What is HDL?

High-Definition Likelihood (HDL) is a likelihood-based method for estimating genetic correlation using GWAS summary statistics. Compared to LD Score regression (LDSC), It reduces the variance of a genetic correlation estimate by about 60%. Here, we provide an R-based computational tool HDL to implement our method. Although HDL is written in R, you can use it with the command line. So no worry if you are not an R user.

What data are required?

• gwas1.df A data frame including GWAS summary statistics of genetic variants for trait 1. The input data frame should include following columns: SNP, SNP ID; A1, effect allele; A2, reference allele; N, sample size; Z, z-score; If Z is not given, alternatively, you may provide: b, estimate of marginal effect in GWAS; se, standard error of the estimates of marginal effects in GWAS. The summary statistics should look like this:

##          SNP A1 A2      N        b       se      Z
## 1  rs3131962  G  A 205475 0.001004 0.004590 0.2187
## 2 rs12562034  A  G 205475 0.005382 0.005011 1.0740
## 3 rs11240779  A  G 205475 0.002259 0.003691 0.6119
## 4 rs57181708  G  A 205475 0.005401 0.005114 1.0562
## 5  rs4422948  G  A 205475 0.005368 0.003604 1.4893
## 6  rs4970383  A  C 205475 0.004685 0.003582 1.3080

• gwas2.df A data frame including GWAS summary statistics of genetic variants for trait 2. The format is the same as gwas1.df.

• The eigenvalues and eigenvectors of LD matrices. For the European-ancestry population, we have computed the LD matrices and their eigen-decomposition from 336,000 Genomic British individuals in UK Biobank. You can download these pre-computed reference files here. Two sets of reference panel are provided:

  • 307,519 QCed UK Biobank Axiom Array SNPs. The size is about 7.5 GB after unzipping.
  • 1,029,876 QCed UK Biobank imputed SNPs. The size is about 31 GB after unzipping. Although it takes more time, using the imputed panel provides more accurate estimates of genetic correlations. Therefore if the GWAS includes most of the HapMap3 SNPs, then we recommend using the imputed reference panel.

Installation

Firstly you need an R (version >= 3.5.2) in your computer. If you have not installed R, you can easily download and install it from https://cran.r-project.org/.

Command line user

In order to download the R package HDL and related scripts, you can clone HDL repository and then install it using HDL.install.R:

git clone https://github.com/zhenin/HDL
cd HDL
Rscript HDL.install.R

R user

You can use the following codes to install HDL from Github:

install.packages("https://github.com/zhenin/HDL/raw/master/HDL_1.0.tar.gz", repos=NULL)

and load the package via:

library(HDL)

Estimating genetic correlation using HDL

To illustrate how to use HDL, we include two cleaned UKB GWAS summary statistics datasets as examples. gwas1.example.rda is for birth weight; gwas2.example.rda is for type 2 diabetes.

Command line user

Next, you can simply run HDL.run.R like below to use HDL:

Rscript HDL.run.R gwas1.df=gwas1.example.rda gwas2.df=gwas2.example.rda LD.path=UKB_SVD_eigen90_extraction output.file=test.Rout

There are several arguments you should pass to HDL. Please note that when you specify arguments, there should not be any space on any side of =.

• Mandatory arguments
  • gwas1.df, the file of the GWAS results for trait 1;
  • gwas2.df, the file of the GWAS results for trait 2;
  • LD.path, the path to the directory where linkage disequilibrium (LD) information is stored.
• Optional arguments
  • output.file, where the log and results should be written. If you do not specify a file, the log will be printed in the console;
  • Nref, the sample size of the reference sample where LD is computed. If the default UK Biobank reference sample is used, Nref = 336000;
  • N0, the number of individuals included in both cohorts. However, the estimated genetic correlation is usually robust against misspecified N0. If not given, the default value is set to the minimum sample size across all SNPs in cohort 1 and cohort 2.

R user

HDL.rg is the function to perform HDL. The arguments for HDL.rg are the same as the above arguments for command-line implementation. Let's have a try with example data as below

data(gwas1.example)
data(gwas2.example)
LD.path <- "/Users/zhengning/Work/HDL/package/UKB_SVD_eigen90_extraction"
res.HDL <- HDL.rg(gwas1.example, gwas2.example, LD.path)
res.HDL

A list is returned with

• rg, the estimation of genetic correlation.
• rg.se, the standard error of estimated genetic correlation.
• P, the Wald test P-value for rg.

For direct R documentation of HDL.rg function, you can use a question mark in R:

?HDL.rg

Reading HDL results

The first section provides version information of HDL package:

*********************************************************************
* High-definition likelihood (HDL)
* Version 1.0
*********************************************************************

The next section gives the proportion of overlap SNPs between GWAS summary statistics and reference panel. A low SNP overlap may lead to poor estimation.

Analysis starts 
307519 out of 307519 (100%) SNPs in reference panel are available in GWAS 1.  
307519 out of 307519 (100%) SNPs in reference panel are available in GWAS 2.  

The following section shows whether the estimated rg is within (-1,1). If not, we switch to an algorithm that guarantees a constrained estimate.

Integrating piecewise results 
Estimated rg is within (-1,1). Continuing computing standard error. 

The last section gives the genetic correlation, its standard error, and P-value based on the Wald test.

Genetic Correlation:  -0.2354 (0.0444)
P:  1.146e-07

For Help

If you have specific questions, you may email the maintainer of HDL via zheng.ning@ki.se.