Normalise the simulated data the same way as the real data

CHANGELOG.md

@@ -1,6 +1,11 @@
 # task_spatial_simulators dev
 
 Bug fixes:
+  - `downstream`: normalise the simulated dataset the same way as the real one,
+    through a new `compute_logcounts()` helper. The simulated side used
+    `log1p(counts)` while the real side used the stored, size-factor normalised
+    `logcounts`. On the positive control, `crosscor_cosine` and
+    `crosscor_mantel` now come out at exactly 1.
   - `scdesign2`: order the input by `spatial_cluster` before simulating.
     `simulate_count_scDesign2()` returns cells grouped per cell type, so the
     coordinates taken from the unsorted input belonged to different spots than

src/helpers/utils.R

@@ -1,3 +1,28 @@
+# log-normalisation
+#
+# Both the real and the simulated dataset must be normalised the exact same way
+# before they are compared, otherwise the metric picks up the difference in
+# normalisation rather than the difference in the data. Do not read the
+# `logcounts` layer directly: on the real dataset it was produced by the dataset
+# loader, on a simulated dataset it may not exist at all.
+compute_logcounts <- function(adata) {
+  requireNamespace("scater", quietly = TRUE)
+  requireNamespace("SingleCellExperiment", quietly = TRUE)
+  requireNamespace("SummarizedExperiment", quietly = TRUE)
+
+  # genes x spots, as expected by SingleCellExperiment
+  counts <- Matrix::t(adata$layers[["counts"]])
+
+  sce <- SingleCellExperiment::SingleCellExperiment(list(counts = counts))
+  sce <- scater::logNormCounts(sce)
+
+  logcounts <- SummarizedExperiment::assay(sce, "logcounts")
+  dimnames(logcounts) <- dimnames(counts)
+
+  # back to spots x genes, matching the anndata layout
+  Matrix::t(logcounts)
+}
+
 # spatial autocorrelation
 generate_moransI <- function(adata) {
   requireNamespace("spots", quietly = TRUE)

src/metrics/downstream/script.R

@@ -33,9 +33,9 @@ ctdeconvolute_rmse <- generate_rmse(real_ct_prop, sim_ct_prop)
 ctdeconvolute_jsd <- generate_jds(real_ct_prop, sim_ct_prop)
 
 cat("spatial autocorrelation evaluation\n")
-counts <- input_simulated_sp$layers[["counts"]]
-logcounts <- log1p(counts)
-input_simulated_sp$layers[["logcounts"]] <- logcounts
+# normalise both datasets the same way, so that Moran's I is comparable
+input_real_sp$layers[["logcounts"]] <- compute_logcounts(input_real_sp)
+input_simulated_sp$layers[["logcounts"]] <- compute_logcounts(input_simulated_sp)
 real_moransI <- generate_moransI(input_real_sp)
 # real_moransI <- input_real_sp$varm$spatial_autocorrelation
 sim_moransI <- generate_moransI(input_simulated_sp)