Analyzing Spatially-Resolved Omics Data

The way cells are organized and interact with each other is important for organ development, disease progression, and treatment response. Recent advances in single-cell sequencing have allowed us to identify many different types of cells and pathways involved in diseases, but this method removes spatial information. Spatially-resolved omics methods preserve this information, but analyzing the resulting data is challenging.

To address this challenge, a new library called Multi-Omics Spatial Networks Analysis (mosna) has been developed to analyze spatially-resolved omics data and identify relevant cellular interactions and tissue organization features. Mosna includes tools for computing statistical estimates of preferential interactions between cell subsets, discovering local cellular niches, and identifying biomarkers of response to treatment. These features can be used to develop machine learning models to predict disease progression and treatment response.

The study uses a library called tysserand to represent tissues as spatial networks, where nodes are cells and edges between cells represent their physical interactions. This representation is compatible with different types of spatially resolved omic methods. Indeed, the study proposes an analysis pipeline of increasing refinement to test hypotheses and assess clinical relevance and predictive power of variables. The pipeline is exemplified using a spatially resolved proteomic data of Cutaneous T-Cell Lymphoma (CTCL) treated with anti-PD-1 immunotherapy.

The study tested whether cell type proportions are predictive of patients' status, but found no statistically significant difference between responders and non-responders. It then hypothesized that an imbalance in the abundances of specific cell types could explain clinical outcomes and generated composed variables by computing ratios of cell type proportions. However, none of these variables had predictive power. Finally, the study computed second order composed variables and obtained 1054 statistically significant variables without FDR correction. When using composed variables with p-values below 0.05 and 0.005 without FDR correction, bi-clustering of patients and composed variables showed a very good blind clustering of patients per response group.

The study found five pairs of cell type interactions that were significantly different between response groups, and a logistic regression model trained on these interactions had perfect performance in predicting response. The Neighbors Aggregation Statistics method was also used to identify local communities or niches of specific cell type proportions or marker levels. However, when aggregating the preferential interactions in samples per patient, the LR model did not perform as well, likely due to a lower number of observations.

So, mosna uses network analysis to identify local niches within tissue samples, and can evaluate the distribution of cell types across niches to predict response to therapy. The approach is compatible with all types of spatially-resolved omics experiments and can be applied to annotated cell types, raw markers data, or any pre-computed attribute per cell.

Mosna is also able to analyze networks reconstructed with segmentation masks, which represent observable direct contacts between cells. The discovery of niches is performed jointly on all samples, allowing for the identification of rarer or smaller spatial patterns that occur across images. The interpretation of these variables is crucial for personalized medicine, and mosna will be a valuable tool for biological and clinical advancements in this area.