Scoring Protein-Ligand Complex Structures by HybridNet

Scoring the binding for a protein-ligand complex structure is a widely-discussed and open problem in structure-based drug design. As deep learning and artificial intelligence continue to rapidly advance, developing deep-learning scoring models is currently an active area of research. Intermolecular-contact features are fast-to-generate and can be efficiently handled by deep-learning models, while they are oversimplified to characterize the binding between a ligand and its target protein. In this work, we have developed the HybridNet model that profiles multi-range intermolecular contacts and deals with the heterogeneous channels of such features using a hybrid deep-learning architecture. The intermolecular-contact profiles keep the simplicity of original features but describe the interactions more deeply. Besides, compared to individual learning architectures and classical scoring models, HybridNet performed more favorably in protein-ligand scoring tasks. The proposed method of featurization and scoring will prospectively benefit related tasks like molecular docking and virtual screening in the long term.