Progress and Controversy in Analysis of Complex Phenotypes Based on Genome-wide Association Statistics

摘要

Recently there has been much interest and great progress in statistical modelling of genome-wide association study test statistics for heritability analyses, genetic correlation estimates and to predict complex phenotypes, correcting for GWAS confounding bias if required. The LDSC model has been widely applied since published in 2015, but it's unrealistic implicit assumption of uniform expected heritability across SNPs is now recognized to have led to poor performance. The model has been further developed (now called S-LDSC), including SNP-specific expected heritabilities that adjust for effects of minor allele fraction, linkage disequilibrium and genome annotations. We have proposed a different approach, implemented in the SumHer software, finding in some settings dramatically different inferences from those based on LDSC. As the models have improved, so have methods for model comparison, including an improved log likelihood approximation and an approach based on leave-one-chromosome-out prediction of summary statistics. We report latest developments, showing that as the models have improved, inferences are converging. However, we also report that the adjustment for GWAS confounding bias is unreliable in all current approaches even if the assumed heritability model is correct, so that summary-statistic analyses should be limited to settings in which the original GWAS adequately adjusted for confounding.