RaMWAS provides a complete toolset for methylome-wide association studies (MWAS). It is primarily designed for data from enrichment-based methylation assays, but can be applied to other methylomic data (e.g. Illumina methylation array) as well as other data types like gene expression and genotype data (see vignette).
RaMWAS includes the following major components (steps):
Most steps of RaMWAS are internally parallelized. This is made possible, in part, by the use of the filematrix package for storing the data and accessing it in parallel jobs.
To install the most recent version of RaMWAS please follow instructions
at GitHub.com
(R 3.3 or newer required).
To install the Bioconductor version of RaMWAS (R 3.4 or newer requred)
use the following commands:
## try http:// if https:// URLs are not supported
if (!requireNamespace("BiocManager", quietly=TRUE))
install.packages("BiocManager")
BiocManager::install("ramwas")
The package vignettes and reference manual can be viewed online and with the following commands.
library(ramwas) # Loads the package
browseVignettes("ramwas") # Opens vignettes
help(package = "ramwas") # Lists package functions
To illustrate the main steps of RaMWAS we first create an artificial data set.
library(ramwas)
dr = paste0(tempdir(), "/simulated_project")
ramwas0createArtificialData(
dir = dr,
nsamples = 20,
nreads = 100e3,
ncpgs = 25e3,
threads = 2)
cat("Project directory:", dr)
## Project directory: /tmp/Rtmpguarmr/simulated_project
Note. The project directory dr
can be set to
a more convenient location when running the code.
The function ramwas0createArtificialData
created the following files
and subdirectories in the project directory.
bams
– directory with 20 BAM filesbam_list.txt
– file with names of all the BAM filescovariates.txt
– file with
age and sex variables for the samples.Simulated_chromosome.rds
- file with genomic locations for all CpGs.Each RaMWAS step accepts parameters in the form of a list. Here is the parameter set we will use for all steps below.
param = ramwasParameters(
dirproject = dr,
dirbam = "bams",
filebamlist = "bam_list.txt",
filecpgset = "Simulated_chromosome.rds",
cputhreads = 2,
scoretag = "MAPQ",
minscore = 4,
minfragmentsize = 50,
maxfragmentsize = 250,
minavgcpgcoverage = 0.3,
minnonzerosamples = 0.3,
filecovariates = "covariates.txt",
modelcovariates = NULL,
modeloutcome = "age",
modelPCs = 0,
toppvthreshold = 1e-5,
bihost = "grch37.ensembl.org",
bimart = "ENSEMBL_MART_ENSEMBL",
bidataset = "hsapiens_gene_ensembl",
biattributes = c("hgnc_symbol","entrezgene","strand"),
bifilters = list(with_hgnc_trans_name = TRUE),
biflank = 0,
cvnfolds = 5,
mmalpha = 0,
mmncpgs = c(5, 10, 50, 100, 500, 1000, 2000)
)
We describe the role of each parameter as they are used below. The complete description of all RaMWAS parameters is available in the parameter vignette.
This step scans all BAM files listed in the filebamlist
file,
it records read start locations,
and calculates a number of quality control metrics.
The BAMs must be located in dirbam
directory.
Reads are filtered by the scoretag
parameter,
which is usually either the “MAPQ” field or “AS” tag in the BAM file
(see BAM file format).
Reads with scores below minscore
are excluded.
The minfragmentsize
and maxfragmentsize
parameters define
the minimum and maximum size of DNA fragments that were sequenced.
Please note, these parameters are not equal to the read length but
instead reflect the length of the DNA fragments that were
extracted and sequenced.
The set of CpGs is loaded from the filecpgset
file.
For more information see the CpG set vignette.
RaMWAS uses cputhreads
CPU cores (parallel jobs) to scan BAMs.
Hard disk speed can be a bottleneck for this step.
