Deep mutational scanning of CYP2C9, CYP2C19, and NUDT15 shows that pharmacogene variant interpretation requires assay-specific functional data.

A pharmacogene missense variant can disrupt protein stability, catalytic competence or substrate handling: these dimensions decouple, and a single pathogenicity score does not say which one is affected, nor whether reduced function is substrate-dependent.

Five deep mutational scanning datasets from MaveDB, comprising 26,198 missense variants across CYP2C9, CYP2C19 and NUDT15, were reanalysed by comparing paired assays that measure different biochemical dimensions. Decoupling between these dimensions dominates the data: 28% of CYP2C9 variants (1,236 of 4,421) decouple catalytic activity from protein abundance, and 48% of NUDT15 variants (1,364 of 2,844) decouple thiopurine sensitivity from stability, with CYP2C9 discordance concentrating at substrate-channel residues. AlphaMissense, taken as a representative general-purpose pathogenicity predictor, scored these variants in line with its clinical training objective rather than the assayed biochemistry: likely-benign scores for 38 of 195 stable-but-dead CYP2C9 variants, likely-pathogenic scores for 140 of 222 destabilized but thiopurine-resistant NUDT15 variants. A supervised ESM-2 sequence baseline reached Pearson r of 0.54-0.72, matching or marginally exceeding AlphaMissense and the ESM1v zero-shot ensemble, with no architectural extension improving on it, and recovered the paired-assay difference only at r = 0.28 for CYP2C9 and 0.43 for NUDT15.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

The practical message is negative and useful: a generic pathogenicity score does not tell you what you need to know about a pharmacogene variant, namely whether loss of function is substrate-dependent. A stable but catalytically dead variant and a destabilized but thiopurine-insensitive variant do not behave the same way in front of a prescription, and AlphaMissense gets both wrong. The limitation is that the work offers no operational alternative: the models tested recover the useful dimension poorly, so annotating a CYP2C9 or NUDT15 variant remains dependent on assay-specific functional measurements.

Analysis by Dr Thibaut Benquey