In TransBioInfoLab/MethReg: Assessing the regulatory potential of DNA methylation regions or sites on gene transcription
```{css, echo = FALSE, eval = T}
.whiteCode {
background-color: black;
border-color: #337ab7 !important;
border: 1px solid;
}
```r
options(width = 100)
knitr::opts_chunk$set(collapse = TRUE, comment = "#>",class.source = "whiteCode")
library(dplyr)
library(sesameData)
Introduction
Transcription factors (TFs) are proteins that facilitate the transcription of
DNA into RNA. A number of recent studies have observed that the binding of TFs
onto DNA can be affected by DNA methylation, and in turn, DNA methylation can
also be added or removed by proteins associated with transcription
factors [@bonder2017disease; @banovich2014methylation; @zhu2016transcription].
To provide functional annotations for differentially methylated regions (DMRs)
and differentially methylated CpG sites (DMS), MethReg performs integrative
analyses using matched DNA methylation and gene expression along with
Transcription Factor Binding Sites (TFBS) data. MethReg evaluates, prioritizes
and annotates DNA methylation regions (or sites) with high regulatory potential
that works synergistically with TFs to regulate target gene expressions,
without any additional ChIP-seq data.
The results from MethReg can be used to generate testable hypothesis on the
synergistic collaboration of DNA methylation changes and TFs in gene regulation.
MethReg can be used either to evaluate regulatory potentials of candidate
regions or to search for methylation coupled TF regulatory processes in the entire genome.
Installation
MethReg is a Bioconductor package and can be installed through BiocManager::install().
if (!"BiocManager" %in% rownames(installed.packages()))
install.packages("BiocManager")
BiocManager::install("MethReg", dependencies = TRUE)
After the package is installed, it can be loaded into R workspace by
library(MethReg)
MethReg workflow
The figure below illustrates the workflow for MethReg.
Given matched array DNA methylation data and RNA-seq gene expression data,
MethReg additionally incorporates TF binding information from ReMap2020 [@remap2020]
or the JASPAR2020 [@JASPAR2020; @fornes2020jaspar] database,
and optionally additional TF-target gene interaction databases,
to perform both promoter and distal (enhancer) analysis.
In the unsupervised mode, MethReg analyzes all CpGs on the Illumina arrays.
In the supervised mode, MethReg analyzes and prioritizes differentially methylated CpGs identified in EWAS.
There are three main steps: (1) create a dataset with triplets of CpGs, TFs that bind near the CpGs,
and putative target genes, (2) for each triplet (CpG, TF, target gene), apply integrative
statistical models to DNA methylation, target gene expression, and TF expression values,
and (3) visualize and interpret results from statistical models to estimate individual
and joint impacts of DNA methylation and TF on target gene expression, as well as
annotate the roles of TF and CpG methylation in each triplet.
The results from
the statistical models will also allow us to identify a list of CpGs that work
synergistically with TFs to influence target gene expression.
png::readPNG("workflow_methReg.png") %>% grid::grid.raster()
Analysis illustration
Input data
For illustration, we will use chromosome 21 data from 38 TCGA-COAD (colon cancer) samples.
Input DNA methylation dataset
The DNA methylation dataset is a matrix or SummarizedExperiment object with
methylation beta or M-values
(The samples are in the columns and methylation regions or probes are in the rows).
If there are potential confounding factors (e.g. batch effect, age, sex) in the dataset,
this matrix would contain residuals from fitting linear regression
instead (see details Section 5 "Controlling effects from confounding variables" below).
Analysis for individual CpGs data
We will analyze all CpGs on chromosome 21 in this vignette.
However, oftentimes, the methylation data can also be, for example,
differentially methylated sites (DMS) or differentially methylated regions (DMRs)
obtained in an epigenome-wide association study (EWAS) study.
```{R warning=FALSE}
data("dna.met.chr21")
```{R}
dna.met.chr21[1:5,1:5]
We will first create a SummarizedExperiment object with the function
make_dnam_se. This function will use the Sesame R/Bioconductor package
to map the array probes into genomic regions. You cen set human genome version
(hg38 or hg19) and the array type ("450k" or "EPIC")
dna.met.chr21.se <- make_dnam_se(
dnam = dna.met.chr21,
genome = "hg38",
arrayType = "450k",
betaToM = FALSE, # transform beta to m-values
verbose = FALSE # hide informative messages
)
dna.met.chr21.se
SummarizedExperiment::rowRanges(dna.met.chr21.se)[1:4,1:4]
Analysis of DMRs
Differentially Methylated Regions (DMRs) associated with phenotypes such
as tumor stage can be obtained from R packages such as
coMethDMR, comb-p, DMRcate and many others.
The methylation levels in multiple CpGs within the DMRs need to be
summarized (e.g. using medians), then the analysis for
DMR will proceed in the same way
as those for CpGs.
Input gene expression dataset
The gene expression dataset is a matrix with log2 transformed and
normalized gene expression values.
If there are potential confounding factors (e.g. batch effect, age, sex) in the dataset,
this matrix can also contain residuals from linear regression instead
(see Section 6 "Controlling effects from confounding variables" below).
The samples are in the columns and the genes are in the rows.
data("gene.exp.chr21.log2")
gene.exp.chr21.log2[1:5,1:5]
We will also create a SummarizedExperiment object for the gene expression data.
This object will contain the genomic information for each gene.
gene.exp.chr21.se <- make_exp_se(
exp = gene.exp.chr21.log2,
genome = "hg38",
verbose = FALSE
)
gene.exp.chr21.se
SummarizedExperiment::rowRanges(gene.exp.chr21.se)[1:5,]
Creating triplet dataset
Creating triplet dataset using distance based approaches and JASPAR2022
In this section, regions refer to the regions where CpGs are located.
Linking a region to a target gene
To evaluate the DNA methylation effect on the expression of a gene,
first we need to define which are the possible affected genes.
For this we initially define if the DNA methylation occurred withing a promoter regions,
defined as 2 kbp upstream and 2 kbp downstream of the transcription start site (TSS),
or in a non-promoter region, also known as distal regions, that could behave like enhancer of the gene expression.