If BAMs are stored on a single rotational hard drive,
using more than 4 parallel jobs may not provide further speed improvements.
ramwas1scanBams(param)
This creates the following subdirectories in the project directory:
qc
– directory with a number of quality control (QC) plots,
one plot per QC metrix per BAM.edit_distance
– plots showing distribution of edit distance,
i.e. number of mismatches between the aligned read and the reference genome.isolated_distance
– plots showing distribution of distances
from read starts to isolated CpGs.maxfragmentsize
basepairs away from any other CpG.matched_length
– plots showing distribution of the length of the
reads aligned to the reference genome.score
– plots showing distribution of the score
(scoretag
parameter).coverage_by_density
– plots showing average CpG score (coverage) as
a function of CpG density.rds_rbam
and rds_qc
– directories with RaMWAS BAM info files
(BAM read start locations) and BAM quality control metrics,
one RDS file per BAM.Here is a coverage_by_density
plot for the simulated data.
It shows higher average CpG score (fragment coverage) for regions
with higher CpG densities, up to the saturation point.
pfull = parameterPreprocess(param)
qc = readRDS(paste0(pfull$dirrqc, "/BAM007.qc.rds"))
plot(qc$qc$avg.coverage.by.density)
Note. If a BAM file has previously been scanned,
it will not be scanned again in subsequent calls of ramwas1scanBams
.
This way Step 1 can be rerun multiple times
to efficiently include additional BAMs (samples) when they become available.
Note. The BAM files are no longer needed after this step and the
RaMWAS BAM info files are 50 to 100 times smaller than the original BAMs.
This step aggregates quality control metrics across all scanned BAM files, produces a number of summary plots and tables, and estimates fragment size distribution.
In practice, multiple BAM files may correspond to the same sample.
For simplicity, in the example here
each BAM corresponds to one sample with the same name.
RaMWAS can be instructed about BAM to sample correspondence via
filebam2sample
or bam2sample
parameters
(See parameter vignette).
ramwas2collectqc(param)
The following files and directories are created in the project directory:
Fragment_size_distribution.txt
– text file with estimated
fragment size distribution.qc/Fragment_size_distribution_estimate.pdf
– plot showing both
estimated fragment size distribution and the distribution of distances
from read starts to isolated CpGs.summary_bams
– by BAMs.summary_bams_in_bam2sample
– by BAMs,
but only those listed in filebam2sample
parameter.summary_by_sample
– by sample.summary_total
– total across all BAMs.filebam2sample
.Summary_QC.txt
– table with a number of numerical QC measures,
in an Excel friendly format.Summary_QC_R.txt
– table with a number of numerical QC measures,
in an R friendly format.qclist.rds
– an R file with list of QC objects.Fig_hist_*.pdf
– histograms of several QC measures across samples.Fig_*.pdf
– PDF files with various QC plots,
one page per BAM or sample,
depending on the directory.After exclusion of BAMs and samples not passing QC, step 2 should be rerun. This ensures that fragment size distribution is estimated using selected data only.
The fragment size distribution estimation plot is presented below. The points indicate the number of reads (y-axis) observed at varying distances from isolated CpGs (x-axis). The red line is the parametic fit for these points. For more information on estimation of fragment size distribution see (Oord et al. 2013).
qc = readRDS(paste0(pfull$dirqc, "/summary_total/qclist.rds"))
frdata = qc$total$hist.isolated.dist1
estimate = as.double(readLines(
con = paste0(pfull$dirproject,"/Fragment_size_distribution.txt")))
plotFragmentSizeDistributionEstimate(frdata, estimate)
This step creates a CpG score matrix (a.k.a. fragment coverage matrix)
for all samples and all CpGs in the CpG set.
The samples are defined by either filebam2sample
or bam2sample
parameter.
The CpGs are defined by the filecpgset
parameter,
see CpG set vignette for more information.
