DMR Approach and Interpretation

DMRichR leverages the statistical algorithms from dmrseq and bsseq, which enable the inference of differentially methylated regions (DMRs). In these smoothing based approaches, CpG sites with higher coverage are given a higher weight and are used to infer the methylation level of neighboring CpGs with lower coverage. This approach favors a larger sample size over a deeper sequencing depth, and focuses on the differences in methylation levels between groups, rather than the absolute levels within a group. Together, these methodologies enable a low-pass WGBS approach (1-5x coverage) that also works well with higher coverage datasets.

The main statistical approach applied by DMRichR::DM.R() is dmrseq::dmrseq(), which identifies DMRs in a two step approach:

  1. DMR Detection: The differences in CpG methylation for the effect of interest are pooled and smoothed to give CpG sites with higher coverage a higher weight, and candidate DMRs with a difference between groups are assembled.
  2. Statistical Analysis: A region statistic for each DMR, which is comparable across the genome, is estimated via the application of a generalized least squares (GLS) regression model with a nested autoregressive correlated error structure for the effect of interest. Then, permutation testing of a pooled null distribution enables the identification of significant DMRs. This approach accounts for both inter-individual and inter-CpG variability across the entire genome.

The main estimate of a difference in methylation between groups is not a fold change but rather a beta coefficient, which is representative of the average effect size; however, it is on the scale of the arcsine transformed differences and must be divided by π (3.14) to be similar to the mean methylation difference over a DMR, which is provided in the percentDifference column. Since the testing is permutation based, it provides empirical p-values as well as FDR corrected q-values.

One of the key differences between dmrseq and other DMR identification packages, like bsseq, is that dmrseq is performing statistical testing on the DMRs themselves rather than testing for differences in single CpGs that are then assembled into DMRs like bsseq::dmrFinder() does. This unique approach helps with controlling the false discovery rate and testing the correlated nature of CpG sites in a regulatory region, while also enabling complex experimental designs. However, since dmrseq::dmrseq() does not provide individual smoothed methylation values, bsseq::BSmooth() is utilized to generate individual smoothed methylation values from the DMRs. Therefore, while the DMRs themselves are adjusted for covariates, the individual smoothed methylation values for these DMRs are not adjusted for covariates.

You can also read my general summary of the drmseq approach on EpiGenie.

Example DMR

Example DMR Each dot represents the methylation level of an individual CpG in a single sample, where the size of the dot is representative of coverage. The lines represent smoothed methylation levels for each sample, either control (blue) or DS (red). Gene and CpG annotations are shown below the plot.

Input

Design Matrix and Covariates

This script requires a basic design matrix to identify the groups and covariates, which should be named sample_info.xlsx and contain header columns to identify the covariates. The first column of this file should be the sample names and have a header labelled as Name. In terms of the testCovariate label (i.e. Group or Diagnosis), it is important to have the label for the experimental samples start with a letter in the alphabet that comes after the one used for control samples in order to obtain results for experimental vs. control rather than control vs. experimental. You can select which specific samples to analyze from the working directory through the design matrix, where pattern matching of the sample name will only select bismark cytosine report files with a matching name before the first underscore, which also means that sample names should not contain underscores. Within the script, covariates can be selected for adjustment. There are two different ways to adjust for covariates: directly adjust values or balance permutations. Overall, DMRichR supports pairwise comparisons with a minimum of 4 samples (2 per a group). For each discrete covariate, you should also aim to have two samples per each grouping level.

Name Diagnosis Age Sex
SRR3537014 Idiopathic_ASD 14 M
SRR3536981 Control 42 F

Cytosine Reports

DMRichR utilizes Bismark cytosine reports, which are genome-wide CpG methylation count matrices that contain all the CpGs in your genome of interest, including CpGs that were not covered in the experiment. The genome-wide cytosine reports contain important information for merging the top and bottom strand of symmetric CpG sites, which is not present in Bismark coverage and bedGraph files. In general, cytosine reports have the following pattern: *_bismark_bt2_pe.deduplicated.bismark.cov.gz.CpG_report.txt.gz. CpG_Me will generate a folder called cytosine_reports after calling the final QC script (please don’t use the cytosine_reports_merged folder for DMRichR). If you didn’t use CpG_Me, then you can use the coverage2cytosine module in Bismark to generate the cytosine reports. The cytosine reports have the following format:

chromosome position strand count methylated count non-methylated C-context trinucleotide context
chr2 10470 + 1 0 CG CGA
chr2 10471 - 0 0 CG CGG
chr2 10477 + 0 1 CG CGA
chr2 10478 - 0 0 CG CGG

