DeepPocket

DeepPocket detects and segments ligand binding sites on protein structures using 3D convolutional neural networks to support structure-based drug design.

Key Features:

• Integration with Fpocket: Combines geometry-based pocket detection from Fpocket with deep learning-based rescoring of identified pockets.
• 3D Convolutional Neural Networks (3D CNNs): Uses 3D CNNs to process volumetric protein data and to perform cavity segmentation on protein surfaces.
• Performance and Generalization: Demonstrates improved detection, ranking, and generalization across multiple ligand binding site datasets compared to existing methods.
• SC6K Dataset: Includes evaluation on the SC6K dataset comprising Protein Data Bank (PDB) structures submitted between 2018-01-01 and 2020-02-28.

Scientific Applications:

• Structure-based drug design: Identifies and characterizes ligand binding pockets on protein structures to support target assessment and lead discovery.
• Druggability assessment: Enables rescoring and ranking of pockets to prioritize druggable and functionally relevant binding sites.
• Lead optimization support: Provides segmented cavity information that can inform ligand fitting and optimization strategies.

Methodology:

DeepPocket uses Fpocket for initial pocket identification and applies 3D convolutional neural networks to rescore and segment detected cavities, with performance evaluated on multiple binding site datasets including SC6K (PDB submissions 2018-01-01 to 2020-02-28).