Detecting signals in pharmacogenomic genome-wide association studies

Abstract

In one common pharmacogenomic scenario, outcome measures are compared for treated and untreated subjects across genotype defined subgroups. The key question is whether treatment benefit (or harm) is particularly strong in certain subgroups, and therefore statistical analysis focuses on the interaction between treatment and genotype. However, genome-wide analysis in such scenarios requires careful statistical thought since, in addition to the usual problems of multiple testing, the marker-defined sample sizes, and therefore power, vary across the individual genotypes being evaluated. The variability in power means the usual practice of using a common p-value threshold across tests has difficulties. The reason is that the use of a fixed threshold, with variable power, implies that the costs of type I and type II errors are varying across tests in a manner which is implicit rather than dictated by the analyst. In this paper we discuss this problem and describe an easily implementable solution based on Bayes factors. We pay particular attention to the specification of priors, which is not a straightforward task. The methods are illustrated using data from a randomized controlled clinical trial in which homocysteine levels are compared in individuals receiving low and high doses of folate supplements and across marker subgroups. The method we describe is implemented in the R computing environment with code available from http://faculty.washington.edu/jonno/cv.html.

Exploring predictive biomarkers from clinical genome-wide association studies via multidimensional hierarchical mixture models

“…Although the biological mechanisms of the harmful effect with low-dose administration are unclear, this result suggests the need for high-dose administration of multivitamin tablets in stroke patients. Another SNP, rs1739317, which has also been reported previously [13] and which reached a suggestive level (p<10 -6 ) using the standard regression model ( Supplementary Fig. S2a), was also detected.…”

“…The ODP detected five independent SNPs (FDR < 5%) that are associated with outcomes. Of these, rs12631354, rs2367209, and rs10017302 are newly detected in this analysis while others have previously been suggested as being associated with homocysteine levels [13]. In particular, rs12631354 and rs10017302 might significantly change the effect of the administration of multivitamin tablets.…”

“…The ODP also detected SNPs that have been suggested as being associated with homocysteine levels [13], as well as rs3736238, which has been previously reported as the most statistically significant SNP [13] and which reached genome-wide significance (p<5×10 -8 ) using the standard regression model with interaction terms (see Section E in the Supplementary Notes and Supplementary Fig. S2a).…”

“…In this study, we used this dataset to investigate SNPs associated with blood homocysteine levels which are strongly associated with the risk of stroke. We took as outcome the difference in blood homocysteine levels between baseline and the first post-baseline measurements, as in the study of Wakefield et al [13]. Association tests were conducted using a linear regression model (see Section C in the Supplementary Notes).…”

A genome-wide investigation into parent-of-origin effects in autism spectrum disorder identifies previously associated genes including SHANK3

“…25,26,30 These genome-wide significance levels are suggested for all sample sizes (and MAFs) and hence do not take into account the power at individual SNPs. 25,31 To take these factors into account, we adopted a Bayesian method proposed by Wakefield 25 to determine an appropriate threshold for identifying noteworthy findings.…”