Enhancers can increase the transcription of genes and are found in different locations
(upstream or downstream of genes, within introns).
Their functional complexity lies in the possibility genes located more distantly than
the neighboring genes and being able to regulate multiple genes [@pennacchio2013enhancers].
Also, enhancer–promoter looping could happen at two sequences within approximately 1 Mb of each other [@pennacchio2013enhancers].
@williamson2011enhancers also highlighted not only that a proportion of enhancers
are situated hundreds to thousands of kilobases from their target genes,
often in large gene-poor regions, but also the promiscuous activity when placed within gene-rich domains.
These promoters and enhancers interactions could be further identified using Chromosome conformation capture techniques such as 3C, 4C, Hi-C.
However, in the lack of this information one could use the position information in the genome to link an enhancer to a candidate target gene.
Such problem is also identified in the GWAS studies, for example,
@brodie2016far found that affected genes are often up to $2 Mbps$ away from the
associated SNP and highlighted that some studies suggested to use a cutoff of $500 Kbps$
since enhancers and repressors may be as distant as $500 Kbps$ from their genes.
The issue of this method is that with a big window in gene-rich regions would map
to several genes, and a small window might not map the gene-poor region, making the decision on the window size very difficult.
Another method was presented by @yao2015inferring which provided
a linkage method based on a fixed quantity of genes upstream and downstream of the enhancers.
MethReg offer two methods for enhancer linking 1) a window-based method similar to the ones in the GWAS studies,
2) a fixed number of genes upstream and downstream of the DNA methylation loci similar to the one suggested by @yao2015inferring,
and one method for promoter linking, which maps to the gene of the promoter region.
The function create_triplet_distance_based provides those three different methods to
link a region to a target gene:
- Mapping the region to the closest gene (
target.method = "genes.promoter.overlap")
- Mapping the region to a specific number of genes upstream down/upstream
of the region (
target.method = "nearby.genes") [@silva2019elmer].
- Mapping the region to all the genes within a window
(default size = 500 kbp around the region, i.e. +- 250 kbp from start or end
of the region) (
target.method = "window") [@reese2019epigenome].
png::readPNG("mapping_target_strategies.png") %>% grid::grid.raster()
For the analysis of probes in gene promoter region, we recommend setting
method = "genes.promoter.overlap", or
method = "closest.gene".
For the analysis of probes in distal regions, we recommend setting either
method = "window" or method = "nearby.genes".
Note that the distal analysis will be more time and resource consuming.
To link regions to TF using JASPAR2022, MethReg uses motifmatchr [@motifmatchr] to scan
these regions for occurrences of motifs in the database. JASPAR2020 is an
open-access database of curated, non-redundant transcription
factor (TF)-binding profiles [@JASPAR2022; @fornes2022jaspar], which contains
more the 500 human TF motifs.
The motif search width of the scanned region is one important parameter.
Although TF recognizes short specific DNA sequence motifs ($6–12 bp$) [@leporcq2020tfmotifview],
the output of a ChIP-seq experiment can include peaks longer than $1000 bp$ [@boeva2016analysis], but
most of the motifs are found $\pm$ $50-75 bp$ from the TF peak center [@heinz2010simple].
Also, recently, it has been shown by @grossman2018positional that TFs have different positional
bindings around nucleosome-depleted regions of DNA, which could range from $\pm200bp$ around
the center of the DNaseI-hypersensitive (DHS) sites defined by the Roadmap Epigenomics project and
@wang2019identification showed that the methylation levels at UM (unmethylated motifs) and MM (methylated Motifs) were also altered within that range.
Since a single CpG is only 1bp, to predict if the methylation at the loci would affect the TF binding site,
we suggest using a motif search window no bigger than $400bp$.
The argument motif.search.window.size will be used to extend the region when scanning
for the motifs, for example, a motif.search.window.size of 50 will add 25 bp
upstream and 25 bp downstream of the original region.
As an example, the following scripts link CpGs with the probes in gene
promoter region (method 1. above)
```{R, message = FALSE, results = "hide"}
triplet.promoter <- create_triplet_distance_based(
region = dna.met.chr21.se,
target.method = "genes.promoter.overlap",
genome = "hg38",
target.promoter.upstream.dist.tss = 2000,
target.promoter.downstream.dist.tss = 2000,
motif.search.window.size = 400,
motif.search.p.cutoff = 1e-08,
cores = 1
)
Alternatively, we can also link each probe with genes within
$500 kb$ window (method 2).
```{R, message = FALSE, results = "hide"}
# Map probes to genes within 500kb window
triplet.distal.window <- create_triplet_distance_based(
region = dna.met.chr21.se,
genome = "hg38",
target.method = "window",
target.window.size = 500 * 10^3,
target.rm.promoter.regions.from.distal.linking = TRUE,
motif.search.window.size = 500,
motif.search.p.cutoff = 1e-08,
cores = 1
)
For method 3, to map probes to 5 nearest upstream and downstream genes:
```{R, message = FALSE, results = "hide"}
Map probes to 5 genes upstream and 5 downstream
triplet.distal.nearby.genes <- create_triplet_distance_based(
region = dna.met.chr21.se,
genome = "hg38",
target.method = "nearby.genes",
target.num.flanking.genes = 5,
target.window.size = 500 * 10^3,
target.rm.promoter.regions.from.distal.linking = TRUE,
motif.search.window.size = 400,
motif.search.p.cutoff = 1e-08,
cores = 1
)
#### Creating triplet dataset using distance based approaches and REMAP2020
Instead of using JASPAR2020 motifs, we will be using REMAP2018 catalogue of
TF peaks which can be access either using the package `ReMapEnrich`
or a most updated version (RemMap2022) is available online at https://remap.univ-amu.fr/download_page
```r
if (!"BiocManager" %in% rownames(installed.packages()))
install.packages("BiocManager")
BiocManager::install("remap-cisreg/ReMapEnrich", dependencies = TRUE)
To download REMAP2018 catalogue (~1Gb) the following functions are used:
```{R, eval = FALSE}
library(ReMapEnrich)
remapCatalog2018hg38 <- downloadRemapCatalog("/tmp/", assembly = "hg38")
remapCatalog <- bedToGranges(remapCatalog2018hg38)
The function `create_triplet_distance_based` will accept any Granges with TF
information in the same format as the `remapCatalog` one.
