Error converting integer scalar during aggregate_neighbors

[tadams1205]

Report

I'm following the tutorial from the CosMX NSCLC dataset, but I'm instead using two tissue array slides of Xenium data. There are 48 tissue punches in the array, and I'm using these for the scvi batch_key and downstream library_key in spatial_neighbors and aggregate_neighbors.

Everything seems fine until I call cc.gr.aggregate_neighbors(), where I make it 12% of the way through when i receive an error message about converting integer scalar.

cc.gr.aggregate_neighbors(adata, n_layers=3, use_rep='X_scVI', out_key='X_cellcharter', sample_key='TMApunch')

12%|███████████████ | 6/48 [00:50<05:56, 8.50s/it]
Traceback (most recent call last):
File "", line 1, in
File "/home/tsa27/.conda/envs/cellcharter-env/lib/python3.12/site-packages/cellcharter/gr/_aggr.py", line 170, in aggregate_neighbors
adj=adata[idxs].obsp[connectivity_key],
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^
File "/home/tsa27/.conda/envs/cellcharter-env/lib/python3.12/site-packages/anndata/_core/aligned_mapping.py", line 164, in getitem
_subset(self.parent_mapping[key], self.subset_idx),
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/tsa27/.conda/envs/cellcharter-env/lib/python3.12/functools.py", line 912, in wrapper
return dispatch(args[0].class)(*args, **kw)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/tsa27/.conda/envs/cellcharter-env/lib/python3.12/site-packages/anndata/_core/index.py", line 200, in _subset_sparse
return a[subset_idx]
~^^^^^^^^^^^^
File "/home/tsa27/.conda/envs/cellcharter-env/lib/python3.12/site-packages/scipy/sparse/_index.py", line 104, in getitem
res = self._get_arrayXarray(row, col)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/tsa27/.conda/envs/cellcharter-env/lib/python3.12/site-packages/scipy/sparse/_compressed.py", line 565, in _get_arrayXarray
csr_sample_values(M, N, self.indptr, self.indices, self.data,
ValueError: could not convert integer scalar

I've looked at the delauney graph layouts for TMAs 5-8 (where it seemingly fails?) and there's nothing particularly strange about them, though there are sometimes parts of the graphs where one or two cells are disconnected from the rest of the graph.

Converting the connectivity matrix into integer class does not change the outcome.

Reducing n_layers=2 does not change the outcome.

This was run with an RTX A5000, I can try other GPUs if that's a factor.

my adata object looks like this:

AnnData object with n_obs × n_vars = 831791 × 479
obs: 'fov', 'AspectRatio', 'CenterX_global_px', 'CenterY_global_px', 'nCount_Xenium', 'nFeature_Xenium', 'TMApunch', 'SampleID', 'PatientID', 'Cohort', 'HistoSlide', 'Disease', 'cellTypeFibrosis', 'cell_ID', 'slide', '_scvi_batch', '_scvi_labels'
var: 'n_counts'
uns: 'spatial_fov', 'log1p', '_scvi_uuid', '_scvi_manager_uuid', 'spatial_neighbors', 'TMApunch_colors'
obsm: 'spatial', 'spatial_fov', 'X_scVI'
layers: 'counts'
obsp: 'spatial_connectivities', 'spatial_distances'

here is my session info:

---

anndata 0.12.7
cellcharter 0.3.5
lightning 2.6.0
matplotlib 3.10.8
numpy 2.3.5
pandas 2.3.3
scanpy 1.11.5
scvi 1.4.1
session_info v1.0.1
squidpy 1.7.0

