1 Using RaMWAS for with other methylation platforms or data types

RaMWAS is primarily designed for studies of methylation measurements from enrichment platforms.

However, RaMWAS can also be useful for the analysis of methylation measurements from other platforms (e.g. Illumina HumanMethylation450K array) or other data types such as gene expression levels or genotype information. RaMWAS can perform several analysis steps on such data including: principal component analysis (PCA), association testing (MWAS, TWAS, GWAS), and multimarker analysis with cross validation using the elastic net.

1.1 Import data from other sources

Without external data source at hand, we show how to create and fill data matrices with artificial data. Importing real data can be done in a similar way, with random data generation replaced with reading data from existing sources.

We create data files in the same format as produced by Step 3 of RaMWAS.

These files include

CpG_locations.* – filematrix with the location of the CpGs / SNPs / gene trascription start sites.
It must have 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 must be sample names.
Each column has data for a single location (CpG / SNP / gene trascription start site). Columns must match rows of the location filematrix.

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")

## /tmp/Rtmp3SAAij/simulated_matrix_data

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

nsamples = 200
nvariables = 100000

For these 200 samples we generate a data frame with age and sex phenotypes and a batch effect covariate.

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))

sample	sex	age	batch
Sample_1	1	71.5	batch1
Sample_2	0	35.8	batch2
Sample_3	1	60.4	batch0
Sample_4	0	64.5	batch1
Sample_5	1	28.4	batch2
Sample_6	0	26.3	batch0

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

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", 1:22)
pander(head(locmat))

chr	position
1	1717
1	2245
1	3591
1	5074
1	5355
1	5565

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 data matrix. We include sex effect in 225 variables and age effect in 16 variables out of each 2000. Each variable is also affected by noise and batch effects.

fm = fm.create(paste0(dr,"/Coverage"), nrow = nsamples, ncol = nvariables)

# Row names of the matrix are set to sample names
rownames(fm) = as.character(covariates$sample)

# The matrix is filled, 2000 variables at a time
byrows = 2000
for( i in seq_len(nvariables/byrows) ){ # i=1
    slice = matrix(runif(nsamples*byrows), nrow = nsamples, ncol = byrows)
    slice[,  1:225] = slice[,  1:225] + covariates$sex / 30 / sd(covariates$sex)
    slice[,101:116] = slice[,101:116] + covariates$age / 10 / sd(covariates$age)
    slice = slice + ((as.integer(factor(covariates$batch))+i) %% 3) / 40
    fm[,(1:byrows) + byrows*(i-1)] = slice
}
close(fm)

1.2 Principal Component Analysis (PCA)

To run PCA with RaMWAS we specify three parameters:

dircoveragenorm – directory with the data matrix
covariates – data frame with covariates
modelcovariates – names of covariates to regress out

param = ramwasParameters(
    dircoveragenorm = dr,
    covariates = covariates,
    modelcovariates = NULL
)

Now we run PCA.

ramwas4PCA(param)

The top several PCs are marginally distinct from the rest.

There are strong correlations between top PCs with sex, age, and batch covariates.
Note, for the categorical covariate (batch) the table shows R² instead of correlations.

# Get the directory with PCA results
pfull = parameterPreprocess(param)
tblcr = read.table(paste0(pfull$dirpca, "/PC_vs_covs_corr.txt"),
                 header = TRUE,
                 sep = "\t")
pander(head(tblcr, 5))

name	sex	age	batch_R2
PC1	-0.0278	-0.0991	0.984
PC2	0.0326	0.0372	0.986
PC3	-0.938	-0.163	0.00167
PC4	0.286	-0.942	0.000988
PC5	0.0126	-0.0021	8.94e-05

The p-values for these correlations and R² show that the top two PCs are correlated with sex and age while a number of other PCs are affected by sample batch effects.

pfull = parameterPreprocess(param)
tblpv = read.table(paste0(pfull$dirpca, "/PC_vs_covs_pvalue.txt"),
                 header = TRUE,
                 sep = "\t")
pander(head(tblpv, 5))

name	sex	age	batch_R2
PC1	0.696	0.163	1.11e-178
PC2	0.647	0.601	1.6e-183
PC3	2.55e-93	0.0211	0.848
PC4	4.06e-05	6.37e-96	0.907
PC5	0.86	0.976	0.991

1.3 PCA with batch regressed out

It is common to regress out batch and lab-technical effects from the data in the analysis.

Let’s regress out batch in our example by changing modelcovariates parameter.

param$modelcovariates = "batch"

ramwas4PCA(param)

The p-values for association between PCs and covariates changed slightly:

# Get the directory with PCA results
pfull = parameterPreprocess(param)
tblpv = read.table(paste0(pfull$dirpca, "/PC_vs_covs_pvalue.txt"),
                 header = TRUE,
                 sep = "\t")
pander(head(tblpv, 5))

name	sex	age	batch_R2
PC1	4.54e-93	0.0185	1
PC2	4.41e-05	1.25e-98	1
PC3	0.86	0.997	1
PC4	0.852	0.584	1
PC5	0.883	0.692	1

Note that the PCs are now orthogonal to the batch effects and thus the corresponding p-values all equal to 1.

1.4 Association testing

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

param$modelcovariates = "batch"
param$modeloutcome = "sex"
param$toppvthreshold = 20

ramwas5MWAS(param)

The QQ-plot shows mild enrichment among a large number of variables, which is consistent with how the data was generated – 22% of variables are affected by sex.

The top finding saved in the text file are:

# Get the directory with testing results
pfull = parameterPreprocess(param)
toptbl = read.table(
                paste0(pfull$dirmwas,"/Top_tests.txt"),
                header = TRUE,
                sep = "\t")
pander(head(toptbl, 5))

chr	position	tstat	pvalue	qvalue
chr9	3952879	5.63	6.12e-08	0.00612
chr2	8054473	5.18	5.46e-07	0.0162
chr9	5871812	5.15	6.45e-07	0.0162
chr1	214380	5.15	6.47e-07	0.0162
chr2	6013918	5.09	8.34e-07	0.0163

1.5 Further steps of RaMWAS pipeline

Steps 6 and 7 of RaMWAS pipeline can also be applied to the data matrix exactly as described in the overview vignette.

1.6 Cleanup

Here we remove all the files created by the code above.

unlink(paste0(dr,"/*"), recursive=TRUE)

Analyzing Data from Other Methylation Platforms or Data Types

2017-10-30

Contents