```{R, eval = FALSE}
#-------------------------------------------------------------------------------
# Triplets promoter using remap
#-------------------------------------------------------------------------------
triplet.promoter.remap <- create_triplet_distance_based(
region = dna.met.chr21.se,
genome = "hg19",
target.method = "genes.promoter.overlap",
TF.peaks.gr = remapCatalog,
motif.search.window.size = 400,
max.distance.region.target = 10^6,
)
Creating triplet dataset using regulon-based approaches
The human regulons from the dorothea database will be used as an example:
if (!"BiocManager" %in% rownames(installed.packages()))
install.packages("BiocManager")
BiocManager::install("dorothea", dependencies = TRUE)
regulons.dorothea <- dorothea::dorothea_hs
regulons.dorothea %>% head
Using the regulons, you can calculate enrichment scores for each TF across
all samples using dorothea and viper.
rnaseq.tf.es <- get_tf_ES(
exp = gene.exp.chr21.se %>% SummarizedExperiment::assay(),
regulons = regulons.dorothea
)
Finally, triplets can be identified using TF-target from regulon databases with the function create_triplet_regulon_based.
```{R, message = FALSE, results = "hide"}
triplet.regulon <- create_triplet_regulon_based(
region = dna.met.chr21.se,
genome = "hg38",
motif.search.window.size = 400,
tf.target = regulons.dorothea,
max.distance.region.target = 10^6 # 1Mbp
)
```{R}
triplet.regulon %>% head
Example of triplet data frame
The triplet is a data frame with the following columns:
target: gene identifier (obtained from row names of the gene expression matrix),regionID: region/CpG identifier (obtained from row names of the DNA methylation matrix)TF: gene identifier (obtained from the row names of the gene expression matrix)
str(triplet.promoter)
triplet.promoter$distance_region_target_tss %>% range
triplet.promoter %>% head
Note that there may be multiple rows for a CpG region, when multiple
target gene and/or TFs are found close to it.
Evaluating the regulatory potential of CpGs (or DMRs)
Because TF binding to DNA can be influenced by (or influences)
DNA methylation levels nearby [@yin2017impact],
target gene expression levels are often resulted from the synergistic effects
of both TF and DNA methylation. In other words, TF activities in gene
regulation is often affected by DNA methylation.
Our goal then is to highlight DNA methylation regions (or CpGs) where
these synergistic DNAm and TF collaborations occur.
We will perform analyses using the 3 datasets described above in Section 3:
- An input DNA methylation matrix
- An input Gene expression matrix
- The created triplet data frame
Analysis using model with methylation by TF interaction
The function interaction_model assess the regulatory impact of
DNA methylation on TF regulation of target genes via the following approach:
considering DNAm values as a binary variable - we define a binary variable
DNAm Group for DNA methylation values (high = 1, low = 0).
That is, samples with the highest DNAm levels (top 25 percent) has high = 1,
samples with lowest DNAm levels (bottom 25 pecent) has high = 0.
Note that in this implementation, only samples with DNAm values in
the first and last quartiles are considered.
$$log_2(RNA target) \sim log_2(TF) + \text{DNAm Group} + log_2(TF) * \text{DNAm Group}$$
```{R interaction_model, message = FALSE, results = "hide", eval = TRUE}
results.interaction.model <- interaction_model(
triplet = triplet.promoter,
dnam = dna.met.chr21.se,
exp = gene.exp.chr21.se,
dnam.group.threshold = 0.1,
sig.threshold = 0.05,
fdr = T,
stage.wise.analysis = FALSE,
filter.correlated.tf.exp.dnam = F,
filter.triplet.by.sig.term = T
)
The output of `interaction_model` function will be a data frame with the following variables:
* `<variable>_pvalue`: p-value for a tested variable (methylation or TF), given the other variables included in the model.
* `<variable>_estimate`: estimated effect for a variable. If estimate > 0, increasing values
of the variable corresponds to increased outcome values (target gene expression).
If estimate < 0, increasing values of the variable correspond to decreased target gene expression levels.
The following columns are provided for the results of fitting **quartile model** to triplet data:
* direct effect of methylation:
+ `RLM_DNAmGroup_pvalue`: p-value for binary DNA methylation variable
+ `RLM_DNAmGroup_estimate`: estimated DNA methylation effect
* direct effect of TF:
+ `RLM_TF_pvalue` : p-value for TF expression
+ `RLM_TF_estimate`: estimated TF effect
* synergistic effects of methylation and TF:
+ `RLM_DNAmGroup:TF_pvalue`: : p-value for DNA methylation by TF interaction
+ `RLM_DNAmGroup:TF_estimate`: estimated DNA methylation by TF interaction effect
```{R}
# Results for quartile model
results.interaction.model %>% dplyr::select(
c(1,4,5,grep("RLM",colnames(results.interaction.model)))
) %>% head
Stratified analysis by high and low DNA methylation levels
For triplets with significant $log_2(TF) × DNAm$ interaction effect identified
above, we can further assess how gene regulation by TF changes when DNAm
is high or low. To this end, the function
stratified_model fits two separate models (see below) to only
samples with the highest DNAm levels (top 25 percent), and then to
only samples with lowest DNAm levels (bottom 25 percent), separately.
$$\text{Stratified Model: } log_2(RNA target) \sim log_2(TF)$$
```{R stratified_model, message = FALSE, warning = FALSE, results = "hide", eval = TRUE}
results.stratified.model <- stratified_model(
triplet = results.interaction.model,
dnam = dna.met.chr21.se,
exp = gene.exp.chr21.se,
dnam.group.threshold = 0.25
)
```{R}
results.stratified.model %>% head
Visualization of data
The functions plot_interaction_model
will create figures to visualize the data,
in a way that corresponds to the linear model we considered above.
