R/ClustMIC.R

In ManonMartin/MBXUCL: R Routines for metabolomic data

### CLUSTERING ##############################################

## Ward et K-means unsupervised clustering MIC criteria: Dunn, Davies-Bouldin,
## Rand and Adjusted-Rand


#' @export ClustMIC
#' @title Unsupervised clustering on (GPL) intensities and associated MIC indexes
#'
#' @description
#' Unsupervised clustering on (GPL) intensities based on Ward and K-Means algorithms.
#' Calculation of MIC statistical criteria of clustering quality:
#' Dunn, Davies-Bouldin, Rand and adjusted-Rand indexes.
#'
#' @param Intensities A numeric matrix of intensities (can be a pTreatGPL object).
#' @param nClust The number of groups to retrieve (donors, mixtures, ...).
#' @param Trcl The real groups' memberships of the samples, true class.
#' @param Dendr Logical argument (TRUE/FALSE) to obtain graphical dendrogram based on the Ward algorithm.
#'
#' @return A list of MIC quality indexes (Dunn, Davies-Bouldin, Rand and adjusted-Rand):
#' \describe{
#'   \item{\code{DunnW}}{Dunn index for Ward clustering}
#'   \item{\code{DunnKM}}{Dunn index for Kmeans clustering}
#'   \item{\code{DBW}}{Davies-Bouldin index for Ward clustering}
#'   \item{\code{DBKM}}{Davies-Bouldin index for Kmeans clustering}
#'   \item{\code{RandW}}{Rand index for Ward clustering}
#'   \item{\code{RandKM}}{Rand index for Kmeans clustering}
#'   \item{\code{AdjRandW}}{Adjusted Rand index for Ward clustering}
#'   \item{\code{AdjRandKM}}{Adjusted Rand index for Kmeans clustering}
#'    }
#' @author Baptiste Feraud, Manon Martin
#'
#' @examples
#' data('HumanSerum')
#' ClustMIC(Intensities = HumanSerumSpectra, nClust = 4, Trcl = ClassHS, Dendr = TRUE)
#'
#'@importFrom proxy dist
#'@importFrom clValid dunn
#'@importFrom phyclust RRand


ClustMIC <- function(Intensities, nClust, Trcl, Dendr = TRUE){

  # checks

  if (missing(Intensities)) {
    warning("Intensities is missing with no default value")
  }

  if (missing(nClust)) {
    warning("nClust is missing with no default value")
  }


  if (sum(Trcl <= 0) > 0) {
    warning("Automatic rewritting of Trcl since values below 1 are not permitted")
    Trcl <- Trcl + (1 - min(Trcl))
  }

  if (!is.numeric(Intensities)) {
    stop(deparse(substitute(Intensities)), " is not numeric.")
  }

  if (!is.numeric(nClust)) {
    stop(deparse(substitute(nClust)), " is not numeric.")
  } else if (length(nClust) > 1) {
    stop(deparse(substitute(nClust)), " has a length > 1.")
  }

  if (missing(Trcl) | sum(Trcl%%1 != 0) > 0) {
    warning("Trcl is missing with no default value or is not an integer")
  }


  names(Trcl) <- 1:length(Trcl)

  if (!is.logical(Dendr)) {
    stop(deparse(substitute(Dendr)), " is not logical.")
  }


  #### WARD
  DistI <- proxy::dist(Intensities, method = "euclidean", diag = FALSE, upper = FALSE)

  WardI <- stats::hclust(DistI, "ward.D2")
  clWI <- stats::cutree(WardI, nClust)


  #### K-MEANS
  clKMI <- stats::kmeans(Intensities, nClust, iter.max = 20, nstart = 10, algorithm = c("Hartigan-Wong",
    "Lloyd", "Forgy", "MacQueen"))


  #### INDICES MIC ##############################

  ################################ Homogeneite des groupes ##

  #### DUNN
  DunnW <- clValid::dunn(DistI, clWI)
  DunnKM <- clValid::dunn(DistI, clKMI$cluster)


  #### DAVIES-BOULDIN
  DBW <- index.DB(Intensities, clWI, DistI, centrotypes = "medoids")
  DBKM <- index.DB(Intensities, clKMI$cluster, DistI, centrotypes = "medoids")


  ################################ Qualite du clustering ##

  #### RAND
  RandW <- phyclust::RRand(Trcl, clWI, lab = NULL)
  RandKM <- phyclust::RRand(Trcl, clKMI$cluster, lab = NULL)


  #### DENDROGRAM ################################

  if (Dendr == TRUE) {
    graphics::plot(WardI, lwd = 1, cex = 0.5)
    stats::rect.hclust(WardI, nClust, border = "red")
  }

  res <- list(DunnW = DunnW, DunnKM = DunnKM, DBW = DBW$DB, DBKM = DBKM$DB, RandW = RandW[[1]],
    RandKM = RandKM[[1]], AdjRandW = RandW[[2]], AdjRandKM = RandKM[[2]])


  table <- data.frame(res)
  return(table)

}