RaMWAS can filter out CpGs with low scores. These CpGs are unmethylated and are unlikely to produce any findings. A CpG is preserved if
minavgcpgcoverage
(default is 0.3).minnonzerosamples
proportion of samples
with nonzero coverageFor each sample, the CpG scores are affected by the number of sequenced DNA fragments, which varies from sample to sample. To remove this variation, the CpG score matrix is normalized to have the same average score for each sample.
ramwas3normalizedCoverage(param)
This step a new creates directory named
coverage_norm_X
in the project directory
(X is the number of samples, see Directory names)
with the following files:
Coverage.*
– filematrix with the CpG scores for all samples and all
CpGs that passed the filtering.CpG_locations.*
– filematrix with the location of the CpGs
that passed the threshold.chr
and position
.CpG_chromosome_names.txt
– file with chromosome names (factor levels)
for the integer column chr
in CpG_locations.*
filematrix.raw_sample_sums.*
– filematrix with total coverage for each
sample before normalization.Note. This step created temporary files in the dirtemp
directory.
This step performs principal component analysis (PCA) on the CpG score matrix.
PCA is capturing the main directions of unmeasured sources of variation in the data. The main goal of PCA is estimation of laboratory technical variations which can be used as covariates in the association analyses. PCA can be performed after regressing out known covariates as they represent measured sources of variation, which we need not estimate.
Additionally, large sample loadings of the top PCs can indicate multidimensional outlying samples. Such sample may be excluded from the analysis.
If measured covariates are regressed out prior to conducting the PCA,
these covariates are loaded from the file filecovariates
.
The file can be comma or tab-separated, with sample names in the first column.
The artificial data set includes a covariate file covariates.txt.
The parameter modelcovariates
names
the covariates regressed out before PCA (NULL
for none).
ramwas4PCA(param)
This step creates sub-directory PCA_X_cvrts_Y
in the directory with
score matrix, where X is the number of covariates regressed out and
Y is a code distinguishing different sets with the same number of covariates
(see Directory names).
The sub-directory includes:
covmat.*
, eigenvalues.*
, and eigenvectors.*
–
filematrices with the sample covariance matrix and
its eigenvalue decomposition.PC_values.txt
– principal components scores.PC_loadings.txt
– sample loadings for the top 20 PCs.PC_plot_covariates_removed.pdf
– plots of principal components scores
(i.e. % variance explained on page 1) and
sample loadings for the top 40 PCs (pages 2+).PC_vs_covariates_corr.txt
– correlations of principal components
with phenotypes/covariates (from filecovariates
file).PC_vs_covariates_pvalue.txt
– p-values for these correlations.The PC plot for artificial data shows one strong component and no outliers in the sample loadings, with first PC clearly capturing sample sex.
eigenvalues = fm.load(paste0(pfull$dirpca, "/eigenvalues"));
eigenvectors = fm.open(
filenamebase = paste0(pfull$dirpca, "/eigenvectors"),
readonly = TRUE);
plotPCvalues(eigenvalues)
plotPCvectors(eigenvectors[,1], 1)
plotPCvectors(eigenvectors[,2], 2)
close(eigenvectors)
The file with correlations shows strong correlation of the top PCs with age and sex.
tblcr = read.table(
file = paste0(pfull$dirpca, "/PC_vs_covs_corr.txt"),
header = TRUE,
sep = "\t")
pander(head(tblcr, 3))
name | age | sex |
---|---|---|
PC1 | -0.28 | 0.997 |
PC2 | 0.723 | 0.0335 |
PC3 | 0.184 | -0.0281 |
The file with p-values indicate statistical significance of these correlations.
Table: Correlations in PC_vs_covariates_corr.txt
file.
tblpv = read.table(
file = paste0(pfull$dirpca, "/PC_vs_covs_pvalue.txt"),
header = TRUE,
sep = "\t")
pander(head(tblpv, 3))
name | age | sex |
---|---|---|
PC1 | 0.231 | 1.62e-21 |
PC2 | 0.000319 | 0.889 |
PC3 | 0.437 | 0.906 |
This step performs tests for association
between normalized CpG scores and the
outcome variable named by modeloutcome
parameter.
The analysis corrects for covariates listed in modelcovariates
parameter and a number of top principal components (modelPCs
).
Tests are performed using the linear regression model if
the outcome variable is numeric, ANOVA if categorical.
All results are saved in a filematrix.
