Determine the causal relationship between NMF-derived Niches 3 and 4

Determine whether the association between Non-negative Matrix Factorization (NMF)-derived Niche 3 and Niche 4 in the FCI-Stable partially directed acyclic graph learned from log-transformed niche composition features of the CosMx colonic mucosa spatial transcriptomics dataset reflects feedback (bidirectional causation) or latent confounding, thereby clarifying the causal directionality and mechanism linking these two niches.

Background

The paper applied the FCI-Stable causal discovery algorithm to three groups of engineered features—niche composition, niche–niche neighborhood enrichment, and niche–gene signals—to reveal potential causal structures linking features and disease status in colonic spatial transcriptomics data. In the causal graph for niche composition features, disease status was connected to Niches 1 and 2, and Niches 3 and 4 appeared upstream of Niche 1.

However, the specific relationship between Niche 3 and Niche 4 was not determined by the algorithm, leaving open whether their association is due to a direct causal influence, bidirectional feedback, or shared unmeasured confounding. Resolving this would refine the mechanistic interpretation of niche interactions and improve the biological interpretability of the causal model.