Example of sHDL

Installation

To install the latest version of sHDL package via Github, run the following code in R:

# install.packages('remotes')
remotes::install_github("now2014/sHDL", ref='main')

The data.table package is also required for this example:

install.packages('data.table')

Data preparation

Pre-computed LD reference panels

There are three publically available pre-computed autosome LD reference panels for the European-ancestry population (335,265 Genomic British individuals in UK Biobank) from HDL:

Panel 1. QCed UK Biobank Axiom Array (307,519 SNPs):

wget -c -t 1 \
  https://www.dropbox.com/s/fuvpwsf6r8tjd6c/UKB_array_SVD_eigen90_extraction.tar.gz?dl=0 \
  --no-check-certificate -O /path/to/UKB_array_SVD_eigen90_extraction.tar.gz

Panel 2. QCed UK Biobank imputed HapMap2 (769,306 SNPs):

wget -c -t 1 \
  https://www.dropbox.com/s/4vuktycxz1an6sp/UKB_imputed_hapmap2_SVD_eigen99_extraction.tar.gz?dl=0 \
  --no-check-certificate -O /path/to/UKB_imputed_hapmap2_SVD_eigen99_extraction.tar.gz

Panel 3. QCed UK Biobank imputed HapMap3 (1,029,876 SNPs):

wget -c -t 1 \
  https://www.dropbox.com/s/6js1dzy4tkc3gac/UKB_imputed_SVD_eigen99_extraction.tar.gz?dl=0 \
  --no-check-certificate -O /path/to/UKB_imputed_SVD_eigen99_extraction.tar.gz

You may download one of the above reference panel according to the SNP density of your GWAS summary statistics, or build your own LD reference panel. Alternatively, you can download the reference panels via Baidu Netdisk referring to this FAQ.

Here, we illustrate the usage of sHDL with the Panel 2. QCed UK Biobank Axiom Array.

Step 01. Download the LD reference panel.

wget -c -t 1 https://www.dropbox.com/s/fuvpwsf6r8tjd6c/UKB_array_SVD_eigen90_extraction.tar.gz?dl=0 \ --no-check-certificate -O /your/path/to/UKB_array_SVD_eigen90_extraction.tar.gz

Step 02. Extract files from the archive.

# checking the md5 = ff3fadd7ea08bd29759b6c652618cd1f
md5sum /your/path/to/UKB_array_SVD_eigen90_extraction.tar.gz

# extraction
tar -xvf /your/path/to/UKB_array_SVD_eigen90_extraction.tar.gz

GWAS summary statistics

You can test with the dataset from HDL directly:

data(gwas1.example, package='HDL') # GWAS summary statistics from the Neale Lab round 2 GWAS of UK Biobank of birth weight.
gwas.df <- gwas1.example[, c('SNP', 'A1', 'A2', 'Z', 'N')]
saveRDS(gwas1.example, file='gwas.df.rds') # save the gwas.df for sHDL

Or you can prepare your data as follows:

Step 01. Download the Height GWAS summary statistics from the GIANT consortium.

wget -c https://portals.broadinstitute.org/collaboration/giant/images/f/f7/GIANT_HEIGHT_YENGO_2022_GWAS_SUMMARY_STATS_EUR.gz

Step 02. Load the GWAS summary statistics into R.

gwas.df <- data.table::fread('/path/to/GIANT_HEIGHT_YENGO_2022_GWAS_SUMMARY_STATS_EUR.gz', fill=TRUE, header=TRUE)

head(gwas.df, 5) # show the first five rows
#           SNPID       RSID CHR    POS EFFECT_ALLELE OTHER_ALLELE EFFECT_ALLELE_FREQ         BETA      SE        P       N
# 1: 1:566875:C:T  rs2185539   1 566875             T            C           0.002800 -0.046315600 0.03930 0.238128  537968
# 2: 1:728951:C:T rs11240767   1 728951             T            C           0.000356  0.167358000 0.12600 0.185025   85591
# 3: 1:734462:A:G rs12564807   1 734462             A            G           0.893000  0.004656900 0.01100 0.672866  112953
# 4: 1:752721:A:G  rs3131972   1 752721             G            A           0.840000  0.000544089 0.00284  0.84811  615932
# 5: 1:754182:A:G  rs3131969   1 754182             G            A           0.865000  0.001333110 0.00185 0.470389 1100634

Step 03. Select & rename 4 required columns for sHDL:

SNP: rsid of variants
A1: effect allele
A2: other allele
Z: GWAS Z-score
N: Sample size

gwas.df$Z <- gwas.df$BETA / gwas.df$SE # calculate the GWAS Z-score
gwas.df <- gwas.df[, c('RSID', 'EFFECT_ALLELE', 'OTHER_ALLELE', 'Z', 'N')]
colnames(gwas.df) <- c('SNP', 'A1', 'A2', 'Z', 'N')

head(gwas.df, 5)
#           SNP A1 A2          Z       N
# 1:  rs2185539  T  C -1.1785140  537968
# 2: rs11240767  T  C  1.3282381   85591
# 3: rs12564807  A  G  0.4233545  112953
# 4:  rs3131972  G  A  0.1915806  615932
# 5:  rs3131969  G  A  0.7206000 1100634

saveRDS(gwas.df, file='gwas.df.rds') # save the gwas.df for sHDL

Genomic annotation

In this example, we will create a binary annotation of all genes based on the gene locations (hg19).

