Fixes

src/methods_cell_type_annotation/rctd/script.R

@@ -43,10 +43,27 @@ filtered_ref <- ref[,colData(ref)$cell_type %in% valid_celltypes]
 
 ref_counts <- assay(filtered_ref, "counts")
 # factor to drop filtered cell types
-colData(filtered_ref)$cell_type <- factor(colData(filtered_ref)$cell_type) 
+colData(filtered_ref)$cell_type <- factor(colData(filtered_ref)$cell_type)
 cell_types <- colData(filtered_ref)$cell_type
 names(cell_types) <- colnames(ref_counts)
 
+# --- Diagnostics: confirm what RCTD actually reads. The reference itself has
+#     ~475/477 panel genes with clear cross-cell-type fold-change, so a "fewer
+#     than 10 DE genes" failure means the data reaching RCTD is degenerate:
+#     cell types collapsed to ~1, non-integer/normalized counts, or few genes
+#     shared with the spatial puck. This block surfaces which. ---
+cat("=== RCTD input diagnostics ===\n")
+cat(sprintf("spatial puck: %d genes x %d cells\n", nrow(counts), ncol(counts)))
+cat(sprintf("reference (>=25 cells/type): %d genes x %d cells, %d cell types\n",
+            nrow(ref_counts), ncol(ref_counts), length(unique(as.character(cell_types)))))
+print(utils::head(sort(table(as.character(cell_types)), decreasing = TRUE), 8))
+cat(sprintf("genes shared reference<->spatial: %d\n",
+            length(intersect(rownames(ref_counts), rownames(counts)))))
+.rcv <- if (methods::is(ref_counts, "sparseMatrix")) ref_counts@x else as.numeric(ref_counts)
+if (length(.rcv)) cat(sprintf("reference counts: min=%.4g max=%.4g mean=%.4g all-integer=%s\n",
+            min(.rcv), max(.rcv), mean(.rcv), isTRUE(all.equal(.rcv, round(.rcv)))))
+cat("==============================\n")
+
 # spacexr::check_cell_types() rejects cell-type names containing '/'
 # (e.g. "Ciliated/secretory cells", "T/NK lineage"). Sanitize the factor levels
 # before building the Reference, and keep a map (safe name -> original name) so

src/methods_expression_correction/split/config.vsh.yaml

@@ -17,6 +17,34 @@ arguments:
     type: boolean
     default: false
     description: Whether to keep cells with 0 counts (may cause errors if set to TRUE)
+  # RCTD's default DE-gene / UMI thresholds are tuned for whole-transcriptome
+  # references; on small iST panels (a few hundred genes, many similar fine-grained
+  # cell types) fold-changes are compressed and RCTD finds "fewer than 10 DE genes".
+  # These relaxed defaults mirror the rctd component so create.RCTD below succeeds.
+  - name: --gene_cutoff
+    type: double
+    default: 0.0
+    description: "Minimum normalized mean expression for a gene to be a platform-effect DE gene (RCTD default 0.000125)."
+  - name: --fc_cutoff
+    type: double
+    default: 0.1
+    description: "Minimum log-fold-change for a gene to be a platform-effect DE gene (RCTD default 0.5)."
+  - name: --gene_cutoff_reg
+    type: double
+    default: 0.0
+    description: "Minimum normalized mean expression for a gene to be a regression DE gene (RCTD default 0.0002)."
+  - name: --fc_cutoff_reg
+    type: double
+    default: 0.1
+    description: "Minimum log-fold-change for a gene to be a regression DE gene (RCTD default 0.75)."
+  - name: --umi_min
+    type: integer
+    default: 20
+    description: "Minimum total UMI per spatial cell to be included (RCTD default 100; iST cells are low-count)."
+  - name: --umi_min_sigma
+    type: integer
+    default: 20
+    description: "Minimum UMI for cells used to fit the platform-effect variance (RCTD default 300)."
 
 resources:
   - type: r_script

src/methods_expression_correction/split/script.R

@@ -12,6 +12,12 @@ par <- list(
   "input_scrnaseq_reference"= "task_ist_preprocessing/resources_test/task_ist_preprocessing/mouse_brain_combined/scrnaseq_reference.h5ad",
   "output" = "task_ist_preprocessing/tmp/split_corrected.h5ad",
   "keep_all_cells" = FALSE,
+  "gene_cutoff" = 0.0,
+  "fc_cutoff" = 0.1,
+  "gene_cutoff_reg" = 0.0,
+  "fc_cutoff_reg" = 0.1,
+  "umi_min" = 20,
+  "umi_min_sigma" = 20
 )
 
 meta <- list(
@@ -67,7 +73,14 @@ cat(sprintf("Number of cores: %s\n", cores))
 
