DeepBindPoc

DeepBindPoc applies deep learning to identify and rank ligand-binding pockets in proteins to prioritize native-like binding sites for structure-based drug design.

Key Features:

• Deep Learning Integration: Employs deep learning to capture physicochemical and spatial information of ligand-binding pockets and automatically extract high-level pocket–ligand interaction features.
• Ligand Information Utilization: Incorporates specific ligand information into the predictive model to improve pocket ranking when ligand data are available.
• Vector Representation via mol2vec: Represents ligands and their associated pockets as vectors using mol2vec to feed a densely connected neural network model.
• Complementary Advantage: Provides complementary ranking capability to existing pocket-detection methods such as fpocket and P2Rank, improving detection of native-like pockets.

Scientific Applications:

• Structure-based drug design: Prioritizes potential binding sites on protein structures for downstream docking and lead optimization.
• Target-site identification for ligand-known proteins: Ranks native-like pockets for proteins with known ligands, with evaluations including datasets containing G-protein Coupled receptors.

Methodology:

Candidate pocket decoys (e.g., from fpocket) and ligands are represented with mol2vec vectors, and a densely connected fully connected layer model is trained on datasets that mimic real-world applications to learn to identify and rank native-like pockets; evaluations included datasets containing G-protein Coupled receptors.