Step 01. Download the gene location information from GeneBreak package.

wget -c https://raw.githubusercontent.com/stefvanlieshout/GeneBreak/master/data/ens.gene.ann.hg19.rda

Step 02. Create a binary annotation of all genes.

load('ens.gene.ann.hg19.rda') # load ens.gene.ann.hg19
LD.path <- '/your/path/to/UKB_array_SVD_eigen90_extraction'

## extract gene locations ##
gene.pos <- unique(ens.gene.ann.hg19[, c(3:5)])
colnames(gene.pos) <- c('chrom', 'start', 'end')
autosomes <- as.character(1:22)
gene.pos <- gene.pos[gene.pos$chrom %in% autosomes, ]
gene.pos$chrom <- as.numeric(gene.pos$chrom)

## extract variants position from LD reference panel ##
bim <- do.call(rbind, lapply(
  sHDL:::list.LD.ref.files(LD.path, suffix='.bim'), data.table::fread, header=FALSE
))
bim <- bim[, c(2, 1, 4)]
colnames(bim) <- c('SNP', 'chrom', 'pos')


## match annotated variants ##
gene2snp <- function(gene.pos, bim, mc.cores=4){
  map.gene <- function(i, gene.pos.chr, bim.chr){
    s <- gene.pos.chr[i, 'start']
    e <- gene.pos.chr[i, 'end']
    return(bim.chr[bim.chr$pos >= s & bim.chr$pos <= e, ]$SNP)
  }
  chroms <- unique(bim$chrom)
  all.snps <- c()
  for(cur.chrom in chroms){
    gene.pos.chr <- gene.pos[gene.pos$chrom == cur.chrom, ]
    bim.chr <- bim[bim$chrom == cur.chrom, ]
    snps <- parallel::mclapply(
      seq_len(nrow(gene.pos.chr)), map.gene,
      gene.pos.chr=gene.pos.chr, bim.chr=bim.chr,
      mc.cores=mc.cores
    )
    snps <- unique(unlist(snps))
    all.snps <- c(all.snps, snps)
  }
  return(all.snps)
}
snps <- gene2snp(gene.pos, bim, 4)

## create the annotation weights for sHDL ##
D <- matrix(1, nrow=length(snps), ncol=1, dimnames=list(snps, 'ENS.gene'))

saveRDS(D, file='D.rds') # save the annotation weights for sHDL

Notice: sHDL will automatically align D to the LD reference panel according to rsid, and set the weights of missing variants in the LD reference panel to ZERO.

Run sHDL

library(sHDL)

LD.path <- '/your/path/to/UKB_array_SVD_eigen90_extraction'
gwas.df <- readRDS('gwas.df.rds')
D <- readRDS('D.rds')

t0 <- Sys.time()
res <- sHDL::sHDL(D, gwas.df, LD.path, mc.cores=4, stepwise=TRUE, lam.cut=1, mode='memory', Dr.path=NULL)
print(Sys.time() - t0)

sHDL will finish in ~2 mins with 4 CPU cores.

Down below message in the console during the analysis (birth weight from HDL) :

Starting sHDL analysis...
307519 out of 307519 (100.00%) SNPs in reference panel are available in the GWAS.
Transfoming z (D) to zr (Dr)...
D is not NULL and Dr.path is not provided. The default path ./sHDL_Dr will be used.
Annotation name(s): [ENS.gene].
Applied `minmax` weight nomalization on [147849 (48.078%)] annotated variants. The theoretical upper boundary for enrichment fold is M / Md = [2.080].
Dr saved to 
Starting optimization...
Optimization done in 112.997 seconds.
item    estimation      se      p       note
time    112.996711730957        NA      NA      seconds
h2      0.357640057353687       0.0752782467546387      2.02504822923168e-06    total heritability
intercept       0.920875135658489       0.0241090513453767      3.25119898245832e-319   intercept
eta     -0.177230823347607      0.039603506139107       7.63649707679419e-06    eta
fold.ENS.gene   1.1880681869999 0.0304909124069927      6.9155713083808e-10     enrichment fold
converged       NA      NA      NA      TRUE
message NA      NA      NA      CONVERGENCE: REL_REDUCTION_OF_F <= FACTR*EPSMCH
Time difference of 2.824235 mins

Results of sHDL

item	estimation	se	p	note
time	112.9967117	NA	NA	seconds
h2	0.3576401	0.0752782	2.0e-06	total heritability
intercept	0.9208751	0.0241091	0.0e+00	intercept
eta	-0.1772308	0.0396035	7.6e-06	eta
fold.ENS.gene	1.1880682	0.0304909	0.0e+00	enrichment fold
converged	NA	NA	NA	TRUE
message	NA	NA	NA	CONVERGENCE: REL_REDUCTION_OF_F <= FACTR*EPSMCH

Notice:

The item time only indicates the time for maximizing log likelihood, NOT including the time for converting z (D) to zr (Dr).
The null hypothesis of p value for the fold parameter is \(fold \ne 1\), NOT \(fold > 1\).