Advancements in forecasting liver cytosol metabolic stability: a comprehensive update on machine learning predictive models.

Motivation: Drug metabolism plays a key role in determining a compound's pharmacokinetics and overall success in development. While cytochrome P450 enzymes have been widely studied, non-cytochrome P450 enzymes, such as aldehyde oxidase, xanthine oxidase, and carboxylesterases, are increasingly recognized as important contributors to drug clearance. Since standard liver microsomal screening fails to include these enzymes, stability assessment using liver cytosols is essential to prevent late-stage development failures. Results: We expanded our dataset to over 7000 cytosol metabolic stability measurements for human, mouse, and rat. We developed and evaluated machine learning models using molecular fingerprints and descriptors, with performance assessed on scaffold-based external test sets. Class imbalance was addressed using stratified bagging which improved identification of minority class compounds. The final models achieved an external balanced accuracy (BACC) of 0.812 for human and 0.658 for mouse. The rat model, built from a dataset with severe minority class scarcity, reached a BACC of 0.724 and the highest area under the curve (0.886). This work improves the BACC of our previous human model by 29.7% and introduces the first publicly available models for mouse and rat. Availability and implementation: These in silico models are freely accessible through the ADME@NCATS platform for early-stage assessment of species-specific cytosol metabolic stability.