An Introduction to gsmr

Overview

The gsmr R-package implements the GSMR (Generalised Summary-data-based Mendelian Randomisation) method to test for putative causal association between a risk factor and a disease using summary-level data from genome-wide association studies (GWAS) (Zhu et al. 2018 Nat. Commun.). The R package is developed by Zhihong Zhu, Zhili Zheng, Futao Zhang and Jian Yang at Institute for Molecular Bioscience, the University of Queensland. Bug reports or questions: [email protected].

Note: The GSMR method has also been implemented in the GCTA software (GCTA-GSMR)

Citation

Zhu, Z. et al. (2018) Causal associations between risk factors and common diseases inferred from GWAS summary data. Nat. Commun. 9, 224 (https://www.nature.com/articles/s41467-017-02317-2).

Installation

The gsmr requires R >= 2.15, you can install it in R by:

# gsmr requires the R-package(s)
install.packages(c('survey'));
# install gsmr
install.packages("http://cnsgenomics.com/software/gsmr/static/gsmr_1.0.6.tar.gz",repos=NULL,type="source")

The gsmr source codes are available in gsmr_1.0.6.tar.gz. Sample data are available in test_data.zip.

This document has been integrated in the gsmr R-package, we can check it by the standard command “?function_name” in R.

Update log

V1.0.6 (gmr_1.0.6.tar.gz PDF, 23 Jan. 2018): Added a function to remove SNPs in high LD.

V1.0.5 (gmr_1.0.5.tar.gz PDF, 13 Dec. 2017): Improved the approximation of the sampling covariance matrix.

V1.0.4 (gsmr_1.0.4.tar.gz PDF, 6 Nov. 2017): Add the bi-directional GSMR analysis. The HEIDI-outlier analysis has been integrated in the GSMR analysis by default.

V1.0.3 (gsmr_1.0.3.tar.gz PDF, 12 Oct. 2017): Add more example data.

Removed the initial versions (8 Nov 2016).

Tutorial

The GSMR analysis only requires summary-level data from GWAS. Here is an example, where the risk factor (x) is LDL cholesterol (LDL-c) and the disease (y) is coronary artery disease (CAD). GWAS summary data for both LDL-c and CAD are available in the public domain (Global Lipids Genetics Consortium et al. 2013, Nature Genetics; Nikpay, M. et al. 2015, Nature Genetics).

1. Prepare data for GSMR analysis

1.1 Load the GWAS summary data

library("gsmr")
data("gsmr")
head(gsmr_data)

##          SNP a1 a2    a1_freq     bzx bzx_se  bzx_pval    bzx_n       bzy
## 1 rs10903129  A  G 0.45001947 -0.0328 0.0037 3.030e-17 169920.0  0.008038
## 2 rs12748152  T  C 0.08087758  0.0499 0.0066 3.209e-12 172987.5  0.013671
## 3 rs11206508  A  G 0.14396988  0.0434 0.0055 2.256e-14 172239.0  0.030222
## 4 rs11206510  C  T 0.19128911 -0.0831 0.0050 2.380e-53 172812.0 -0.074519
## 5 rs10788994  T  C 0.18395430  0.0687 0.0049 8.867e-41 172941.9  0.038267
## 6   rs529787  G  C 0.19713099 -0.0553 0.0052 8.746e-24 161969.0  0.001707
##      bzy_se     bzy_pval  bzy_n
## 1 0.0092442 0.3845651000 184305
## 2 0.0185515 0.4611690000 184305
## 3 0.0141781 0.0330400000 184305
## 4 0.0133438 0.0000000234 184305
## 5 0.0118752 0.0012711000 184305
## 6 0.0135491 0.8997431000 184305

dim(gsmr_data)

## [1] 189  12

This is the input format for the GSMR analysis. In this data set, there are 189 near-independent SNPs associated with LDL-c at a genome-wide significance level (i.e. p < 5e-8).

