Pharmacogenomic Artificial Intelligence beyond Ancestry Labels through Population-Aware Representation, Transportability, and Uncertainty Download PDF

Pharmacogenomic artificial intelligence seeks to connect molecular variation with treatment selection, dosing, efficacy, and toxicity, yet its development commonly depends on ancestry categories that compress heterogeneous genetic, environmental, clinical, and social information into a small number of labels. Such labels may describe aspects of study composition or support inequity auditing, but they are inadequate stand-alone representations of pharmacogene haplotypes, structural variation, linkage disequilibrium, admixture, environmental exposure, treatment context, or individual source-to-target similarity. This article develops a proposed population-aware pharmacogenomic representation–transportability–uncertainty framework for organizing these non-equivalent determinants without treating ancestry as a biological shortcut. The framework separates molecular pharmacogenomic representation, population-genetic structure, health and environmental context, treatment and endpoint definition, transportability assessment, calibration, uncertainty characterization, and equity governance. It argues that equitable treatment prediction cannot be established through a single discrimination metric, pooled performance estimate, or subgroup comparison. Instead, claims of usefulness should be conditional on representation adequacy, target-population support, external validation, probability calibration, uncertainty reliability, subgroup harm assessment, and clearly bounded decision use. Ancestry, race, and ethnicity may remain relevant as transparently defined contextual variables, but they should not substitute for measured biological or social determinants. The proposed framework is conceptual rather than empirically validated and does not establish clinical utility, causal treatment effects, regulatory acceptability, or deployment readiness. Its principal implication is that pharmacogenomic artificial intelligence should be evaluated as a context-dependent pharmaceutical evidence system whose reliability depends on what is represented, where predictions are transferred, how uncertainty is communicated, and whether errors and benefits are distributed equitably.