DeepBindBC

DeepBindBC classifies native-like protein-ligand complexes (PLCs) from docking decoys to improve hit selection in early-stage virtual drug screening.

Key Features:

• Deep Learning Model: DeepBindBC employs a ResNet deep learning architecture to learn implicit features that distinguish binding from non-binding interactions.
• Incorporation of Non-Binding Data: The model is trained using both binding and non-binding interaction data to enhance predictive reliability.
• Detailed Atom Type Representation: An enriched atom type representation is used to capture physicochemical features important for binding classification.
• Comparative Performance: DeepBindBC outperformed Autodock Vina, Pafnucy, and DLSCORE across three DUD.E testing sets.
• Experimental Validation: The method identified a novel human pancreatic α-amylase binder that was experimentally validated with a dissociation constant (K_a) of 1.0 × 10^5 M by fluorescence spectral experiments.
• Integration with Hybrid Pipelines: DeepBindBC can be integrated with DFCNN, Autodock Vina docking, and pocket molecular dynamics simulation in virtual screening workflows.

Scientific Applications:

• Early-stage Virtual Screening: Identification of native-like PLCs among docking decoys to improve hit selection during early drug discovery.
• Lead Identification and Validation: Improvement of virtual screening reliability to support efficient lead compound identification and validation.

Methodology:

DeepBindBC uses a ResNet deep learning architecture trained with both binding and non-binding examples using an enriched atom type representation, was evaluated against Autodock Vina, Pafnucy, and DLSCORE on three DUD.E testing sets, and can be combined with DFCNN, Autodock Vina docking, and pocket molecular dynamics simulation.