use sparse count matrices

R/shrinkBins.R

@@ -61,11 +61,12 @@ shrinkBins <- function(
   }
 
   # get the prior means
-  prior.means <- prior.means %||% getGlobalMeans(
-    obj = original.x,
-    targets = targets,
-    assay = assay
-  )
+  prior.means <- prior.means %||%
+    getGlobalMeans(
+      obj = original.x,
+      targets = targets,
+      assay = assay
+    )
 
   is.atac_or_rna <- assay %in% c("atac", "rna")
   input.fun <- if (jse) {
@@ -84,7 +85,7 @@ shrinkBins <- function(
 
   # bin the input
   bin.mat <- getBinMatrix(
-    mat = as.matrix(cbind(input.assay, prior.means)),
+    mat = cbind(input.assay, prior.means),
     genloc = rowRanges(x),
     chr = chr,
     res = res,

R/transformTFIDF.R

@@ -7,12 +7,8 @@
 #'
 #' @param mat n x p input matrix (n = samples/cells; p = rna counts)
 #' @param scale.factor Scaling factor for the term-frequency (TF)
-#' @param count.min The minimum expression count used for TF-IDF. Binarizes
-#' when `count.min` = 0 and `count.max` = 1.
-#' @param count.max The maximum expression count used for TF-IDF. Binarizes
-#' when `count.min` = 0 and `count.max` = 1.
-#' binarizes the matrix. A `cap` value greater than 1 will cap counts at that
-#' value.
+#' @param binarize Whether to binarize the input matrix: any value > 0 is set
+#' to 1
 #'
 #' @return A TF-IDF transformed matrix of the same dimensions as the input
 #'
@@ -26,9 +22,7 @@
 #' tfidf <- transformTFIDF(mat)
 #'
 #' @export
-transformTFIDF <- function(mat, scale.factor = 1e5, count.min = 0, count.max = 1) {
-  stopifnot("'count.min' must be less than 'count.max'" = count.min < count.max)
-
+transformTFIDF <- function(mat, scale.factor = 1e5, binarize = FALSE) {
   if (!is(mat, "matrix") & !is(mat, "Matrix")) {
     stop("Input needs to be a matrix.")
   }
@@ -41,25 +35,17 @@ transformTFIDF <- function(mat, scale.factor = 1e5, count.min = 0, count.max = 1
     mat.capped <- t(Matrix(mat, sparse = TRUE))
   }
 
-  # constrain the matrix
-  mat.capped@x <- .constrain(mat.capped@x, count.min, count.max)
+  if (binarize) {
+    mat.capped@x <- .binarize(mat.capped@x)
+  }
   tf <- t(t(mat.capped) / Matrix::colSums(mat.capped)) # compute term-frequency
   tf@x <- log1p(tf@x * scale.factor) # scale
   idf <- log(1 + ncol(mat.capped) / Matrix::rowSums(mat.capped)) # inverse-document frequency smooth
-  tfidf <- .tfidf(tf, idf) # transform
-
-  # cast back to a matrix since things like UMAP don't like sparse matrices
-  as.matrix(t(tfidf))
+  .tfidf(tf, idf) # transform
 }
 
-# binarize when lower is 0 and upper is 1, constrain otherwise
-.constrain <- function(v, lower, upper) {
-  if (lower == 0 & upper == 1) {
-    v[v > 0] <- 1
-  } else {
-    v[v < lower] <- 0
-    v[v > upper] <- upper
-  }
+.binarize <- function(v) {
+  v[v > 0] <- 1
   v
 }
 

tests/testthat/test-transformTFIDF.R

@@ -5,22 +5,14 @@ test_that("transformTFIDF", {
     "Input needs to be a matrix",
     fixed = TRUE
   )
-  expect_error(
-    transformTFIDF(0:10, count.min = 1, count.max = 0),
-    "'count.min' must be less than 'count.max'"
-  )
 })
 # }}}
 
 # .constrain {{{
-test_that(".constrain", {
-  expect_equal(
-    compartmap:::.constrain(0:10, lower = 0, upper = 1),
-    c(0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1)
-  )
+test_that(".binarize", {
   expect_equal(
-    compartmap:::.constrain(0:10, lower = 2, upper = 5),
-    c(0, 0, 2, 3, 4, 5, 5, 5, 5, 5, 5)
+    compartmap:::.binarize(seq(0, 10, by = 0.5)),
+    c(0, rep(1, 20))
   )
 })
 # }}}

vignettes/compartmap.Rmd

@@ -270,7 +270,7 @@ data("k562_scrna_raw", package = "compartmap")
 k562_scrna_chr14_tfidf <- transformTFIDF(assay(k562_scrna_se_chr14))
 
 # Add back the TF-IDF counts to the object in the counts slot
-assay(k562_scrna_se_chr14, "counts") <- t(k562_scrna_chr14_tfidf)
+assay(k562_scrna_se_chr14, "counts") <- k562_scrna_chr14_tfidf
 
 # Compute chromatin domains at the group level
 k562_scrna_chr14_raw_domains <- scCompartments(k562_scrna_se_chr14,