A set of optimized 3D-MoRSE descriptors for molecular representation

Abstract

Accurate early-stage prediction of molecular properties and toxicity is critical for reducing cost and attrition in drug discovery. Here, we develop and evaluate an optimized 3D-MoRSE (OPT3D) molecular representation that introduces a tunable distance-scale factor to better capture informative interatomic distance regimes. With simple traditional machine-learning models, for example, random forest (RF), OPT3D achieves an average (10 times) test root mean squared error (RMSE) of 0.942 (ESOL), 0.940 (Lipophilicity), and 2.108 (FreeSolv), and AUC of 0.824 (BACE), 0.878 (BBBP), 0.621 (SIDER), 0.828 (Tox21), and 0.717 (ToxCast). These results are competitive with recent state-of-the-art predictions that rely on more complex architectures. The prediction performance of this set of descriptors can be further improved by more advanced models: stacked ensembling further reduces regression errors and maintains strong classification performance, and the adaptive checkpointing with specialization neural model can also improve performance relative to simple traditional machine-learning models. Performed perturbation-based tests also show that the OPT3D descriptor exhibits strong robustness under moderate noise. Our results highlight that developing high-quality molecular representations is as important as model innovation, which has been intensively pursued but has yielded limited performance gains.