 # Run the algorithm
 cat("Running RCTD\n")
-myRCTD <- create.RCTD(puck, reference, max_cores = cores)
+# Relaxed DE-gene / UMI thresholds (see config): RCTD's defaults find "fewer than
+# 10 DE genes" on small iST panels with many fine-grained cell types.
+myRCTD <- create.RCTD(
+  puck, reference, max_cores = cores,
+  gene_cutoff = par$gene_cutoff, fc_cutoff = par$fc_cutoff,
+  gene_cutoff_reg = par$gene_cutoff_reg, fc_cutoff_reg = par$fc_cutoff_reg,
+  UMI_min = par$umi_min, UMI_min_sigma = par$umi_min_sigma
+)
 myRCTD <- run.RCTD(myRCTD, doublet_mode = "doublet")
 
 # Get the "spot_class" annotation from RCTD

src/methods_transcript_assignment/segger/config.vsh.yaml

@@ -45,69 +45,62 @@ engines:
   #       (cudf, cuspatial, torch). Develop/test on a CUDA host; `viash test`
   #       will not run on a machine without an NVIDIA GPU.
   - type: docker
-    # NGC 26.06 ships numpy 2.1 and a torch (2.13) compiled against numpy 2.x, so
-    # the torch<->numpy bridge (used by the `segger segment` subprocess) stays
-    # alive alongside the numpy-2.x task stack (spatialdata>=0.7.3 hard-requires
-    # numpy>=2 via multiscale-spatial-image -> xarray-dataclass). The older 25.06
-    # image shipped numpy-1.x torch, which cannot coexist with spatialdata>=0.7.3.
-    image: nvcr.io/nvidia/pytorch:26.06-py3
+    # segger is fundamentally a RAPIDS application (cudf + cuspatial + cugraph +
+    # cuml). Bolting that full stack onto the NGC PyTorch base made RAPIDS's UCX
+    # (cugraph -> raft_dask) collide with the image's HPC-X UCX and abort with heap
+    # corruption (SIGABRT, exit 134) that no pip pin could fix. So we base on the
+    # official RAPIDS image (ships cudf/cugraph/cuml/cuspatial 25.04 + a consistent,
+    # working UCX/dask stack) and layer torch (GNN) + the spatialdata/anndata I/O
+    # stack + segger on top. numpy 2.0.2 / pandas 2.2.3 come from the base.
+    #
+    # NOTE: this image runs as the NON-root 'rapids' user, so `type: apt` will NOT
+    # work in a viash build (needs root). Everything below installs via conda/pip
+    # only (both write to the rapids-owned conda env) — do NOT add apt steps.
+    image: rapidsai/base:25.04-cuda12.8-py3.12
     setup:
-      # libgl1 + libglib2.0-0: segger's writer imports segger.io -> cv2 (opencv),
-      # which needs libGL.so.1 / libgthread at import; without them the run crashes
-      # at write time with "libGL.so.1: cannot open shared object file".
-      - type: apt
-        packages: [procps, git, libxcb1, libgl1, libglib2.0-0]
+      # git (for the segger + openproblems github installs) via conda: apt is
+      # unavailable as non-root and the RAPIDS base ships no git.
+      - type: docker
+        run: conda install -y -c conda-forge git
       - type: python
         github:
           - openproblems-bio/core#subdirectory=packages/python/openproblems
-      - type: python
-        packages:
-          - "spatialdata>=0.7.3"
-          - "anndata>=0.12.0"
-          - "zarr>=3.0.0"
-          - geopandas
-          - shapely
-          - "rasterio>=1.3"
-          - scikit-image
-      # cuspatial is imported by segger.geometry.conversion and is NOT shipped in
-      # the 26.06 image (unlike 25.06). Install from the NVIDIA index (its
-      # libcudf-cu12/libcuspatial-cu12 deps are not on PyPI); it pulls a matching
-      # cudf-cu12. Unpinned -> resolves to the newest cuda-12 build (25.4, the last
-      # cuda-12 cuspatial-cu12 on the index).
-      - type: docker
-        run: |
-          pip install --no-cache-dir --extra-index-url=https://pypi.nvidia.com cuspatial-cu12
-      # segger imports torch_geometric (at import time via _patches), lightning
-      # (segger segment -> Trainer) and torch_scatter (models: scatter_max) but
-      # declares NONE of them in its pyproject, so the github install below does
-      # not pull them. torch_geometric/lightning are pure Python. torch_scatter has
-      # no prebuilt wheel for the NGC custom torch build, so compile against it
-      # (--no-build-isolation). The build host has no GPU, so force a CUDA build for
-      # the target archs (A100=8.0, A10/A40=8.6, L4/L40=8.9, H100=9.0).
