R/KLdist.matrix.R

In bioDist: Different distance measures

setGeneric("KLdist.matrix", function(x, ...) standardGeneric("KLdist.matrix"))


setMethod("KLdist.matrix", signature=signature("matrix"), 
          function(x,gridsize=NULL, symmetrize=FALSE, diag=FALSE, upper=FALSE)
{ 
   x <- as.matrix(x)
   nc <- ncol(x)
   nr <- nrow(x)
   clist <- vector("list", length=nr)
   me <- .Machine$double.eps
   
   interpfunc <- function(x,y,...) {
	f <- function(w) approx(x,y, w, yleft=0, yright=0)$y
	class(f) <- "dfun"
	f
   }

   datRange <- matrix(ncol=2,nrow=nr)
   binWidth <- vector(length=nr)
   bins <- list()
   binCounts <-list()
   
   for( i in 1:nr){
        if(nc >10000)  dat <- sample(x[i,],nc*0.1)
        else           dat <- x[i,]
        datRange[i,] <- range(dat)
		binWidth[i] <- dpih(dat, gridsize=if(is.null(gridsize))
				max(401,length(dat)/10)
						else gridsize)
        bins[[i]] <- seq(datRange[i,1]-0.1, datRange[i,2]+0.1+binWidth[i], 
						 by=binWidth[i])
        binCounts[[i]] <- KernSmooth:::linbin(x[i,],bins[[i]],truncate=T)/
                                    (nc*binWidth[i])
	}   

   appfun <- function(x,bins,binCounts,binWidth,i,j)
    {   
       step <- min(binWidth[[i]],binWidth[[j]])
       combRange <- datRange[c(i,j),]
       supp <- c(min(combRange[,1]),max(combRange[,2]))
       p<- seq(from= supp[1], to =supp[2], by= step)
       f <- interpfunc(bins[[i]],binCounts[[i]])
       g <- interpfunc(bins[[j]],binCounts[[j]])
       dist<-sum(log((f(p)+me)/(g(p)+me))*f(p))*step
       if(symmetrize)
       {
          dist <- (dist +  sum(log((g(p)+me)/(f(p)+me))*g(p))*step)/2
       }  
      return(dist )   
    }
   
   rvec<-rep(NA, nr*(nr-1)/2)
   ct <- 1
   for(i in 1:(nr-1)){ 
       for(j in (i+1):nr) {
           rvec[ct] <- appfun(x,bins,binCounts,binWidth,i,j)
           ct <- ct+1	   
       }
   }
   attributes(rvec) <- list(Size = nr, Labels = row.names(x),
                            Diag = diag, Upper = upper, methods =
                            "KLdist", class = "dist")
   rvec

} )


setMethod("KLdist.matrix", signature=signature("eSet"),
function(x,gridsize=NULL, symmetrize=FALSE, diag=FALSE, 
         upper=FALSE, sample=TRUE)  {
    if( sample ) ep <- t(exprs(x)) else ep <- exprs(x)
    KLdist.matrix(ep, gridsize=gridsize, symmetrize=symmetrize, diag, upper)
})


setMethod("KLdist.matrix",signature=signature("list"), 
function(x,gridsize=NULL,symmetrize = FALSE, diag = FALSE,upper=FALSE){
  n <- length(x)
  clist <- vector("list", length=n)
  me <- .Machine$double.eps
  
  interpfunc <- function(x,y,...) {
      f <- function(w) approx(x,y, w, yleft=0, yright=0)$y
      class(f) <- "dfun"
      f
  }

  datRange <- matrix(ncol=2,nrow=n)
  binWidth <- vector(length=n)
  bins <- list()
  binCounts <-list()

