Computational variant predictors for pharmacogenomics: from evaluation of single alleles to assessment of adverse drug reactions to antidepressants

Currently, pharmacogenetics relies on partially annotated star alleles, leaving novel variants and complex haplotypes uninterpretable. Computational scoring frameworks could overcome these limitations. Here, we comprehensively evaluated the ability of existing (CADD, FATHMM-XF, PROVEAN, MutationAssessor, SIFT, PhyloP100, APF, APF2) and novel (PharmGScore and PharmMLScore) variant effect predictors to assess pharmacogenetic alleles in multiple scenarios. Altogether we analyzed 541 PharmVar alleles, high‑throughput CYP2C9 and CYP2C19 mutational maps, and 200 642 UK Biobank exomes linked with health records containing antidepressant treatment outcomes. Many evaluated tools, especially ensemble frameworks, matched or exceeded star allele classifications (ROC‑AUC up to 0.85 for allele definitions, 0.95 in vitro; TPR up to 0.99 for exomes) and accurately predicted severe antidepressant adverse events for carriers of deleterious variants in CYP2C19 (OR 1.20-1.35). Our findings show that computational predictors deliver star allele accuracy while overcoming their limitations. With additional validation, computational tools could enhance clinical decision frameworks by enabling continuous scoring, incorporating previously unknown variants, and providing genome-wide applicability.