SNP: the genetic instrument
a1: effect allele
a2: the other allele
a1_freq: frequency of a1
bzx: the effect size of a1 on risk factor
bzx_se: standard error of bzx
bzx_pval: p value for bzx
bzx_n: per-SNP sample size of GWAS for the risk factor
bzy: the effect size of a1 on disease
bzy_se: standard error of bzy
bzy_pval: p value for bzy
bzy_n: per-SNP sample size of GWAS for the disease

1.2 Estimate the LD correlation matrix

# Save the genetic variants and effect alleles in a text file using R
write.table(gsmr_data[,c(1,2)], "gsmr_example_snps.allele", col.names=F, row.names=F, quote=F)
# Extract the genotype data from a GWAS dataset using GCTA
gcta64 --bfile gsmr_example --extract gsmr_example_snps.allele --update-ref-allele gsmr_example_snps.allele --recode --out gsmr_example

Note: the two steps above guarantee that the LD correlations are calculated based on the effect alleles for the SNP effects.

# Estimate LD correlation matrix using R
snp_coeff_id = scan("gsmr_example.xmat.gz", what="", nlines=1)
snp_coeff = read.table("gsmr_example.xmat.gz", header=F, skip=2)

# Match the SNP genotype data with the summary data
snp_id = Reduce(intersect, list(gsmr_data$SNP, snp_coeff_id))
gsmr_data = gsmr_data[match(snp_id, gsmr_data$SNP),]
snp_order = match(snp_id, snp_coeff_id)
snp_coeff_id = snp_coeff_id[snp_order]
snp_coeff = snp_coeff[, snp_order]

# Calculate the LD correlation matrix
ldrho = cor(snp_coeff)

# Check the size of the correlation matrix and double-check if the order of the SNPs in the LD correlation matrix is consistent with that in the GWAS summary data
colnames(ldrho) = rownames(ldrho) = snp_coeff_id

dim(ldrho)

## [1] 189 189

# Show the first 5 rows and columns of the matrix  
ldrho[1:5,1:5]

##              rs10903129    rs12748152   rs11206508  rs11206510    rs10788994
## rs10903129  1.000000000 -0.0045378845  0.008066621 -0.01372112 -0.0234447102
## rs12748152 -0.004537884  1.0000000000 -0.006687181  0.00445927  0.0003629201
## rs11206508  0.008066621 -0.0066871806  1.000000000 -0.21125757  0.0512593434
## rs11206510 -0.013721120  0.0044592696 -0.211257567  1.00000000 -0.1842706205
## rs10788994 -0.023444710  0.0003629201  0.051259343 -0.18427062  1.0000000000

Note: all the analyses implemented in this R-package only require the summary data (e.g. “gsmr_data”) and the LD correlation matrix (e.g. “ldrho”) listed above.

2. Standardization

This is an optional process. If the risk factor was not standardised in GWAS, the effect sizes can be scaled using the method below. Note that this process requires allele frequencies, z-statistics and sample size. After scaling, bzx is interpreted as the per-allele effect of a SNP on the exposure in standard deviation units.

snpfreq = gsmr_data$a1_freq             # allele frequencies of the SNPs
bzx = gsmr_data$bzx     # effects of the instruments on risk factor
bzx_se = gsmr_data$bzx_se       # standard errors of bzx
bzx_n = gsmr_data$bzx_n          # GWAS sample size for the risk factor
std_zx = std_effect(snpfreq, bzx, bzx_se, bzx_n)    # perform standardisation
gsmr_data$std_bzx = std_zx$b    # standardized bzx
gsmr_data$std_bzx_se = std_zx$se    # standardized bzx_se
head(gsmr_data)

##          SNP a1 a2    a1_freq     bzx bzx_se  bzx_pval    bzx_n       bzy
## 1 rs10903129  A  G 0.45001947 -0.0328 0.0037 3.030e-17 169920.0  0.008038
## 2 rs12748152  T  C 0.08087758  0.0499 0.0066 3.209e-12 172987.5  0.013671
## 3 rs11206508  A  G 0.14396988  0.0434 0.0055 2.256e-14 172239.0  0.030222
## 4 rs11206510  C  T 0.19128911 -0.0831 0.0050 2.380e-53 172812.0 -0.074519
## 5 rs10788994  T  C 0.18395430  0.0687 0.0049 8.867e-41 172941.9  0.038267
## 6   rs529787  G  C 0.19713099 -0.0553 0.0052 8.746e-24 161969.0  0.001707
##      bzy_se     bzy_pval  bzy_n     std_bzx  std_bzx_se
## 1 0.0092442 0.3845651000 184305 -0.03055942 0.003447252
## 2 0.0185515 0.4611690000 184305  0.04713698 0.006234550
## 3 0.0141781 0.0330400000 184305  0.03829018 0.004852442
## 4 0.0133438 0.0000000234 184305 -0.07181919 0.004321251
## 5 0.0118752 0.0012711000 184305  0.06149455 0.004386074
## 6 0.0135491 0.8997431000 184305 -0.04695042 0.004414868