  nc <- unlist(lapply(x,length))
  for( i in 1:n){
        if(nc[i] >10000){
			  dat <- sample(x[[i]],nc[i]*0.1)
		}else{
			  dat <- x[[i]]
		}
        if(!all(is.na(dat))){
			datRange[i,] <- range(dat)
			binWidth[i] <- dpih(dat,gridsize=if(is.null(gridsize))	max(401,length(dat)/10) else gridsize)
        	bins[[i]] <- seq(datRange[i,1]-0.1, datRange[i,2]+0.1+binWidth[i], 
						 by=binWidth[i])
        	binCounts[[i]] <- KernSmooth:::linbin(x[[i]],bins[[i]],truncate=T)/
                                   (nc[i]*binWidth[i])
	     }
  }   
 
  distfun <- function(x,bins,binCounts,binWidth,i,j)
  {   
       step <- min(binWidth[i],binWidth[j])
       combRange <- datRange[c(i,j),]
       supp <- c(min(combRange[,1]),max(combRange[,2]))
       p<- seq(from= supp[1], to =supp[2], by= step)
       f <- interpfunc(bins[[i]],binCounts[[i]])
       g <- interpfunc(bins[[j]],binCounts[[j]])
       dist<-sum(log((f(p)+me)/(g(p)+me))*f(p))*step

       if(symmetrize)
       {
	    dist <- (dist +  sum(log((g(p)+me)/(f(p)+me))*g(p))*step)/2
       }
       return(dist)
  }
     

  ans<-rep(NA, n*(n-1)/2)
  ct <- 1
  for(i in 1:(n-1))
      for(j in (i+1):n) {
	  if(!is.na(x[[i]]) && !is.na(x[[j]]))
              	ans[ct] <- distfun(x,bins,binCounts,binWidth,i,j)
	  else
	 	ans[ct]=NA
		ct <- ct+1
	  }
    attributes(ans) <- list(Size = n, Labels = names(x),
                        Diag = diag, Upper = upper, 
			methods ="KLdist", 
                        class = "dist")
          
          ans
})


# # # ## tentative "list" method for unequal sized samples (added by
# # # ## Deepayan Sarkar)
# # # 
# # # 
# # # setMethod("KLdist.matrix",
# # #           signature=signature("list"), 
# # #           function(x, 
# # #                    discretize = TRUE, nbin = 10,
# # #                    symmetrize = FALSE,
# # #                    diag = FALSE, upper=FALSE)
# # #       {
# # #           n <- length(x)
# # #           clist <- vector("list", length=n)
# # #           me <- .Machine$double.eps
# # #           
# # #           ##note: we combine x and y before binning, to make sure we span
# # #           ##   the range of the data, and we add machine epsilon to 
# # #           ##   protect against +/- Inf; this could use some work.
# # #           distfun <- function(x, y)
# # #           {
# # #               ## not clear what should be done if exactly one of x and y is a factor
# # #               if (discretize && !is.factor(x))
# # #               {
# # #                   breaks.x <- hist(c(x,y), breaks = nbin, plot = FALSE)$breaks
# # #                   temp1 <- table(cut(y, breaks.x, include.lowest = TRUE)) / length(y)
# # #                   ## temp1 <- temp1 + me
# # #                   temp2 <- table(cut(x, breaks.x, include.lowest = TRUE)) / length(x)
# # #                   ## temp2 <- temp2 + me
# # #                   sum( ifelse(temp2 > 0, log(temp2 / (temp1 + me)) * temp2, 0) , na.rm = TRUE)
# # #               }
# # #               else
# # #               {
# # #                   levs <- sort(unique(c(x, y)))
# # #                   tabx <- table(factor(x, levels = levs)) / length(x)
# # #                   taby <- table(factor(y, levels = levs)) / length(y)
# # #                   sum(ifelse(tabx > 0, log(tabx / (taby + me)) * tabx, 0), na.rm = TRUE)
# # #               }
# # #           }
# # #           ans <- matrix(NA, n, n)
# # #           for(i in seq_len(n))
# # #               for(j in seq_len(n))
# # #               {
# # #                   ans[i, j] <- distfun(x[[i]], x[[j]])
# # #               }
# # #           if(symmetrize) ans <- t(ans) + ans
# # #           if (!is.null(names(x)))
# # #               rownames(ans) <- colnames(ans) <- names(x)
# # #           ans
# # #       })
# # #