Before running the executable, ensure you have the following project directory tree structure for the cytosine reports and design matrix:

├── Project
│   ├── sample1_bismark_bt2.deduplicated.bismark.cov.gz.CpG_report.txt.gz
│   ├── sample2_bismark_bt2.deduplicated.bismark.cov.gz.CpG_report.txt.gz
│   ├── sample_info.xlsx

Running DMRichR

DMRichR::DM.R() accepts the following variables:

  1. genome Select either: hg38, hg19, mm10, mm9, rheMac10, rheMac8, rn6, danRer11, galGal6, bosTau9, panTro6, dm6, canFam3, susScr11, TAIR10, or TAIR9. It is also possible to add other genomes with BSgenome, TxDb, and org.db databases by modifying DMRichR::annotationDatabases().
  2. coverage CpG coverage cutoff for all samples, 1x is the default and minimum value.
  3. perGroup Percent of samples per a group for CpG coverage cutoff, values range from 0 to 1. 0.75 (75%) is the default.
  4. minCpGs Minimum number of CpGs for a DMR, 5 is default.
  5. maxPerms Number of permutations for the DMR analysis, 10 is default. The total number of permutations should not exceed the number of samples.
  6. maxBlockPerms Number of permutations for the block analysis, 10 is default. The total number of permutations should not exceed the number of samples.
  7. cutoff The cutoff value for the single CpG coefficient utilized to discover testable background regions. Values range from 0 to 1 and 0.05 (5%) is the default. If you get too many DMRs you should try 0.1 (10%).
  8. testCovariate Covariate to test for significant differences between experimental and control (i.e. Diagnosis).
  9. adjustCovariate Adjust covariates that are continuous (i.e. Age) or discrete with two or more factor groups (i.e. Sex). More than one covariate can be adjusted for (c("Sex, "Age") for R and single brackets with the ; delimiter ('Sex;Age') for command line.
  10. matchCovariate Covariate to balance permutations, which is meant for two-group factor covariates in small sample sizes in order to prevent extremely unbalanced permutations. Only one two-group factor can be balanced (i.e. Sex). Note: This will not work for larger sample sizes (> 500,000 permutations) and is not needed for them as the odds of sampling an extremely unbalanced permutation for a covariate decreases with increasing sample size. Furthermore, we generally do not use this in our analyses, since we prefer to directly adjust for sex.
  11. cores The number of cores to use, 20 is the default. The RAM requirements depend on the number of samples and coverage, which is typically between 16 to 128 GB.
  12. GOfuncR A logical (TRUE or FALSE) indicating whether to run a GOfuncR gene ontology analysis. This is our preferred GO method; however, it is time consuming when there is a large number of DMRs.
  13. sexCheck A logical (TRUE or FALSE) indicating whether to run an analysis to confirm the sex listed in the design matrix based on the ratio of the coverage for the Y and X chromosomes. The sex chromosomes will also be removed from downstream analyses if both sexes are detected. This argument assumes there is a column in the design matrix named “Sex” [case sensitive] with Males coded as either “Male”, “male”, “M”, or “m” and Females coded as “Female”, “female”, “F”, or “f”.
  14. EnsDb A logical (TRUE or FALSE) indicating whether to use Ensembl transcript annotations instead of the default Bioconductor annotations, which are typically from UCSC. These annotations may allow DMRs for non-model organism genomes (i.e. rheMac10) to be mapped to substantially more genes, which will improve DMReport and gene ontology results.
  15. cellComposition A logical (TRUE or FALSE) indicating whether to run an analysis to estimate cell composition in adult whole blood samples. The analysis will only run for hg38 and hg19. This is an experimental feature and requires follow up comparisons with similar array-based papers to confirm accuracy. Use at your own risk.

