Merge pull request #126 from biomage-org/2399-low-pc-clustering

[BIOMAGE-2399] - lower clustering dimensions for low dimensional PCAs
hms-dbmi-cellenics · Feb 15, 2023 · 44db180 · 44db180
2 parents 46e82e1 + 24898dc
commit 44db180
Show file tree

Hide file tree

Showing 2 changed files with 55 additions and 2 deletions.
diff --git a/pipeline-runner/R/qc-helpers.R b/pipeline-runner/R/qc-helpers.R
@@ -195,7 +195,16 @@ getClusters <- function(clustering_method, resolution, data) {
   } else {
     graph_name <- paste0(Seurat::DefaultAssay(data), "_snn")
     if (!graph_name %in% names(data)) {
-      data <- Seurat::FindNeighbors(data, k.param = 20, annoy.metric = "cosine", verbose = FALSE, reduction = active.reduction)
+      # number of dimensions used must be lte to available dimensions
+      dims <- 1:min(10, length(data@reductions[[active.reduction]]))
+      data <-
+        Seurat::FindNeighbors(
+          data,
+          annoy.metric = "cosine",
+          reduction = active.reduction,
+          dims = dims,
+          verbose = FALSE,
+        )
     }
     data <- Seurat::FindClusters(data, resolution = resolution, verbose = FALSE, algorithm = algorithm, random.seed = RANDOM_SEED)
   }
@@ -216,7 +225,12 @@ getSNNiGraph <- function(data, active.reduction) {
 
   # if doesn't exist, run SNN
   if (!snn_name %in% names(data)) {
-    data <- Seurat::FindNeighbors(data, reduction = active.reduction)
+    dims <- 1:min(10, length(data@reductions[[active.reduction]]))
+    data <-
+      Seurat::FindNeighbors(data,
+                            reduction = active.reduction,
+                            dims = dims,
+                            verbose = FALSE)
   }
 
   # convert Seurat Graph object to igraph

diff --git a/pipeline-runner/tests/testthat/test-qc-7-embed_and_cluster.R b/pipeline-runner/tests/testthat/test-qc-7-embed_and_cluster.R
@@ -266,3 +266,42 @@ test_that("embed_and_cluster works", {
   withr::defer(unlink(cell_sets_bucket, recursive = TRUE))
 
 })
+
+
+test_that("runClusters does not crash with less than 10 dimensions available", {
+  algos <- c("louvain", "leiden")
+  scdata <- mock_scdata()
+  expected_keys <- c("cluster", "cell_ids")
+  resolution <- 0.8
+
+  # remove all pre-existing reductions and calculate low-PC PCA
+  scdata <- Seurat::DietSeurat(scdata, scale.data = T)
+  scdata <- suppressWarnings(Seurat::RunPCA(scdata, assay = "RNA", npcs = 2, verbose = F))
+
+  for (algo in algos) {
+    res <- runClusters(algo, resolution, scdata)
+    expect_equal(names(res), expected_keys)
+  }
+})
+
+
+test_that("getClusters uses the default value of 10 if there are enough PCs available",{
+  algos <- c("louvain", "leiden")
+  scdata <- mock_scdata()
+  resolution <- 0.8
+
+  # remove all pre-existing reductions and calculate low-PC PCA
+  scdata <- Seurat::DietSeurat(scdata, scale.data = T)
+  scdata@commands <- list()
+  scdata <- suppressWarnings(Seurat::RunPCA(scdata, assay = "RNA", npcs = 20, verbose = F))
+
+  for (algo in algos) {
+    clustered_scdata <- getClusters(algo, resolution, scdata)
+    if (algo == "louvain") expect_equal(clustered_scdata@commands$FindNeighbors.RNA.pca$dims, 1:10)
+    # difficult to test in leiden, so test internal state as proxy
+    if (algo == "leiden") expect_true("seurat_clusters" %in% names(clustered_scdata@meta.data))
+  }
+})
+
+
+