Fixes

src/methods_transcript_assignment/fastreseg/config.vsh.yaml

@@ -46,24 +46,45 @@ engines:
           - gdal-bin
           - libgdal-dev
           - r-base
-      - type: r
-        packages:
-          - devtools
-          - terra
-          - Matrix
-          - dbscan
-          - igraph
-          - matrixStats
-          - codetools
-          - data.table
-          # Add other CRAN packages here
-        github:
-          # Add GitHub packages here if needed
-          - Nanostring-Biostats/FastReseg
-      #- type: python
-      #  github:
-      #    - openproblems-bio/core#subdirectory=packages/python/openproblems
-    __merge__: 
+          # FastReseg's R dependencies as prebuilt Debian binaries. Installing
+          # these via apt avoids compiling terra/igraph/sf/spatstat/stringi/etc.
+          # from source (extremely slow, especially under amd64 emulation on
+          # arm64 hosts). Only concaveman + GiottoUtils/GiottoClass + FastReseg
+          # itself still build from source below; their heavy deps are already
+          # present as binaries here.
+          - r-cran-remotes
+          - r-cran-terra
+          - r-cran-igraph
+          - r-cran-sf
+          - r-cran-matrix
+          - r-cran-mass
+          - r-cran-e1071
+          - r-cran-dbscan
+          - r-cran-data.table
+          - r-cran-dplyr
+          - r-cran-ggplot2
+          - r-cran-reshape2
+          - r-cran-stringr
+          - r-cran-fs
+          - r-cran-lmtest
+          - r-cran-matrixstats
+          - r-cran-geometry
+          - r-cran-spatstat.geom
+      - type: docker
+        # Install FastReseg from GitHub. upgrade = "never" so remotes reuses the
+        # apt-installed binary deps above instead of re-fetching/recompiling them
+        # from CRAN; only the unpackaged deps (concaveman, GiottoUtils,
+        # GiottoClass) and FastReseg itself are compiled from source.
+        run:
+          - Rscript -e 'options(warn = 2); remotes::install_github("Nanostring-Biostats/FastReseg", upgrade = "never", repos = "https://cran.rstudio.com")'
+      - type: python
+        # `plankton` (pulled in by txsim.run_ssam for ssam cell annotation in
+        # input.py) still does `from matplotlib.cm import get_cmap`, removed in
+        # matplotlib 3.9. Pin < 3.9 so run_ssam imports. This is a local setup
+        # step, so it runs AFTER the __merge__'d spatialdata/txsim partials (which
+        # otherwise pull matplotlib >= 3.9) and wins. Mirrors methods_cell_type_annotation/ssam.
+        pypi: [planktonspace, "matplotlib<3.9"]
+    __merge__:
       - /src/base/setup_txsim_partial.yaml
       - /src/base/setup_spatialdata_partial.yaml
   - type: native

