In lamortenera/bamsignals: Extract read count signals from bam files

BiocStyle::markdown()

Introduction to the bamsignals package

The goal of the bamsignals package is to load count data from bam files as easily and quickly as possible. A typical workflow without the bamsignals package requires to firstly load all reads in R (e.g. using the Rsamtools package), secondly process them and lastly convert them into counts. The bamsignals package optimizes this workflow by merging these steps into one using efficient C code, which makes the whole process easier and faster. Additionally, bamsignals comes with native support for paired end data.

Loading toy data

We will use the following libraries (which are all required for installing bamsignals).

library(GenomicRanges)
library(Rsamtools)
library(bamsignals)

In the following we will use a sorted and indexed bam file and a gene annotation.

bampath <- system.file("extdata", "randomBam.bam", package="bamsignals")
genes <- 
  get(load(system.file("extdata", "randomAnnot.Rdata", package="bamsignals")))
genes

The chromosome names in the bam file and those in the GenomicRanges object need to match. Additionally, the bam file needs to be sorted and indexed. Note that bamsignals requires the bam index to be named like bam file with ".bai" suffix.

#sequence names of the GenomicRanges object
seqinfo(genes)
#sequence names in the bam file
bf <- Rsamtools::BamFile(bampath)
seqinfo(bf)
#checking if there is an index
file.exists(gsub(".bam$", ".bam.bai", bampath))

Counting reads in given ranges with bamCount()

Basic counting

Let's count how many reads map to the promoter regions of our genes. Using the bamCount() function, this is straightforward.

proms <- GenomicRanges::promoters(genes, upstream=100, downstream=100)
counts <- bamCount(bampath, proms, verbose=FALSE)
str(counts)

The object counts is a vector of the same length as the number of ranges that we are analyzing, the i-th count corresponds to the i-th range.

Accounting for fragment length

With the bamCount() function a read is counted in a range if the 5' end of the read falls in that range. This might be appropriate when analyzing DNase I hypersensitivity tags, however for ChIP-seq data the immunoprecipitated protein is normally located downstream with respect to the 5' end of the sequenced reads. To correct for that, it is possible to count reads with a strand-specific shift, i.e. reads will be counted in a range if the shifted 5' end falls in that range. Note that this shift will move reads mapped to the positive strand to the right and reads mapped to the negative strand to the left with respect to the reference orientation. The shift should correspond approximately to half of the average length of the fragments in the sequencing experiment.

counts <- bamCount(bampath, proms, verbose=FALSE, shift=75)
str(counts)

Counting on each strand separately

Sometimes it is necessary to consider the two genomic strands separately. This is achieved with the ss option (separate strands, or strand-specific), and depends also on the strand of the GenomicRanges object.

strand(proms)
counts <- bamCount(bampath, proms, verbose=FALSE, ss=TRUE)
str(counts)

Now counts is a matrix with two rows, one for the sense strand, the other for the antisense strand. Note that the sense of a read is decided also by the region it falls into, so if both the region and the read are on the same strand the read is counted as a sense read, otherwise as an antisense read.

Read profiles for each region with bamProfile()

If you are interested in counting how many reads map to each base pair of your genes, the bamProfile() function might save you a day.

sigs <- bamProfile(bampath, genes, verbose=FALSE)
sigs

The CountSignals class is a read-only container for count vectors. Conceptually it is like a list of vectors, and in fact it can be immediately converted to that format.

#CountSignals is conceptually like a list
lsigs <- as.list(sigs)
stopifnot(length(lsigs[[1]])==length(sigs[1]))
#sapply and lapply can be used as if we were using a list
stopifnot(all(sapply(sigs, sum) == sapply(lsigs, sum)))

Similarly as for the bamCount function, the CountSignals object has as many elements (called signals) as there are ranges, and the i-th signal corresponds to the i-th range.

stopifnot(all(width(sigs)==width(genes)))

Counting on each strand separately

As for the bamCount() function, also with bamProfile() the reads can be counted for each strand separately

sssigs <- bamProfile(bampath, genes, verbose=FALSE, ss=TRUE)
sssigs

Now each signal is a matrix with two rows.

str(sssigs[1])
#summing up the counts from the two strands is the same as using ss=FALSE
stopifnot(colSums(sssigs[1])==sigs[1])
#the width function takes into account that now the signals are strand-specific
stopifnot(width(sssigs)==width(sigs))
#the length function does not, a strand-specific signal is twice as long
stopifnot(length(sssigs[1])==2*length(sigs[1]))

