mosna is a Python package for spatial omics data analysis, designed to extract clinically relevant features from single-cell spatial measurements (transcriptomics, proteomics, multiplexed imaging). MOSNA reconstructs spatial networks, computes interaction and neighborhood statistics, and trains predictive models integrating clinical outcomes.
MOSNA’s workflow progresses through increasing analytical complexity, starting with broad hypotheses on cell proportions, then moving to interaction-based metrics (aka "assortativity"), and finally discovering spatial neighboroods (also called niches), leveraging machine learning at each step for predictive modeling.
The pipeline stages:
| Stage | Description |
|---|---|
| 1. Proportions | Quantify global cell-type frequencies |
| 2. Composed Variables | Ratios of proportions |
| 3. Assortativity and mixing matrices | Preferential interaction between cell types |
| 4. Neighbors Aggregation Statistics (NAS) | Cell neighborhoods (niches) from local omics data |
conda create -n mosna-gpu -c rapidsai -c conda-forge -c nvidia -c pytorch \
rapids=23.04.01 python=3.10 cuda-version=11.2 pytorch=1.12.1 scanpy
conda activate mosna-gpu
pip install ipykernel ipywidgets tysserand scipy==1.13
cd /path/to/mosna
pip install -e .conda create -n mosna -c conda-forge python=3.10 scanpy
conda activate mosna
pip install ipykernel ipywidgets tysserand scipy==1.13
pip install -e .- MOSNA computes proportions of cell types per sample.
- Performs statistical tests (e.g., Mann-Whitney U) to identify differences between groups (e.g., responder vs non-responder), correcting with Benjamini-Hochberg.
- Computes first-order (cell-type ratios) and higher-order ratios.
- Enables detection of enriched features not obvious from absolute counts.
- Uses tysserand to build spatial cell networks (e.g., Delaunay triangulation).
- Computes mixing matrices and assortativity coefficients (AC) for categorical attributes (cell types or marker positivity).
- Applies network attribute randomization and z‑scoring to account for cell type proportions bias.
- Identifies interacting cell-type pairs predictive of response (e.g., neutrophil self-interactions, CD8–B cell, etc...).
- Aggregates features (omics data) of each cell and its neighbors to compute local statistics (means, variances, proportions, etc...).
- Clusters these aggregated features (e.g., UMAP + k‑means) to define niches (local microenvironments).
- At each stage, MOSNA allows training elastic‑net penalized logistic regression (for binary outcome) or Cox Proportional Hazard (for survival) models with cross-validation.
- It reports peformance metrics, model coefficients, and identifies top predictive features.
- Typically, as feature sets become more refined (e.g., cell types interactions or niche proportions), predictive power increases.
- The core methodology and pipeline are published in bioRxiv, demonstrating its use on a spatial proteomics dataset of Cutaneous T-Cell Lymphoma (CTCL), generated with the CODEX technology with binary response, and on a a spatial proteomics dataset of breast cancer generated with the IMC technology, with survival (right-censored) data.
- MOSNA leverages tysserand for fast network construction applicable to any spatial omics data.
- Contributions are welcome via GitHub pull requests and issues.
- Licensed under GPL‑3.0, involving open-access reuse conditions.
MOSNA provides:
- A reproducible, interpretable pipeline for spatial network analysis.
- Multiple layers of feature computation: proportions → interactions → niches.
- Predictive modeling tightly integrated with exploratory analysis.
- Support for diverse spatial omics platforms such as CODEX, MERFISH, CyCIF, Visium, etc.
Explore the example notebooks to get hands-on experience and adapt workflows to your own spatial omics datasets.