Merge pull request #8 from rcannood/devel

Refactor SCORPIUS code

DESCRIPTION

@@ -1,7 +1,7 @@
 Package: SCORPIUS
 Type: Package
 Title: Inferring Developmental Chronologies from Single-Cell RNA Sequencing Data
-Version: 1.0.1
+Version: 1.0.2.9001
 Date: 2018-06-12
 Description: An accurate and easy tool for performing trajectory inference on
   single cells using single-cell RNA sequencing data. In addition, SCORPIUS
@@ -34,13 +34,12 @@ Depends:
   R (>= 3.0.0)
 Imports:
   dplyr,
-  fitdistrplus,
+  dynutils,
   grDevices,
   ggplot2 (>= 2.0),
   magrittr,
   MASS,
   mclust,
-  Rcpp,
   pbapply,
   pheatmap,
   princurve (>= 2.0.2),
@@ -50,13 +49,11 @@ Imports:
   reshape2,
   splines,
   stats,
-  testthat,
   tidyr,
   TSP,
   utils
 Suggests:
   knitr,
-  rmarkdown
-LinkingTo:
-  Rcpp
+  rmarkdown,
+  testthat
 

NAMESPACE

@@ -15,10 +15,7 @@ export(infer_initial_trajectory)
 export(infer_trajectory)
 export(knn)
 export(knn_distances)
-export(outlier_filter)
-export(outlierness)
 export(reduce_dimensionality)
-export(reduce_dimensionality_landmarked)
 export(reverse_trajectory)
 export(scale_quantile)
 export(scale_uniform)
@@ -29,7 +26,8 @@ importFrom(RColorBrewer,brewer.pal)
 importFrom(TSP,TSP)
 importFrom(TSP,insert_dummy)
 importFrom(TSP,solve_TSP)
-importFrom(fitdistrplus,fitdist)
+importFrom(dynutils,correlation_distance)
+importFrom(dynutils,euclidean_distance)
 importFrom(grDevices,hcl)
 importFrom(magrittr,"%$%")
 importFrom(magrittr,"%<>%")
@@ -61,4 +59,3 @@ importFrom(stats,setNames)
 importFrom(tidyr,gather)
 importFrom(tidyr,spread)
 importFrom(utils,head)
-useDynLib(SCORPIUS)

R/SCORPIUS.R

@@ -3,14 +3,11 @@
 #' SCORPIUS orders single cells with regard to an implicit timeline,
 #' such as cellular development or progression over time.
 #'
-#' @section Outlier functions:
-#' \code{\link{outlierness}} \code{\link{outlier_filter}}
-#'
 #' @section Distance functions:
-#' \code{\link{correlation_distance}}, \code{\link{euclidean_distance}}, \code{\link{knn}}
+#' \code{\link{correlation_distance}}, \code{\link{euclidean_distance}}
 #'
 #' @section Dimensionality Reduction functions:
-#' \code{\link{reduce_dimensionality}}, \code{\link{reduce_dimensionality_landmarked}}, \code{\link{scale_uniform}}, \code{\link{scale_quantile}}
+#' \code{\link{reduce_dimensionality}}, \code{\link{scale_uniform}}, \code{\link{scale_quantile}}
 #'
 #' @section Trajectory Inference functions:
 #' \code{\link{infer_trajectory}}, \code{\link{infer_initial_trajectory}}, \code{\link{reverse_trajectory}}, \code{\link{gene_importances}}, \code{\link{extract_modules}}