+      # torch for the GNN. Standard PyPI cu124 wheel (numpy-2 compatible) coexists
+      # with the RAPIDS cuda-12 libs. Unlike the NGC custom torch, torch_scatter has
+      # a PREBUILT wheel for this torch (the pyg index) — no CUDA compile needed.
       - type: docker
         run:
+          - pip install --no-cache-dir torch==2.6.0 --index-url https://download.pytorch.org/whl/cu124
           - pip install --no-cache-dir torch_geometric lightning
-          - FORCE_CUDA=1 TORCH_CUDA_ARCH_LIST="8.0;8.6;8.9;9.0" pip install --no-cache-dir --no-build-isolation torch_scatter
-      # Pin to a specific commit: segger's CLI flags AND output schema change on
-      # trunk (0.1.0 renamed --prediction-expansion-ratio -> --prediction-graph-
-      # buffer-ratio and dropped the `keep` column). script.py targets THIS commit's
-      # contract, so bump this deliberately and re-check `segger segment --help`
-      # and data/writer.py before doing so.
+          - pip install --no-cache-dir torch_scatter -f https://data.pyg.org/whl/torch-2.6.0+cu124.html
+      # I/O stack. spatialdata pinned <0.8 (resolves to 0.7.1): 0.7.2+ require
+      # dask>=2026.x, but RAPIDS 25.04 hard-pins dask==2025.2.0. 0.7.1 accepts dask
+      # 2025.2.0; pin dask/distributed so spatialdata can't drag them up (which would
+      # break RAPIDS's dask-cudf/cugraph). We validated sd.transform works on 2025.2.0.
+      - type: docker
+        run:
+          - pip install --no-cache-dir "spatialdata>=0.7,<0.8" "zarr>=3.0.0" geopandas shapely "rasterio>=1.3" scikit-image "dask==2025.2.0" "distributed==2025.2.0"
+      # Pin segger to a commit: its CLI flags AND output schema change on trunk
+      # (0.1.0 renamed --prediction-expansion-ratio -> --prediction-graph-buffer-
+      # ratio and dropped the `keep` column). script.py targets THIS commit; bump
+      # deliberately and re-check `segger segment --help` and data/writer.py first.
       - type: python
         github: dpeerlab/segger@0233cf62eb177b6e44e49318037705550765a010
-      # cudf 25.4 (cuda-12; there is no cuspatial-cu13, so we are stuck on this
-      # RAPIDS) hard-requires pandas<2.2.4. But segger pulls the modern single-cell
-      # stack — anndata 0.13 and scanpy 1.12 — which BOTH require pandas>=2.3
-      # (anndata 0.13's zarr reader calls pd.StringDtype(na_value=...), an API added
-      # in pandas 2.3), so under the pinned pandas 2.2.3 reading ANY AnnData/zarr
-      # table crashes with "StringDtype.__init__() got an unexpected keyword
-      # argument 'na_value'". Pin pandas back into cudf's range AND hold anndata /
-      # scanpy at their last pandas-2.2-compatible majors (anndata 0.12.x needs
-      # pandas>=2.1; scanpy 1.11.x needs pandas>=1.5) as the LAST step so they win
-      # over the deps above. numpy stays 2.x. See NOTES.md for the full conflict.
+      # Final step, must be LAST (after segger's github install pulls its deps):
+      #  - cudf 25.4 needs pandas<2.2.4; but anndata 0.13 / scanpy 1.12 (pulled by
+      #    segger) need pandas>=2.3 (anndata 0.13's zarr reader uses
+      #    pd.StringDtype(na_value=...), a pandas-2.3 API) -> reading any zarr table
+      #    crashes under pandas 2.2.3. Hold anndata/scanpy at their last pandas-2.2
+      #    majors (anndata 0.12.x: pandas>=2.1; scanpy 1.11.x: pandas>=1.5), and
+      #    re-assert spatialdata<0.8 + dask==2025.2.0 in case a dep nudged them.
+      #  - swap opencv-python -> opencv-python-headless: segger pulls opencv-python,
+      #    whose cv2 needs libGL.so.1 (a system lib we cannot apt-install as non-root);
+      #    the headless build provides the same cv2 without it. See NOTES.md.
       - type: docker
         run:
-          - pip install --no-cache-dir "pandas>=2.0,<2.2.4" "anndata>=0.12,<0.13" "scanpy>=1.11,<1.12"
+          - pip install --no-cache-dir "pandas>=2.0,<2.2.4" "anndata>=0.12,<0.13" "scanpy>=1.11,<1.12" "spatialdata>=0.7,<0.8" "dask==2025.2.0" "distributed==2025.2.0"
+          - pip uninstall -y opencv-python || true
+          - pip install --no-cache-dir opencv-python-headless
   - type: native
 
 runners: