Challenges in Transferable Prediction of Solvation Free Energy: A Comparative Analysis of Molecular Representations and Machine Learning Methods

In-silico prediction of physicochemical properties such as solvation free energy is crucial for efficient drug discovery. However, accurate prediction remains challenging due to complexities inherent in molecular representations and model transferability. This study systematically evaluates the influence of different molecular representations, namely descriptor-based, fingerprint-based and graph-based, on the predictive performance and transferability of supervised machine learning (ML) models. Using three diverse datasets (MNSol, FreeSolv, and CombiSolv), we compared classical regression techniques (XGBoost, Random Forest, Support Vector Regression, Kernel Ridge Regression) against deep learning models, specifically the Chemically Interpretable Graph Interaction Network (CIGIN). Our findings indicate that while traditional models with interpretable descriptors provide insights into the important features, their transferability is limited by dataset size and chemical diversity. Molecular fingerprints show improved performance, and a Multilayer Perceptron (MLP) Regressor demonstrates better regularization with high-dimensional fingerprints compared to traditional models. The graph-based CIGIN model exhibits strong performance and chemical interpretability but faces challenges in generalizing to novel chemical entities absent in the training data, showing increased errors for molecules with long hydrocarbon chains or polyol moieties. This research highlights the critical interplay between data quality, molecular representation, and model choice in achieving accurate and transferable predictions of molecular properties, underscoring the need for further refinement in handling novel chemical space and incorporating physics-informed features.