R/clustering-kmeans.R

In pRoloc: A unifying bioinformatics framework for spatial proteomics

.optim.kmeans <- function(k, X, 
                          criterion = c("BIC", "AIC"),
                          ...) {
  ## value must a named vector
  criterion <- match.arg(criterion, several.ok = TRUE) 
  .var <- function(x) 
    sum(x^2 - mean(x)^2)/length(x)
  
  N <- nrow(X)
  P <- ncol(X)
  cl <- kmeans(X, centers = k, ...)
  Y <- cl$cluster
  Nc <- cl$size
  K <- length(Nc)

  ## per cluster variance - using N instead of N-1 for clusters of size 1
  Vc <- matrix(0, nrow = P, ncol = K)
  rownames(Vc) <- colnames(X)
  colnames(Vc) <- 1:K
  for (k in 1:K) {
    for (p in 1:P) {
      Vc[p, k] <- .var(X[Y == k, p]) 
    }
  }
  ## total variance
  V <- matrix(apply(X, 2, var), ncol = K, nrow = P)
  LL <- -Nc * colSums(log(Vc + V)/2) 
  .BIC <- -2 * sum(LL) + 2 * K * P * log(N)
  .AIC <- -2 * sum(LL) + 4 * K * P
  ans <- c(BIC = .BIC, AIC = .AIC)
  if (length(criterion) == 1) 
    ans <- switch(criterion,
                  BIC = ans["BIC"],
                  AIC = ans["AIC"])
  return(ans)
}

setMethod("kmeansClustering",
          signature(object = "MSnSet",
                    params = "missing"),
          function(object, ...) {
            cl <- kmeans(exprs(object), ...)
            fData(object)$kmeans <- cl$cluster
            ## possibly update processingData
            if (validObject(object))
              return(object)
          })

setMethod("kmeansClustering",
          signature(object = "matrix",
                    params = "missing"),
          function(object, ...) {
            kmeans(object, ...)
          })

setMethod("kmeansClustering",
          signature(object = "MSnSet",
                    params = "ClustRegRes"), 
          function(object, params, criterion = "BIC"){
            k <- getParams(params, criterion)
            if (length(params@algoparams) == 0) {
              cl <- kmeans(exprs(object),
                           centers = getParams(params, criterion))
            } else {
              cl <- kmeans(exprs(object),
                           centers = getParams(params, criterion),
                           as.pairlist(params@algoparams))
            }
            fData(object)$kmeans <- cl$cluster
            if (validObject(object))
              return(object)
          })


setMethod("kmeansOptimisation",
          signature(object = "MSnSet"), 
          function(object, k = 1:20, ...) {
            x <- lapply(k, .optim.kmeans, exprs(object), ...)
            .criteria <- names(x[[1]])
            x <- do.call(rbind, x)
            x <- cbind(k, x)
            ans <- new("ClustRegRes",
                       algorithm = "kmeans",
                       criteria = .criteria,
                       parameters = list(k = k),
                       results = x,
                       algoparams = list(...))
            if (validObject(ans))
              return(ans)
          })

## - compare/contrast 2 clusters - could be 2 fcols
## - estimate parameters - use 1 fcol

## setGeneric("kmeansOptimisation",
##             function(object, fcol, ...) standardGeneric("kmeansOptimisation"))
## setMethod("kmeansOptimisation",
##           signature(object = "MSnSet",
##                     fcol = "character"),
##           function(object, model, ...) {
##             ## compare results of cl and fData(object)[, fcol]
##             return(NULL)
##           })