Predicting CYP450 Induction and PXR Activation with Machine Learning Approaches: Challenges and Opportunities

Abstract Many drugs increase the activity of cytochrome P450 (CYP450) enzymes, an effect known as induction, which is a major cause of harmful drug–drug interactions (DDIs). Predicting induction from chemical structure would help identify DDI risks early in compound development. Yet, unlike many other drug properties, CYP450 induction is rarely modeled, because the data needed to train models are scarce and expensive to produce. Using an industry-scale data set covering induction of three CYP450 enzymes (CYP3A4, CYP2B6, CYP1A2) and PXR activation, we benchmark four machine learning models across six molecular representations and compare performance against models trained on a curated public data set. On the internal data, the best models separated inducers from noninducers moderately well (MCC 0.35–0.46) but could not predict the magnitude of induction. Almost no model or representation performed significantly better than the others, pointing to the data rather than the method as the limiting factor. The same pattern was observed for public data, where a large public PXR set reached this range, whereas a small, heterogeneous public CYP3A4 set failed where the data are scarcer and noisier. We then show that the shared biology of these end points carries a usable signal. A model trained on one end point partly predicted the others, and this transfer followed the known regulatory relationships between PXR and the CYP450 enzymes. Hence, these biological links can serve as useful inductive biases for future models. Finally, we found that the models we trained are already reasonably well calibrated, which, combined with their classification performance, makes them useful for prioritizing compounds that are within the applicability domain of the models.