Deep mutational scanning of CYP2C9, CYP2C19, and NUDT15 shows that pharmacogene variant interpretation requires assay-specific functional data.
TL;DR
This dimensionality dominates the data, and a supervised ESM-2 sequence baseline was benchmarked against the ESM1v zero-shot ensemble and AlphaMissense under position-based 5-fold cross-validation, together with three architectural extensions: AlphaFold structural features, multi-task learning across paired assays, and contact-graph neural networks.
Abstract
Pharmacogene missense variants can disrupt protein stability, catalytic competence, or substrate handling through distinct mechanisms. General-purpose predictors estimate clinical pathogenicity as a single scalar, whereas pharmacogene interpretation requires knowing which biochemical dimension a variant perturbs, since that determines whether reduced function is substrate-dependent. Five deep mutational scanning datasets comprising 26,198 missense variants across CYP2C9, CYP2C19, and NUDT15 were assembled from MaveDB. Paired assays showed that this dimensionality dominates the data: 28% of CYP2C9 variants (1,236 of 4,421) decoupled catalytic activity from abundance, and 48% of NUDT15 variants (1,364 of 2,844) decoupled thiopurine sensitivity from stability, with CYP2C9 discordance concentrating at substrate-channel residues. AlphaMissense, a representative general-purpose pathogenicity predictor, scored these classes in line with its clinical training objective rather than the assayed biochemistry, assigning likely-benign scores to 38 of 195 stable-but-dead CYP2C9 variants and likely-pathogenic scores to 140 of 222 destabilized but thiopurine-resistant NUDT15 variants. To test whether this dimensionality is recoverable, a supervised ESM-2 sequence baseline was benchmarked against the ESM1v zero-shot ensemble and AlphaMissense under position-based 5-fold cross-validation, together with three architectural extensions: AlphaFold structural features, multi-task learning across paired assays, and contact-graph neural networks. The baseline reached Pearson r of 0.54-0.72, matching or marginally exceeding both comparators, and no extension improved upon it. Trained directly on each assay, it nonetheless recovered the paired-assay difference at r = 0.28 for CYP2C9 and 0.43 for NUDT15, separating discordant variants at AUROC 0.60 and 0.51. Pharmacogene interpretation therefore requires assay-specific, substrate-aware functional measurements rather than a single generic score.
