Answers are grounded in this paper and cite their sources.
Machine learning (ML) has become a powerful tool for predicting molecular physicochemical properties. It finds applications in various research and development sectors, such as materials science, pharmaceutical chemistry, and environmental science. However, systematic comparisons between different types of properties remain limited. In this study, we developed two structured datasets: a logP dataset containing 1117 molecules and a pKb dataset containing 1268 molecules. For logP, each molecule is represented by 623 molecular descriptors generated exclusively by RDKit/Mordred, while a combination of 150 quantum chemistry descriptors from DFT calculations and molecular fingerprints derived from RDKit is used for pKb. Several ML algorithms were evaluated using an identical workflow, and the relevance of the descriptors was analyzed using SHAP, followed by feature pruning based on correlation. For the logP dataset, the LightGradBoost model achieved an R 2 of 0.94, an RMSE of 0.31, and an MAE of 0.42 on the independent test set, accurately reproducing experimental logP values in the range of −11.6 to 1.58. For pKb prediction, Random Forest (RF) proved most accurate, with an MAE of 1.69 and an RMSE of 1.68, with predicted values covering the entire range of experimental pKb values (−37 to 29.2). Our results indicate that, while RDKit/Mordred descriptors can predict logP with high accuracy, pKb remains a more challenging property to model, even when incorporating high-level DFT descriptors. The study therefore proposes a unified framework for the comparative evaluation of cross-property machine learning models and highlights the influence of the type of descriptor and the choice of algorithm on performance for chemically distinct properties.
No takes yet. Share an insight, caveat, or question.
Baikété et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: