R/fragmentLenDetect.R

In nucleR: Nucleosome positioning package for R

#' Fragments length detection from single-end sequencing samples
#'
#' When using single-ended sequencing, the resulting partial sequences map only
#' in one strand, causing a bias in the coverage profile if not corrected. The
#' only way to correct this is knowing the average size of the real fragments.
#' `nucleR` uses this information when preprocessing single-ended sequences.
#' You can provide this information by your own (usually a 147bp length is a
#' good aproximation) or you can use this method to automatically guess the
#' size of the inserts.
#'
#' This function shifts one strand downstream one base by one from `min.shift`
#' to `max.shift`. In every step, the correlation on a random position of
#' length `window` is checked between both strands. The maximum correlation is
#' returned and averaged for `samples` repetitions.
#'
#' The final returned length is the best shift detected plus the width of the
#' reads. You can increase the performance of this function by reducing the
#' `samples` value and/or narrowing the shift range. The `window` size has
#' almost no impact on the performance, despite a to small value can give
#' biased results.
#'
#' @param reads Raw single-end reads [ShortRead::AlignedRead] or
#'   [GenomicRanges::GRanges] format)
#' @param samples Number of samples to perform the analysis (more = slower but
#'   more accurate)
#' @param window Analysis window. Usually there's no need to touch this
#'   parameter.
#' @param min.shift,max.shift Minimum and maximum shift to apply on the strands
#'   to detect the optimal fragment size. If the range is too big, the
#'   performance decreases.
#' @param as.shift If TRUE, returns the shift needed to align the middle of the
#'   reads in opposite strand. If FALSE, returns the mean inferred fragment
#'   length.
#' @param mc.cores If multicore support, maximum number of cores allowed to
#'   use.
#'
#' @return Inferred mean lenght of the inserts by default, or shift needed to
#'   align strands if `as.shift=TRUE`.
#'
#' @author Oscar Flores \email{oflores@@mmb.pcb.ub.es}
#' @keywords attribute
#' @rdname fragmentLenDetect
#'
#' @examples
#' library(GenomicRanges)
#' library(IRanges)
#'
#' # Create a sinthetic dataset, simulating single-end reads, for positive and
#' # negative strands
#' # Positive strand reads
#' pos <- syntheticNucMap(nuc.len=40, lin.len=130)$syn.reads
#' # Negative strand (shifted 147bp)
#' neg <- IRanges(end=start(pos)+147, width=40)
#' sim <- GRanges(
#'     seqnames="chr1",
#'     ranges=c(pos, neg),
#'     strand=c(rep("+", length(pos)), rep("-", length(neg)))
#' )
#'
#' # Detect fragment lenght (we know by construction it is really 147)
#' fragmentLenDetect(sim, samples=50)
#' # The function restricts the sampling to speed up the example
#'
#' @export
#'
setGeneric(
    "fragmentLenDetect",
    function(reads, samples=1000, window=5000, min.shift=1, max.shift=100,
            mc.cores=1, as.shift=FALSE)
        standardGeneric("fragmentLenDetect")
)

#' @rdname fragmentLenDetect
#' @importFrom stats runif cor
#' @importMethodsFrom BiocGenerics strand width
#' @importMethodsFrom IRanges coverage
#' @importMethodsFrom ShortRead chromosome
setMethod(
    "fragmentLenDetect",
    signature(reads="AlignedRead"),
    function (reads, samples=1000, window=1000, min.shift=1, max.shift=100,
              mc.cores=1, as.shift=FALSE)
    {

        # Randomly select regions in the available chromosome bounds
        chrSample <- as.character(sample(chromosome(reads), samples))
        chrsLen <- sapply(
            levels(chromosome(reads)),
            function(x)
                max(position(reads[chromosome(reads) == x]))
        )
        position <- round(runif(samples, max=chrsLen[chrSample] - window))
        dd <- data.frame(chrSample, position)

        # For each sampled region, look for the shift with higher correlation
        # between strands
        shiftPos <- function (i) {
            chr <- as.character(dd[i, "chrSample"])
            sta <- as.numeric(dd[i, "position"])
            end <- sta + window
            rea <- reads[
                chromosome(reads) == chr &
                position(reads) > sta &
                position(reads) < end
            ]
            if (length(rea) == 0) {
                return (numeric(0))  # Discard uncovered regions
            }

            cpos <- try(
                as.vector(
                    coverage(rea[as.character(strand(rea)) == "+"])[[1]]
                )[sta:end],
                silent=TRUE
            )
            cneg <- try(
                as.vector(
                    coverage(rea[as.character(strand(rea)) == "-"])[[1]]
                )[sta:end],
                silent=TRUE
            )

