This R package is the reference implementation of the BSmooth algorithm for analyzing whole-genome bisulfite sequencing (WGBS) data. This package does not contain alignment software, which is available from . This package is not intended for analysis of single-loci bisulfite sequencing (typically Sanger bisulfite sequencing or pyro bisulfite sequencing).
The package has been used to analyze capture bisulfite sequencing data. For this type of data, the most useful parts of the package is its data-handling functions. The BSmooth algorithm itself may not be useful for a standard capture bisulfite sequencing experiment, since it relies heavily on smoothing, which again requires that we have measured methylation in bigger regions of the genome.
The BSmooth algorithm is described in detail in (Hansen, Langmead, and Irizarry 2012). It was applied to human cancer data in (Hansen, Timp, et al. 2011) and we have also used it to analyze data from Rhesus Macaque (Tung et al. 2012). Specifically, the algorithm uses smoothing to get reliable semi-local methylation estimates from low-coverage bisulfite sequencing. After smoothing it uses biological replicates to estimate biological variation and identify differentially methylated regions (DMRs). The smoothing portion could be used even on a single sample, but we believe that variation between individuals is an important aspect of DNA methylation and needs to be accounted for, see also (Hansen, Wu, et al. 2011) for a relevant discussion.
The main focus for our development has been the analysis of CpG methylation in humans, and we have successfully used it in other primates (Tung et al. 2012). It is highly likely that the approach will work in non-human organisms, but care must be taken: the smoothing parameters (which we have selected based on human data) should be evaluated carefully. Furthermore, it may not be suitable at all for organisms with vastly different (from humans) methylation structures.
With respect to non-CpG methylation, the situation is mixed. We have never used the algorithm to analyze non-CpG methylation, but it should be straightforward to do so. However, the data structures used in the current code might not conveniently scale from the 28.2M CpGs in the human genome to the roughly 2x585M Cs (it may well be possible to do an separate analysis for each chromosome). This should be less of an issue for organisms with smaller genomes. We are considering changing these underlying data structures to allow for easier analysis of non-CpG methylation in humans.
The package requires that all the data is loaded into system memory. By ``data’’ we do not mean the individual reads (which is big for a whole-genome experiment). Instead, what we need are summarized data given us the number of reads supporting methylation as well as the total number of reads at each loci. Focusing on human data, we need to work with objects with a maximum of 28.2 million entries, per sample (since there are roughly 28.2 millions CpGs in the human genome). This provides us with an upper limit on the data object.
Based on this, the 8 samples from (Hansen, Timp, et al. 2011) including the smoothed values, take up roughly 1.2GB of RAM, meaning an analysis can easily be done with 8GB of RAM. In order to improve speed, the package allows for easy parallel processing of samples/chromosomes. This will require multiple copies of the data object for each core used, increasing the memory usage substantially to perhaps even 32GB. This can be avoided by processing the samples sequentially at the cost of speed.
On a 64GB node, the 8 samples from (Hansen, Timp, et al. 2011) takes roughly one hour to process in parallel using 8 cores (using much less than 64GB of RAM in total). This does not including parsing the data from the alignment output files.
Because not all methylation happens at CpG sites, we have tried to use the term ``methylation loci’’ instead of CpG. We use this term to refer to a single-base methylation site.
Some standard terms from the DNA methylation field: differentially methylated region (DMR), hyper methylation (methylation that is higher in one condition/sample than in another), hypo methylation (methylation that is lower in one condition/sample than in another), and finally differentially methylated position (DMP) referring to a single loci.
If you use this package, please cite our BSmooth paper (Hansen, Langmead, and Irizarry 2012).
The package assumes that the following data has been extracted from alignments:
We can have a look at some data from (Lister et al. 2009), specifically chromosome 22 from the IMR90 cell line.
