GENEOnet: a breakthrough in protein binding pocket detection using group equivariant non-expansive operators

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Drug Discovery

Background:

• Accurate protein binding pocket identification is crucial for structure-based virtual screening and drug discovery.
• Existing methods often face challenges with model complexity, parameter interpretability, and computational cost.

Purpose of the Study:

• To introduce GENEOnet, a novel machine learning model for volumetric protein pocket detection.
• To leverage Group Equivariant Non-Expansive Operators (GENEOs) for simplified model complexity and enhanced domain knowledge integration.
• To provide a more explainable and computationally efficient alternative for identifying potential drug binding sites.

Main Methods:

• GENEOnet utilizes Group Equivariant Non-Expansive Operators (GENEOs) for protein pocket detection.
• The model processes protein structures by converting empty space into a 3D grid of voxels.
• It identifies and ranks potential binding pockets based on voxel output values.

Main Results:

• GENEOnet demonstrates robust performance, even with limited training data (200 proteins).
• It surpasses state-of-the-art methods, achieving a [Formula: see text] score of 0.764 on the PDBbind test set, outperforming P2Rank (0.702).
• Case studies, like ABL1 kinase, show excellent agreement between GENEOnet predictions and experimental binding sites.

Conclusions:

• GENEOnet offers a powerful, efficient, and explainable approach to protein pocket detection.
• The model's reduced parameter count lowers training costs and enhances interpretability.
• GENEOnet represents a significant advancement for structure-based drug discovery and virtual screening applications.