Abstract
Accurately predicting solvation free energy and understanding its physical determinants are essential for studying solute behavior in solution. This work employs advanced machine learning techniques to enhance predictive accuracy and extract insights into the solvation free energy of small molecules. Traditional machine learning approaches, compared to deep learning, are lightweight and require fewer computational resources. Our analysis identifies molecular geometry and topology as critical factors in predicting alchemical free energy, aligning with the theory that surface tension is a key determinant, while highlighting the role of charge distribution in improving force field designs for molecular dynamics. We propose an improved machine learning scheme that integrates K-nearest neighbors for feature processing, ensemble modeling, and dimensionality reduction. This scheme achieves a mean unsigned error of 0.53 kcal/mol on the FreeSolv data set using only two-dimensional features without pretraining on large databases, offering substantial accuracy improvements. This lightweight approach provides a viable alternative to computationally intensive deep learning models and holds promise for broad applications in chemical predictions.