Findings passing the toppvthreshold
p-value
threshold are recorded in a text file.
ramwas5MWAS(param)
This step creates a sub-directory in the PCA directory named
Testing_X_Y_PCs
, where X is the name of the outcome variable and Y
is the number of top PCs included as covariates (see Directory names).
The sub-directory includes:
QQ_plot.pdf
– QQ-plot with a confidence band and inflation factor lambda
(median of the chi-squared test statistics
divided by the expected median of the chi-squared distribution under null).Top_tests.txt
– text file with the top findings.Stats_and_pvalues.*
– filematrix with MWAS results.
The columns include test statistic, p-value, and q-value
(calculated using Benjamini-Hochberg Procedure (Benjamini and Hochberg 1995)).
The rows match the CpGs of the coverage matrix.DegreesOfFreedom.txt
– file with the numbers of degrees of freedom for
the t- or F test used for testing.For the simulated data the QQ-plot for age shows moderate deviation from the diagonal for many CpGs, suggesting many markers with small effects. This is consistent with how the data was generated – there is weak signal in 1% of CpGs.
RaMWAS also creates a Manhattan plot with matching Y axis.
mwas = getMWASandLocations(param)
layout(matrix(c(1,2), 1, 2, byrow = TRUE), widths=c(1,2.2))
qqPlotFast(mwas$`p-value`)
man = manPlotPrepare(
pvalues = mwas$`p-value`,
chr = mwas$chr,
pos = mwas$start,
chrmargins = 0)
manPlotFast(man)
RaMWAS can annotate top findings using data from
biomaRt
.
The parameters include:
bihost
– biomart host site.bimart
– BioMart database name.bidataset
– data set.biattributes
– are attributes of interest.bifilters
– list of filters (if any).biflank
– indicates the maximum allowed distance from the
CpG to the annotation element.For detailed description of these parameters and R script for listing allowed values see parameter vignette
Here we annotate top findings with human gene symbols, gene ids, and strand information.
ramwas6annotateTopFindings(param)
The function updates the Top_tests.txt
file in the MWAS directory.
chr | position | tstat | pvalue | qvalue | hgnc_symbol | entrezgene | strand |
---|---|---|---|---|---|---|---|
chr1 | 15,975,530 | 8.771 | 6.446x10-8 | 0.022 | DDI2/DDI2 | 84301/6248 | 1/1 |
chr1 | 15,975,533 | 8.568 | 9.097x10-8 | 0.022 | DDI2/DDI2 | 84301/6248 | 1/1 |
chr1 | 15,418,248 | -7.654 | 4.571x10-7 | 0.071 | KAZN | 23254 | 1 |
While it is important to find individual CpGs associated with a phenotype, we can often achieve better predictive power by combining information from multiple CpGs.
We take an approach similar to the one proposed by (Horvath 2013) for predicting biological age from methylation data. In particular, we combine signal across multiple CpGs using the elastic net model (Tibshirani et al. 2012). To avoid overfitting and correctly estimate the predictive power of the model we use k-fold cross-validation. Within the cross-validation procedure, for each training set of samples we perform MWAS, select top MWAS sites, train the elastic net, and make predictions for the test samples. The set of predictions is recorded as the MRS.
RaMWAS predicts the outcome variable (modeloutcomes
parameter)
using top mmncpgs
CpGs from the MWAS on the training set of samples.
If mmncpgs
is a vector of several values,
each number of CpGs is tested separately.
The elastic net parameter alpha can be set via mmalpha
parameter.
The number of folds cvnfolds
in the K-fold cross validation is 10 by default.
ramwas7riskScoreCV(param)
This step creates CV_X_folds
sub-directory in MWAS directory,
where X is the number of folds in k-fold cross validation
(see Directory names).
