Machine Learning Prediction of Hydration Free Energy with Physically Inspired Descriptors

Area of Science:

• Computational chemistry
• Molecular modeling
• Machine learning

Background:

• Accurate prediction of hydration free energies is crucial for drug discovery and chemical process design.
• Traditional methods often rely on atomistic details or molecular fingerprints, limiting their applicability to novel chemical spaces.

Purpose of the Study:

• To develop novel machine learning models for predicting molecular hydration free energies.
• To create descriptors that do not require atom-, bond-, or geometry-specific input features.
• To assess the performance of these models against existing methods.

Main Methods:

• Physically inspired molecular descriptors were developed, focusing on properties like dipole moment, polarizability, electrostatic potential distribution, solvent accessible surface area, and shape.
• Machine learning regression models were trained on the FreeSolv database (approximately 600 samples).
• Model performance was evaluated against traditional approaches and fingerprint-based methods.

Main Results:

• The developed machine learning models demonstrated superior performance in predicting hydration free energies compared to most traditional and fingerprint-based methods.
• The models successfully predicted hydration free energies for new compounds containing previously unseen elements or fragments.
• Analysis of descriptor importance and prediction outliers was performed.

Conclusions:

• The proposed descriptor set and machine learning models offer a robust and generalizable approach for predicting hydration free energies.
• This method overcomes limitations of traditional approaches, enabling predictions for novel chemical entities.
• The findings contribute to more efficient molecular design and property prediction in computational chemistry.