Pharmacogenomics holds the promise of personalized drug efficacy optimization and drug toxicity minimization. Much of the research conducted to date, however, suffers from an ascertainment bias towards European participants. Here, we leverage publicly available, whole genome sequencing data collected from global populations, evolutionary characteristics, and annotated protein features to construct a new in silico machine learning pharmacogenetic identification method called XGB-PGX. When applied to pharmacogenetic data, XGB-PGX outperformed all existing prediction methods and identified over 2000 new pharmacogenetic variants. While there are modest pharmacogenetic allele frequency distribution differences across global population samples, the most striking distinction is between the relatively rare putatively neutral pharmacogene variants and the relatively common established and newly predicted functional pharamacogenetic variants. Our findings therefore support a focus on individual patient pharmacogenetic testing rather than on clinical presumptions about patient race, ethnicity, or ancestral geographic residence. We further encourage more attention be given to the impact of common variation on drug response and propose a new 'common treatment, common variant' perspective for pharmacogenetic prediction that is distinct from the types of variation that underlie complex and Mendelian disease. XGB-PGX has identified many new pharmacovariants that are present across all global communities; however, communities that have been underrepresented in genomic research are likely to benefit the most from XGB-PGX's in silico predictions.
Keywords: adaptation; human evolution; machine learning; pharmacogenomic.