Ensemble Molecular Surface Electrostatic Potential Descriptors as Informative Features for Active Learning

Active learning ideally needs informative and inexpensive features that provide reliable uncertainty estimates. General interaction properties function (GIPF) descriptors, which encode molecular surface electrostatic potential, are informative but usually computed by density functional approximations on individual structures, which is not only costly but also neglects conformational flexibility. In this work, we compute GIPF descriptors efficiently based on force-field point charges over conformer ensembles via an automated pipeline starting from SMILES strings. We benchmark our approach against inexpensive cheminformatics descriptors on three experimental datasets, covering surfactants, peptide catalysts, and supramolecular assemblies. We evaluate prediction accuracy, screening efficiency, and uncertainty calibration. Our results demonstrate that force-field-based GIPF descriptors are complementary to cheminformatics descriptors and lead to better predictions. Additionally, we find that ensemble-based GIPF descriptors improve uncertainty calibration significantly. Accordingly, we recommend force-field-based GIPF descriptors, evaluated over inexpensive conformer ensembles, as affordable and scalable featurization method for active learning campaigns in chemistry.