OctSurf

OctSurf represents protein–ligand molecular surfaces using an octree-based hierarchical voxelization to enable protein–ligand binding affinity prediction.

Key Features:

• Octree-Based Representation: Recursively partitions 3D space of protein–ligand pockets into eight octants and subdivides only those containing van der Waals surface points until a predefined depth.
• Memory Efficiency: Maintains unoccupied octants without subdivision and processes only non-empty leaf octants to reduce memory usage versus uniform voxel grids.
• Surface Approximation: Uses non-empty leaf octants to approximate molecular surfaces of protein pockets and bound ligands, preserving chemical and geometric features needed for affinity prediction.
• Integration with 3D-CNNs: Provides octant-based surface representations as input to 3D convolutional neural networks, including VGG and ResNet architectures, and restricts convolutions to the smallest octants to minimize computation.
• Application in Binding Affinity Prediction: Enables high-resolution analysis of protein–ligand interactions to predict binding affinities relevant to drug discovery.

Scientific Applications:

• Binding Affinity Prediction: Predicts protein–ligand binding affinities using octree-derived surface representations combined with 3D-CNN models.
• Drug Discovery and Molecular Interaction Analysis: Accelerates identification and evaluation of potential therapeutic compounds by enabling efficient high-resolution modeling of complex protein–ligand interactions.

Methodology:

Implements an octree structure that recursively subdivides pocket space based on van der Waals surface points to produce non-empty leaf octants, uses those octants as input to 3D-CNNs (VGG and ResNet) while restricting convolutions to the smallest octants, and demonstrates reduced disk storage and computational load compared to conventional voxel-based 3D-CNN approaches.