It contains:
fold_*
– directories with MWAS on 90% of the data,
each with different 10% of the samples excluded from the analysis.NA
) outcome variable are not included in these MWAS.MMCVN_prediction_folds=10_CpGs=*.txt
–
table with true outcome and cross-validation prediction.MMCVN_prediction_folds=10_CpGs=*.pdf
–
scatter-plot of true outcome vs. cross-validation prediction.Prediction_alpha=0.000000.pdf
– scatter plot of
outcome-prediction correlations as a function of the
number of CpGs used for prediction.For the simulated data we get moderately good prediction using just 50 top CpGs.
cv = readRDS(paste0(pfull$dircv, "/rds/CpGs=000050_alpha=0.000000.rds"))
plotPrediction(
param = pfull,
outcome = cv$outcome,
forecast = cv$forecast,
cpgs2use = 50,
main = "Prediction success (EN on coverage)")
The correlation of prediction with actual age is maximal when we use 100 top CpGs in elastic net.
cl = readRDS(sprintf("%s/rds/cor_data_alpha=%f.rds",
pfull$dircv,
pfull$mmalpha))
plotCVcors(cl, pfull)
Point mutated CpG sites, called CpG-SNPs (Shoemaker et al. 2010), are interesting sites for human diseases because in addition to the sequence variation they may show individual differences in DNA methylation. RaMWAS can perform CpG-SNP analyses if SNP data from the same subjects/samples is also available. These tests are performed using a regression model that assesses whether the case-control differences are proportional to the number of CpG sites (Oord et al. 2015).
This type of analysis is explained in the CpG-SNP vignette.
The names of CpG score/PCA/MWAS/MRS directories can be recovered
by first calling parameterPreprocess
function.
pfull = parameterPreprocess(param)
cat("CpG score directory:", "\n", pfull$dircoveragenorm,"\n")
## CpG score directory:
## /tmp/Rtmpguarmr/simulated_project/coverage_norm_20
cat("PCA directory:", "\n", pfull$dirpca, "\n")
## PCA directory:
## /tmp/Rtmpguarmr/simulated_project/coverage_norm_20/PCA_00_cvrts
cat("MWAS directory:", "\n", pfull$dirmwas, "\n")
## MWAS directory:
## /tmp/Rtmpguarmr/simulated_project/coverage_norm_20/PCA_00_cvrts/Testing_age_0_PCs
cat("MRS directory:", "\n", pfull$dircv, "\n")
## MRS directory:
## /tmp/Rtmpguarmr/simulated_project/coverage_norm_20/PCA_00_cvrts/Testing_age_0_PCs/CV_05_folds
cat("CpG-SNP directory:", "\n", pfull$dirSNPs, "\n")
## CpG-SNP directory:
## /tmp/Rtmpguarmr/simulated_project/coverage_norm_20/Testing_wSNPs_age_0cvrts_0PCs
sessionInfo()
## R version 4.0.3 (2020-10-10)
## Platform: x86_64-pc-linux-gnu (64-bit)
## Running under: Ubuntu 18.04.5 LTS
##
## Matrix products: default
## BLAS: /home/biocbuild/bbs-3.12-bioc/R/lib/libRblas.so
## LAPACK: /home/biocbuild/bbs-3.12-bioc/R/lib/libRlapack.so
##
## locale:
## [1] LC_CTYPE=en_US.UTF-8 LC_NUMERIC=C
## [3] LC_TIME=en_US.UTF-8 LC_COLLATE=C
## [5] LC_MONETARY=en_US.UTF-8 LC_MESSAGES=en_US.UTF-8
## [7] LC_PAPER=en_US.UTF-8 LC_NAME=C
## [9] LC_ADDRESS=C LC_TELEPHONE=C
## [11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C
##
## attached base packages:
## [1] stats graphics grDevices utils datasets