3. GSMR analysis

This is the main analysis of this R-package. It uses SNPs associated with the risk factor (e.g. at p < 5e-8) as the instruments to test for putative causal effect of the risk factor on the disease. The analysis involves a step that uses the HEIDI-outlier approach to remove SNPs that have effects on both the risk factor and the disease because of pleiotropy.

bzx = gsmr_data$std_bzx    # SNP effects on the risk factor
bzx_se = gsmr_data$std_bzx_se    # standard errors of bzx
bzx_pval = gsmr_data$bzx_pval   # p-values for bzx
bzy = gsmr_data$bzy    # SNP effects on the disease
bzy_se = gsmr_data$bzy_se    # standard errors of bzy
bzy_pval = gsmr_data$bzy_pval    # p-values for bzy
n_ref = 7703    # Sample size of the reference sample
gwas_thresh = 5e-8    # GWAS threshold to select SNPs as the instruments for the GSMR analysis
heidi_outlier_thresh = 0.01    # HEIDI-outlier threshold
nsnps_thresh = 10   # the minimum number of instruments required for the GSMR analysis
heidi_outlier_flag = T    # flag for HEIDI-outlier analysis
ld_r2_thresh = 0.1    # LD r2 threshold to remove SNPs in high LD
ld_fdr_thresh = 0.05   # FDR threshold to remove the chance correlations between the SNP instruments
gsmr_results = gsmr(bzx, bzx_se, bzx_pval, bzy, bzy_se, ldrho, snp_coeff_id, n_ref, heidi_outlier_flag, gwas_thresh, heidi_outlier_thresh, nsnps_thresh, ld_r2_thresh, ld_fdr_thresh)    # GSMR analysis 
cat("The estimated effect of the exposure on outcome: ",gsmr_results$bxy)

## The estimated effect of the exposure on outcome:  0.4082179

cat("Standard error of bxy: ",gsmr_results$bxy_se)

## Standard error of bxy:  0.02294163

cat("P-value for bxy: ", gsmr_results$bxy_pval)

## P-value for bxy:  7.898847e-71

cat("Indexes of the SNPs used in the GSMR analysis: ", gsmr_results$used_index[1:5], "...")

## Indexes of the SNPs used in the GSMR analysis:  1 2 3 5 6 ...

cat("Number of SNPs with missing estimates in the summary data: ", length(gsmr_results$na_snps))

## Number of SNPs with missing estimates in the summary data:  0

cat("Number of non-significant SNPs: ", length(gsmr_results$weak_snps))

## Number of non-significant SNPs:  38

cat("Number of SNPs in high LD ( LD rsq >", ld_r2_thresh, "): ", length(gsmr_results$linkage_snps))

## Number of SNPs in high LD ( LD rsq > 0.1 ):  2

cat("Number of pleiotropic outliers: ", length(gsmr_results$pleio_snps))

## Number of pleiotropic outliers:  12

4. HEIDI-outlier analysis

The estimate of causal effect of risk factor on disease can be biased by pleiotropy (Zhu et al. 2018 Nat. Commun). This is an analysis to detect and eliminate from the analysis instruments that show significant pleiotropic effects on both risk factor and disease. The HEIDI-outlier analysis requires bzx (effect of genetic instrument on risk factor), bzx_se (standard error of bzx), bzx_pval (p-value of bzx), bzy (effect of genetic instrument on disease), bzy_se (standard error of bzy) and ldrho (LD matrix of instruments). Note that similar to that in the GSMR analysis above, the LD matrix can be estimated from a reference sample with individual-level genotype data.

