Scaling Graph Neural Networks to Large Proteins.

Area of Science:

• Computational chemistry
• Biophysics
• Machine learning

Background:

• Graph neural networks (GNNs) are emerging as accurate models for predicting molecular energies and forces.
• Optimizing computational efficiency is crucial for applying GNNs to large biomolecular systems in molecular dynamics.
• Existing GNN benchmarks lack large, biologically relevant protein datasets.

Purpose of the Study:

• Introduce a new dataset, DISPEF, for benchmarking GNNs on large proteins.
• Evaluate GNN performance on implicit solvation free energies, a challenging many-body problem.
• Develop a novel, efficient, and transferable GNN architecture for protein modeling.

Main Methods:

• Created DISPEF dataset with 207,454 proteins (up to 12,499 atoms) featuring diverse chemical environments.
• Benchmarked seven GNN architectures on DISPEF, focusing on long-range interactions.
• Developed and tested the novel multiscale architecture, Schake.

Main Results:

• DISPEF provides a stringent test for GNN expressiveness, especially for implicit solvation free energies.
• Accounting for long-range interactions is vital for GNN model transferability.
• The Schake architecture achieves transferable and computationally efficient energy and force predictions for large proteins.

Conclusions:

• The DISPEF dataset and Schake architecture advance GNN applications in protein modeling.
• Efficient and accurate GNNs are critical for large-scale biomolecular simulations.
• This work provides valuable tools for the development of machine learning force fields.