Fig. 1: Exploring cell state covariation in local niches with NiCo.

The schematic illustrates the three modules of the NiCo pipeline. Annotations: query imaging-based spatial transcriptomics data are annotated by label transfer from reference scRNA-seq data using a soft mutual nearest neighbor approach to derive anchors followed by iterative annotation of non-anchors. Interactions: For each cell type a logistic regression classifier is trained to predict the cell type identity from the niche composition. Predicted coefficients for each cell type are informative on predictive cell types within the niche and a cell type neighborhood graph is derived from the regression coefficients. Covariations: First, latent variables are inferred for each cell type to capture cell state variability using non-negative matrix factorization (NMF). The gene-by-factor matrix is learned simultaneously from reference and query data; if the cell segmentation is imperfect as indicated, e.g., by “spill over” of cell type-specific markers, it can be learned only from reference data and transferred to the spatial modality. Factors can be associated with full transcriptome information based on gene-factor correlations derived from the scRNA-seq data. A ridge regression infers the dependence of each factor of the central cell type on all factors of the niche-cell types. Significance and magnitude of the regression coefficients indicating factor covariation in co-localized cell types, can be inspected in a dot plot. Covarying factors are interrogated for enriched pathways and ligand-receptor pairs to functionally interpret niche interactions. Created in BioRender. Grün, D. (2024) https://BioRender.com/v74f094.