Abstract
Predictive models hold considerable promise in enabling the faster discovery of safe, efficacious therapeutics. To better understand and improve the performance of small molecule predictive models for drug discovery, we conduct multiple experiments with deep learning and traditional machine learning approaches, leveraging our large internal datasets as well as publicly available datasets. The experiments include assessing performance on random, temporal, and reverse-temporal data ablation tasks, as well as tasks testing model extrapolation to different property spaces. We identify factors that contribute to higher performance of predictive models built using graph neural networks compared to traditional methods such as XGboost and random forest. These insights were successfully used to develop a scaling relationship that explains 81% of the variance in model performance across various assays and data regimes. This relationship can be used to estimate the performance of models for ADMET (absorption, distribution, metabolism, excretion, and toxicity) endpoints, as well as for drug discovery assay data more broadly. The findings offer guidance for further improving model performance in drug discovery.