            # Avoid unpaired strands (ie, all strands being "+")
            if (class(cpos) == "try-error" | class(cneg) == "try-error") {
                return(numeric(0))
            }

            cpos[is.na(cpos)] <- 0
            cneg[is.na(cneg)] <- 0

            x <- sapply(
                min.shift:max.shift,
                function(s)
                    cor(cpos[1:(window + 1 - s)], cneg[s:window])
            )
            res <- which(x == max(x))[1]

            # We only shifted the negative strand, but in real case both
            # strands will be shifted the half of this amount
            res <- res / 2

            #Discard NAs
            if (is.na(res) | !is.numeric(res)) {
                return(numeric(0))
            }

            return(res + min.shift - 1)
        }

        shift <- round(mean(unlist(.xlapply(
            1:nrow(dd),
            shiftPos,
            mc.cores=mc.cores
        ))))

        kk <- lapply(1:nrow(dd), shiftPos)

        #Fragment length is the shift * 2 + the length of the read
        fragLen <- shift * 2 + width(reads)[1]

        if (as.shift) {
            return(shift)
        } else {
            return(fragLen)
        }

    }
)

#' @rdname fragmentLenDetect
#' @importFrom stats runif cor
#' @importMethodsFrom BiocGenerics strand width
#' @importMethodsFrom GenomeInfoDb seqnames
setMethod(
    "fragmentLenDetect",
    signature(reads="GRanges"),
    function (reads, samples=1000, window=1000, min.shift=1, max.shift=100,
              mc.cores=1, as.shift=FALSE)
    {

        # Calculate the whole coverage here saves cpu and memory later for big
        # genomes. This improves a lot the performance on big genomes if the
        # sampling value is big. For small datasets could be less efficient
        # than obvious approach, but it is worth it
        .doCover <- function (x)
            lapply(
                list(pos="+", neg="-"),
                function (s)
                    coverage(x[strand(x) == s, ])[[1]]
            )

        chrs <- levels(seqnames(reads))
        splt <- split(reads, chrs)
        cover <- lapply(splt, .doCover)

        # Randomly select regions in the available chromosome bounds
        count <- sapply(splt, length)
        probs <- count / sum(count)
        chrSample <- sample(chrs, samples, replace=TRUE, prob=probs)
        chrLength <- sapply(
            cover,
            function (x)
                min(length(x[["pos"]]), length(x[["neg"]]))
        )
        position <- round(runif(samples, max=chrLength[chrSample] - window))
        dd <- data.frame(chrSample, position)

        # For each sampled region, look for the shift with higher correlation
        # between strands
        shiftPos <- function (i) {
            chr <- as.character(dd[i, "chrSample"])
            sta <- as.numeric(dd[i, "position"])
            end <- sta + window

            cpos <- try(as.vector(cover[[chr]][["pos"]][sta:end]), silent=TRUE)
            cneg <- try(as.vector(cover[[chr]][["neg"]][sta:end]), silent=TRUE)

            if (class(cpos) == "try-error" | class(cneg) == "try-error") {
                return(NA)
            }
            if (sum(cpos) == 0 | sum(cneg) == 0) {
                return(NA)
            }

            cpos[is.na(cpos)] <- 0
            cneg[is.na(cneg)] <- 0

            x <- sapply(
                min.shift:max.shift,
                function (s)
                    cor(cpos[1:(window + 1 - s)], cneg[s:window])
            )
            res <- which(x == max(x, na.rm=TRUE))[1]

            # We only shifted the negative strand, but both strands will be
            # shifted the half of this amount
            res <- res / 2

            # Discard NAs
            if (is.na(res) | !is.numeric(res)) {
                return(numeric(0))
            }

            return(res + min.shift - 1)
        }

        shift <- round(mean(unlist(.xlapply(
            1:nrow(dd),
            shiftPos,
            mc.cores=mc.cores
        )), na.rm=TRUE))

        # Fragment length is the shift * 2 + the length of the read
        fragLen <- shift * 2 + width(reads)[1]

        if (as.shift) {
            return (shift)
        } else {
            return (fragLen)
        }
    }
)