Skip to main content
Back to timeline
bioRxivSource publication:

CIRCA splits cells from their niches into identifiable programs, reading pathology-linked interactions without preset gene sets

Synopsis

Spatial transcriptomics measures gene expression in each cell and its neighbors, but conventional nonlinear models learn non-identifiable latents, so several representations can explain the same observations. CIRCA treats a central cell and its neighborhood as two feature groups and uses grouped causal representation learning to jointly learn identifiable cell- and niche-level transcriptional programs and their directed interactions under explicit assumptions, optionally adding pathology measurements. Across seven cell-resolution datasets, CIRCA had the highest biological accuracy in six and the highest overall score in all five multi-section datasets; in an Alzheimer's mouse model, adding amyloid plaque pathology resolved four region-specific plaque niches.

Source-provided article image: CIRCA learns identifiable cell- and niche-level programs and their interactions with tissue pathology from spatial transcriptomics
Figure 1 · bioRxiv · Page 39

Interpretation

CIRCA treats each cell and its local neighborhood in spatial transcriptomics as two separately sampled but interacting feature groups, yielding both cell-level and niche-level transcriptional programs plus the directed interaction structure linking them. Earlier spatial methods combined gene expression with spatial proximity to define spatial domains and cellular niches, or modeled how the microenvironment influences cellular expression, but few jointly learned the transcriptional programs themselves and their interactions; routes relying on pathway and cell-cell communication databases are limited to curated knowledge, and their autoencoder-derived programs are not guaranteed to be unique. On seven datasets spanning multiple technologies and gene panels from 33 to 6,000 genes, compared with five methods (BANKSY, SEDR, GraphST, Novae and NicheCompass), CIRCA achieved the highest biological accuracy in six of seven datasets and ranked second in the remaining one, and achieved the highest overall score in all five multi-section datasets.

Without predefined gene sets or interaction databases, the model yields reproducible, interpretable transcriptional programs and uses Integrated Gradients to attribute each latent factor to genes and pathways. Conventional nonlinear representation learning such as autoencoders is non-identifiable: even with unlimited data, multiple distinct latent representations can explain the same observations, so learned factors need not map stably onto specific biological programs. On CosMx data from an Alzheimer's mouse model, 39 of 50 cell-level factors and 44 of 50 niche-level factors replicated in at least two of three independent training runs, with mean absolute Spearman correlations of 0.70 and 0.72 among replicated factors; both pathology-associated factors were recovered in all three runs, with a mean absolute correlation of 0.98.

Perspective

The demonstrated scope is seven cell-resolution spatial transcriptomics datasets used for benchmarking plus two biological applications: CosMx data from an Alzheimer's mouse model, comprising hippocampal sections from four 18-month-old AppNL-G-F knock-in mice, and human colorectal cancer Visium HD data, comprising three tumor sections and two normal adjacent tissue sections. The intended reader is a researcher who needs to extract interpretable programs and microenvironment structure from spatial transcriptomics.

The identifiability guarantees rely on assumptions about the data-generating process that may not hold in real tissue; even when they hold, theoretical identifiability does not guarantee accurate recovery of latent factors and interactions in finite samples, so inferred interactions cannot be confirmed as causal relationships in real data. Systematic within-section technical variation may also be captured by latent factors. Many available datasets contain samples from relatively few individuals, limiting assessment of associations between factors or spatial niches and patient-level phenotypes and the generalizability of the findings; the Alzheimer's analysis is restricted to a mouse model, so the identified plaque-associated programs and their regional differences may not fully generalize to human pathophysiology.