Tracing Molecular Lipophilicity from Chemical Structure Using Interpretable Machine Learning

Abstract

Molecular lipophilicity is a key physicochemical property governing partitioning behavior and influencing compound optimization. This study developed an interpretable machine-learning framework to trace synthetic lipophilicity from chemically meaningful structural descriptors. A controlled dataset of 1,500 molecules was analyzed using molecular size, polarity, hydrogen-bonding, flexibility, ring, elemental-composition, refractivity, and aromaticity descriptors. Pearson and Spearman analyses characterized descriptor-lipophilicity relationships, while Elastic Net, Random Forest, Gradient Boosting, XGBoost, and support vector regression were compared using five-fold cross-validation and an independent 20% test subset. Model performance was evaluated with R2, RMSE, and MAE, and molecular determinants were interpreted using permutation importance and SHAP values. Polar surface area and hydrogen-bonding descriptors showed strong inverse relationships with synthetic LogP, whereas carbon content and refractivity were positively associated. The best cross-validated model retained strong predictive performance on held-out molecules, demonstrating that nonlinear structure-property relationships were recoverable from interpretable descriptors. SHAP analysis identified a compact set of chemically coherent features that dominated predictions and agreed with the principal statistical trends. The findings demonstrate a transparent proof-of-concept strategy for connecting molecular structure with lipophilicity prediction. Experimental datasets and scaffold-based external validation are required before application to real compound-design decisions.