It requires the output from the function interaction_model (a dataframe),
the DNA methylation matrix and the gene expression matrix as input.
```{R plot_interaction_model, eval = TRUE, message = FALSE, results = "hide", warning = FALSE}
plots <- plot_interaction_model(
triplet.results = results.interaction.model[1,],
dnam = dna.met.chr21.se,
exp = gene.exp.chr21.se,
dnam.group.threshold = 0.25
)
```{R, fig.height = 8, fig.width = 13, eval = TRUE, fig.cap = "An example output from MethReg."}
plots
The first row of the figures shows pairwise associations between DNA methylation,
TF, and target gene expression levels.
The second row of the figures shows how much TF activity on target gene expression
levels vary varies by DNA methylation levels.
When TF by methylation interaction is significant (Section 4.1),
we expect the association between TF and target gene expression
to vary depending on whether DNA methylation is low or high.
In this example, when DNA methylation is low, target gene expression is relatively
independent of the amount of TF available. On the other hand, when the DNA methylation
level is high, more abundant TF corresponds to increased gene expression (an activator TF).
One possibility is that DNA methylation might enhance TF binding in this case.
This is an example where DNA methylation and TF work synergistically to affect target gene expression.
While the main goal of MethReg is to prioritize methylation CpGs, also
note that without stratifying by DNA methylation, the overall TF-target
effects (p = 0.971) are not as significant as the association in stratified
analysis in high methylation samples (p = 0.0096).
This demonstrates that by additionally modeling DNA methylation,
we can also nominate TF – target associations that might have been missed otherwise.
Note that because of the small sample size (only 38 samples) included in this example for illustration, the P-value for high methylation samples (p = 0.096)
is only marginally significant.
In real data analysis, we
expect MethReg to work well with at least 100 matched samples measured
with both methylations and gene expressions,
and we recommend using a more stringent significance threshold (i.e., FDR < 0.05).
See details in our published paper (Silva et al. 2022, PMID: 35100398).
Results interpretation
Shown below are some expected results from fitting Models 1 & 2 described in
Section 4.1 above, depending on TF binding preferences. Please note that
there can be more possible scenarios than those listed here, therefore,
careful evaluation of the statistical models and visualization of data as
described in Section 4 are needed to gain a good understanding of the
multi-omics data.
png::readPNG("scenarios.png") %>% grid::grid.raster()
Controlling effects from confounding variables
Both gene expressions and DNA methylation levels can be affected by age, sex,
shifting in cell types, batch effects and other confounding (or covariate) variables.
In this section, we illustrate analysis workflow that reduces confounding effects,
by first extracting the residual data with the function get_residuals,
before fitting the models discussed above in Section 4.
The get_residuals function will use gene expression (or DNA methylation data)
and phenotype data as input. To remove confounding effects in gene expression data,
we use the get_residuals function which extract residuals after fitting the
following model for gene expression data:
$$log_2(RNA target) \sim covariate_{1} + covariate_{2} + ... + covariate_{N}$$
or the following model for methylation data:
$$methylation.Mvalues \sim covariate_{1} + covariate_{2} + ... + covariate_{N}$$
```{R residuals, results = "hide", eval = FALSE}
data("gene.exp.chr21.log2")
data("clinical")
metadata <- clinical[,c("sample_type","gender")]
gene.exp.chr21.residuals <- get_residuals(gene.exp.chr21, metadata) %>% as.matrix()
```{R, eval = FALSE}
gene.exp.chr21.residuals[1:5,1:5]
```{R, results = "hide", eval = FALSE}
data("dna.met.chr21")
dna.met.chr21 <- make_se_from_dnam_probes(
dnam = dna.met.chr21,
genome = "hg38",
arrayType = "450k",
betaToM = TRUE
)
dna.met.chr21.residuals <- get_residuals(dna.met.chr21, metadata) %>% as.matrix()
```{R, eval = FALSE}
dna.met.chr21.residuals[1:5,1:5]
The models described in Section 4.1 can then be applied to these residuals
data using the interaction_model function:
```{R, message = FALSE, results = "hide", eval = FALSE}
results <- interaction_model(
triplet = triplet,
dnam = dna.met.chr21.residuals,
exp = gene.exp.chr21.residuals
)
# Calculating enrichment scores
## Using dorothea and viper
This example shows how to use dorothea regulons and viper to calculate
enrichment scores for each TF across all samples.
```{R}
regulons.dorothea <- dorothea::dorothea_hs
regulons.dorothea %>% head
```{R, message = FALSE, results = "hide"}
rnaseq.tf.es <- get_tf_ES(
exp = gene.exp.chr21.se %>% SummarizedExperiment::assay(),
regulons = regulons.dorothea
)
```{R}
rnaseq.tf.es[1:4,1:4]
Using dorothea and GSVA
```{R, message = FALSE, results = "hide"}
regulons.dorothea <- dorothea::dorothea_hs
regulons.dorothea$tf <- MethReg:::map_symbol_to_ensg(
gene.symbol = regulons.dorothea$tf,
genome = "hg38"
)
regulons.dorothea$target <- MethReg:::map_symbol_to_ensg(
gene.symbol = regulons.dorothea$target,
genome = "hg38"
)
split_tibble <- function(tibble, col = 'col') tibble %>% split(., .[, col])
regulons.dorothea.list <- regulons.dorothea %>% na.omit() %>%
split_tibble('tf') %>%
lapply(function(x){x[[3]]})
```{R, message = FALSE, results = "hide", eval = FALSE}
library(GSVA)
rnaseq.tf.es.gsva <- gsva(
expr = gene.exp.chr21.se %>% SummarizedExperiment::assay(),
gset.idx.list = regulons.dorothea.list,
method = "gsva",
kcdf = "Gaussian",
abs.ranking = TRUE,
min.sz = 5,
max.sz = Inf,
parallel.sz = 1L,
mx.diff = TRUE,
ssgsea.norm = TRUE,
verbose = TRUE
)
Session information
{R,size = 'tiny'}
sessionInfo()
Bibliography
TransBioInfoLab/MethReg documentation built on July 28, 2023, 9:17 p.m.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
```{css, echo = FALSE, eval = T} .whiteCode { background-color: black; border-color: #337ab7 !important; border: 1px solid; }
```r options(width = 100) knitr::opts_chunk$set(collapse = TRUE, comment = "#>",class.source = "whiteCode") library(dplyr)
library(sesameData)
Introduction
Transcription factors (TFs) are proteins that facilitate the transcription of DNA into RNA. A number of recent studies have observed that the binding of TFs onto DNA can be affected by DNA methylation, and in turn, DNA methylation can also be added or removed by proteins associated with transcription factors [@bonder2017disease; @banovich2014methylation; @zhu2016transcription].