R Example

DMRichR::DM.R() only requires a minimum of 2 arguments. Your working directory should contain the contain the genome-wide cytosine reports and design matrix.

DM.R <- function(genome = "hg38",
                 testCovariate = "Diagnosis")

However, there are total of 15 supported arguments, which have reasonable defaults:

DM.R <- function(genome = c("hg38", "hg19", "mm10", "mm9", "rheMac10",
                            "rheMac8", "rn6", "danRer11", "galGal6",
                            "bosTau9", "panTro6", "dm6", "susScr11",
                            "canFam3", "TAIR10", "TAIR9"),
                 coverage = 1,
                 perGroup =  0.75,
                 minCpGs =  5,
                 maxPerms =  10,
                 maxBlockPerms = 10,
                 cutoff = 0.05,
                 testCovariate = testCovariate,
                 adjustCovariate = NULL,
                 matchCovariate = NULL,
                 cores = 20,
                 GOfuncR = TRUE,
                 sexCheck = FALSE,
                 EnsDb = FALSE,
                 cellComposition = FALSE)

If your analysis contains both male and female samples, you should adjust for sex via adjustCovariate = "Sex" and it’s highly recommended to set sexCheck = TRUE. If you find that you’re getting too many DMRs, you can reduce the total number by setting cutoff = 0.1.

Command Line Example

Below is an example of how to use the executable R script version in the exec folder on command line. You will have to modify the path to the DM.R script in the call.

call="Rscript \
--vanilla \
/path/to/scripts/DM.R \
--genome hg38 \
--coverage 1 \
--perGroup '0.75' \
--minCpGs 5 \
--maxPerms 10 \
--maxBlockPerms 10 \
--cutoff '0.05' \
--testCovariate Diagnosis \
--adjustCovariate 'Sex;Age' \
--sexCheck TRUE \
--GOfuncR TRUE \
--EnsDb FALSE \
--cores 20"

echo $call
eval $call

UC Davis Example

If you are using the cluster at UC Davis, the following commands can be used to execute DM.R from your login node (i.e. epigenerate). This should be called from the working directory that contains the cytosine reports.

module load R/3.6.3
module load homer

call="nohup \
Rscript \
--vanilla \
/share/lasallelab/programs/DMRichR/DM.R \
--genome hg38 \
--coverage 1 \
--perGroup '0.75' \
--minCpGs 5 \
--maxPerms 10 \
--maxBlockPerms 10 \
--cutoff '0.05' \
--testCovariate Diagnosis \
--adjustCovariate 'Sex;Age' \
--sexCheck TRUE \
--GOfuncR TRUE \
--cores 20 \
--EnsDb FALSE \
> DMRichR.log 2>&1 &"

echo $call
eval $call 
echo $! > save_pid.txt

You can then check on the job using tail -f DMRichR.log and ⌃ Control + c to exit the log view. You can cancel the job from the project directory using cat save_pid.txt | xargs kill. You can also check your running jobs using ps -ef | grep, which should be followed by your username i.e. ps -ef | grep blaufer. Finally, if you still see leftover processes in htop, you can cancel all your processes using pkill -u, which should be followed by your username i.e. pkill -u blaufer.

Alternatively, the executable can also be submitted to the cluster using the shell script via sbatch DM.R.sh.

Workflow and Output

This workflow carries out the following steps:

1) Preprocess Cytosine Reports

DMRichR::processBismark() will load the genome-wide cytosine reports, assign the metadata from the design matrix, and filter the CpGs for equal coverage between the testCovariate as well as any discrete adjustCovariates. There is also an option to confirm the sex of each sample. The end result of this function is a class bsseq object (bs.filtered) that contains the methylated and total count data for each CpG.

2) Blocks

The bsseq object is used to identify large blocks (> 5 kb in size) of differential methylation via dmrseq::dmrseq() by using a different smoothing approach than the DMR calling, which “zooms out”. It will increase the minimum CpG cutoff by 2x when compared to the DMR calling. In addition to bed files and excel spreadsheets with the significant blocks (sigBlocks) and background blocks (blocks), plots of the blocks will be generated by dmrseq::plotDMRs() and an html report with gene annotations are also generated through DMRichR::annotateRegions() and DMRichR::DMReport().