Let's summarize this with a plot

xlab <- "offset from start of the region"
ylab <- "counts per base pair (negative means antisense)"
main <- paste0("read profile of the region ", 
  seqnames(genes)[1], ":", start(genes)[1], "-", end(genes)[1])
plot(sigs[1], ylim=c(-max(sigs[1]), max(sigs[1])), ylab=ylab, xlab=xlab, 
  main=main, type="l")
lines(sssigs[1]["sense",], col="blue")
lines(-sssigs[1]["antisense",], col="red")
legend("bottom", c("sense", "antisense", "both"), 
  col=c("blue", "red", "black"), lty=1)

Regions of the same width

In case our ranges have all the same width, a CountSignals object can be immediately converted into a matrix, or an array, with the alignSignals function

#The promoter regions have all the same width
sigs <- bamProfile(bampath, proms, ss=FALSE, verbose=FALSE)
sssigs <- bamProfile(bampath, proms, ss=TRUE, verbose=FALSE)

sigsMat <- alignSignals(sigs)
sigsArr <- alignSignals(sssigs)

The last dimension of the resulting array (or matrix) represents the different ranges, the second-last one represents the base pairs in each region, and in the strand-specific case, the first-one represents the strand of the signal. This can be changed by using the t() function (for matrices) or aperm() (for arrays).

#the dimensions are [base pair, region]
str(sigsMat)
#the dimensions are [strand, base pair, region]
str(sigsArr)

stopifnot(all(sigsMat == sigsArr["sense",,] + sigsArr["antisense",,]))

Computing the average read profile at promoters in proms is now straightforward

avgSig <- rowMeans(sigsMat)
avgSenseSig <- rowMeans(sigsArr["sense",,])
avgAntisenseSig <- rowMeans(sigsArr["antisense",,])
ylab <- "average counts per base pair"
xlab <- "distance from TSS"
main <- paste0("average profile of ", length(proms), " promoters")
xs <- -99:100
plot(xs, avgSig, ylim=c(0, max(avgSig)), xlab=xlab, ylab=ylab, main=main,
  type="l")
lines(xs, avgSenseSig, col="blue")
lines(xs, avgAntisenseSig, col="red")
legend("bottom", c("sense", "antisense", "both"), 
  col=c("blue", "red", "black"), lty=1)

Binning counts

Very often it is better to count reads mapping to small regions instead of single base pairs. Bins are small non-overlapping regions of fixed size tiling a larger region. Instead of splitting your regions of interest into bins, it is easier and much more efficient to provide the binsize option to bamProfile().

binsize <- 20
binnedSigs <- bamProfile(bampath, proms, binsize=binsize, verbose=FALSE)
stopifnot(all(width(binnedSigs)==ceiling(width(sigs)/binsize)))
binnedSigs

In case the ranges' widths are not multiples of the bin size, a warning will be issued and the last bin in those ranges will be smaller than the others (where "last" depends on the orientation of the region).

Binning means considering a signal at a lower resolution.

avgBinnedSig <- rowMeans(alignSignals(binnedSigs))
#the counts in the bin are the sum of the counts in each base pair
stopifnot(all.equal(colSums(matrix(avgSig, nrow=binsize)),avgBinnedSig))
#let's plot it
ylab <- "average counts per base pair"
plot(xs, avgSig, xlab=xlab, ylab=ylab, main=main, type="l")
lines(xs, rep(avgBinnedSig, each=binsize)/binsize, lty=2)
legend("topright", c("base pair count", "bin count"), lty=c(1, 2))

Read coverage with bamCoverage()

Instead of counting the 5' end of each read, you may want to count how many reads overlap each base pair, you should check out the bamCoverage() function which gives you a smooth read coverage profile by considering the whole read length and not just the 5' end:

covSigs <- bamCoverage(bampath, genes, verbose=FALSE)
puSigs <- bamProfile(bampath, genes, verbose=FALSE)
xlab <- "offset from start of the region"
ylab <- "reads per base pair"
main <- paste0("read coverage and profile of the region ", seqnames(genes)[1],
  ":", start(genes)[1], "-", end(genes)[1])
plot(covSigs[1], ylim=c(0, max(covSigs[1])), ylab=ylab, xlab=xlab, main=main,
  type="l")
lines(puSigs[1], lty=2)
legend("topright", c("covering the base pair", "5' end maps to the base pair"), 
  lty=c(1,2))

