R/dot-sparsify.R

In stefpeschel/NetCoMi: Network Construction and Comparison for Microbiome Data

#' @keywords internal
.sparsify <- function(assoMat,
                      countMat,
                      sampleSize,
                      measure,
                      measurePar,
                      assoType,
                      sparsMethod,
                      thresh,
                      alpha,
                      adjust,
                      lfdrThresh,
                      trueNullMethod,
                      nboot,
                      assoBoot = NULL,
                      softThreshType,
                      softThreshPower,
                      softThreshCut,
                      kNeighbor,
                      knnMutual,
                      cores,
                      verbose,
                      logFile = NULL,
                      seed = NULL) {
  
  if (sparsMethod == "threshold") {
    
    assoNew <- assoMat
    
    if (assoType == "dissimilarity") {
      if (thresh < 0) warning("'thresh' smaller than 0. No edges selected.")
      assoNew[assoNew > thresh] <- Inf
      
    } else {
      if (thresh > 1) warning("'thresh' not in [0,1]. No edges selected.")
      if (thresh < 0) warning("'thresh' not in [0,1]. All edges selected.")
      assoNew[abs(assoNew) < thresh] <- 0
    }
    
  } else if (sparsMethod %in% c("t-test", "bootstrap")) {   #hypothesis testing
    
    assoVec <- assoMat[lower.tri(assoMat)]
    
    if (sparsMethod == "t-test") {
      
      if (is.null(countMat)) {
        df <- sampleSize - 2
      } else {
        df <- nrow(countMat) - 2
      }
      
      tstat <- (assoVec * sqrt(df)) / sqrt(1 - assoVec^2)
      pvals <- 2 * (1 - stats::pt(abs(tstat), df))
      pvals[pvals] <- 1
      
    } else {   #bootstrap procedure
      
      if (is.null(countMat)) {
        stop("Count matrix needed for sparsification via bootstrapping.")
      }
      
      bootOut <- .boottest(countMat = countMat,
                           assoMat = assoMat,
                           nboot = nboot,
                           measure = measure,
                           measurePar = measurePar,
                           cores = cores,
                           logFile = logFile,
                           verbose = verbose,
                           seed = seed,
                           assoBoot = assoBoot)
      
      pvalsBoot <- bootOut$pvals
    
      pvals <- pvalsBoot[lower.tri(pvalsBoot)]
      
      assoBoot <- bootOut$assoBoot
    }
    
    # adjust for multiple testing and identify links
    if (adjust == "none") {
      assoVec[pvals > alpha] <- 0
      
    } else {
      
      if (verbose %in% 2:3) {
        message("\nAdjust for multiple testing via '", adjust, "' ... ",
                appendLF = FALSE)
      }
      
      verb.tmp <- ifelse(verbose == 3, TRUE, FALSE)
      pvals.adj <- multAdjust(pvals = pvals, adjust = adjust,
                              trueNullMethod = trueNullMethod,
                              verbose = verb.tmp)
      
      if (verbose %in% 2:3) {
        message("Done.")
      }
      
      assoVec[pvals.adj > alpha] <- 0
    }
    
    assoNew <- assoMat
    assoNew[lower.tri(assoNew)] <- assoVec
    assoNew[upper.tri(assoNew)] <- t(assoNew)[upper.tri(t(assoNew))]
    
  } else if (sparsMethod == "softThreshold") {
    
    if (softThreshType == "signed") {
      simMat <- 0.5 + 0.5 * assoMat
    } else if (softThreshType == "unsigned") {
      simMat <- abs(assoMat)
    } else {
      simMat <- assoMat
      simMat[simMat < 0] <- 0
    }
    
    # power for soft thresholding
    if (is.null(softThreshPower)) {
      
      diag(assoMat) <- 0
      
      verboseTmp <- ifelse(verbose == 3, 5, 2)
      
      params <- list(simMat,
                     dataIsExpr = FALSE,
                     networkType = softThreshType,
                     RsquaredCut = softThreshCut,
                     verbose = verboseTmp,
                     corOptions = list(use = 'p',
                                       nThreads = 1))
      
      if (verbose == 3) {
        softthresh1 <- .suppress_warnings(do.call(WGCNA::pickSoftThreshold, 
                                                  params), 
                           startsWith, "executing")
        
      } else {
        invisible(capture.output(
          softthresh1 <- .suppress_warnings(do.call(WGCNA::pickSoftThreshold, 
                                                    params), 
                                            startsWith, "executing")
        ))
      }
      
      power <-  softthresh1$powerEstimate
      
      if (is.na(power)) {
        message(paste0("\n No soft-thresholding power with R^2 above ", 
                       softThreshCut, ". Power set to 1."))
        power <- 1
      }
      
      if (verbose >= 2) {
        message("\n Estimated soft-thresholding power: ", power)
      }
      
      
    } else {
      power <- softThreshPower
    }
    
    assoNew <- simMat^power
    
  } else if (sparsMethod == "knn") {
    
    knngraph <- cccd::nng(dx = assoMat, k = kNeighbor, mutual = knnMutual)
    knnMat <- as.matrix(igraph::as_adjacency_matrix(knngraph))
    
    if (!knnMutual) {
      knnMat.tmp <- knnMat
      for (i in 1:nrow(knnMat)) {
        knnMat[i, ] <- knnMat.tmp[i, ] | knnMat.tmp[, i]
        knnMat[, i] <- knnMat[i, ]
      }
    }
    
    assoNew <- assoMat
    assoNew[knnMat == 0] <- Inf
    diag(assoNew) <- 0
    
  } else {
    assoNew <- assoMat
  }
  
  #-----------------------------------
  if (sparsMethod != "softThreshold") {
    power <- NULL
    simMat <- NULL
  }
  
  output <- list(assoNew = assoNew,
                 power = power,
                 simMat = simMat,
                 assoBoot = assoBoot)
  
  return(output)
}