Boosting Protein Graph Representations through Static-Dynamic Fusion

Pengkang Guo, Bruno Correia, Pierre Vandergheynst, Daniel Probst

Abstract: Machine learning for protein modeling faces significant challenges due to proteins' inherently dynamic nature, yet most graph-based machine learning methods rely solely on static structural information. Recently, the growing availability of molecular dynamics trajectories provides new opportunities for understanding the dynamic behavior of proteins; however, computational methods for utilizing this dynamic information remain limited. We propose a novel graph representation that integrates both static structural information and dynamic correlations from molecular dynamics trajectories, enabling more comprehensive modeling of proteins. By applying relational graph neural networks (RGNNs) to process this heterogeneous representation, we demonstrate significant improvements over structure-based approaches across three distinct tasks: atomic adaptability prediction, binding site detection, and binding affinity prediction. Our results validate that combining static and dynamic information provides complementary signals for understanding protein-ligand interactions, offering new possibilities for drug design and structural biology applications.

Lay Summary:

Proteins are essential molecules in living organisms that constantly move and often change shape to perform their functions, like enzymes breaking down food or motor proteins moving materials within cells. Most current artificial intelligence methods for studying proteins only look at their static 3D structures — like taking a single photograph of a dancer instead of watching the entire performance.

We developed a new framework that combines both the static structure of proteins with information about how they move over time, captured through computer simulations called molecular dynamics. Think of it as creating a movie of protein motion rather than just a snapshot. Our approach creates a new way to represent this combined information, then uses specialized AI models called relational graph neural networks to process it.

We tested our framework on three important tasks: predicting how flexible different parts of proteins are, identifying where drugs might bind, and estimating how strongly drugs stick to proteins. Consistently, our method that combines structural and motion information significantly outperformed approaches using structure alone.

This advance could accelerate drug discovery by helping scientists better understand how proteins work and how to design medicines that interact with them more effectively.

Application-Driven Machine Learning: This submission is on Application-Driven Machine Learning.

Link To Code: https://github.com/PKGuo/protein-static-dynamic-fusion.git

Primary Area: Deep Learning->Graph Neural Networks

Keywords: Graph Neural Networks, Protein Modeling, Molecular Dynamics, Heterogeneous Graph

Submission Number: 13452