Recent attempts at incorporating deep learning into structure-based virtual screening have focused on training models on protein-ligand interactions with known crystal structures. The PDBBind dataset has binding affinity data for less than 20K complexes, and models trained on it often fail to generalize to new targets. In fact, such models often perform just as well when trained with only ligand information. The CrossDocked dataset adds data for protein-ligand complexes without known binding affinities but does not introduce new affinity data. ChEMBL, on the other hand, contains a wealth of binding affinity information with no known binding poses. We introduce BigBind, a dataset which maps ChEMBL activity data to known protein binding pockets from CrossDocked. This dataset has 851K ligand binding affinities and 3D pocket structures. After augmenting this dataset with an equal number of putative inactives for each target, we train a message passing neural network (MPNN) model to classify actives from inactives. The resulting model achieves an AUC of 0.76 on BigBind’s test set, while a model using only the ligand can barely do better than random (AUC 0.55). We then test this model on both LIT-PCBA and a custom virtual screening benchmark. While the model does not achieve superior enrichment overall when compared to gnina and vina, its performance is largely uncorrelated with the performance of docking. Thus it shows promise as a means of quickly filtering out compounds prior to docking.

BigBind: Leveraging nonstructural data for structure-based virtual screening

Where thousands of chemistry professionals meet to share ideas and advance scientific and technical knowledge. Spring 2023 abstracts, posters, and presentations will launch at the end of March.

ACS Spring 2023

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.