To provide functional annotations for differentially methylated regions (DMRs)
and differentially methylated CpG sites (DMS), MethReg performs integrative
analyses using matched DNA methylation and gene expression along with
Transcription Factor Binding Sites (TFBS) data. MethReg evaluates, prioritizes
and annotates DNA methylation regions (or sites) with high regulatory potential
that works synergistically with TFs to regulate target gene expressions,
without any additional ChIP-seq data.
The results from MethReg can be used to generate testable hypothesis on the
synergistic collaboration of DNA methylation changes and TFs in gene regulation.
MethReg can be used either to evaluate regulatory potentials of candidate
regions or to search for methylation coupled TF regulatory processes in the entire genome.
Installation
MethReg is a Bioconductor package and can be installed through BiocManager::install().
if (!"BiocManager" %in% rownames(installed.packages())) install.packages("BiocManager") BiocManager::install("MethReg", dependencies = TRUE)
After the package is installed, it can be loaded into R workspace by
library(MethReg)
MethReg workflow
The figure below illustrates the workflow for MethReg. Given matched array DNA methylation data and RNA-seq gene expression data, MethReg additionally incorporates TF binding information from ReMap2020 [@remap2020] or the JASPAR2020 [@JASPAR2020; @fornes2020jaspar] database, and optionally additional TF-target gene interaction databases, to perform both promoter and distal (enhancer) analysis.
In the unsupervised mode, MethReg analyzes all CpGs on the Illumina arrays. In the supervised mode, MethReg analyzes and prioritizes differentially methylated CpGs identified in EWAS.
There are three main steps: (1) create a dataset with triplets of CpGs, TFs that bind near the CpGs, and putative target genes, (2) for each triplet (CpG, TF, target gene), apply integrative statistical models to DNA methylation, target gene expression, and TF expression values, and (3) visualize and interpret results from statistical models to estimate individual and joint impacts of DNA methylation and TF on target gene expression, as well as annotate the roles of TF and CpG methylation in each triplet.
The results from the statistical models will also allow us to identify a list of CpGs that work synergistically with TFs to influence target gene expression.
png::readPNG("workflow_methReg.png") %>% grid::grid.raster()
Analysis illustration
Input data
For illustration, we will use chromosome 21 data from 38 TCGA-COAD (colon cancer) samples.
Input DNA methylation dataset
The DNA methylation dataset is a matrix or SummarizedExperiment object with methylation beta or M-values (The samples are in the columns and methylation regions or probes are in the rows). If there are potential confounding factors (e.g. batch effect, age, sex) in the dataset, this matrix would contain residuals from fitting linear regression instead (see details Section 5 "Controlling effects from confounding variables" below).
Analysis for individual CpGs data
We will analyze all CpGs on chromosome 21 in this vignette.
However, oftentimes, the methylation data can also be, for example, differentially methylated sites (DMS) or differentially methylated regions (DMRs) obtained in an epigenome-wide association study (EWAS) study.
```{R warning=FALSE} data("dna.met.chr21")
```{R}
dna.met.chr21[1:5,1:5]
We will first create a SummarizedExperiment object with the function
make_dnam_se. This function will use the Sesame R/Bioconductor package
to map the array probes into genomic regions. You cen set human genome version
(hg38 or hg19) and the array type ("450k" or "EPIC")
dna.met.chr21.se <- make_dnam_se( dnam = dna.met.chr21, genome = "hg38", arrayType = "450k", betaToM = FALSE, # transform beta to m-values verbose = FALSE # hide informative messages )
dna.met.chr21.se SummarizedExperiment::rowRanges(dna.met.chr21.se)[1:4,1:4]
Analysis of DMRs
Differentially Methylated Regions (DMRs) associated with phenotypes such
as tumor stage can be obtained from R packages such as
coMethDMR, comb-p, DMRcate and many others.
The methylation levels in multiple CpGs within the DMRs need to be
summarized (e.g. using medians), then the analysis for
DMR will proceed in the same way
as those for CpGs.
Input gene expression dataset
The gene expression dataset is a matrix with log2 transformed and normalized gene expression values. If there are potential confounding factors (e.g. batch effect, age, sex) in the dataset, this matrix can also contain residuals from linear regression instead (see Section 6 "Controlling effects from confounding variables" below).
The samples are in the columns and the genes are in the rows.
data("gene.exp.chr21.log2") gene.exp.chr21.log2[1:5,1:5]
We will also create a SummarizedExperiment object for the gene expression data. This object will contain the genomic information for each gene.
gene.exp.chr21.se <- make_exp_se( exp = gene.exp.chr21.log2, genome = "hg38", verbose = FALSE ) gene.exp.chr21.se SummarizedExperiment::rowRanges(gene.exp.chr21.se)[1:5,]
Creating triplet dataset
Creating triplet dataset using distance based approaches and JASPAR2022
In this section, regions refer to the regions where CpGs are located.
Linking a region to a target gene
To evaluate the DNA methylation effect on the expression of a gene, first we need to define which are the possible affected genes. For this we initially define if the DNA methylation occurred withing a promoter regions, defined as 2 kbp upstream and 2 kbp downstream of the transcription start site (TSS), or in a non-promoter region, also known as distal regions, that could behave like enhancer of the gene expression.
