R/BMfinder.R

In vasp: Quantification and Visualization of Variations of Splicing in Population

#' Discover bimodal distrubition features
#'
#' Find bimodal distrubition features and
#' divide the samples into 2 groups by k-means clustering.
#'
#' @param x a numeric matrix with feature rows and sample columns, e.g.,
#' splicing score matrix from \code{\link{spliceGenome}} or
#' \code{\link{spliceGene}} function.
#' @param p.value p.value threshold for bimodal distrubition test
#' @param maf minor allele frequency threshold in k-means clustering
#' @param miss missing grouping rate threshold in k-means clustering
#' @param fold fold change threshold between the two groups
#' @param log whether the scores are to be logarithmic. If TRUE, all the scores
#' are log2 tranformed before k-means clustering: x = log2(x+1).
#' @param cores threads to be used. This value is passed
#'  to \bold{?mclapply} in \bold{parallel} package
#' @return a matrix with feature rows and sample columns.
#' @details The matrix contains 1, 2 and NA, and
#' values of 'x' in group 2 are larger than group 1.
#' @importFrom parallel mclapply detectCores
#' @importFrom stats qnorm sd
#' @export BMfinder
#' @examples
#' data(rice.bg)
#' score<-spliceGene(rice.bg,'MSTRG.183',junc.type='score')
#' score<-round(score,2)
#' as<-BMfinder(score,cores=1) # 4 bimodal distrubition features found
#'
#' ##compare
#' as
#' score[rownames(score)%in%rownames(as),]

BMfinder <- function(x, p.value = 0.01, maf = 0.05, miss = 0.05, fold = 2, 
                    log = FALSE, cores = detectCores() - 1) {
    if (!methods::is(x, "matrix") || !is.numeric(x)) {
        stop("x must be a numeric matrix!")
    }
    qnorm.value <- qnorm(1 - p.value)
    ind <- colnames(x)
    if (log) 
        x <- log2(x + 1)
    if (ncol(x) < 30) {
        warning("sample size < 30, the result should be further tested!")
    }
    cat("---BMfinder:", nrow(x), "features *", ncol(x), "samples\n")
    res <- mclapply(seq_len(nrow(x)), function(i) {
        R <- x[i, ]
        if (all(R == R[1], na.rm = TRUE)) 
            return(rep(NA, length(ind)))
        na.id <- which(is.na(R))
        if (length(na.id) > 0) 
            R[na.id] <- mean(R, na.rm = TRUE)
        R <- sort(R)
        Rcluster <- cluster::pam(R, 2, cluster.only = TRUE)
        d <- split(R, Rcluster)
        d1 <- d[[1]]
        d2 <- d[[2]]
        m1 <- mean(d1)
        m2 <- mean(d2)
        s1 <- sd(d1)
        s2 <- sd(d2)
        z1 <- (d1 - m2)/s2
        z2 <- (d2 - m1)/s1
        Rcluster[names(d1[abs(z1) <= qnorm.value])] <- NA
        Rcluster[names(d2[abs(z2) <= qnorm.value])] <- NA
        Rcluster <- Rcluster[ind]
        if (length(na.id) > 0) 
            Rcluster[na.id] <- NA
        tab <- table(Rcluster)
        if (length(tab) < 2 | min(tab)/sum(tab) < maf) 
            return(rep(NA, length(ind)))
        fc <- mean(x[i, Rcluster == 2], na.rm = TRUE)/mean(x[i, Rcluster == 
            1], na.rm = TRUE)
        if (fc >= fold) {
            return(Rcluster)
        } else {
            return(rep(NA, length(ind)))
        }
    }, mc.cores = cores)
    res <- do.call("rbind", res)
    dimnames(res) <- dimnames(x)
    res <- subset(res, rowSums(is.na(res)) <= length(ind) * miss)
    cat("---Completed:  a total of", nrow(res), 
        "bimodal distrubition features found!", "\n")
    res
}