A Large-Scale Dataset and Benchmark: Do Protein-Ligand Models Learn Binding Sites or Just Binding Likelihood?
Researchers have introduced InteractBind, a new large-scale dataset and benchmark designed to evaluate protein-ligand models in computational drug discovery. This dataset, comprising around 100,000 protein-ligand pairs, focuses on assessing whether models can accurately localize binding sites and identify specific non-covalent interactions, rather than just predicting general binding likelihood. Initial evaluations of eight existing models revealed that while they perform well in predicting binding, their ability to localize binding sites is limited, with significant variation across different interaction types. InteractBind aims to encourage the development of more interpretable and physically grounded protein-ligand models. AI
IMPACT Establishes a new benchmark for evaluating protein-ligand models, pushing for greater interpretability and physical grounding in drug discovery.