CoSMIM

Covariance of Spatial Multi-scale Integrated Microenvironment: a training-free representation for spatial domain identification.

Many spatial domain methods describe the neighbourhood of a cell by its mean gene expression. CoSMIM describes it by the gene–gene covariance of the neighbourhood at several spatial scales (the cell and its 5, 15 and 45 nearest neighbours). The scales are fused with weights learned for each section, the fused matrix is shrunk towards a scaled identity, and the matrix square root is used as a Euclidean feature. Standard PCA and Leiden clustering of this feature give the spatial domains; nothing is trained.

Fig. 1: Overview of CoSMIM.

Install

pip install "cosmim[metrics]"   # [metrics] adds squidpy, needed for Moran's I and Geary's C

To use the package versions the paper's results were computed with:

git clone https://github.com/barondu/cosmim.git
cd cosmim
conda env create -f environment.yml
conda activate cosmim
pip install -e .

CoSMIM needs no GPU.

Quick start

import scanpy as sc
import cosmim

adata = sc.read_h5ad("data/D4_STARmap_VisualCortex/STARmap_20180505_BY3_1k.h5ad")
result = cosmim.run(adata, annotation_key="Region", strategy="svg_only", n_top_hvg=64,
                    n_pcs=30, n_neighbors=200)
result.obs["cosmim"]                                           # the spatial domains
cosmim.calc_all_metrics(result, gt_key="Region", pred_key="cosmim")   # the nine metrics

run clusters a section into as many domains as annotation_key has annotated domains. On a section without an annotation, give the number of domains (max_labels=...), or call the steps directly and give a Leiden resolution:

adata_raw, adata_hvg = cosmim.spatial_preprocess_pipeline(adata, strategy="svg_hvg", n_top_hvg=64)
features, weights = cosmim.compute_cosmim(adata_hvg, k=[5, 15, 45], entropy_reg=10, alpha=0.1)
adata_raw.obs["cosmim"] = cosmim.run_leiden_partition(features, resolution=1.0, n_pcs=30,
                                                      n_neighbors=50)

Parameters

Tutorials

The notebooks were run with the versions in environment.yml on Linux (aarch64). On other machines and library builds a few cells can change cluster, so the scores can differ from the paper.

Benchmark

Fig. 2: Benchmarking of 16 spatial domain identification methods on six datasets (22 sections). Sixteen methods on dataset1–6, scored with eight metrics. CoSMIM has the highest overall score (summed rank 121 of 128; the second method has 106). Tutorial 2 computes the ranks and the statistical tests from the per-section scores.

Data

The seven datasets of the paper. Datasets 1–6 are those distributed with SDMBench (Yuan et al., Nat. Methods 21, 712–722 (2024)), whose published scores of eleven methods are part of the benchmark table.

Citation

The preprint is on bioRxiv: https://doi.org/10.64898/2026.10.20.760373

@article{du2026cosmim,
  title   = {Multi-scale microenvironment covariance for unsupervised spatial domain identification},
  author  = {Du, Bairui and Kiryu, Hisanori},
  journal = {bioRxiv},
  year    = {2026},
  doi     = {10.64898/2026.10.20.760373}
}

License

The code is under the BSD 3-Clause License (see LICENSE). The data files keep the terms of their sources listed above; the figures are those of the paper.