R/data.R

@@ -31,8 +31,7 @@
 #' dataset <- generate_dataset(type = "poly", num_genes = 500, num_samples = 1000, num_groups = 4)
 #'
 #' ## Reduce dimensionality and infer trajectory with SCORPIUS
-#' dist <- correlation_distance(dataset$expression)
-#' space <- reduce_dimensionality(dist, ndim=2)
+#' space <- reduce_dimensionality(dataset$expression, correlation_distance, ndim = 2)
 #' traj <- infer_trajectory(space)
 #'
 #' ## Visualise
@@ -121,8 +120,7 @@ generate_dataset <- function(type = c("splines", "polynomial"), num_samples = 40
 #' data("ginhoux")
 #'
 #' ## Reduce dimensionality and infer trajectory with SCORPIUS
-#' dist <- correlation_distance(ginhoux$expression)
-#' space <- reduce_dimensionality(dist)
+#' space <- reduce_dimensionality(ginhoux$expression, correlation_distance)
 #' traj <- infer_trajectory(space)
 #'
 #' ## Visualise

R/dimensionality_reduction.R

@@ -2,11 +2,11 @@
 #'
 #' @description \code{reduce_dimensionality} performs an eigenanalysis of the given dissimilarity matrix and returns coordinates of the samples represented in an \code{ndim}-dimensional space.
 #'
-#' @usage
-#' reduce_dimensionality(dist, ndim, rescale=TRUE)
-#'
-#' @param dist A numeric matrix, data frame or "\code{dist}" object.
-#' @param ndim The number of dimensions in the new space.
+#' @param x a numeric matrix
+#' @param dist_fun the distance function to be used; must have exactly two arguments, namely dist_fun(x, y).
+#' @param ndim the maximum dimension of the space which the data are to be represented in; must be in {1, 2, \ldots, n-1}.
+#' @param landmark_method Must be "naive" for now. Other landmark methods will be supported in the future.
+#' @param num_landmarks the number of landmarks to be selected.
 #' @param rescale A logical indicating whether or not the returned space should be rescaled and centered.
 #'
 #' @return A matrix containing the coordinates of each sample, represented in an \code{ndim}-dimensional space.
@@ -22,22 +22,115 @@
 #' dataset <- generate_dataset(type = "poly", num_genes = 500, num_samples = 1000, num_groups = 4)
 #'
 #' ## Reduce the dimensionality of this dataset
-#' dist <- correlation_distance(dataset$expression)
-#' space <- reduce_dimensionality(dist, ndim = 2)
+#' space <- reduce_dimensionality(dataset$expression, correlation_distance, ndim = 2)
 #'
 #' ## Visualise the dataset
 #' draw_trajectory_plot(space, progression_group=dataset$sample_info$group_name)
-reduce_dimensionality <- function(dist, ndim = 3, rescale = TRUE) {
+reduce_dimensionality <- function(x, dist_fun, ndim = 3, landmark_method = c("naive", "none"), num_landmarks = 1000, rescale = T) {
   # input check
-  if (!is.matrix(dist) && !is.data.frame(dist) && class(dist) != "dist")
-    stop(sQuote("dist"), " must be a numeric matrix, data frame or a ", sQuote("dist"), " object")
-  if (class(dist) == "dist")
-    dist <- as.matrix(dist)
-  if (!is.finite(ndim) || round(ndim) != ndim || length(ndim) != 1 || ndim < 1 || ndim >= nrow(dist))
-    stop(sQuote("ndim"), " must be a whole number and 1 <= ndim <= nrow(dist)-1")
-
-  space <- stats::cmdscale(dist, k = ndim)
-  if (rescale) space <- scale_uniform(space)
