DeepSurf

DeepSurf predicts potentially druggable ligand binding sites on protein structures using a surface-based deep learning approach.

Key Features:

• Surface-Based Representation: Uses 3D voxelized grids placed on the protein surface to capture local spatial information relevant to potential binding sites.
• Deep Learning Architecture: Applies deep neural network models to the surface-centered voxel representations to detect patterns indicative of ligand binding.
• Training and Evaluation: Trained on the scPDB database and evaluated on three independent testing datasets, outperforming other deep learning methods and exhibiting competitive performance relative to traditional non-data-driven approaches.

Scientific Applications:

• Drug discovery and development: Identification of potentially druggable pockets to inform structure-based design of therapeutic agents.
• Research and benchmarking: Comparative evaluation of binding-site prediction methods in academic research and pharmaceutical settings.

Methodology:

Represents protein surfaces with 3D voxelized grids placed on the protein surface, applies deep learning models to those representations, and trains and evaluates models using the scPDB database and three testing datasets.