Fig. 8

The average rank assigned to all ligand atoms attaining a score above a specified threshold on the PDBBind test set. Each line corresponds to the performance obtained by a PointVS model trained on a different training set. The model trained on the unbiased ZINC dataset outperforms the models trained on the real-world DUD-E and LIT-PCBA datasets, illustrating that ligand-specific biases hamper the ability of virtual screening models to identify the most important functional groups