Improving power for robust trans-ethnic meta-analysis of rare and low-frequency variants with a partitioning approach

While genome-wide association studies have discovered numerous bona fide variants that are associated with common diseases and complex traits; these variants tend to be common in the population and explain only a small proportion of the phenotype variance. The search for the missing heritability has thus switched to rare and low-frequency variants, defined as <5% in the population, but which are expected to have a bigger impact on phenotypic outcomes. The rarer nature of these variants coupled with the curse of testing multiple variants across the genome meant that large sample sizes will still be required despite the assumption of bigger effect sizes. Combining data from multiple studies in a meta-analysis will continue to be the natural approach in boosting sample sizes. However, the population genetics of rare variants suggests that allelic and effect size heterogeneity across populations of different ancestries is likely to pose a greater challenge to trans-ethnic meta-analysis of rare variants than to similar analyses of common variants. Here, we introduce a novel method to perform trans-ethnic meta-analysis of rare and low-frequency variants. The approach is centered on partitioning the studies into distinct clusters using local inference of genomic similarity between population groups, with the aim to minimize both the number of clusters and between-study heterogeneity in each cluster. Through a series of simulations, we show that our approach either performs similarly to or outperforms conventional and recently introduced meta-analysis strategies, particularly in the presence of allelic heterogeneity.