Cohort-specific imputation of gene expression improves prediction of warfarin dose for African Americans

Abstract

Background: Genome-wide association studies are useful for discovering genotype-phenotype associations but are limited because they require large cohorts to identify a signal, which can be population-specific. Mapping genetic variation to genes improves power and allows the effects of both protein-coding variation as well as variation in expression to be combined into "gene level" effects.

Methods: Previous work has shown that warfarin dose can be predicted using information from genetic variation that affects protein-coding regions. Here, we introduce a method that improves dose prediction by integrating tissue-specific gene expression. In particular, we use drug pathways and expression quantitative trait loci knowledge to impute gene expression-on the assumption that differential expression of key pathway genes may impact dose requirement. We focus on 116 genes from the pharmacokinetic and pharmacodynamic pathways of warfarin within training and validation sets comprising both European and African-descent individuals.

Results: We build gene-tissue signatures associated with warfarin dose in a cohort-specific manner and identify a signature of 11 gene-tissue pairs that significantly augments the International Warfarin Pharmacogenetics Consortium dosage-prediction algorithm in both populations.

Conclusions: Our results demonstrate that imputed expression can improve dose prediction and bridge population-specific compositions. MATLAB code is available at https://github.com/assafgo/warfarin-cohort.

Keywords: African Americans; International Warfarin Pharmacogenetics Consortium; Pharmacogenomics; Warfarin dose.

Fig. 1

Illustration of the use of SNPs, measured in GWAS, to impute expression of drug-associated genes

Fig. 2

Illustration of the feature construction and signature selection methods. First, gene expression is imputed by regression models from cis-SNPs (a). Then, a signature is learned by regressing the drug response on the imputed expression features (b)

Fig. 3

R² results of the predicted unexplained variance in warfarin dose by the IWPC algorithm for the EUR (a) and AA (b) validation cohorts. Represented are the signatures (dark blue), random signatures (red), and signatures on shuffled data (light blue) as the background models for the AA and EUR signatures. EUR and AA in parentheses are the training cohort for the signature; G generic imputation method, CS cohort-specific imputation