R/tsne.R

In farrellja/URD: URD

#' Calculate tSNE projection of data
#' 
#' Calculates a spectral tSNE representation of the data, based either on PCA or diffusion map reductions.
#' 
#' @importFrom Rtsne Rtsne
#' 
#' @param object An URD object
#' @param dim.use (Character) Whether to calculate the tSNE projection from PCs (\code{"pca"}) or diffusion components (\code{"dm"}). Default is from PCs.
#' @param which.dims (Numeric vector) Which PCs (or diffusion components) to use. Defaults to the significant PCs. (We don't estimate the significant diffusion components, so if using diffusion map reduction, probably should explicitly specify this.)
#' @param perlexity (Numeric) Perplexity parameter for the tSNE
#' @param theta (Numeric) Speed/accuracy trade-off for Barnes-Hut approximation of tSNE (0-1, 0 is exact tSNE, higher is less accurate, default is 0.5)
#' @param max_iter (Numeric) Number of iterations to perform
#' @param verbose (Logical) Should Rtsne print progress updates?
#' 
#' @return An URD object with tSNE coordinates stored in the \code{@@tsne.y} slot.
#' 
#' @examples 
#' # Set seed to get reproducible reduction, since tSNE is stochastic.
#' set.seed(18)
#' 
#' # Calculate tSNE on PCA (the default)
#' object <- calcTsne(object, perplexity = 30, theta = 0.5)
#' 
#' # Calculate tSNE on 18 DCs from diffusion map.
#' object <- calcTsne(object, dim.use="dm", which.dims=1:18, perplexity=30)
#' 
#' @export
calcTsne <- function(object, dim.use=c("pca", "dm"), which.dims=which(object@pca.sig), perplexity=30, theta=0.5, max_iter=1000, verbose=FALSE) {
  if (length(dim.use) > 1) dim.use <- dim.use[1]
  if (dim.use == "pca") {
    if (length(which.dims) == 0) {
      stop ("No PCs are marked as significant. Specify dimensions to use (which.dims=).")
    }
    cells.use <- rownames(object@pca.scores)
    tsne.result <- Rtsne(as.matrix(object@pca.scores[,which.dims]), dims=2, pca=FALSE, perplexity=perplexity, theta=theta, max_iter=max_iter, verbose=verbose)
  } else if (dim.use == "dm") {
    cells.use <- intersect(colnames(object@logupx.data), rownames(object@dm.obj[[1]]@eigenvectors))
    tsne.result <- Rtsne(as.matrix(object@dm.obj[[1]]@eigenvectors[cells.use,which.dims]), dims=2, pca=FALSE, perplexity=perplexity, theta=theta, max_iter=max_iter, verbose=verbose)
  } else {
    stop("dim.use must be either pca or dm")
  }
  object@tsne.y <- data.frame(tsne.result$Y)
  names(object@tsne.y) <- c("tSNE1", "tSNE2")
  rownames(object@tsne.y) <- cells.use
  return(object)
}