PIL 12.1.0
absl 2.3.1
affine 2.4.0
attr 25.4.0
backports NA
certifi 2026.01.04
charset_normalizer 3.4.4
click 8.3.1
cloudpickle 3.1.2
cuda 12.9.4
cycler 0.12.1
cython_runtime NA
dask 2024.11.2
dask_image NA
datashader 0.18.2
dateutil 2.9.0.post0
docrep 0.3.2
donfig 0.8.1.post1
fsspec 2026.1.0
geopandas 1.1.2
google_crc32c NA
h5py 3.15.1
idna 3.11
imagecodecs 2025.11.11
imageio 2.37.2
jaraco NA
jinja2 3.1.6
joblib 1.5.3
kiwisolver 1.4.9
lazy_loader 0.4
legacy_api_wrap NA
lightning_fabric 2.6.0
lightning_utilities 0.15.2
llvmlite 0.46.0
markupsafe 3.0.3
matplotlib_scalebar 0.9.0
ml_collections 1.1.0
more_itertools 10.8.0
mpl_toolkits NA
mpmath 1.3.0
mudata 0.3.2
multipledispatch 0.6.0
multiscale_spatial_image 2.0.3
natsort 8.4.0
networkx 3.6.1
numba 0.63.1
numcodecs 0.16.5
ome_zarr NA
opt_einsum 3.4.0
packaging 25.0
pathlib_abc NA
patsy 1.0.2
pkg_resources NA
platformdirs 4.5.1
pooch v1.8.2
pyarrow 23.0.0
pyct 0.6.0
pygments 2.19.2
pyparsing 3.3.2
pyproj 3.7.2
pyro 1.9.1
pytorch_lightning 2.6.0
pytz 2025.2
rasterio 1.5.0
requests 2.32.5
rich NA
scipy 1.17.0
seaborn 0.13.2
shapely 2.1.2
six 1.17.0
skimage 0.26.0
sklearn 1.8.0
sknw 0.1
sparse 0.17.0
spatial_image 1.2.3
spatialdata 0.6.1
spatialdata_plot 0.2.13
statsmodels 0.14.6
sympy 1.14.0
testing NA
threadpoolctl 3.6.0
tifffile 2026.1.14
tlz 1.1.0
toolz 1.1.0
torch 2.10.0+cu128
torchgen NA
torchgmm NA
torchmetrics 1.8.2
tqdm 4.67.1
triton 3.6.0
typing_extensions NA
upath 0.3.8
urllib3 2.6.3
validators 0.35.0
xarray 2025.12.0
xarray_dataclass 3.0.0
xarray_schema 0.0.3
xrspatial 0.5.2
yaml 6.0.3
zarr 3.1.5
zoneinfo NA

Python 3.12.12 | packaged by conda-forge | (main, Oct 22 2025, 23:25:55) [GCC 14.3.0]
Linux-4.18.0-553.87.1.el8_10.x86_64-x86_64-with-glibc2.28

Session information updated at 2026-01-21 20:43

Thank you for your time!

Version information

No response

[tadams1205]

The problem was with one specific sample (TMApunch #18). It is the only sample that I cannot plot with with sq.pl.spatial_scatter - which returns the same "integer scalar" error, and it's the only sample that gives me this error when I try to subset the index from the connectivity matrix a la:

for i, sample in enumerate(samples):
    try:
        idxs = adata.obs['TMApunch'] == sample
        test = adata[idxs].obsp[connectivity_key]
        print(f"{i}: Sample {sample} - OK")
    except Exception as e:
        print(f"{i}: Sample {sample} - FAILED: {e}")
        corrupted_samples.append(sample)

0: Sample 1 - OK
1: Sample 2 - OK
2: Sample 3 - OK
3: Sample 4 - OK
4: Sample 5 - OK
5: Sample 6 - OK
6: Sample 7 - OK
7: Sample 8 - OK
8: Sample 9 - OK
9: Sample 10 - OK
10: Sample 11 - OK
11: Sample 12 - OK
12: Sample 13 - OK
13: Sample 14 - OK
14: Sample 15 - OK
15: Sample 16 - OK
16: Sample 17 - OK
17: Sample 18 - FAILED: could not convert integer scalar
18: Sample 19 - OK
19: Sample 20 - OK
20: Sample 21 - OK
21: Sample 22 - OK
22: Sample 23 - OK
23: Sample 24 - OK
24: Sample 25 - OK
25: Sample 26 - OK
26: Sample 27 - OK
27: Sample 28 - OK
28: Sample 29 - OK
29: Sample 30 - OK
30: Sample 31 - OK
31: Sample 32 - OK
32: Sample 33 - OK
33: Sample 34 - OK
34: Sample 35 - OK
35: Sample 36 - OK
36: Sample 37 - OK
37: Sample 38 - OK
38: Sample 39 - OK
39: Sample 40 - OK
40: Sample 41 - OK
41: Sample 42 - OK
42: Sample 43 - OK
43: Sample 44 - OK
44: Sample 45 - OK
45: Sample 46 - OK
46: Sample 47 - OK
47: Sample 48 - OK

The spatial coordinates, raw gene expression values and scVI output values for this sample seem totally fine as far as I can tell. The problem is specifically in the connectivity data output when squidpy is run on this sample, which always seems to be corrupted or something.

This particular sample is representative of a unique disease in our study, the properties of which make it an outlier in terms of having the highest total number of cells (47,711) and the highest degree of cell density. Perhaps this causes the Delauney triangulation some problems?

Removing this sample upstream of generating the connectivity matrix solves my problem. While I ultimately would like to include this sample in my Cell Charter analysis, I also understand if the problem is outside of this Cell Charter's scope and you want to close this issue and let me try bothering the squidpy repo maintainers instead.