Hydration-Free Energies for Small Molecules With Physics-Based Descriptors: Graph Neural Network With...

Area of Science:

• Computational Chemistry
• Machine Learning
• Chemical Physics

Background:

• Hydration-free energy (HFE) is crucial for understanding chemical and biological solvation processes.
• Traditional HFE computation methods (e.g., molecular dynamics) are computationally expensive.
• Existing machine learning models for HFE often lack interpretability and rely on complex features.

Purpose of the Study:

• To develop an accurate and interpretable machine learning model for predicting the hydration-free energy (HFE) of small molecules.
• To integrate physical descriptors into a graph neural network (GNN) for enhanced interpretability.
• To benchmark the GNN model against classical machine learning and existing GNN approaches.

Main Methods:

• A graph neural network (GNN) model was developed using the FreeSolv dataset.
• The GNN integrates graph representations of solute and solvent molecules with a cross-attention mechanism.
• Six global molecular descriptors (electrostatic energy, PSA, logP, H-bond donors/acceptors, rotatable bonds) were incorporated for interpretability.

Main Results:

• The attention-based GNN achieved a mean absolute error (MAE) of 0.54 kcal/mol and a root mean square error (RMSE) of 0.75 kcal/mol.
• This represents a 23% MAE and 36% RMSE improvement over the best-performing baseline.
• Ablation studies indicated electrostatic energy and polar surface area (PSA) as the most critical descriptors for prediction accuracy.

Conclusions:

• The developed GNN model offers a significant improvement in accuracy and interpretability for HFE prediction.
• The model's transparency in feature importance facilitates understanding of solvation processes.
• This framework is suitable for large-scale, data-driven investigations of solvation-free energies.