A Large-Scale Dataset and Benchmark: Do Protein-Ligand Models Learn Binding Sites or Just Binding Likelihood?

In three linesInteractBind, a dataset of ~100k protein-ligand pairs with benchmark, evaluates whether models localize binding sites or merely predict binding likelihood. Eight tested models show strong binary prediction but weak binding-site localization, revealing gaps in physical interpretability.

Summary generated by Claude — human-verified