Joint Analysis of Methylation and Genotype Data
In ramwas: Fast Methylome-Wide Association Study Pipeline for Enrichment Platforms

library(knitr)
library(pander)
suppressPackageStartupMessages(library(ramwas))
panderOptions("digits", 3)
opts_chunk$set(fig.width = 6, fig.height = 6)
# opts_chunk$set(eval=FALSE)
dr = "D:/temp/"

Statistical model for Joint Analysis of Methylation and Genotype Data

Single nucleotide polymorphisms (SNPs) can create and destroy CpGs. As methylation occurs mostly at CpGs, such CpG-SNPs can directly affect methylation measurements.

Recall that enrichment-based methylation methods measure total methylation in a vicinity of a CpG. By creating or destroying a CpG, CpG-SNPs introduce a variation in the total methylation in a vicinity of the CpG which can greatly reduce our power to detect case-control differences.

RaMWAS can account for a possible effect of CpG-SNPs by testing for joint significance of $\beta_1$ and $\beta_2$ the following model:

$$\mu_i = \beta_0 + outcome * \beta_1 + {outcome} * {SNP}_i * \beta_2 + {SNP}_i * \beta_3 + \gamma * { cvrt} + \epsilon$$

where

$\mu_i$ -- methylation measurement for $i$-th CpG.
$outcome$ -- phenotype of interest.
$SNP_i$ -- the SNP values (0/1/2 or dosages for imputed genotype) at the $i$-th CpG.
$cvrt$ -- covariates and the principal components.
$\epsilon$ -- noise.

Input data

For CpG-SNPs analysis RaMWAS requires the usual input (see steps 4 and 5) with an additional SNP matrix.

The SNP data must have the same dimensions as the CpG score matrix, i.e. it must be available for the same set of samples and the same set of locations. Data preparation may include finding the closest SNP for every CpG and exclusion of CpGs without any SNPs in vicinity.

Create data matrices for CpG-SNP analysis

To illustrate this type of analysis we produce the following artificial files.

CpG_locations.* -- filematrix with the location of the SNP-CpGs.\ It has two columns with integer values -- chromosome number and location (chr and position).
CpG_chromosome_names.txt -- file with chromosome names (factor levels) for the integer column chr in the location filematrix.
Coverage.* -- filematrix with the data for all samples and all locations.\ Each row has data for a single sample. Row names are sample names.\ Each column has data for a single location. Columns match rows of the location filematrix.
SNPs.* -- filematrix with genotype data, matching the coverage matrix.

First, we load the package and set up a working directory. The project directory dr can be set to a more convenient location when running the code.

library(ramwas)

# work in a temporary directory
dr = paste0(tempdir(), "/simulated_matrix_data")
dir.create(dr, showWarnings = FALSE)
cat(dr,"\n")

Let the sample data matrix have 200 samples and 100,000 variables.

nsamples = 200
nvariables = 100000

For these r nsamples samples we generate a data frame with age and sex phenotypes and a batch effect covariate.

set.seed(18090212)

covariates = data.frame(
    sample = paste0("Sample_",seq_len(nsamples)),
    sex = seq_len(nsamples) %% 2,
    age = runif(nsamples, min = 20, max = 80),
    batch = paste0("batch",(seq_len(nsamples) %% 3))
)
pander(head(covariates))

Next, we create the genomic locations for 100,000 variables.

set.seed(18090212)

temp = cumsum(sample(20e7 / nvariables, nvariables, replace = TRUE) + 0)
chr      = as.integer(temp %/% 1e7) + 1L
position = as.integer(temp %% 1e7)

locmat = cbind(chr = chr, position = position)
chrnames = paste0("chr", 1:10)
pander(head(locmat))

Now we save locations in a filematrix and create a text file with chromosome names.\

fmloc = fm.create.from.matrix(
            filenamebase = paste0(dr, "/CpG_locations"),
            mat = locmat)
close(fmloc)
writeLines(con = paste0(dr, "/CpG_chromosome_names.txt"), text = chrnames)

Finally, we create methylation and SNP matrices and populate them.

set.seed(18090212)

fmm = fm.create(paste0(dr,"/Coverage"), nrow = nsamples, ncol = nvariables)
fms = fm.create(paste0(dr,"/SNPs"), nrow = nsamples, ncol = nvariables,
                size = 1, type = "integer")

# Row names of the matrices are set to sample names
rownames(fmm) = as.character(covariates$sample)
rownames(fms) = as.character(covariates$sample)

# The matrices are filled, 2000 variables at a time
byrows = 2000
for( i in seq_len(nvariables/byrows) ){ # i=1
    ind = (1:byrows) + byrows*(i-1)

    snps = rbinom(n = byrows * nsamples, size = 2, prob = 0.2)
    dim(snps) = c(nsamples, byrows)
    fms[,ind] = snps

    slice = double(nsamples*byrows)
    dim(slice) = c(nsamples, byrows)
    slice[,  1:225] = slice[,  1:225] + covariates$sex / 50 / sd(covariates$sex)
    slice[,101:116] = slice[,101:116] + covariates$age / 16 / sd(covariates$age)
    slice = slice +
                ((as.integer(factor(covariates$batch))+i) %% 3) / 200 +
                snps / 1.5 +
                runif(nsamples*byrows) / 2
    fmm[,ind] = slice
}
close(fms)
close(fmm)

SNP-CpG analysis

Let us test for association between CpG scores and and the sex covariate (modeloutcome parameter) correcting for batch effects (modelcovariates parameter). Save top 20 results (toppvthreshold parameter) in a text file.

param = ramwasParameters(
    dircoveragenorm = dr,
    covariates = covariates,
    modelcovariates = "batch",
    modeloutcome = "sex",
    toppvthreshold = 20,
    fileSNPs = "SNPs"
)

# Bioconductor requires limit of 2 parallel jobs
param$cputhreads = 2

The CpG-SNP analysis:

ramwasSNPs(param)

The QQ-plot shows better enrichment with significant p-values.

pfull = parameterPreprocess(param)
mwas = getMWAS(pfull$dirSNPs)
qqPlotFast(mwas$`p-value`)
title("QQ-plot for CpG-SNP analysis")

For comparison, we also perform the usual MWAS for these CpGs without regard for SNPs.

ramwas5MWAS(param)

The QQ-plot shows much weaker signal for the standard MWAS.

mwas = getMWAS(param)
qqPlotFast(mwas$`p-value`)
title(pfull$qqplottitle)

The top finding are saved in the text files Top_tests.txt for both analyses:

# Get the directory with testing results
toptbl = read.table(
                paste0(pfull$dirSNPs, "/Top_tests.txt"),
                header = TRUE,
                sep = "\t")
pander(head(toptbl,10))

Note that CpG-SNP analysis tests for joint significance of $\beta_1$ and $\beta_2$ and thus uses F-test, while regular MWAS uses t-test.

pfull = parameterPreprocess(param)
toptbl = read.table(
                paste0(pfull$dirmwas, "/Top_tests.txt"),
                header = TRUE,
                sep = "\t")
pander(head(toptbl,10))