Using Deep Graph Neural Networks Improves Physics-Based Hydration Free Energy Predictions Even for Molecules Outside...

Area of Science:

• Computational Chemistry
• Biomolecular Simulations
• Machine Learning in Science

Background:

• Accurate computational water models are essential for atomistic simulations of biomolecules.
• Physics-based solvation models often have residual errors in hydration free energy (HFE) prediction.
• Deep neural networks (DNNs) can struggle with out-of-distribution data, limiting their standalone application.

Purpose of the Study:

• To evaluate a hybrid framework combining physics-based models and DNNs for HFE prediction.
• To assess the framework's performance on out-of-distribution data and unseen molecular scaffolds.
• To determine the generalizability and limitations of using DNNs as postprocessing corrections.

Main Methods:

• A decoupled framework integrating classical physics-based solvation models with DNNs was developed.
• Graph neural network architectures were employed to generalize predictions.
• The framework was evaluated using multiple dataset splits, including out-of-distribution HFEs.

Main Results:

• Physics + DNN models consistently improved predictions from physics models, particularly for out-of-distribution data.
• For in-distribution data, DNN corrections improved accuracy, achieving root-mean-square error (RMSE) below 1 kcal/mol.
• Model accuracy improved when molecules with high experimental uncertainty were excluded.

Conclusions:

• Combining physics-based models with DNNs offers a practical and generalizable strategy for enhancing HFE prediction accuracy.
• DNNs can serve as effective independent postprocessing corrections, overcoming limitations of standalone physics or ML models.
• The hybrid approach shows significant potential for improving biomolecular simulations.