## An object of type 'BSseq' with
## 494728 methylation loci
## 2 samples
## has not been smoothed
## All assays are in-memoryThe genomic positions are stored as a GRanges object
GRanges are general genomic regions; we represent a single
base methylation loci as an interval of width 1 (which may seem a bit
strange, but there are good reasons for this). For example, the first 4
loci in the Lister data are
## GRanges object with 4 ranges and 0 metadata columns:
## seqnames ranges strand
## <Rle> <IRanges> <Rle>
## [1] chr22 14430632 *
## [2] chr22 14430677 *
## [3] chr22 14430687 *
## [4] chr22 14430702 *
## -------
## seqinfo: 1 sequence from an unspecified genome; no seqlengthsWe also have the M and Cov matrices
## r1 r2
## [1,] 17 20
## [2,] 4 20
## [3,] 6 19
## [4,] 2 4## r1 r2
## [1,] 18 23
## [2,] 11 28
## [3,] 10 25
## [4,] 8 21Since CpG methylation is symmetric on the two strands of a chromosome, we aggregated reads on the forward and reverse strand to form a single value, and we assume the genomic position points to the C of the CpG. It is not crucial in any way to do this, one may easily analyze each strand separately, but CpG methylation is symmetric and this halves the number of loci.
How to input your methylation data into this data structure (called a
BSseq object) is described in a section below. We also have
a section on how to operate on these types of objects.
An analysis typically consists of the following steps.
BSmooth().BSmooth.tstat(). This converts the BSseq
object into a BSseqTstat object.dmrFinder(), returning a simple
data.frame.It is usually a good idea to look at the smoothed data either before
or after identifying DMRs. This can be done using the functions
plotRegion() and plotManyRegions().
We also have functions for assessing goodness of fit for binomial and
poison models; this is like to be of marginal interest to most users.
See the man page for binomialGoodnessOfFit().
We also allow for easy computation of Fisher’s exact test for each
loci, using the function fisherTests().
Objects of class BSseq contains a GRanges
object with the genomic locations. This GRanges object can
be obtained by granges(). A number of standard
GRanges methods works directly on the BSseq
object, such as start(), end(),
seqnames() (chromosome names) etc.
These objects also contains a phenoData object for
sample pheno data. Useful methods are sampleNames(),
pData().
Finally, we have methods such as dim(),
ncol() (number of columns; number of samples),
nrow() (number of rows; number of methylation loci).
Objects can be subsetted using two indicies BSseq[i,j]
with the first index being methylation loci and the second index being
samples. Another very useful way of subsetting the object is by using
the method subsetByOverlaps(). This selects all methylation
loci inside a set of genomic intervals (there is a difference between
first and second argument and either can be BSseq or
GRanges).
Examples:
## [1] 14430632 14430677 14430687 14430702## factor-Rle of length 4 with 1 run
## Lengths: 4
## Values : chr22
## Levels(1): chr22## [1] "r1" "r2"## DataFrame with 2 rows and 1 column
## Rep
## <character>
## r1 replicate1
## r2 replicate2## [1] 494728 2## An object of type 'BSseq' with
## 6 methylation loci
## 1 samples
## has not been smoothed
## All assays are in-memory## An object of type 'BSseq' with
## 67082 methylation loci
## 2 samples
## has not been smoothed
## All assays are in-memoryWe have a number of functions for manipulating one or more
BSseq objects.
BSseq() instantiates an object of class
BSseq. Genomic locations are passed in, either as a
GRanges object (argument gr) or as chromosome
and location vectors (arguments chr and pos).
The arguments M and Cov accepts matrices, and
it is possible to directly give it a phenoData object.
M <- matrix(0:8, 3, 3)
Cov <- matrix(1:9, 3, 3)
BStmp <- BSseq(chr = c("chr1", "chrX", "chr1"), pos = 1:3,
M = M, Cov = Cov, sampleNames = c("A1","A2", "B"))A BSseq object may be ordered by
orderBSseq(). This ensures that data from a single
chromosome appears in an ordered, contiguous block. There is also the
possibility for specifying the chromosome order (this is less
important). The smoothing functions assumes that the underlying
BSseq has been ordered.
