Assessing the impact of data harmonization on human liver microsomal stability prediction model performance

Metabolic stability, often assessed using human liver microsomes (HLM), is a key determinant of drug clearance and pharmacokinetics (PK), yet experimental evaluation of the vast chemical space is resource intensive. Machine learning-based quantitative structure-activity relationship (QSAR) models offer a cost-effective strategy, but most publicly available HLM predictors rely heavily on heterogeneous ChEMBL-derived data, where variability in protocols may impact experimental readouts and calculated metabolic stability values. In this study, we systematically evaluate the impact of data heterogeneity versus homogeneity on model performance for HLM stability prediction. Three standardized datasets from NCATS, Biogen, and ChEMBL and combinations of these datasets were used to train k-nearest neighbor (KNN), support vector machine classifiers (SVC), extreme gradient boosting (XGBoost), and Chemprop models. Generalizability was assessed against two external test sets (Astra-Zeneca and Polaris). Our results provide insights into how dataset source and composition influence predictive performance and highlight the trade-offs between leveraging large, heterogeneous datasets and smaller, homogeneous ones in developing robust, generalizable models for metabolic stability predictions.