Physics-Based Machine Learning to Predict Hydration Free Energies for Small Molecules with a Minimal Number of...

Area of Science:

• Computational Chemistry
• Molecular Modeling
• Machine Learning Applications

Background:

• Hydration Free Energy (HFE) is crucial in chemistry and biology.
• Classical molecular dynamics simulations for HFE are computationally intensive.
• Existing machine learning (ML) models for HFE lack interpretability.

Purpose of the Study:

• Develop a physics-based ML model for accurate and interpretable HFE prediction.
• Apply the model to the FreeSolv database of small molecule HFE.
• Identify key molecular descriptors influencing HFE.

Main Methods:

• Utilized six interpretable, physics-based descriptors.
• Included electrostatic energy (via Generalized Born model), polar surface area, logP, hydrogen bond acceptors/donors, and rotatable bonds.
• Employed machine learning algorithms like random forest and extreme gradient boosting.

Main Results:

• Achieved a mean absolute error of 0.74 kcal/mol for HFE prediction.
• Demonstrated the model's accuracy and full interpretability.
• Identified electrostatics, polar surface area, and hydrogen bonding as primary HFE drivers.

Conclusions:

• The developed physics-based ML model provides an accurate and interpretable method for HFE calculation.
• This approach offers a significant advantage over traditional, less interpretable ML models.
• The model's interpretability aids in understanding the physical factors governing HFE.