Methods for collapsing multiple rare variants in whole-genome sequence data

Genetic Analysis Workshop 18 provided whole-genome sequence data in a pedigree-based sample and longitudinal phenotype data for hypertension and related traits, presenting an excellent opportunity for evaluating analysis choices. We summarize the nine contributions to the working group on collapsing methods, which evaluated various approaches for the analysis of multiple rare variants. One contributor defined a variant prioritization scheme, whereas the remaining eight contributors evaluated statistical methods for association analysis. Six contributors chose the gene as the genomic region for collapsing variants, whereas three contributors chose nonoverlapping sliding windows across the entire genome. Statistical methods spanned most of the published methods, including well-established burden tests, variance-components-type tests, and recently developed hybrid approaches. Lesser known methods, such as functional principal components analysis, higher criticism, and homozygosity association, and some newly introduced methods were also used. We found that performance of these methods depended on the characteristics of the genomic region, such as effect size and direction of variants under consideration. Except for MAP4 and FLT3, the performance of all statistical methods to identify rare casual variants was disappointingly poor, providing overall power almost identical to the type I error. This poor performance may have arisen from a combination of (1) small sample size, (2) small effects of most of the causal variants, explaining a small fraction of variance, (3) use of incomplete annotation information, and (4) linkage disequilibrium between causal variants in a gene and noncausal variants in nearby genes. Our findings demonstrate challenges in analyzing rare variants identified from sequence data.