src/methods_transcript_assignment/fastreseg/input.py

@@ -7,6 +7,61 @@
 import txsim as tx
 import anndata as ad
 import argparse
+from scipy.sparse import csr_matrix
+
+
+def generate_adata(input_spots, cell_id_col="cell", gene_col="Gene"):
+    """Aggregate per-transcript assignments into a cell x gene AnnData.
+
+    Reimplementation of txsim.preprocessing.generate_adata: txsim's version is
+    incompatible with anndata>=0.12 (it passes the removed `dtype=` argument to
+    AnnData() and uses per-row item assignment `adata[cell, :] = ...`, which
+    anndata no longer supports). This fills the count matrix directly and
+    produces an identical output structure (X, layers['raw_counts'],
+    obs/var stats, uns['spots'/'pct_noise']). Kept in sync with the copy in
+    src/methods_count_aggregation/basic_count_aggregation/script.py.
+    """
+    spots = input_spots.copy()
+    pct_noise = sum(spots[cell_id_col] <= 0) / len(spots[cell_id_col])
+    spots_raw = spots.copy()  # kept in uns['spots']; 0 (background) -> None
+    spots_raw.loc[spots_raw[cell_id_col] == 0, cell_id_col] = None
+    spots = spots[spots[cell_id_col] > 0]
+
+    cell_ids = pd.unique(spots[cell_id_col])
+    genes = pd.unique(spots[gene_col])
+    cell_pos = {c: i for i, c in enumerate(cell_ids)}
+
+    # Populate the count matrix + centroids per cell (no AnnData item assignment).
+    # feature_name may be categorical, so value_counts() can return unobserved
+    # categories; reindex to `genes` (fill 0) to align to the var order, mirroring
+    # txsim's `.reindex(var_names, fill_value=0)`.
+    X = np.zeros((len(cell_ids), len(genes)), dtype=np.float32)
+    centroid_x = np.zeros(len(cell_ids))
+    centroid_y = np.zeros(len(cell_ids))
+    for cell_id, grp in spots.groupby(cell_id_col, sort=False):
+        row = cell_pos[cell_id]
+        cts = grp[gene_col].value_counts().reindex(genes, fill_value=0)
+        X[row, :] = cts.values
+        centroid_x[row] = grp["x"].mean()
+        centroid_y[row] = grp["y"].mean()
+
+    adata = ad.AnnData(csr_matrix(X))
+    adata.obs["cell_id"] = cell_ids
+    adata.obs_names = [f"{i:d}" for i in cell_ids]
+    adata.var_names = genes
+    adata.obs["centroid_x"] = centroid_x
+    adata.obs["centroid_y"] = centroid_y
+
+    adata.uns["spots"] = spots_raw
+    adata.uns["pct_noise"] = pct_noise
+    adata.layers["raw_counts"] = adata.X.copy()
+
+    adata.obs["n_counts"] = np.ravel(adata.layers["raw_counts"].sum(axis=1))
+    adata.obs["n_genes"] = adata.layers["raw_counts"].getnnz(axis=1)
+    adata.var["n_counts"] = np.ravel(adata.layers["raw_counts"].sum(axis=0))
+    adata.var["n_cells"] = adata.layers["raw_counts"].getnnz(axis=0)
+    return adata
+
 
 def parse_arguments():
     parser = argparse.ArgumentParser(
@@ -70,8 +125,12 @@ def parse_arguments():
 print('Transforming transcripts coordinates', flush=True)
 transcripts = sd.transform(sdata['transcripts'], to_coordinate_system='global')
 
-transcripts_df = transcripts.compute()
-transcripts_df.rename(columns = {'feature_name': 'target', 
+# .copy() is required: for a single-partition points object, transcripts.compute()
+# returns a reference to the dask graph's backing frame, so an in-place rename here
+# would corrupt it and the later transcripts.compute() (passed to run_ssam) would
+# yield renamed columns that no longer match the dask _meta -> "Metadata mismatch".
+transcripts_df = transcripts.compute().copy()
+transcripts_df.rename(columns = {'feature_name': 'target',
 'transcript_id': 'UMI_transID', 'cell_id': 'UMI_cellID'}, inplace = True)
 
 transcripts_df = transcripts_df.loc[:, ['target', 'x', 'y', 'z', 'UMI_transID', 'UMI_cellID']]
@@ -82,7 +141,7 @@ def parse_arguments():
 df = sdata['transcripts'].compute()
 df.feature_name = df.feature_name.astype(str)
 
-adata_sp = tx.preprocessing.generate_adata(df, cell_id_col='cell_id', gene_col='feature_name') #TODO: x and y refers to a specific coordinate system. Decide which space we want to use here. (probably should be handled in the previous assignment step)
+adata_sp = generate_adata(df, cell_id_col='cell_id', gene_col='feature_name') #TODO: x and y refers to a specific coordinate system. Decide which space we want to use here. (probably should be handled in the previous assignment step)
 adata_sp.layers['counts'] = adata_sp.layers['raw_counts']
 del adata_sp.layers['raw_counts']
 adata_sp.var["gene_name"] = adata_sp.var_names

src/methods_transcript_assignment/fastreseg/orchestrator.sh

@@ -1,5 +1,10 @@
 #!/bin/bash
 
+# Fail loudly: without this, a crash in any sub-step (input.py / script.R /
+# output.py) is swallowed and Viash only reports the generic "required output
+# file is missing", hiding the real Python/R traceback.
+set -eo pipefail
+
 ## VIASH START
 # The following code has been auto-generated by Viash.
 par_input_ist='resources_test/task_ist_preprocessing/mouse_brain_combined/raw_ist.zarr'