3) DMRs

The bsseq object is used to call DMRs through dmrseq::dmrseq(). The DMRs typically range in size from a several hundred bp to a few kb. In addition to bed files and excel spreadsheets with the significant DMRs (sigRegions) and background regions (regions), plots of the DMRs will be generated by DMRichR::plotDMRs2() and an html report with gene annotations are also generated through DMRichR::annotateRegions() and DMRichR::DMReport().

4) Smoothed Individual Methylation Values

Since dmrseq::dmrseq() smooths the differences between groups, it isn’t possible to get individual smoothed methylation values for downstream analyses and visualization. Therefore, the bsseq object is smoothed using bsseq::BSmooth() to create a new bsseq object (bs.filtered.bsseq) with individual smoothed methylation values.

5) ChromHMM and Reference Epigenome Enrichments

Enrichment testing from the chromHMM core 15-state chromatin state model and the related 5 core histone modifications from Roadmap epigenomics 127 reference epigenomes is performed using the LOLA package through the DMRichR::chromHMM() and DMRichR::roadmap() functions. The results are also plotted on a heatmap by DMRichR::chromHMM_heatmap() and DMRichR::roadmap_heatmap(). This is currently restricted to the UC Davis cluster due to requiring large external databases; however, an advanced user can download the databases and modify the functions to refer to their local copy.

6) Transcription Factor Motif Enrichments

DMRichR::prepareHOMER() creates a directory with bed files for DMRichR::HOMER() to run an enrichment analysis for known transcription factor motifs via HOMER. Analyses will be run for all DMRs as well for the hypermethylated and hypomethylated DMRs. The script is set to perform the enrichment testing relative to background regions, accommodate for percent CpG content in the normalization, and to analyze the exact coordinates of the DMRs. The analysis will only run if HOMER’s findMotifsGenome.pl is in the path and the genome of interest has been installed by configureHomer.pl.

7) Global Methylation Analyses and Plots

DMRichR::globalStats() uses the smoothed bsseq object to test for differences in global and chromosomal methylation with the same adjustments as the DMR calling, where it generates an excel spreadsheet with the results and input. CpG island statistics are also generated for almost all genomes through through DMRichR::getCpGs(). Additionally, a number of plots of smoothed methylation values are generated. The values for the plots are extracted by functions that build on bsseq::getMeth(). DMRichR::windows() obtains methylation values for different window sizes (20 Kb is the default), DMRichR::CGi() obtains methylation values for all CpG islands, and DMRichR::CpGs() will obtain individual CpG values. The above functions are used to generate 3 types of plots. PCA plots are made through DMRichR::PCA(), which uses ggbiplot. The ellipses in the PCAs represent the 68% confidence interval, which is 1 standard deviation from the mean for a normal distribution. DMRichR::densityPlot() generates density plots for the mean of each group. Finally, Glimma::glMDSPlot() is used to generate interactive MDS plots.

8) DMR Heatmap

DMRichR::smoothPheatmap() uses pheatmap::pheatmap() to generate a heatmap of the results with annotations for discrete covariates. The heatmap shows the hierarchical clustering of Z-scores for the non-adjusted percent smoothed individual methylation values for each DMR, where the Z-score corresponds to the number of standard deviations from the mean value of each DMR.

9) DMR Annotations and DMRichments

DMRichR::DMRichCpG() and DMRichR::DMRichGenic() will obtain and perform enrichment testing for CpG and gene region (genic) annotations, respetively. The results are plotted using DMRichR::DMRichPlot() and DMRichR::DMparseR() enables plots with facets based on DMR directionality. The above approach for genic annotations is similair to what is typically preformed for arrays and other sequencing approaches, where certain terms are given priority over others if DMRs overlap with more than one.

10) Manhattan Plot

DMRichR::Manhattan() will take the output of DMRichR::annotateRegions() and use it to generate a Manhattan plot through CMplot::CMplot().