Enhancers can increase the transcription of genes and are found in different locations (upstream or downstream of genes, within introns). Their functional complexity lies in the possibility genes located more distantly than the neighboring genes and being able to regulate multiple genes [@pennacchio2013enhancers]. Also, enhancer–promoter looping could happen at two sequences within approximately 1 Mb of each other [@pennacchio2013enhancers]. @williamson2011enhancers also highlighted not only that a proportion of enhancers are situated hundreds to thousands of kilobases from their target genes, often in large gene-poor regions, but also the promiscuous activity when placed within gene-rich domains.
These promoters and enhancers interactions could be further identified using Chromosome conformation capture techniques such as 3C, 4C, Hi-C. However, in the lack of this information one could use the position information in the genome to link an enhancer to a candidate target gene. Such problem is also identified in the GWAS studies, for example, @brodie2016far found that affected genes are often up to $2 Mbps$ away from the associated SNP and highlighted that some studies suggested to use a cutoff of $500 Kbps$ since enhancers and repressors may be as distant as $500 Kbps$ from their genes. The issue of this method is that with a big window in gene-rich regions would map to several genes, and a small window might not map the gene-poor region, making the decision on the window size very difficult. Another method was presented by @yao2015inferring which provided a linkage method based on a fixed quantity of genes upstream and downstream of the enhancers.
MethReg offer two methods for enhancer linking 1) a window-based method similar to the ones in the GWAS studies, 2) a fixed number of genes upstream and downstream of the DNA methylation loci similar to the one suggested by @yao2015inferring, and one method for promoter linking, which maps to the gene of the promoter region.
The function create_triplet_distance_based provides those three different methods to
link a region to a target gene:
- Mapping the region to the closest gene (
target.method = "genes.promoter.overlap") - Mapping the region to a specific number of genes upstream down/upstream
of the region (
target.method = "nearby.genes") [@silva2019elmer]. - Mapping the region to all the genes within a window
(default size = 500 kbp around the region, i.e. +- 250 kbp from start or end
of the region) (
target.method = "window") [@reese2019epigenome].
png::readPNG("mapping_target_strategies.png") %>% grid::grid.raster()
For the analysis of probes in gene promoter region, we recommend setting
method = "genes.promoter.overlap", or
method = "closest.gene".
For the analysis of probes in distal regions, we recommend setting either
method = "window" or method = "nearby.genes".
Note that the distal analysis will be more time and resource consuming.
To link regions to TF using JASPAR2022, MethReg uses motifmatchr [@motifmatchr] to scan
these regions for occurrences of motifs in the database. JASPAR2020 is an
open-access database of curated, non-redundant transcription
factor (TF)-binding profiles [@JASPAR2022; @fornes2022jaspar], which contains
more the 500 human TF motifs.
The motif search width of the scanned region is one important parameter. Although TF recognizes short specific DNA sequence motifs ($6–12 bp$) [@leporcq2020tfmotifview], the output of a ChIP-seq experiment can include peaks longer than $1000 bp$ [@boeva2016analysis], but most of the motifs are found $\pm$ $50-75 bp$ from the TF peak center [@heinz2010simple]. Also, recently, it has been shown by @grossman2018positional that TFs have different positional bindings around nucleosome-depleted regions of DNA, which could range from $\pm200bp$ around the center of the DNaseI-hypersensitive (DHS) sites defined by the Roadmap Epigenomics project and @wang2019identification showed that the methylation levels at UM (unmethylated motifs) and MM (methylated Motifs) were also altered within that range. Since a single CpG is only 1bp, to predict if the methylation at the loci would affect the TF binding site, we suggest using a motif search window no bigger than $400bp$.
The argument motif.search.window.size will be used to extend the region when scanning
for the motifs, for example, a motif.search.window.size of 50 will add 25 bp
upstream and 25 bp downstream of the original region.
As an example, the following scripts link CpGs with the probes in gene promoter region (method 1. above)
```{R, message = FALSE, results = "hide"}
triplet.promoter <- create_triplet_distance_based(
region = dna.met.chr21.se,
target.method = "genes.promoter.overlap",
genome = "hg38",
target.promoter.upstream.dist.tss = 2000,
target.promoter.downstream.dist.tss = 2000,
motif.search.window.size = 400,
motif.search.p.cutoff = 1e-08,
cores = 1
)
Alternatively, we can also link each probe with genes within $500 kb$ window (method 2). ```{R, message = FALSE, results = "hide"} # Map probes to genes within 500kb window triplet.distal.window <- create_triplet_distance_based( region = dna.met.chr21.se, genome = "hg38", target.method = "window", target.window.size = 500 * 10^3, target.rm.promoter.regions.from.distal.linking = TRUE, motif.search.window.size = 500, motif.search.p.cutoff = 1e-08, cores = 1 )
For method 3, to map probes to 5 nearest upstream and downstream genes:
```{R, message = FALSE, results = "hide"}
Map probes to 5 genes upstream and 5 downstream
triplet.distal.nearby.genes <- create_triplet_distance_based(
region = dna.met.chr21.se,
genome = "hg38",
target.method = "nearby.genes",
target.num.flanking.genes = 5,
target.window.size = 500 * 10^3,
target.rm.promoter.regions.from.distal.linking = TRUE,
motif.search.window.size = 400,
motif.search.p.cutoff = 1e-08,
cores = 1
)
#### Creating triplet dataset using distance based approaches and REMAP2020
Instead of using JASPAR2020 motifs, we will be using REMAP2018 catalogue of
TF peaks which can be access either using the package `ReMapEnrich`
or a most updated version (RemMap2022) is available online at https://remap.univ-amu.fr/download_page
```r
if (!"BiocManager" %in% rownames(installed.packages()))
install.packages("BiocManager")
BiocManager::install("remap-cisreg/ReMapEnrich", dependencies = TRUE)
To download REMAP2018 catalogue (~1Gb) the following functions are used:
```{R, eval = FALSE} library(ReMapEnrich) remapCatalog2018hg38 <- downloadRemapCatalog("/tmp/", assembly = "hg38") remapCatalog <- bedToGranges(remapCatalog2018hg38)
The function `create_triplet_distance_based` will accept any Granges with TF
information in the same format as the `remapCatalog` one.
