925 crystallographic binding events and 601 affinity measurements for EV-A71 2A protease, designed to train and stress-test structure-based machine learning

Protein-ligand prediction has become a central proving ground for AI in drug discovery. Models such as AlphaFold3, Boltz and OpenFold3 report striking accuracy on internal benchmarks, but careful re-evaluation has been less flattering. The Runs N’ Poses analysis showed that cofolding accuracy collapses as binding-site similarity to the training data decreases, suggesting the field has been measuring memorization rather than generalization. PoseBusters revealed that many AI-predicted poses also fail basic chemistry checks despite passing RMSD cutoffs. Beneath both findings lies the same issue: the Protein Data Bank is biased toward highly characterized target classes, and curated derivatives such as PLINDER have brought the limits of public training data to the forefront of community attention.

The OpenBind consortium, a Diamond Light Source-led partnership funded by the UK Department for Science, Innovation and Technology, has just released its first dataset to address that gap: 925 crystallographic binding events from 699 compounds on Enterovirus A71 2A protease, paired with 601 KD measurements from a Creoptix WAVE biosensor. The target, chosen with the ASAP antiviral discovery consortium, lies in a region of protein-ligand space dissimilar to that observed in PDB structures available before the AlphaFold3 training cutoff. The team is explicit about what they are trying to fix.

“We believe that the next generation of structure-based machine learning methods requires better experimental data, not just new architectures.”OpenBind Consortium, 2026

The reference benchmarks make the case. GNINA redocking succeeds on 85% of follow-on compounds, but rigid cross-docking into apo or AlphaFold receptors falls below 5%, defeated by a single binding-site loop. Cofolding partially recovers performance, with Protenix and OpenFold3-p2 leading but still trailing redocking. The most arresting result: fine-tuning OpenFold3-p2 on the fragment-bound structures alone lifts cofolding success from 36% to 76%, approaching redocking accuracy without the cognate receptor in hand. Affinity prediction is harder still. Structure-aware models, including Boltz-2, struggle to outperform a molecular-weight baseline on the EV-A71 series, a sobering reminder of how much remains between observed binding modes and measured potency.

The release is CC0 licensed, with benchmarks on GitHub, structures on Fragalysis and Zenodo, and full protocols on protocols.io. OpenBind frames it as the first output of an experimental pipeline that can be repeated across many targets. For anyone building structure-based AI, it provides a stress test that current methods cannot pass by memorization alone, and a reminder that the bottleneck on AI-driven drug discovery may sit in the wet lab as much as the architecture.