11) Gene Ontology Enrichments

Gene ontology enrichments are performed separately for all DMRs, the hypermethylated DMRs and the hypomethylated DMRs. All results all saved as excel spreadsheets. There are three approaches used, which are based on R programs that interface with widely used tools:

A) rGREAT

rGREAT enables the GREAT approach, which works for hg38, hg19, mm10, and mm9. It performs testing based on genomic coordinates and relative to the background regions. It is set to use the “oneClosest” rule.

B) GOfuncR

GOfuncR enables the FUNC approach, which works for all genomes and is our preferred method. It utilizes genomic coordinates and performs permutation based enrichment testing for the DMRs relative to the background regions, which accommodates for gene length bias. By default, DMRichR::GOfuncR() will only map regions to genes if they are between 5 kb upstream and 1 downstream.

C) enrichR

enrichR enables the Enrichr approach, which is based on gene symbols and uses the closest gene to a DMR. It works for all mammalian genomes. While it doesn’t utilize genomic coordinates or background regions, it offers a number of extra databases.

rrvgo and Plots

Finally, DMRichR::slimGO() will take the significant results from of all tools and slim them using rrvgo. DMRichR::GOplot() will then plot the top slimmed significant terms. This approach reduces the redundancy of closely related terms and allows for a more comprehensive overview of the top ontologies.

12) Machine Learning

DMRichR::methylLearn() utilizes random forest and support vector machine algorithms from Boruta and sigFeature in a feature selection approach to identify the most informative DMRs based on individual smoothed methylation values. It creates an excel spreadsheet and an html report of the results along with a heatmap.

13) Cell Composition Estimation

The epigenome is defined by its ability to create cell type specific differences. Therefore, when assaying heterogenous sample sources, it is standard for array-based methylation studies to estimate cell type composition and adjust for it in their model. While this is a standard for array-based studies, it is a significant challenge for WGBS studies due to differences in the nature of the data and the lack of appropriate reference sets and methods. In order to address this, we offer two approaches, both of which provide statistics and plots through DMRichR::CCstats() and DMRichR::CCplot(). However, it must be said that, unlike the rest of DMRichR, this is an experimental feature that you need to further investigate by comparing to array studies that are similar to yours.

A) The Houseman Method

The Houseman method is a standard for arrays and we have adapted it to work with WGBS data. The workflow will convert the smoothed bsseq object to a matrix of beta values for all EPIC array probes. It will then estimate cell composition using the IDOL reference CpGs in a modified Houseman method via DMRichR::Houseman(). If you use the results from this method you should also cite: 1, 2, 3, and 4.

B) The methylCC Method

methylCC is designed to be technology independent by identifying DMRs that define cell types. The workflow uses bumphunter() to find cell type specific DMRs in an array reference database and then examines those regions within your dataset. In this case, it has been modified to utilize the FlowSorted.Blood.EPIC reference dataset and quantile normalization. If you use the results from this method you should also cite: 1 and 2.

14) RData

The output from the main steps is saved in the RData folder so that it can be loaded for custom analyses or to resume an interrupted run:

settings.RData contains the parsed command line options given to DMRichR as well as the annotation database variables. These variables are needed for many of the DMRichR functions, and if you need to reload them, you should also run DMRichR::annotationDatabases(genome) after, since some of the annotation databases have temporary pointers.

bismark.RData contains bs.filtered, which is a bsseq object that contains the filtered cytosine report data and the metadata from sample_info.xlsx in the pData.

Blocks.RData contains blocks, which is a GRanges object of the background blocks. This can be further filtered to produce the sigBlocks object if significant blocks are present.

DMRs.RData contains regions and sigRegions, which are GRanges objects with the background regions and DMRs, respectively.

bsseq.RData contains bs.filtered.bsseq, which is a bsseq object that has been smoothed by bsseq::BSmooth() and is used for the individual methylation values (but not the DMR or block calling by dmrseq, which uses a different smoothing approach).

machineLearning.RData contains methylLearnOutput, which is the output from the machine learning feature selection.

cellComposition.RData contains the output from the cell composition estimation analyses. HousemanCC is from the Houseman method, while methylCC and ccDMRs are from the methylCC method.