```{R, eval = FALSE}
#-------------------------------------------------------------------------------
# Triplets promoter using remap
#-------------------------------------------------------------------------------
triplet.promoter.remap <- create_triplet_distance_based(
region = dna.met.chr21.se,
genome = "hg19",
target.method = "genes.promoter.overlap",
TF.peaks.gr = remapCatalog,
motif.search.window.size = 400,
max.distance.region.target = 10^6,
)
Creating triplet dataset using regulon-based approaches
The human regulons from the dorothea database will be used as an example:
if (!"BiocManager" %in% rownames(installed.packages())) install.packages("BiocManager") BiocManager::install("dorothea", dependencies = TRUE)
regulons.dorothea <- dorothea::dorothea_hs regulons.dorothea %>% head
Using the regulons, you can calculate enrichment scores for each TF across all samples using dorothea and viper.
rnaseq.tf.es <- get_tf_ES( exp = gene.exp.chr21.se %>% SummarizedExperiment::assay(), regulons = regulons.dorothea )
Finally, triplets can be identified using TF-target from regulon databases with the function create_triplet_regulon_based.
```{R, message = FALSE, results = "hide"}
triplet.regulon <- create_triplet_regulon_based(
region = dna.met.chr21.se,
genome = "hg38",
motif.search.window.size = 400,
tf.target = regulons.dorothea,
max.distance.region.target = 10^6 # 1Mbp
)
```{R}
triplet.regulon %>% head
Example of triplet data frame
The triplet is a data frame with the following columns:
target: gene identifier (obtained from row names of the gene expression matrix),regionID: region/CpG identifier (obtained from row names of the DNA methylation matrix)TF: gene identifier (obtained from the row names of the gene expression matrix)
str(triplet.promoter) triplet.promoter$distance_region_target_tss %>% range triplet.promoter %>% head
Note that there may be multiple rows for a CpG region, when multiple target gene and/or TFs are found close to it.
Evaluating the regulatory potential of CpGs (or DMRs)
Because TF binding to DNA can be influenced by (or influences) DNA methylation levels nearby [@yin2017impact], target gene expression levels are often resulted from the synergistic effects of both TF and DNA methylation. In other words, TF activities in gene regulation is often affected by DNA methylation.
Our goal then is to highlight DNA methylation regions (or CpGs) where these synergistic DNAm and TF collaborations occur. We will perform analyses using the 3 datasets described above in Section 3:
- An input DNA methylation matrix
- An input Gene expression matrix
- The created triplet data frame
Analysis using model with methylation by TF interaction
The function interaction_model assess the regulatory impact of
DNA methylation on TF regulation of target genes via the following approach:
considering DNAm values as a binary variable - we define a binary variable
DNAm Group for DNA methylation values (high = 1, low = 0).
That is, samples with the highest DNAm levels (top 25 percent) has high = 1,
samples with lowest DNAm levels (bottom 25 pecent) has high = 0.
Note that in this implementation, only samples with DNAm values in the first and last quartiles are considered.
$$log_2(RNA target) \sim log_2(TF) + \text{DNAm Group} + log_2(TF) * \text{DNAm Group}$$
```{R interaction_model, message = FALSE, results = "hide", eval = TRUE} results.interaction.model <- interaction_model( triplet = triplet.promoter, dnam = dna.met.chr21.se, exp = gene.exp.chr21.se, dnam.group.threshold = 0.1, sig.threshold = 0.05, fdr = T, stage.wise.analysis = FALSE, filter.correlated.tf.exp.dnam = F, filter.triplet.by.sig.term = T )
The output of `interaction_model` function will be a data frame with the following variables:
* `<variable>_pvalue`: p-value for a tested variable (methylation or TF), given the other variables included in the model.
* `<variable>_estimate`: estimated effect for a variable. If estimate > 0, increasing values
of the variable corresponds to increased outcome values (target gene expression).
If estimate < 0, increasing values of the variable correspond to decreased target gene expression levels.
The following columns are provided for the results of fitting **quartile model** to triplet data:
* direct effect of methylation:
+ `RLM_DNAmGroup_pvalue`: p-value for binary DNA methylation variable
+ `RLM_DNAmGroup_estimate`: estimated DNA methylation effect
* direct effect of TF:
+ `RLM_TF_pvalue` : p-value for TF expression
+ `RLM_TF_estimate`: estimated TF effect
* synergistic effects of methylation and TF:
+ `RLM_DNAmGroup:TF_pvalue`: : p-value for DNA methylation by TF interaction
+ `RLM_DNAmGroup:TF_estimate`: estimated DNA methylation by TF interaction effect
```{R}
# Results for quartile model
results.interaction.model %>% dplyr::select(
c(1,4,5,grep("RLM",colnames(results.interaction.model)))
) %>% head
Stratified analysis by high and low DNA methylation levels
For triplets with significant $log_2(TF) × DNAm$ interaction effect identified
above, we can further assess how gene regulation by TF changes when DNAm
is high or low. To this end, the function
stratified_model fits two separate models (see below) to only
samples with the highest DNAm levels (top 25 percent), and then to
only samples with lowest DNAm levels (bottom 25 percent), separately.
$$\text{Stratified Model: } log_2(RNA target) \sim log_2(TF)$$
```{R stratified_model, message = FALSE, warning = FALSE, results = "hide", eval = TRUE} results.stratified.model <- stratified_model( triplet = results.interaction.model, dnam = dna.met.chr21.se, exp = gene.exp.chr21.se, dnam.group.threshold = 0.25 )
```{R}
results.stratified.model %>% head
Visualization of data
The functions plot_interaction_model
will create figures to visualize the data,
in a way that corresponds to the linear model we considered above.
