Fix anno norm bug

CHANGELOG.md

@@ -7,6 +7,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## Unreleased
 
+### Fixes
+- Fixed a bug where running `AnnotateCells(..., method = "nmf")` could unintentionally propagate normalization changes from an internal temporary `Seurat` object to the returned object.
+
 ## [0.17.0] - 2026-03-23
 
 ### Added

NAMESPACE

@@ -279,6 +279,7 @@ importFrom(R6,R6Class)
 importFrom(SeuratObject,"Key<-")
 importFrom(SeuratObject,"LayerData<-")
 importFrom(SeuratObject,"VariableFeatures<-")
+importFrom(SeuratObject,AddMetaData)
 importFrom(SeuratObject,Assays)
 importFrom(SeuratObject,Cells)
 importFrom(SeuratObject,CreateAssay5Object)

R/annotate.R

@@ -336,34 +336,44 @@ AnnotateCells <- function(
 ) {
   expect_RcppML()
 
-  # Apply normalization
+  # When normalization is required for NMF: normalize the reference in place, and build
+  # a separate minimal Seurat from the query counts for NormalizeData.
   if (!skip_normalization) {
     if (verbose && check_global_verbosity()) {
       cli::cli_alert_info("Normalizing reference and query data using method {.str {normalization_method}}")
     }
-    reference <-
-      Seurat::NormalizeData(
-        reference,
-        normalization.method = normalization_method,
-        margin = 2,
-        verbose = verbose
-      )
-    object <-
-      Seurat::NormalizeData(
-        object,
-        normalization.method = normalization_method,
-        margin = 2,
-        verbose = verbose
-      )
+    reference <- Seurat::NormalizeData(
+      reference,
+      assay = reference_assay,
+      normalization.method = normalization_method,
+      margin = 2,
+      verbose = verbose
+    )
+    query_counts <- SeuratObject::LayerData(object, assay = query_assay, layer = "counts")
+    query_for_nm <- SeuratObject::CreateSeuratObject(
+      counts = query_counts,
+      assay = query_assay,
+      meta.data = data.frame(row.names = colnames(query_counts))
+    )
+    query_for_nm <- Seurat::NormalizeData(
+      query_for_nm,
+      assay = query_assay,
+      normalization.method = normalization_method,
+      margin = 2,
+      verbose = verbose
+    )
   }
 
   if (verbose && check_global_verbosity()) {
     cli::cli_alert_info("Fetching overlapping data from reference and query assays.")
   }
 
-  # Get data
   reference_data <- SeuratObject::LayerData(reference, assay = reference_assay)
-  target_data <- SeuratObject::LayerData(object, assay = query_assay)
+  target_data <- if (!skip_normalization) {
+    SeuratObject::LayerData(query_for_nm, assay = query_assay)
+  } else {
+    SeuratObject::LayerData(object, assay = query_assay)
+  }
 
   # Subset data to use overlapping markers
   shared_markers <- intersect(rownames(reference_data), rownames(target_data))

R/pixelatorR-package.R

@@ -176,6 +176,7 @@
 #' @importFrom R6 R6Class
 #' @importFrom rlang on_load
 #' @importFrom rlang run_on_load
+#' @importFrom SeuratObject AddMetaData
 #' @importFrom SeuratObject Assays
 #' @importFrom SeuratObject Cells
 #' @importFrom SeuratObject CreateAssay5Object

tests/testthat/test-AnnotateCells.R

@@ -24,7 +24,7 @@ reference <- SeuratObject::CreateSeuratObject(
 
 test_that("annotate_cells works as expected", {
   # NMF default
-  expect_no_error(seur <- AnnotateCells(
+  expect_no_error(seur_anno <- AnnotateCells(
     seur,
     reference,
     query_assay = "PNA",
@@ -35,7 +35,7 @@ test_that("annotate_cells works as expected", {
     verbose = FALSE
   ))
   expect_equal(
-    seur$cell_type %>% unname(),
+    seur_anno$cell_type %>% unname(),
     c(
       "Mono CD16+", "pDC", "CD4T", "CD4T", "CD4T", "Mono CD16+",
       "pDC", "CD4T", "CD4T", "CD4T", "Mono CD16+", "pDC", "CD4T", "CD4T",
@@ -63,6 +63,40 @@ test_that("annotate_cells works as expected", {
       "CD4T", "CD4T", "Mono CD16+", "Unknown", "CD4T", "CD4T", "CD4T"
     )
   )
+
+  # Test that normalization is preserved
+  reference <- read_pbmc_reference()
+  seur_sub <-
+    seur %>%
+    Seurat::NormalizeData(normalization.method = "CLR", margin = 2) %>%
+    # Z score scaling
+    Seurat::ScaleData() %>%
+    Seurat::RunPCA(npcs = 3, features = rownames(seur)) %>%
+    Seurat::FindNeighbors(reduction = "pca", dims = 1:3) %>%
+    Seurat::FindClusters(resolution = 0.5) %>%
+    subset(features = head(rownames(seur), 50))
+
+
+  # NMF default
+  expect_no_error(seur_anno <- AnnotateCells(
+    seur_sub,
+    reference,
+    reference_groups = c("celltype.l1", "celltype.l2"),
+    # Summarise annotations by cluster, adding an extra column with majority vote
+    summarize_by_column = "seurat_clusters",
+    query_assay = "PNA",
+    reference_assay = "PNA",
+    method = "nmf"
+  ))
+
+  expect_equal(
+    LayerData(seur_anno, "counts"),
+    LayerData(seur_sub, "counts")
+  )
+  expect_equal(
+    LayerData(seur_anno, "data"),
+    LayerData(seur_sub, "data")
+  )
 })