Publications

The following publications utilize DMRichR:

Laufer BI*, Neier KE*, Valenzuela AE, Yasui DH, Lein PJ, LaSalle JM. Genome-Wide DNA Methylation Profiles of Neurodevelopmental Disorder Genes in Mouse Placenta and Fetal Brain Following Prenatal Exposure to Polychlorinated Biphenyls. bioRxiv preprint, 2021. doi: 10.1101/2021.05.27.446011

Laufer BI*, Gomez JA*, Jianu JM, LaSalle, JM. Stable DNMT3L Overexpression in SH-SY5Y Neurons Recreates a Facet of the Genome-Wide Down Syndrome DNA Methylation Signature. Epigenetics & Chromatin, 2021. doi: 10.1186/s13072-021-00387-7

Maggio AG, Shu HT, Laufer BI, Hwang H, Bi C, Lai Y, LaSalle JM, Hu VW. Impact of exposures to persistent endocrine disrupting compounds on the sperm methylome in regions associated with neurodevelopmental disorders. medRxiv preprint, 2021. doi: 10.1101/2021.02.21.21252162

Mordaunt CE, Jianu JM, Laufer BI, Zhu Y, Dunaway KW, Bakulski KM, Feinberg JI, Volk HE, Lyall K, Croen LA, Newschaffer CJ, Ozonoff S, Hertz-Picciotto I, Fallin DM, Schmidt RJ, LaSalle JM. Cord blood DNA methylome in newborns later diagnosed with autism spectrum disorder reflects early dysregulation of neurodevelopmental and X-linked genes. Genome Medicine, 2020. doi: 10.1186/s13073-020-00785-8

Laufer BI, Hwang H, Jianu JM, Mordaunt CE, Korf IF, Hertz-Picciotto I, LaSalle JM. Low-Pass Whole Genome Bisulfite Sequencing of Neonatal Dried Blood Spots Identifies a Role for RUNX1 in Down Syndrome DNA Methylation Profiles. Human Molecular Genetics, 2020. doi: 10.1093/hmg/ddaa218

Murat El Houdigui S, Adam-Guillermin C, Armant O. Ionising Radiation Induces Promoter DNA Hypomethylation and Perturbs Transcriptional Activity of Genes Involved in Morphogenesis during Gastrulation in Zebrafish. International Journal of Molecular Sciences, 2020. doi: 10.3390/ijms21114014

Wöste M, Leitão E, Laurentino S, Horsthemke B, Rahmann S, Schröder C. wg-blimp: an end-to-end analysis pipeline for whole genome bisulfite sequencing data. BMC Bioinformatics, 2020. doi: 10.1186/s12859-020-3470-5

Lopez SJ, Laufer BI, Beitnere U, Berg E, Silverman JL, Segal DJ, LaSalle JM. Imprinting effects of UBE3A loss on synaptic gene networks and Wnt signaling pathways. Human Molecular Genetics, 2019. doi: 10.1093/hmg/ddz221

Vogel Ciernia A*, Laufer BI*, Hwang H, Dunaway KW, Mordaunt CE, Coulson RL, Yasui DH, LaSalle JM. Epigenomic convergence of genetic and immune risk factors in autism brain. Cerebral Cortex, 2019. doi: 10.1093/cercor/bhz115

Laufer BI, Hwang H, Vogel Ciernia A, Mordaunt CE, LaSalle JM. Whole genome bisulfite sequencing of Down syndrome brain reveals regional DNA hypermethylation and novel disease insights. Epigenetics, 2019. doi: 10.1080/15592294.2019.1609867

Acknowledgements

The development of this program was suppourted by a Canadian Institutes of Health Research (CIHR) postdoctoral fellowship [MFE-146824] and a CIHR Banting postdoctoral fellowship [BPF-162684]. Hyeyeon Hwang developed methylLearn() and the sex checker for processBismark(). Charles Mordaunt developed getBackground() and plotDMRs2() as well as the CpG filtering approach in processBismark(). I would also like to thank Keegan Korthauer, Matt Settles, and Ian Korf, and Janine LaSalle for invaluable discussions related to the bioinformatic approaches utilized in this repository.