EGGNet, a Generalizable Geometric Deep Learning Framework for Protein Complex Pose Scoring.

Area of Science:

• Computational biology
• Biophysics
• Machine learning

Background:

• Predicting molecule-protein interactions is crucial for drug development.
• Existing methods often treat protein-protein interactions (PPIs) and drug-target interactions (DTIs) separately.
• This isolation limits the generalization of computational models for noncovalent interactions.

Purpose of the Study:

• To develop a unified computational framework for predicting interactions between proteins and diverse molecules (small molecules, peptides, proteins).
• To improve the generalization capabilities of predictive models across different types of protein interactions.
• To accelerate drug discovery and protein engineering through enhanced computational prediction.

Main Methods:

• Developed Equivariant Graph of Graphs neural network (EGGNet), a geometric deep learning (GDL) framework.
• Utilized a graph of graphs (GoG) representation at atomic resolution.
• Employed a multiresolution equivariant graph neural network incorporating atom- and residue-level biophysical interactions.

Main Results:

• EGGNet demonstrates competitive performance on protein-small-molecule binding affinity prediction (80.2% top 1 success rate on CASF-2016).
• Achieved high accuracy on a synthetic protein interface prediction task (88.4% AUC).
• Successfully handles predictions for small molecules, synthetic peptides, and natural proteins interacting with a target protein.

Conclusions:

• EGGNet provides a generalized GDL framework for various protein interaction prediction tasks.
• The model's ability to integrate diverse molecular types and biophysical interactions enhances predictive power.
• This approach holds significant potential for accelerating structure-based drug development and protein engineering.