From 38550ffbec697e589607a92683b770ba751d997d Mon Sep 17 00:00:00 2001
From: Robrecht Cannoodt <rcannood@gmail.com>
Date: Tue, 28 Jul 2026 14:31:32 +0200
Subject: [PATCH 1/2] normalise the simulated data the same way as the real
 data

`downstream` built the simulated `logcounts` with `log1p(counts)`, but read the
real dataset's stored `logcounts` layer for the other side of the comparison.
That layer is `log2(counts / sizeFactor + 1)` -- verified exactly against the
shipped dataset -- so the two sides differed by a whole transform and Moran's I
partly measured that rather than the data.

Both sides now go through a shared `compute_logcounts()`, which reproduces the
stored layer to 9e-16, so the real side is unchanged and only the simulated
side moves. On the positive control, where the simulated data is the real data,
crosscor_cosine and crosscor_mantel go from 0.95 to exactly 1.
---
 src/helpers/utils.R             | 25 +++++++++++++++++++++++++
 src/metrics/downstream/script.R |  6 +++---
 2 files changed, 28 insertions(+), 3 deletions(-)

diff --git a/src/helpers/utils.R b/src/helpers/utils.R
index b9ec2703..c4b46ea7 100644
--- a/src/helpers/utils.R
+++ b/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)
diff --git a/src/metrics/downstream/script.R b/src/metrics/downstream/script.R
index 51980152..a8d59f43 100644
--- a/src/metrics/downstream/script.R
+++ b/src/metrics/downstream/script.R
@@ -34,9 +34,9 @@ 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)

From db076bb041d54325f7e2129834d5f04daa2eba71 Mon Sep 17 00:00:00 2001
From: Robrecht Cannoodt <rcannood@gmail.com>
Date: Tue, 28 Jul 2026 14:31:32 +0200
Subject: [PATCH 2/2] update changelog

---
 CHANGELOG.md | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index e3468069..be1c40dc 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,5 +1,14 @@
 # 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.
+
+# task_spatial_simulators dev
+
 Bug fixes:
   - `run_benchmark`: raise `uns_length_cutoff` from 15 to 50, so that
     `extract_uns_metadata` no longer drops the `metric_ids` of components that