-  colnames(space) <- paste("Comp", seq_len(ncol(space)), sep = "")
-  space
+  if (!is.matrix(x) && !is.data.frame(x))
+    stop(sQuote("x"), " must be a numeric matrix or data frame")
+  if (!is.function(dist_fun)) {
+    stop(sQuote("dist_fun"), " must be a function(x, y) {...}")
+  }
+  if (!is.finite(ndim) || round(ndim) != ndim || length(ndim) != 1 || ndim < 1 || ndim >= nrow(x))
+    stop(sQuote("ndim"), " must be a whole number and 1 <= ndim <= nrow(x)-1")
+
+  landmark_method <- match.arg(landmark_method)
+
+  if (landmark_method == "none") {
+    space <- stats::cmdscale(dist_fun(x), k = ndim)
+    if (rescale) space <- scale_uniform(space)
+    colnames(space) <- paste("Comp", seq_len(ncol(space)), sep = "")
+    space
+  } else {
+    lm_out <- landmark_selection(
+      x = x,
+      dist_fun = dist_fun,
+      landmark_method = landmark_method,
+      num_landmarks = num_landmarks
+    )
+
+    cmd_out <- cmdscale_withlandmarks(
+      dist_lm = lm_out$dist_lm,
+      dist_2lm = lm_out$dist_2lm,
+      ndim = ndim,
+      rescale = rescale
+    )
+    attr(cmd_out, "landmarks") <- lm_out$ix_lm
+    cmd_out
+  }
+}
+
+landmark_selection <- function(x, dist_fun, landmark_method, num_landmarks) {
+  if (nrow(x) < num_landmarks) {
+    num_landmarks <- nrow(x)
+  }
+  switch(
+    landmark_method,
+    "naive" = {
+      ix_lm <- sample.int(nrow(x), num_landmarks)
+      dist_lm <- dist_fun(x[ix_lm,,drop=FALSE], x[ix_lm,,drop=FALSE])
+      dist_2lm <- dist_fun(x[ix_lm,,drop=FALSE], x)
+      list(ix_lm = ix_lm, dist_lm = dist_lm, dist_2lm = dist_2lm)
+    },
+    {
+      stop("landmark_method must be ", sQuote("naive"), ".")
+    }
+  )
+}
+
+cmdscale_withlandmarks <- function(dist_lm, dist_2lm, ndim = 3, rescale = T) {
+  d <- dist_lm
+  if (anyNA(d))
+    stop("NA values not allowed in 'd'")
+
+  x <- as.matrix(d^2)
+  storage.mode(x) <- "double"
+  if (nrow(x) != ncol(x))
+    stop("distances must be result of 'dist' or a square matrix")
+
+  rn <- rownames(x)
+  rn_all <- colnames(dist_2lm)
+  n <- as.integer(nrow(x))
+  N <- as.integer(ncol(dist_2lm))
+
+  if((ndim <- as.integer(ndim)) > n - 1 || ndim < 1)
+    stop("'ndim' must be in {1, 2, ..  n - 1}")
+
+  # double center data
+  mu_n <- rowMeans(x)
+  mu <- mean(x)
+  x_dc <- x - rep(mu_n, n) - rep(mu_n, each = n) + mu
+
+  # classical MDS on landmarks
+  e <- eigen(-x_dc/2, symmetric = TRUE)
+  ev <- e$values[seq_len(ndim)]
+  evec <- e$vectors[, seq_len(ndim), drop = FALSE]
+  ndim1 <- sum(ev > 0)
+  if (ndim1 < ndim) {
+    warning(gettextf("only %d of the first %d eigenvalues are > 0", ndim1, ndim), domain = NA)
+    evec <- evec[, ev > 0,  drop = FALSE]
+    ev <- ev[ev > 0]
+  }
+  Slm <- evec * rep(sqrt(ev), each=n)
+
+  # distance-based triangulation
+  points_inv <- evec / rep(sqrt(ev), each=n)
+  S <- (-t(dist_2lm - rep(mu_n, each = N))/2) %*% points_inv
+
+  # clean up dimension names
+  dimnames(Slm) <- list(rn, paste0("Comp", seq_len(ndim)))
+  dimnames(S) <- list(rn_all, paste0("Comp", seq_len(ndim)))
+
+  # rescale if necessary
+  if (rescale) {
+    Slm <- scale_uniform(Slm)
+    S <- scale_uniform(S)
+  }
+
+  # output
+  attr(S, "landmark_space") <- Slm
+
+  S
 }