and is positioned as the largest public single-target experimental structure–affinity dataset to date. 2 Dataset scale and composition: 925 crystallographic binding events from 699 compounds, with grating-coupled interferometry (GCI; waveRAPID) affinity measurements for 601 compounds. After curation, 881 binding events remain for structural benchmarking (79 fragments, 802 follow-ons), and a QC-filtered affinity benchmark contains 490 compounds. 3 A key design choice is “campaign coherence”: structures come from an initial crystallographic fragment screen followed by follow-on soaking, and affinities are measured in a consistent biophysical setup. This reduces heterogeneity typical of retrospective datasets aggregated across assays, conditions, and curation standards. 4 The release is intended to test generalization beyond near-neighbor leakage. Using the Runs N’ Poses ligand–pocket similarity metric against PLINDER structures up to 30 Sep 2021 (AlphaFold3 cutoff), curated complexes show only moderate similarity (median ~40; most <50), a regime where prior work observed substantial cofolding performance drops. 5 Docking benchmarks show receptor-state selection dominates outcomes. With cognate holo receptors (redocking), GNINA reaches ~85% Top-25 success, but cross-docking into apo-like receptors (experimental apo 8POA or AF2 apo model) collapses due to a loop near the pocket adopting a partially occluding conformation. 6 Fragment-informed cross-docking partially rescues performance: selecting a receptor from the most chemically similar fragment-bound complex improves pose recovery (e.g., GNINA Top-25 RMSD-valid ~38%), but stringent interaction recovery remains low (Success ~14%), highlighting that fragments may not fully encode the pocket geometry needed for larger follow-ons. 7 Cofolding methods (AlphaFold3, Boltz-1/2, OpenFold3-p2, Protenix-v1, RosettaFold3) mitigate rigid receptor mismatch by allowing pocket adaptation, improving over apo cross-docking. However, Top-1 performance is much lower than Top-25, indicating that generating a correct pose and ranking it first are still distinct bottlenecks; an oracle ensemble across methods suggests complementary failure modes and room for better ranking/selection strategies. 8 Target-specific adaptation is a central takeaway: fine-tuning OpenFold3-p2 on just the curated fragment-bound structures (n=79) substantially improves follow-on pose prediction and virtual screening, with the biggest gains under stricter criteria that require both RMSD accuracy and protein–ligand interaction recovery (LDDT-PLI). 9 Affinity prediction remains difficult even with experimental structures. Across 490 QC-filtered compounds, structure-based methods (e.g., Boltz-2, GNINA, smina, AEV-PLIG, AQAffinity) show only modest correlation overall, and simple baselines are competitive (molecular weight gives the highest Spearman; mean pKD yields the lowest RMSE). The authors emphasize this as a reminder to benchmark against physicochemical baselines in coherent medicinal-chemistry series. 10 Virtual screening benchmark: 2,084 follow-on compounds (566 crystallographic binders; 1,518 “suspected non-binders” defined by no observed binding in technically successful soaks). Docking gives moderate discrimination (smina ROC-AUC 0.624; GNINA CNN VS 0.665). Base OpenFold3-p2 pair ipTM ranking is near-random (ROC-AUC ~0.53), but fragment-fine-tuned OpenFold3-p2 reaches ROC-AUC 0.76, suggesting fragment-derived structural data can make model confidence more informative for hit prioritization on related chemistry. 💻Code: github.com/OpenBind-Consorti… 📜Paper: biorxiv.org/content/10.64898… #OpenBind #StructuralBiology #DrugDiscovery #ProteinLigand #Cofolding #Docking #VirtualScreening #BindingAffinity #OpenScience #Benchmarking