BigBind: Learning from Nonstructural Data for Structure-Based Virtual Screening

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Deep learning models for protein-ligand binding prediction are crucial for structure-based virtual screening.
• Existing models trained on PDBBind data struggle with generalization due to dataset size limitations.
• ChEMBL database offers extensive activity data but lacks binding pose information.

Purpose of the Study:

• To introduce BigBind, a novel dataset mapping ChEMBL activity data to protein structures.
• To develop and evaluate Banana, a neural network model for classifying active/inactive compounds using BigBind.
• To enhance the efficiency and accuracy of virtual screening for drug discovery.

Main Methods:

• Creation of the BigBind dataset by integrating ChEMBL activity data with CrossDocked protein structures, including 583K ligand activities and 3D binding pocket structures.
• Augmentation of BigBind with an equal number of putative inactive compounds for each target protein.
• Development of the Banana (basic neural network for binding affinity) model for binary classification of ligand activity.

Main Results:

• Banana achieved an AUC of 0.72 on the BigBind test set, outperforming a ligand-only model (AUC 0.59).
• Banana demonstrated competitive performance on the LIT-PCBA benchmark (median EF1% 1.81).
• Banana operated 16,000 times faster than molecular docking with Gnina, showcasing significant computational efficiency.

Conclusions:

• The BigBind dataset provides a valuable resource for training deep learning models in drug discovery.
• The Banana model shows promise for improving the accuracy and speed of virtual screening.
• These advancements are expected to significantly enhance the outcomes of prospective virtual screening tasks.