Disease association with frequented regions of genotype graphs

Complex diseases, with many associated genetic and environmental factors, are a challenging target for genomic risk assessment. Genome-wide association studies (GWAS) associate disease status with, and compute risk from, individual common variants, which can be problematic for diseases with many interacting or rare variants. In addition, GWAS typically employ a reference genome which is not built from the subjects of the study, whose genetic background may differ from the reference and whose genetic characterization may be limited. We present a complementary method based on disease association with collections of genotypes, called frequented regions, on a pangenomic graph built from subjects genomes. We introduce the pangenomic genotype graph, which is better suited than sequence graphs to human disease studies. Our method draws out collections of features, across multiple genomic segments, which are associated with disease status. We show that the frequented regions method consistently improves machine-learning classification of disease status over GWAS classification, allowing incorporation of rare or interacting variants. Notably, genomic segments that have few or no variants of genome-wide signif-icance (p < 5 x 10-8) provide much-improved classification with frequented regions, encouraging their application across the entire genome. Frequented regions may also be utilized for purposes such as choice of treatment in addition to prediction of disease risk.