It requires the output from the function interaction_model (a dataframe),
the DNA methylation matrix and the gene expression matrix as input.
```{R plot_interaction_model, eval = TRUE, message = FALSE, results = "hide", warning = FALSE} plots <- plot_interaction_model( triplet.results = results.interaction.model[1,], dnam = dna.met.chr21.se, exp = gene.exp.chr21.se, dnam.group.threshold = 0.25 )
```{R, fig.height = 8, fig.width = 13, eval = TRUE, fig.cap = "An example output from MethReg."}
plots
The first row of the figures shows pairwise associations between DNA methylation, TF, and target gene expression levels.
The second row of the figures shows how much TF activity on target gene expression levels vary varies by DNA methylation levels. When TF by methylation interaction is significant (Section 4.1), we expect the association between TF and target gene expression to vary depending on whether DNA methylation is low or high.
In this example, when DNA methylation is low, target gene expression is relatively independent of the amount of TF available. On the other hand, when the DNA methylation level is high, more abundant TF corresponds to increased gene expression (an activator TF). One possibility is that DNA methylation might enhance TF binding in this case. This is an example where DNA methylation and TF work synergistically to affect target gene expression.
While the main goal of MethReg is to prioritize methylation CpGs, also note that without stratifying by DNA methylation, the overall TF-target effects (p = 0.971) are not as significant as the association in stratified analysis in high methylation samples (p = 0.0096). This demonstrates that by additionally modeling DNA methylation, we can also nominate TF – target associations that might have been missed otherwise.
Note that because of the small sample size (only 38 samples) included in this example for illustration, the P-value for high methylation samples (p = 0.096)
is only marginally significant.
In real data analysis, we
expect MethReg to work well with at least 100 matched samples measured
with both methylations and gene expressions,
and we recommend using a more stringent significance threshold (i.e., FDR < 0.05).
See details in our published paper (Silva et al. 2022, PMID: 35100398).
Results interpretation
Shown below are some expected results from fitting Models 1 & 2 described in Section 4.1 above, depending on TF binding preferences. Please note that there can be more possible scenarios than those listed here, therefore, careful evaluation of the statistical models and visualization of data as described in Section 4 are needed to gain a good understanding of the multi-omics data.
png::readPNG("scenarios.png") %>% grid::grid.raster()
Controlling effects from confounding variables
Both gene expressions and DNA methylation levels can be affected by age, sex,
shifting in cell types, batch effects and other confounding (or covariate) variables.
In this section, we illustrate analysis workflow that reduces confounding effects,
by first extracting the residual data with the function get_residuals,
before fitting the models discussed above in Section 4.
The get_residuals function will use gene expression (or DNA methylation data)
and phenotype data as input. To remove confounding effects in gene expression data,
we use the get_residuals function which extract residuals after fitting the
following model for gene expression data:
$$log_2(RNA target) \sim covariate_{1} + covariate_{2} + ... + covariate_{N}$$
or the following model for methylation data:
$$methylation.Mvalues \sim covariate_{1} + covariate_{2} + ... + covariate_{N}$$
```{R residuals, results = "hide", eval = FALSE} data("gene.exp.chr21.log2") data("clinical") metadata <- clinical[,c("sample_type","gender")]
gene.exp.chr21.residuals <- get_residuals(gene.exp.chr21, metadata) %>% as.matrix()
```{R, eval = FALSE}
gene.exp.chr21.residuals[1:5,1:5]
```{R, results = "hide", eval = FALSE} data("dna.met.chr21") dna.met.chr21 <- make_se_from_dnam_probes( dnam = dna.met.chr21, genome = "hg38", arrayType = "450k", betaToM = TRUE ) dna.met.chr21.residuals <- get_residuals(dna.met.chr21, metadata) %>% as.matrix()
```{R, eval = FALSE}
dna.met.chr21.residuals[1:5,1:5]
The models described in Section 4.1 can then be applied to these residuals
data using the interaction_model function:
```{R, message = FALSE, results = "hide", eval = FALSE} results <- interaction_model( triplet = triplet, dnam = dna.met.chr21.residuals, exp = gene.exp.chr21.residuals )
# Calculating enrichment scores ## Using dorothea and viper This example shows how to use dorothea regulons and viper to calculate enrichment scores for each TF across all samples. ```{R} regulons.dorothea <- dorothea::dorothea_hs regulons.dorothea %>% head
```{R, message = FALSE, results = "hide"} rnaseq.tf.es <- get_tf_ES( exp = gene.exp.chr21.se %>% SummarizedExperiment::assay(), regulons = regulons.dorothea )
```{R}
rnaseq.tf.es[1:4,1:4]
Using dorothea and GSVA
```{R, message = FALSE, results = "hide"} regulons.dorothea <- dorothea::dorothea_hs regulons.dorothea$tf <- MethReg:::map_symbol_to_ensg( gene.symbol = regulons.dorothea$tf, genome = "hg38" ) regulons.dorothea$target <- MethReg:::map_symbol_to_ensg( gene.symbol = regulons.dorothea$target, genome = "hg38" ) split_tibble <- function(tibble, col = 'col') tibble %>% split(., .[, col]) regulons.dorothea.list <- regulons.dorothea %>% na.omit() %>% split_tibble('tf') %>% lapply(function(x){x[[3]]})
```{R, message = FALSE, results = "hide", eval = FALSE}
library(GSVA)
rnaseq.tf.es.gsva <- gsva(
expr = gene.exp.chr21.se %>% SummarizedExperiment::assay(),
gset.idx.list = regulons.dorothea.list,
method = "gsva",
kcdf = "Gaussian",
abs.ranking = TRUE,
min.sz = 5,
max.sz = Inf,
parallel.sz = 1L,
mx.diff = TRUE,
ssgsea.norm = TRUE,
verbose = TRUE
)
Session information
{R,size = 'tiny'}
sessionInfo()
Bibliography
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Embedding an R snippet on your website
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.