Homology-Based Variant-Effect Predictors Break Down on Cytochrome P450 Pharmacogenes

Cytochrome P450 (CYP) enzymes metabolize roughly three-quarters of clinically used drugs, and genetic variation in these enzymes is a leading source of interindividual differences in drug response. Predicting a variant's functional effect is therefore critical, yet the consequences of most CYP variants remain unknown. Many state-of-the-art variant-effect predictors rest on a homology-based paradigm that scores variants by evolutionary conservation - an assumption that pharmacogenes, including CYPs, violate. Focusing on human CYPs, we show that homology-based models fail systematically. AlphaMissense (AM) assigns variants to its "ambiguous" class at nearly twice the proteome-wide rate across six clinically important CYPs, and within that class its scores are essentially uninformative about measured CYP2C9 deep mutational scan (DMS) activity (ρ=-0.130). Using catalytic activity and cellular abundance from the same CYP2C9 DMS, we show that AM is largely an abundance signal: it tracks activity well for destabilizing, low-abundance variants (ρ=-0.568) but far more weakly for abundance-independent variants (ρ=-0.357; Fisher r-to-z z=-8.99, p<1e-18). We further hypothesized that non-homology-based features (sequence position, substitution chemistry, distance to the substrate-recognition sites, and secondary structure) might account for where AM errs, but their explanatory power is weak. Using CYP2C9 DMS activity as ground truth, we built a k-nearest-neighbors model over ESM-2 embeddings and ensembled it with AM and ESM-2 masked marginal probability, raising the correlation between predicted and measured activity within the ambiguous class from ρ=0.080 for AM alone to 0.399, and reaching an overall ρ=0.721 that exceeds every out-of-the-box predictor we tested (ρ=0.523 to 0.686). However, neither this improved ensemble model nor any other method agrees significantly with clinical annotations. Drawing on evidence that a variant's effect can depend on the drug, we hypothesize that substrate identity is the key missing feature in current models, and that predicting function for these multi-substrate enzymes may require redefining function as substrate-conditioned.