Add method to pick the best matrix class for count

 - Avoid integer overflow when trying to convert a
`matrix` with more than 2^31-1 non-null values to
a sparse matrix
 - Prefer R base dense `matrix` class when less than
35% of zeros (smaller object size than `dgCMatrix`)

Author: fspecque

DESCRIPTION

@@ -52,7 +52,8 @@ Imports:
     scales,
     cli,
     pbapply,
-    matrixStats
+    matrixStats,
+    Rcpp
 Suggests: 
     SeuratData,
     lisi,
@@ -67,6 +68,7 @@ Collate:
     'Combat.R'
     'Harmony.R'
     'MNN.R'
+    'RcppExports.R'
     'Scanorama.R'
     'SeuratIntegrate-package.R'
     'bbknn.R'
@@ -104,3 +106,6 @@ BugReports: https://github.com/cbib/Seurat-Integrate/issues
 Remotes: SeuratData=github::satijalab/seurat-data,
   github::theislab/kBET,
   lisi=github::immunogenomics/LISI
+LinkingTo: 
+    Rcpp,
+    RcppArmadillo

NAMESPACE

@@ -108,6 +108,7 @@ importFrom(Matrix,spMatrix)
 importFrom(Matrix,sparse.model.matrix)
 importFrom(Matrix,sparseMatrix)
 importFrom(Matrix,t)
+importFrom(Rcpp,sourceCpp)
 importFrom(RhpcBLASctl,blas_get_num_procs)
 importFrom(RhpcBLASctl,blas_set_num_threads)
 importFrom(RhpcBLASctl,omp_get_num_procs)
@@ -338,3 +339,4 @@ importFrom(utils,object.size)
 importFrom(utils,setTxtProgressBar)
 importFrom(vctrs,vec_cast)
 importFrom(vctrs,vec_ptype2)
+useDynLib(SeuratIntegrate, .registration = TRUE)

NEWS.md

@@ -1,5 +1,8 @@
 # SeuratIntegrate (development version)
 
+* The most suited matrix format is automatically chosen for corrected counts
+output by integration methods (should be dense matrix most of the time)
+
 * Add support for scGraph metric (https://doi.org/10.1101/2024.04.02.587824)
 
 * Improved speed of `CreateIntegrationGroups` for non-SCT assay in unambiguous cases

R/Combat.R

@@ -145,7 +145,7 @@ CombatIntegration <- function(
   colnames(corrected.mat) <- colnames(data)
   output.list <- list()
   output.list[[reconstructed.assay]] <- CreateAssayObject(
-    data = as(corrected.mat, "sparseMatrix"),
+    data = choose_matrix_format(corrected.mat),
     key = key.assay)
   if(use.scaled) {
     output.list[[reconstructed.assay]] <- SetAssayData(

R/MNN.R

@@ -111,7 +111,7 @@ MNNIntegration <- function(
 
   # Add reconstructed matrix (gene x cell)
   reconstructed_assay <- CreateAssayObject(
-    data = as(object = assay(x = out), Class = "sparseMatrix"),
+    data = choose_matrix_format(mat = assay(x = out)),
   )
   # Add variable features
   VariableFeatures(object = reconstructed_assay) <- features

R/RcppExports.R

@@ -0,0 +1,7 @@
+# Generated by using Rcpp::compileAttributes() -> do not edit by hand
+# Generator token: 10BE3573-1514-4C36-9D1C-5A225CD40393
+
+n_zeros_mat <- function(mat) {
+    .Call(`_SeuratIntegrate_n_zeros_mat`, mat)
+}
+

R/Scanorama.R

@@ -220,7 +220,7 @@ ScanoramaIntegration <- function(
   }
   if (cntcor) {
     output.list[[reconstructed.assay]] <- CreateAssayObject(
-      data = t( as(cnts.scano, "sparseMatrix") )
+      data = t( choose_matrix_format(cnts.scano) )
     )
     VariableFeatures(object = output.list[[reconstructed.assay]]) <- features
   }

R/SeuratIntegrate-package.R

@@ -2,9 +2,11 @@
 "_PACKAGE"
 
 ## usethis namespace: start
+#' @import rlang
+#' @importFrom Rcpp sourceCpp
 #' @importFrom rlang %||% check_installed abort
 #' @importFrom SeuratObject %iff%
-#' @import rlang
+#' @useDynLib SeuratIntegrate, .registration = TRUE
 ## usethis namespace: end
 
 #' @name integration-method

R/utils.R

@@ -741,6 +741,20 @@ could.be.connectivity.Matrix <- function(object, check.symmetry = T) {
     (! check.symmetry || isSymmetric(object))
   return(res)
 }
+#' @keywords internal
+#' @noRd
+choose_matrix_format <- function(mat) {
+  max_int_32bit <- 2^31 - 1
+  if (all(c(ncol(mat), nrow(mat)) <= max_int_32bit)) { # should always be TRUE
+    n_0s <- n_zeros_mat(mat = mat)
+    if ( (n_0s > max_int_32bit) | (n_0s/length(mat) < .35) ) {
+      mat <- as.matrix(mat)
+    } else {
+      as.dgcmatrix(mat)
+    }
+  }
+  return(mat)
+}
 
 
 # Creates data.frame with cell group assignments for integration

src/.gitignore

@@ -0,0 +1,3 @@
+*.o
+*.so
+*.dll