Advanced bamsignals filtering options

Exclude ambiguous reads with the mapq argument

Most mapping software (e.g. bwa, bowtie2) stores information about mapping confidence in the MAPQ field of a bam file. In general, it is recommended to exclude reads with bad mapping quality because their mapping location is ambiguous. In bowtie2, a mapping quality of 20 or less indicates that there is at least a 1 in 20 chance that the read truly originated elsewhere. In that case, the mapq argument is a lower bound on MAPQ:

counts.all <- bamCount(bampath, proms, verbose=FALSE)
summary(counts.all)
counts.mapq <- bamCount(bampath, proms, mapq=20, verbose=FALSE)
summary(counts.mapq)

Filter reads with the filteredFlag argument

Analogously to the MAPQ field, every bam contains a SAMFLAG field where the mapping software or the post-processing software (e.g. picard) stores information on the read. See Decoding SAM flags for explanation. For instance, a SAMFLAG of 1024 indicates a optical duplicate. We would like to filter out optical duplicate reads with filteredFlag=1024 from the read counts with MAPQ >= 19 to get a higher confidence on the results:

counts.mapq.noDups <- bamCount(bampath, proms, mapq=20, filteredFlag=1024, verbose=FALSE)
summary(counts.mapq.noDups)

Paired End Data

Paired end data handling with the paired.end argument

All bamsignals methods (bamCount(), bamProfile() and bamCoverage()) discussed above support dealing with paired end sequencing data. Considering only one read avoids counting both reads in read pair which may bias downstream analysis. The argument paired.end can be set to ignore (treat like single end), filter (consider 5'-end of first read in a properly aligned pair, i.e. SAMFLAG=66) or midpoint (consider the midpoint of an aligned fragment). Please note, that the strand of the first read in a pair defines the strand of fragment.

#5' end falls into regions defined in `proms`
counts.pe.filter <- bamCount(bampath, proms, paired.end="filter", verbose=FALSE)
summary(counts.pe.filter)

#fragment midpoint falls into regions defined in `proms`
counts.pe.midpoint <- bamCount(bampath, proms, paired.end="midpoint", verbose=FALSE)
summary(counts.pe.midpoint)

#counts are similar but not identical
cor(counts.pe.filter, counts.pe.midpoint)

For bamCoverage(), paired.end=="midpoint" is not defined. However, paired.end=="extend" computes "how many fragments cover each base pair" (as opposed to "how many reads cover each base pair" in the single end case). This is done by utilizing the actual length of a fragment is stored in the TLEN field of the paired end bam file. The result is a very smooth coverage plot:

covSigs <- bamCoverage(bampath, genes, paired.end="extend", verbose=FALSE)
puSigs <- bamProfile(bampath, genes, paired.end="midpoint", verbose=FALSE)
xlab <- "offset from start of the region"
ylab <- "reads per base pair"
main <- paste0("Paired end whole fragment coverage and fragment midpoint profile\n", 
  "of the region ", seqnames(genes)[1], ":", start(genes)[1], "-",
  end(genes)[1])
plot(covSigs[1], ylim=c(0, max(covSigs[1])), ylab=ylab, xlab=xlab, main=main,
  type="l")
lines(puSigs[1], lty=2)
legend("topright", c("covering the base pair", "fragment midpoint maps to the base pair"), 
  lty=c(1,2))

Filtering fragments with the tlenFilter argument

In paired end data, the actual fragment length can be inferred from the distance between two read mates. This information is then stored in the TLEN field of a bam file. One might need to filter for fragments within a certain "allowed" size, e.g. mono-nucleosomal fragments in ChIP-seq.

counts.monoNucl <- bamCount(bampath, genes, paired.end="midpoint", tlenFilter=c(120,170), verbose=FALSE)
summary(counts.monoNucl)

#Coverage of mononucleosomal fragments
covSigs.monoNucl <- bamCoverage(bampath, genes, paired.end="extend", tlenFilter=c(120,170), verbose=FALSE)
xlab <- "offset from start of the region"
ylab <- "reads per base pair"
main <- paste0("Paired end whole fragment coverage for\n", 
  "of the region ", seqnames(genes)[1], ":", start(genes)[1], "-",
  end(genes)[1])
plot(covSigs[1], ylim=c(0, max(covSigs[1])), ylab=ylab, xlab=xlab, main=main,
  type="l")
lines(covSigs.monoNucl[1], lty=3)
legend("topright", c("all fragment sizes", "mononucleosomal fragments only"), 
  lty=c(1,3))

There are many more use cases for tlenFilter, e.g. count only long range reads in ChIA-PET or HiC data or profile only very small fragments in ChIP-exo/nexus data.

lamortenera/bamsignals documentation built on Oct. 9, 2021, 10:03 p.m.