The HEIDI-outlier analysis has been integrated in the GSMR analysis above (with the heidi_outlier_flag and heidi_outlier_thresh flags). It can also be performed separately following the example below.

heidi_results = heidi_outlier(bzx, bzx_se, bzx_pval, bzy, bzy_se, ldrho, snp_coeff_id, n_ref, gwas_thresh, heidi_outlier_thresh, nsnps_thresh, ld_r2_thresh, ld_fdr_thresh) # perform HEIDI-outlier analysis
cat("Number of SNPs in high LD ( LD rsq >", ld_r2_thresh, "): ", length(gsmr_results$linkage_snps))

## Number of SNPs in high LD ( LD rsq > 0.1 ):  2

cat("Number of pleiotropic outliers: ", length(heidi_results$pleio_snps))

## Number of pleiotropic outliers:  12

filtered_index = heidi_results$remain_index
filtered_gsmr_data = gsmr_data[filtered_index,]   # select data passed HEIDI-outlier filtering
filtered_snp_id = snp_coeff_id[filtered_index]   # select SNPs that passed HEIDI-outlier filtering
dim(filtered_gsmr_data)

## [1] 137  14

# Number of SNPs in the gmsr_data with bzx_pval < 5e-8
dim(gsmr_data[gsmr_data$bzx_pval < 5e-8, ])

## [1] 151  14

In the example above, 14 SNPs are filtered out by HEIDI-outlier.

5. Bi-directional GSMR analysis

The script below runs bi-directional GSMR analyses, i.e. a forward-GSMR analysis as described above and a reverse-GSMR analysis that uses SNPs associated with the disease (e.g. at p < 5e-8) as the instruments to test for putative causal effect of the disease on risk factor.

gsmr_results = bi_gsmr(bzx, bzx_se, bzx_pval, bzy, bzy_se, bzy_pval, ldrho, snp_coeff_id, n_ref, heidi_outlier_flag, gwas_thresh, heidi_outlier_thresh, nsnps_thresh, ld_r2_thresh, ld_fdr_thresh)    # GSMR analysis 
cat("Effect of risk factor on disease: ",gsmr_results$forward_bxy)

## Effect of risk factor on disease:  0.4082179

cat("Standard error of bxy in the forward-GSMR analysis: ",gsmr_results$forward_bxy_se)

## Standard error of bxy in the forward-GSMR analysis:  0.02294163

cat("P-value of bxy in the forward-GSMR analysis: ", gsmr_results$forward_bxy_pval)

## P-value of bxy in the forward-GSMR analysis:  7.898847e-71

cat("Effect of disease on risk factor: ",gsmr_results$reverse_bxy)

## Effect of disease on risk factor:  -0.02376614

cat("Standard error of bxy in the reverse-GSMR analysis: ",gsmr_results$reverse_bxy_se)

## Standard error of bxy in the reverse-GSMR analysis:  0.00958462

cat("P-value of bxy in the reverse-GSMR analysis: ", gsmr_results$reverse_bxy_pval)

## P-value of bxy in the reverse-GSMR analysis:  0.01315254

6. Visualization

effect_col = colors()[75]
vals = c(bzx[filtered_index]-bzx_se[filtered_index], bzx[filtered_index]+bzx_se[filtered_index])
xmin = min(vals); xmax = max(vals)
vals = c(bzy[filtered_index]-bzy_se[filtered_index], bzy[filtered_index]+bzy_se[filtered_index])
ymin = min(vals); ymax = max(vals)
par(mar=c(5,5,4,2))
plot(bzx[filtered_index], bzy[filtered_index], pch=20, cex=0.8, bty="n", cex.axis=1.1, cex.lab=1.2,
        col=effect_col, xlim=c(xmin, xmax), ylim=c(ymin, ymax),
        xlab=expression(LDL~cholesterol~(italic(b[zx]))),
        ylab=expression(Coronary~artery~disease~(italic(b[zy]))))
abline(0, gsmr_results$forward_bxy, lwd=1.5, lty=2, col="dim grey")

nsnps = length(bzx[filtered_index])
for( i in 1:nsnps ) {
    # x axis
    xstart = bzx[filtered_index [i]] - bzx_se[filtered_index[i]]; xend = bzx[filtered_index[i]] + bzx_se[filtered_index[i]]
    ystart = bzy[filtered_index[i]]; yend = bzy[filtered_index[i]]
    segments(xstart, ystart, xend, yend, lwd=1.5, col=effect_col)
    # y axis
    xstart = bzx[filtered_index[i]]; xend = bzx[filtered_index[i]] 
    ystart = bzy[filtered_index[i]] - bzy_se[filtered_index[i]]; yend = bzy[filtered_index[i]] + bzy_se[filtered_index[i]]
    segments(xstart, ystart, xend, yend, lwd=1.5, col=effect_col)
}