## GRanges object with 3 ranges and 0 metadata columns:
## seqnames ranges strand
## <Rle> <IRanges> <Rle>
## [1] chr1 1 *
## [2] chrX 2 *
## [3] chr1 3 *
## -------
## seqinfo: 2 sequences from an unspecified genome; no seqlengths## GRanges object with 3 ranges and 0 metadata columns:
## seqnames ranges strand
## <Rle> <IRanges> <Rle>
## [1] chr1 1 *
## [2] chr1 3 *
## [3] chrX 2 *
## -------
## seqinfo: 2 sequences from an unspecified genome; no seqlengthschrSelectBSseq() performs the often useful task of
selecting one or more chromosomes and can also order the output. In case
order = TRUE, the output is ordered according to the order
of the seqnames argument.
## An object of type 'BSseq' with
## 2 methylation loci
## 3 samples
## has not been smoothed
## All assays are in-memoryOf course, in this case the function does little, since
BS.chr22 already only contains data from chromosome 22.
combine() combines two BSseq objects in the
following way: the samples for the return objects is the union of the
samples from the two objects, and the methylation loci are the union of
the two methylation loci. The two objects do not need to have measured
the same loci (in the example below, BStmp has data on
chromosome 1 and X).
## Warning in .merge_two_Seqinfo_objects(x, y): The 2 combined objects have no sequence levels in common. (Use
## suppressWarnings() to suppress this warning.)
## Warning in .merge_two_Seqinfo_objects(x, y): The 2 combined objects have no sequence levels in common. (Use
## suppressWarnings() to suppress this warning.)## GRanges object with 6 ranges and 0 metadata columns:
## seqnames ranges strand
## <Rle> <IRanges> <Rle>
## [1] chr1 1 *
## [2] chr1 3 *
## [3] chrX 2 *
## [4] chr22 14430632 *
## [5] chr22 14430677 *
## [6] chr22 14430687 *
## -------
## seqinfo: 3 sequences from an unspecified genome; no seqlengths## A1 A2 B r1 r2
## [1,] 1 4 7 0 0
## [2,] 3 6 9 0 0
## [3,] 2 5 8 0 0
## [4,] 0 0 0 18 23
## [5,] 0 0 0 11 28
## [6,] 0 0 0 10 25collapseBSseq performs the often useful task of adding
several columns of a BSseq object. This is often used in
the beginning of an analysis where each column may correspond to a lane
and several such columns represents the data for a single biological
sample.
## An object of type 'BSseq' with
## 3 methylation loci
## 2 samples
## has not been smoothed
## All assays are in-memoryCoverage, either Cov or M values, are obtained by
getCoverage(), using the type argument:
## r1 r2
## [1,] 18 23
## [2,] 11 28
## [3,] 10 25
## [4,] 8 21## r1 r2
## [1,] 17 20
## [2,] 4 20
## [3,] 6 19
## [4,] 2 4This will return a – possibly very large – matrix. It is also
possible to get region based coverage by using the regions
argument to the function. This argument is either a
data.frame (with columns chr,
start and end) or a GRanges
object. Let us do an example
regions <- GRanges(seqnames = c("chr22", "chr22"),
ranges = IRanges(start = 1.5 * 10^7 + c(0,200000),
width = 1000))
getCoverage(BS.chr22, regions = regions, what = "perRegionTotal")## r1 r2
## [1,] 30 38
## [2,] NA NAWhen what is perRegionTotal the return
value is the total coverage of each region (and note that regions
without coverage return NA). Similarly,
perRegionAverage yields the average coverage in the region.
However, it is often useful to get the actual values, like
## [[1]]
## [,1] [,2]
## [1,] 21 30
## [2,] 2 8
## [3,] 7 0
##
## [[2]]
## NULLThis is the default behaviour, and it returns a list where each
element corresponds to a region. Note that regions with no coverage gets
a NULL.