methods base
##
## other attached packages:
## [1] ramwas_1.14.0 filematrix_1.3 pander_0.6.3 knitr_1.30
## [5] BiocStyle_2.18.0
##
## loaded via a namespace (and not attached):
## [1] MatrixGenerics_1.2.0 Biobase_2.50.0
## [3] httr_1.4.2 bit64_4.0.5
## [5] splines_4.0.3 foreach_1.5.1
## [7] assertthat_0.2.1 askpass_1.1
## [9] BiocManager_1.30.10 stats4_4.0.3
## [11] BiocFileCache_1.14.0 blob_1.2.1
## [13] GenomeInfoDbData_1.2.4 Rsamtools_2.6.0
## [15] yaml_2.2.1 progress_1.2.2
## [17] pillar_1.4.6 RSQLite_2.2.1
## [19] lattice_0.20-41 glue_1.4.2
## [21] digest_0.6.27 GenomicRanges_1.42.0
## [23] XVector_0.30.0 htmltools_0.5.0
## [25] Matrix_1.2-18 XML_3.99-0.5
## [27] pkgconfig_2.0.3 biomaRt_2.46.0
## [29] magick_2.5.0 bookdown_0.21
## [31] zlibbioc_1.36.0 purrr_0.3.4
## [33] BiocParallel_1.24.0 tibble_3.0.4
## [35] openssl_1.4.3 generics_0.0.2
## [37] IRanges_2.24.0 ellipsis_0.3.1
## [39] SummarizedExperiment_1.20.0 BiocGenerics_0.36.0
## [41] survival_3.2-7 magrittr_1.5
## [43] crayon_1.3.4 memoise_1.1.0
## [45] evaluate_0.14 xml2_1.3.2
## [47] tools_4.0.3 prettyunits_1.1.1
## [49] hms_0.5.3 lifecycle_0.2.0
## [51] matrixStats_0.57.0 stringr_1.4.0
## [53] S4Vectors_0.28.0 glmnet_4.0-2
## [55] DelayedArray_0.16.0 AnnotationDbi_1.52.0
## [57] Biostrings_2.58.0 compiler_4.0.3
## [59] GenomeInfoDb_1.26.0 rlang_0.4.8
## [61] grid_4.0.3 RCurl_1.98-1.2
## [63] iterators_1.0.13 rappdirs_0.3.1
## [65] bitops_1.0-6 rmarkdown_2.5
## [67] codetools_0.2-16 curl_4.3
## [69] DBI_1.1.0 R6_2.4.1
## [71] GenomicAlignments_1.26.0 dplyr_1.0.2
## [73] bit_4.0.4 KernSmooth_2.23-17
## [75] shape_1.4.5 stringi_1.5.3
## [77] parallel_4.0.3 Rcpp_1.0.5
## [79] vctrs_0.3.4 tidyselect_1.1.0
## [81] dbplyr_1.4.4 xfun_0.18
Benjamini, Y., and Y. Hochberg. 1995. “Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing.” Journal of the Royal Statistical Society. Series B (Methodological). JSTOR, 289–300.
Horvath, Steve. 2013. “DNA Methylation Age of Human Tissues and Cell Types.” Genome Biology 14 (10). BioMed Central:3156.
Oord, Edwin JCG van den, Jozsef Bukszar, Gábor Rudolf, Srilaxmi Nerella, Joseph L McClay, Lin Y Xie, and Karolina A Aberg. 2013. “Estimation of Cpg Coverage in Whole Methylome Next-Generation Sequencing Studies.” BMC Bioinformatics 14 (1). BioMed Central:50.
Oord, Edwin JCG van den, Shaunna L Clark, Lin Ying Xie, Andrey A Shabalin, Mikhail G Dozmorov, Gaurav Kumar, Vladimir I Vladimirov, et al. 2015. “A Whole Methylome Cpg-Snp Association Study of Psychosis in Blood and Brain Tissue.” Schizophrenia Bulletin. MPRC, sbv182.
Shoemaker, Robert, Jie Deng, Wei Wang, and Kun Zhang. 2010. “Allele-Specific Methylation Is Prevalent and Is Contributed by Cpg-Snps in the Human Genome.” Genome Research 20 (7). Cold Spring Harbor Lab:883–89.
Tibshirani, Robert, Jacob Bien, Jerome Friedman, Trevor Hastie, Noah Simon, Jonathan Taylor, and Ryan J Tibshirani. 2012. “Strong Rules for Discarding Predictors in Lasso-Type Problems.” Journal of the Royal Statistical Society: Series B (Statistical Methodology) 74 (2). Wiley Online Library:245–66.