High-definition likelihood inference of genetic correlations across human complex traits

Abstract

A general objective of genetic studies is to understand the genetic basis of complex traits such as height, body mass index (BMI), disease endpoints, etc. Such researches have been facilitated due to the completion of the human genome project and developments of high-throughput technologies. With the help of high-throughput genotyping and sequencing technologies, the information on millions of genetic markers can be measured for each individual. The most widely used strategy to detect the associations between genetic variants and a complex trait is genome-wide association study (GWAS). Because the genetic architecture of most complex traits is highly polygenic, the signal to noise ratio is usually tiny. Thus, especially in human populations, GWAS often requires large samples to obtain sufficient power. Unfortunately, given the restrictions on sharing individual-level data, it is often not feasible to pool data from different cohorts. Despite that, in each cohort, it is possible to report and share GWAS summary statistics, such as sample sizes, allele frequencies, estimates of genetic effect sizes, and their standard errors for the genetic markers across the genome. Therefore one recent focus in statistical methodology development for genetic studies has been on meta-analysis techniques using summary-level data. The objective of this thesis is to develop novel statistical genetics methods based on GWAS summary statistics and to apply these methods to better understand the genetic architecture underlying complex traits. In Study I, we developed a Selection Operator for JOint analyzing multiple SNPs (SOJO). We mathematically proved and empirically showed that the least absolute shrinkage and selection operator (LASSO) could be achieved using GWAS summary-level data. Compared to the stepwise selection procedures, SOJO performs better in variable selection. SOJO is useful for detecting additional variants with independent effects and assessing the magnitude of allelic heterogeneity within loci. In Study II, we developed a High-Definition Likelihood (HDL) method to improve the accuracy in genetic correlation estimation using GWAS summary statistics. Compared to the stateof-the-art method LD Score regression (LDSC), HDL achieves higher statistical power to detect genetic correlations between phenotypes by fully accounting for linkage disequilibrium (LD) information across the genome. In Study III, we introduced a four-level strategy for replication of loci detected by multi-trait GWAS methods. The four methods provide different degrees of replication strength, useful for providing additional evidence when a locus has been discovered and replicated by multivariate analysis of variance (MANOVA) or other multi-trait methods. The replication methods only require summary association statistics and are straightforward to be applied to multi-trait GWAS analyses. In Study IV, using GWAS summary statistics, we developed a method named Genetic Correlation Contrast for Causality (G3C) as a more robust test for the existence and di...

Comparison of methods for estimating genetic correlation between complex traits using GWAS summary statistics

“…We considered four genetic correlation estimation methods in our comparisons, including REML 6 which requires individual-level data, and three summary-statistics-based methods, LDSC 3 , GNOVA 4 , and HDL 5 . We included REML in the analyses to evaluate the loss of estimation precision when using summary statistics only.…”

“…In this section, the genotype data used to generate the phenotypes were also used as the LD reference panel. To implement the software of HDL, we used its provided reference panel based on 307,519 SNPs of 336,000 white British individuals from the UKBB Axiom Array data 5 . We note that both LDSC and GNOVA can use customized reference panels, hence we took the SNPs overlapping between quality controlled UKBB SNPs and UKBB Axiom Array SNPs in the HDL reference panel to construct the reference panel for LDSC and GNOVA.…”

“…Although LDSC is the most widely used method in terms of genome-wide genetic correlation estimation, Lu et al 4 used numerical and theoretical illustrations to show that the GNOVA estimator is more accurate. A recently proposed method, high-definition likelihood (HDL) 5 , also outperformed LDSC in simulations. Although simulations were conducted to investigate the performance of these methods in each original publication, the simulation settings were largely restricted to those that could demonstrate the merits of each method.…”

“…In the past 15 years, thousands of genome-wide association studies (GWAS) have been conducted, successfully identifying tens of thousands of single-nucleotide polymorphisms (SNPs) associated with complex human traits and diseases 2 . Beside association mapping for individual traits, methods have been developed to estimate genetic correlation based on linear mixed models (LMM) which use individual genotype data of independent subjects from GWAS as the design matrix [3][4][5][6] . Compared with traditional family-based approaches 7,8 , these GWAS-based methods do not need to collect related samples.…”

“…Genetic correlation estimation methods can be classified as methods requiring individual-level data 6,[10][11][12][13] and methods that use GWAS summary statistics as input [3][4][5][14][15][16][17][18][19] . Restricted maximum likelihood (REML) is the most common approach among individual-level-data-based methods where genetic correlation is estimated as one of the (co)variance component parameters of LMM.…”