Abstract
High-quality data on protein-ligand complex structures and binding affinities are crucial for structure-based drug design. Existing datasets often lack diversity and quantity, limiting the comprehensive understanding of protein-ligand interactions. Here, we present BindingNet v2, an expanded dataset comprising 689,796 modeled protein-ligand binding complexes across 1794 protein targets. Constructed using an enhanced template-based modeling workflow from BindingNet v1, it incorporates pharmacophore and molecular shape similarities. BindingNet v2's effectiveness in binding pose generation was evaluated, showing an improved generalization ability of Uni-Mol model for novel ligands. The success rate on the PoseBusters dataset increased from 38.55% with the PDBbind dataset alone to 64.25% with augmenting BindingNet v2. Coupled with physics-based refinement, the success rate rose to 74.07%, passing PoseBusters validity checks. These results highlight the value of larger, diverse datasets in enhancing the accuracy and reliability of deep learning models for binding pose prediction.