The first OpenBind release: An open experimental structure-affinity dataset and benchmark for structure-based AI

OpenBind's first release provides the largest single-target experimental structure-affinity dataset with 925 binding events for enteroviral 2A protease, advancing structure-based AI.

• Why it matters: High-quality datasets linking protein-ligand structures with binding affinities are crucial for developing accurate machine learning models in drug discovery, yet such comprehensive data remain scarce.
• What they did: The team compiled and publicly released a dataset combining crystallographic structures and affinity measurements from 699 compounds, and evaluated structure prediction, binding affinity, and virtual screening methods on this data.
• The result: Fine-tuning OpenFold3-p2 on early fragment structures improved pose prediction and virtual screening, highlighting how experimental data can enhance model accuracy and support antiviral research.