OctSurf: Efficient hierarchical voxel-based molecular surface representation for protein-ligand affinity prediction

Area of Science:

• Computational chemistry
• Structural biology
• Machine learning

Background:

• Predicting protein-ligand binding affinity is crucial for drug discovery.
• Voxel-based 3D Convolutional Neural Networks (CNNs) are effective but computationally expensive.
• High memory and computation costs limit the resolution and scalability of volumetric CNNs.

Purpose of the Study:

• To develop a memory-efficient 3D representation for protein-ligand interactions.
• To accelerate 3D convolutional operations for binding affinity prediction.
• To enable the use of higher spatial resolutions in volumetric CNNs.

Main Methods:

• Implemented OctSurf, an octree-based surface representation for protein binding pockets and ligands.
• OctSurf recursively partitions space, focusing computation on relevant surface octants.
• Applied VGG and ResNet CNN architectures to the OctSurf representation for binding affinity prediction.

Main Results:

• OctSurf significantly reduces memory consumption compared to voxel representations at equivalent resolutions.
• Restricting convolutions to relevant octants alleviates computational overhead.
• Experimental results demonstrate the disk storage and computational efficiency of OctSurf.

Conclusions:

• OctSurf offers a scalable and efficient alternative to voxel-based methods for 3D CNNs in computational chemistry.
• This novel representation facilitates more accurate and faster prediction of protein-ligand binding affinity.
• The method shows promise for advancing drug discovery through improved molecular modeling techniques.