The first OpenBind release: An open experimental structure-affinity dataset and benchmark for structure-based AI
Jochem NeleniD, Omeir KhaniD, E. Charles AdamsiD, Jasmin Cara AschenbrenneriD, Warren Athol ThompsoniD, Ali EbrahimiD, Eda ÇapkıniD, Cédric ValléeiD, OpenBind, Elizabeth J. ShottoniD, Ed GriffeniD, John D. ChoderaiD, Charlotte Mary DeaneiD, F. von DelftiD, Mohammed AlQuraishiiD, Fergus ImrieiD
Abstract
High-quality experimental datasets that link protein-ligand structures with binding affinity data are essential for developing and evaluating structure-based machine learning methods. To help address this need, we established OpenBind as an open-science initiative to generate large-scale experimental datasets for structure-based AI and molecular discovery. Here, we describe the first public OpenBind release, which, to the best of our knowledge, is the largest public single-target experimental structure-affinity dataset. The dataset focuses on enteroviral 2A protease, comprising 925 crystallographic binding events from 699 compounds and associated affinity measurements for 601 compounds. It combines structures from an initial fragment screen and follow-on molecules, together with affinity data, linking experimentally determined protein-ligand binding modes to biophysical measurements within a coherent antiviral discovery campaign. We used this dataset to evaluate protein-ligand structure prediction, binding-affinity prediction, and virtual screening using representative structure-based methods, including docking and cofolding. This exposed several challenges that are central to practical structure-based modelling: docking performance depends strongly on binding-pocket conformation, poses are difficult to rank, and structure-based affinity prediction remains challenging. Fine-tuning OpenFold3-p2 on the fragment-screen structures substantially improved pose prediction and virtual screening for related follow-on compounds, demonstrating how early-stage experimental structures can support target-specific model adaptation.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 1
- Organism
- Assay
- Research Focus
- Approach
In-depth summary
Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (1 Sept 2026) · CC BY 4.0
A public crystallographic structure and affinity dataset for enteroviral 2A protease (925 binding events, 699 compounds), with docking, cofolding and screening baselines. Fragment fine-tuning of OpenFold3-p2 improves follow-on predictions.
Existing structure-affinity resources are described as retrospective assemblies from many public sources, or as padded with predicted structures, which brings noise and heterogeneous assay conditions. This release instead comes from one coherent fragment-to-follow-on campaign on a single target, with structures and affinities generated consistently.
Key results
ShownRedocking works well; cross-docking into apo structures drops success substantially for all methods, apparently mainly from receptor-state mismatch (a loop partly occludes the pocket). Fragment-informed receptors recover only part of the lost performance.
GNINA Top-25 success about 85% on redocking (802 follow-on events). Fragment-informed cross-docking with GNINA: 38% RMSD-valid but 14% meeting the stricter Success criterion. Five fragment receptors: Top-25 RMSD-valid 54%, Success 23%. · 3.2 Docking benchmarks; 2.3 Computational benchmarking and analysis…; Docking benchmarks
ShownCofolding beats rigid apo cross-docking, but producing a correct pose and ranking it first remain separate problems. Performance drops for all methods at Top-1.
Protenix-v1 was strongest at Top-25, followed by OpenFold3-p2 and AlphaFold3; Boltz-1, Boltz-2 and RosettaFold3 were lower. An oracle best-of-all cofolding ensemble scored substantially higher than any single method. Docking kept more of its Top-25 performance at Top-1 than most cofolding methods. · 3.3 Cofolding benchmarks; Cofolding benchmarks
SuggestedFine-tuning OpenFold3-p2 on 79 fragment-bound structures improved pose prediction for follow-on compounds. It also made the model's interface confidence useful for virtual screening, whereas the base model had little signal.
Virtual screening ROC-AUC (566 binders vs 1,518 suspected non-binders): fine-tuned 0.76 vs base 0.53; GNINA (CNN VS) 0.665; smina 0.624; molecular-weight baseline 0.539. Pose gains were largest under the stricter RMSD plus LDDT-PLI criterion (figure only). · 3.4 OpenFold3 fine-tuning; 3.6 Virtual screening benchmark; OpenFold3 fine-tuning; Virtual screening benchmark
ShownAffinity prediction remained challenging: a constant mean-pKD predictor and a molecular-weight baseline matched or beat structure-based methods. Authors suggest potency may correlate with compound elaboration or size, and caution against reading this as molecular weight being a sufficient model.
On 490 compounds, a molecular-weight baseline had the highest Spearman correlation and a constant mean-pKD predictor had the lowest RMSE. Boltz-2 and GNINA had the best rank correlations among structure-based methods; smina, AEV-PLIG and Boltz-2 had the lowest RMSE among them. · 3.5 Affinity prediction benchmarks; Affinity prediction benchmark; Affinity benchmark construction
ShownThe curated complexes are only moderately similar to public structures from before the AlphaFold3 training cutoff, and most follow-on compounds trace to a few fragments.
