Hierarchical geometric deep learning enables scalable analysis of molecular dynamics

Area of Science:

• Computational chemistry
• Biophysics
• Machine learning for science

Background:

• Molecular dynamics (MD) simulations provide detailed atomic trajectories but analyzing them is challenging without predefined features.
• Graph neural networks (GNNs) show promise for learning features directly from MD data.
• Existing GNNs struggle with long-range interactions or have high computational costs for large systems.

Purpose of the Study:

• To develop a novel GNN architecture for efficient and accurate analysis of MD simulations.
• To overcome limitations of current GNNs in capturing long-range interactions and managing computational resources.
• To enable analysis of large-scale biomolecular systems at atomic resolution.

Main Methods:

• A hierarchical GNN scheme was developed to aggregate local atomic information.
• This approach reduces memory and runtime requirements compared to traditional GNNs.
• The method was applied to simulations of protein-nucleic acid complexes.

Main Results:

• The hierarchical GNN successfully analyzed large protein-nucleic acid complexes (thousands of residues) at full atomic resolution on single GPUs within minutes.
• For systems with hundreds of residues, the approach maintained or improved performance and interpretability while reducing computational cost.
• The method demonstrated efficient handling of long-range interactions without quadratic scaling.

Conclusions:

• The hierarchical GNN offers a computationally efficient solution for analyzing complex molecular dynamics simulations.
• This advancement facilitates the study of large biomolecular systems, previously intractable for detailed analysis.
• The approach enhances both the performance and interpretability of GNNs in computational chemistry and biophysics.