Methylation estimates can be obtained in the same way, using the
function getMeth(). If type is set to
raw the function returns simple single-loci methylation
estimates (which are M/Cov). To get smoothed estimates, the
BSseq object needs to have been smoothed using Bsmooth, and
type set to smooth (default). The
getCoverage(0, getMeth() has a
regions and a what argument. For
getMeth() the what argument can be
perBase or perRegion (the later is really the
per-region average methylation). Additionally, confidence intervals can
be computed using a method taking possibly low coverage loci into
account as well as loci where the methylation percentage might be close
to 0 or 1 (Agresti and Coull 1998).
Currently, confidence intervals cannot be computed for region-level
summary estimates. Examples
## [[1]]
## [,1] [,2]
## [1,] 0.1904762 0.1666667
## [2,] 1.0000000 0.7500000
## [3,] 0.1428571 NaN
##
## [[2]]
## NULL## [[1]]
## NULLIt is straightforward to read output files (summarized evidence) from
the BSmooth alignment suite, using read.bsmooth(). This
function takes as argument a number of directories, usually
corresponding to samples, with the alignment output. Sometimes, one
might want to read only certain chromosomes, which can be accomplished
by the seqnames argument. Also, the default values of
qualityCutoff = 20 ensures that we do not use methylation
evidence with a base quality strictly lower than 20 (since we may not be
confident whether the read really supports methylation or not). The
function read.bsmooth() allows for both gzipped and
non-gzipped input files. It is faster to read gzipped files, so we
recommend gzipping post-alignment.
During the development of BSmooth we experimented with a number of
different output formats. These legacy formats can be read with
read.umtab() and read.umtab2().
We are interested in adding additional parsers to the package; if your favorite alignment software is not supported, feel free to get in touch.
In general, we need summarized methylation evidence. In short, for each methylation loci, we need to know how many reads cover the loci and how many of those reads support methylation.
As an example, consider the Lister data. The files posted online looks like
> head mc_imr90_r1_22
assembly position strand class mc h
22 14430632 + CG 9 10
22 14430633 - CG 8 8
22 14430677 + CG 1 1
22 14430678 - CG 3 10
22 14430688 - CG 6 10
22 14430703 - CG 2 8
22 14431244 + CG 5 10
22 14431245 - CG 5 11For these files, the evidence is split by strand. It is often easiest
to read in one sample at a time, instantiate a BSseq object
and then use combine() and collapseBSseq() to
combine the samples, since these functions deal objects that have
different sets of CpGs. In the Lister data, note that the position is
the position of the “C”, so basically, if you want to combine the
evidence from both strands, CpGs on the “-” strand needs to have 1
subtracted from their position. A full script showing how these files
are used to create BS.chr22 can be found in
inst/scripts/get\_BS.chr22.R in the bsseq
package. The path of this file may be found by
Computing smoothed methylation levels is simple. We subset
`BS.chr22 so we only smooth 1 sample, for run-time
purpose
## An object of type 'BSseq' with
## 494728 methylation loci
## 1 samples
## has been smoothed with
## BSmooth (ns = 70, h = 1000, maxGap = 100000000)
## All assays are in-memoryA number of arguments pertains to using multiple cores. This is
useful if you have multiple samples or chromosomes.
mc.cores tells BSmooth() how many cores to use
and parallelBy describes whether the parallelization is
over samples or chromosomes. If the parallelization is over samples,
each core will compute the smoothing of the entire genome. And if it is
over samples, each cores will smooth all samples for one or more
chromosomes. The appropriate choice depends on the number of samples and
cpu cores available. If you have the exact same number of samples as
cores, the fastest is parallelBy = "sample". If you have
less samples than cores, the fastest is
parallelBy = "chromosome". The argument
mc.preshedule should not need to be changed (unless perhaps
if a small value of maxGap is uses); see the man page for
parallel::mclapply. Note that setting mc.cores
to a value greater than 1 is not support on MS Windows due to a
limitation of the operating system.