Median ligand-pocket similarity about 40 (scale 0-100) across 881 curated events, with almost all below 50. Fragment x0926a was the nearest fragment for 507 of 611 follow-on compounds (83.0%). · 3.1 Similarity to existing structural datasets; Structural benchmark construction; Similarity analysis against public structural…
Methods, data and reproducibility
Crystallographic fragment soaking and follow-on compound soaking produced protein-ligand structures, and grating-coupled interferometry gave affinities. After quality control, the authors benchmarked docking, cofolding, fine-tuning, affinity prediction and virtual screening on the curated data.
- Crystallographic soaking campaign: 1,146 unique fragments and 3,245 unique follow-on compounds tested, giving 113 fragment and 812 follow-on binding events · virus (Coxsackievirus A16 2A protease, strain G10, surrogate for EV-A71) · n = 925 binding events from 699 compounds
- Curated structural benchmark after PoseBusters and crystal-packing checks (79 fragment and 802 follow-on complexes) · virus (Coxsackievirus A16 2A protease) · n = 881 events, 677 compounds
- Grating-coupled interferometry affinities (KD, ka, kd), quality-controlled subset used for the affinity benchmark · virus (Coxsackievirus A16 2A protease) · n = 601 compounds measured; 490 in benchmark
- Virtual screening set: crystallographic binders vs suspected non-binders among follow-on compounds · virus (Coxsackievirus A16 2A protease) · n = 2,084 compounds (566 binders, 1,518 suspected non-binders)
- Methods & tools
- GNINA v1.3.2, smina, DiffDock-L v1.1.3, AlphaFold3, Boltz-2, OpenFold3-p2, Protenix-v1, AEV-PLIG, PoseBusters v0.6.5, OpenStructure v2.11.1, PanDDA2, Creoptix WAVEsystem
- Data & accessions
- https://fragalysis.diamond.ac.uk/viewer/react/preview/target/A71EV2A/tas/lb42888-1https://zenodo.org/records/20026661https://zenodo.org/records/20798528
- Software & environment
- Versions given for GNINA v1.3.2, DiffDock-L v1.1.3, PoseBusters v0.6.5 and OpenStructure v2.11.1. No container or requirements file is mentioned.
- Experimental conditions
- Fine-tuning OpenFold3-p2: 1,000 steps, about 8 h on 8 NVIDIA GH200 GPUs.
- Reproducibility gaps
- Assay quality-control criteria, docking parameters and wet-lab protocols are in the Supplementary Information and protocols.io, not in the main text.
Limitations (5)
- All results come from one protein target and one discovery campaign, so the ranking of methods and the docking-versus-cofolding balance may not generalise. · stated by the authors
- Suspected non-binders are compounds with no observed crystallographic binding in a technically successful soak, not biochemically confirmed non-binders. · stated by the authors
- Fine-tuning and evaluation data come from the same campaign, where many follow-on compounds derive from the fragment hits, so transfer across chemical series or targets is unestablished. · stated by the authors
- The oracle best-of cofolding result uses many more predictions than a practical setting, so it is only an upper bound. · stated by the authors
- The fine-tuning gains on pose prediction are described only qualitatively and in figures; the text gives no numeric success rates or replicate variability for the base and fine-tuned models. · noted in this summary
Who should read:Read this if you develop or benchmark docking, cofolding or affinity models and want an open single-target structure-affinity dataset with baselines.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- GatorAffinity: Boosting Protein-Ligand Binding Affinity Prediction with Large-Scale Synthetic Structural Data
Jinhang Wei, Yupu Zhang, … Yanjun Li·bioRxiv·Sep 29, 2025
- SAIR: Enabling deep learning for protein-ligand lnteractions with a synthetic structural dataset
Pablo Lemos, Zane Beckwith, … Martin Ganahl·bioRxiv·Jun 17, 2025
- A Folding-Docking-Affinity framework for protein-ligand binding affinity prediction
Ming-Hsiu Wu, Ziqian Xie, Degui Zhi·Communications ChemistryPeer-reviewed·Apr 7, 2025
- Physical priors improve performance of structure-based binding affinity models
Benjamin Kaminow, Alexander Matthew Payne, … Sukrit Singh·bioRxiv·Sep 18, 2026
- Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction
Saro Passaro, Gabriele Corso, … Regina Barzilay·bioRxiv·Jun 14, 2025
- ProAffinity-GNN: A Novel Approach to Structure-based Protein-Protein Binding Affinity Prediction via a Curated Dataset and Graph Neural Networks
Zhiyuan Zhou, Yueming Yin, … Yuguang Mu·Journal of Chemical Information and ModelingPeer-reviewed·Nov 19, 2024