Electron iso-density surfaces (EIS) provide quantum mechanically rigorous descriptions of molecular size governing intermolecular interactions and binding affinity. We developed machine learning models predicting EIS surface area from molecular descriptors using 288 diverse organic molecules. Eight algorithms were evaluated including Ridge regression, Gaussian Process Regression, Random Forest, and XGBoost using 5-fold cross-validation. Gaussian Process Regression achieved best test set performance ( R 2 = 0 . 9896 ), with mean absolute errors below 4.0 Å 2 (3% uncertainty). SHAP analysis revealed atomic count, molecular weight, and connectivity dominate predictions, enabling computationally efficient property estimation with chemical interpretability. • Machine learning models predict electron iso-density surface areas with exceptional accuracy (R 2 = 0.9896) from readily computed molecular descriptors. • Gaussian Process Regression outperforms seven other algorithms including tree-based ensembles and neural networks, with mean absolute errors below 4.0 Å 2 (3% prediction uncertainty). • SHAP interpretability analysis reveals that molecular architecture descriptors (atomic count, molecular weight, connectivity) dominate surface area predictions. • The approach enables computationally efficient prediction of quantum mechanical properties while maintaining chemical interpretability, accelerating high-throughput molecular screening.
Ho et al. (Fri,) studied this question.