Adapt-Mix: learning local genetic correlation structure improves summary statistics-based analyses

Abstract

Motivation: Approaches to identifying new risk loci, training risk prediction models, imputing untyped variants and fine-mapping causal variants from summary statistics of genome-wide association studies are playing an increasingly important role in the human genetics community. Current summary statistics-based methods rely on global ‘best guess’ reference panels to model the genetic correlation structure of the dataset being studied. This approach, especially in admixed populations, has the potential to produce misleading results, ignores variation in local structure and is not feasible when appropriate reference panels are missing or small. Here, we develop a method, Adapt-Mix, that combines information across all available reference panels to produce estimates of local genetic correlation structure for summary statistics-based methods in arbitrary populations.Results: We applied Adapt-Mix to estimate the genetic correlation structure of both admixed and non-admixed individuals using simulated and real data. We evaluated our method by measuring the performance of two summary statistics-based methods: imputation and joint-testing. When using our method as opposed to the current standard of ‘best guess’ reference panels, we observed a 28% decrease in mean-squared error for imputation and a 73.7% decrease in mean-squared error for joint-testing.Availability and implementation: Our method is publicly available in a software package called ADAPT-Mix available at https://github.com/dpark27/adapt_mix.Contact: noah.zaitlen@ucsf.edu

Transethnic Genetic-Correlation Estimates from Summary Statistics

“…Second, we develop an approach that uses only summary-level GWAS data to estimate genetic correlation across populations. Like other recent methods based on summary statistics, [10][11][12][13][14][15][16][17][18][19][20][21] our approach supplements summary association data with linkage disequilibrium (LD) information from external reference panels, avoids privacy concerns, and is scalable to hundreds of thousands of individuals and millions of markers. Unlike traditional approaches that focus on the similarity of GWAS results, [22][23][24][25][26] we use the entire spectrum of GWAS associations while accounting for LD to avoid filtering correlated SNPs.…”

Dissecting the genetics of complex traits using summary association statistics

“…Imputation is traditionally performed using individual-level data, which requires substantial computational resources and can be logistically cumbersome when new reference panels become available, particularly for large consortia combining data from multiple studies. As an alternative to imputation using individual-level data, approaches have been developed to perform imputation directly at the level of summary statistics 12–18 (providing an alternative to other multivariate tests 19,20 ). The key insight of these approaches is that LD induces correlations between z-scores, which can be modeled using a multi-variate normal (MVN) distribution with variance equal to the LD correlation matrix 21 (an adjustment in the LD computation is needed for z-scores estimated using mixed models 22 ).…”

“…A simple approach to regularization is to set all correlations between distal SNPs to zero, based on a fixed distance threshold 7 or approximately independent LD blocks inferred from the data 23 . An alternative is to specify a prior distribution and compute Bayesian posteriors 12 ; data can be combined across multiple ancestry reference panels to further boost accuracy 17,18 . Singular value decomposition based approaches for LD regularization have also been proposed in other contexts 10 .…”

Genetics and Genomics of Coronary Artery Disease

“…This effort has identified 38 million variants and after selecting variants with minor allele frequency (MAF) > 0.005 that were detectable in over 60 % of the 48 GWA studies, 8.6 million SNPs and 836,000 indels were included in the meta-analysis, which represents to date the largest collection of variants tested for association with CAD. One of the limitations of the study is that the majority (77 %) of the participants were of European ancestry, suggesting that new CAD loci functional in other racial/ethnic groups remain to be discovered [4]. …”