For a more detailed discussion of the analysis tools, read the companion vignette ``Analyzing WGBS with the bsseq package’’, which also finding DMRs and plotting them.
Fisher’s exact tests may be efficiently computed using the function
fisherTests()`.
Binomial and poisson goodness of fit tests statistics may be computed
using binomialGoodnessOfFit() and
poissonGoodnessOfFit().
## R version 4.4.2 (2024-10-31)
## Platform: x86_64-pc-linux-gnu
## Running under: Ubuntu 24.04.1 LTS
##
## Matrix products: default
## BLAS: /usr/lib/x86_64-linux-gnu/openblas-pthread/libblas.so.3
## LAPACK: /usr/lib/x86_64-linux-gnu/openblas-pthread/libopenblasp-r0.3.26.so; LAPACK version 3.12.0
##
## 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
##
## time zone: Etc/UTC
## tzcode source: system (glibc)
##
## attached base packages:
## [1] stats4 stats graphics grDevices utils datasets methods
## [8] base
##
## other attached packages:
## [1] bsseq_1.43.1 SummarizedExperiment_1.37.0
## [3] Biobase_2.67.0 MatrixGenerics_1.19.0
## [5] matrixStats_1.4.1 GenomicRanges_1.59.1
## [7] GenomeInfoDb_1.43.2 IRanges_2.41.2
## [9] S4Vectors_0.45.2 BiocGenerics_0.53.3
## [11] generics_0.1.3 BiocStyle_2.35.0
##
## loaded via a namespace (and not attached):
## [1] rjson_0.2.23 xfun_0.49
## [3] bslib_0.8.0 rhdf5_2.51.1
## [5] lattice_0.22-6 bitops_1.0-9
## [7] rhdf5filters_1.19.0 tools_4.4.2
## [9] curl_6.0.1 parallel_4.4.2
## [11] R.oo_1.27.0 Matrix_1.7-1
## [13] BSgenome_1.75.0 data.table_1.16.4
## [15] sparseMatrixStats_1.19.0 lifecycle_1.0.4
## [17] GenomeInfoDbData_1.2.13 compiler_4.4.2
## [19] Rsamtools_2.23.1 Biostrings_2.75.3
## [21] statmod_1.5.0 munsell_0.5.1
## [23] codetools_0.2-20 permute_0.9-7
## [25] htmltools_0.5.8.1 sys_3.4.3
## [27] buildtools_1.0.0 sass_0.4.9
## [29] RCurl_1.98-1.16 yaml_2.3.10
## [31] crayon_1.5.3 jquerylib_0.1.4
## [33] R.utils_2.12.3 BiocParallel_1.41.0
## [35] DelayedArray_0.33.3 cachem_1.1.0
## [37] limma_3.63.2 abind_1.4-8
## [39] gtools_3.9.5 locfit_1.5-9.10
## [41] digest_0.6.37 restfulr_0.0.15
## [43] maketools_1.3.1 fastmap_1.2.0
## [45] grid_4.4.2 colorspace_2.1-1
## [47] cli_3.6.3 SparseArray_1.7.2
## [49] S4Arrays_1.7.1 XML_3.99-0.17
## [51] DelayedMatrixStats_1.29.0 UCSC.utils_1.3.0
## [53] scales_1.3.0 rmarkdown_2.29
## [55] XVector_0.47.0 httr_1.4.7
## [57] R.methodsS3_1.8.2 beachmat_2.23.5
## [59] HDF5Array_1.35.2 evaluate_1.0.1
## [61] knitr_1.49 BiocIO_1.17.1
## [63] rtracklayer_1.67.0 rlang_1.1.4
## [65] Rcpp_1.0.13-1 glue_1.8.0
## [67] BiocManager_1.30.25 jsonlite_1.8.9
## [69] R6_2.5.1 Rhdf5lib_1.29.0
## [71] GenomicAlignments_1.43.0 zlibbioc_1.52.0