- The paper introduces a rigorous framework for defining tissue graph counterfactuals, demonstrating that spatial interventions improve counterfactual queries.
- It presents Cellina, a dual-latent graph variational autoencoder that separates cell identity and spatial context using adversarial training and contrastive loss.
- Empirical results on colorectal cancer and mouse brain datasets reveal state-of-the-art performance and enhanced biological interpretability over existing benchmarks.
Querying Counterfactuals on Tissue Graphs with Supervised Disentanglement: Technical Analysis
This work introduces a rigorous definition of tissue graph counterfactuals, conceptualizing them as spatial interventions that operate at two mechanistic levels: (1) edge perturbations, altering the cellular neighborhood topology; and (2) node perturbations, modifying the gene expression states of spatial neighbors. This formalization advances the field by unifying previously disparate approaches—some treating spatial context as a mere covariate, others ignoring heterogeneity in neighbor composition. The proposed framework generalizes interventional settings in spatial omics and enables precise queries on how a cell's transcriptome would adapt to alternative microenvironments.
Cellina: Architecture and Supervised Disentanglement
Cellina is a dual-latent graph variational autoencoder employing explicit supervised disentanglement. The model infers, for each cell v, two independent latent codes: an intrinsic code z (cell identity) and an extrinsic code s (spatial context), both with standard normal priors.
The cell-intrinsic representation z is anchored via a cell-type classification loss, constraining it to encode only cell-specific information while adversarially minimizing spatial-domain recoverability. This is realized by including an adversarial domain discriminator, whose gradients are reversed during encoder training to encourage domain-invariance of z. Neighbor-driven information is isolated in s, which is estimated through two alternatives: a mean-aggregated MLP of neighbor expressions (base Cellina), or a local subgraph GATv2 (Cellina-GAT) with an explicit contrastive loss that leverages graph-defined neighborhood structures and domain identities. The output gene expression is decoded via a negative binomial likelihood, consistent with single-cell count data modeling standards.
Supervised disentanglement is critical. The paper demonstrates that fully unsupervised factorization of latent structure is non-identifiable, leading to leakage of domain context into the cell-identity code. Injecting biological supervision rectifies this, producing independent axes for precise counterfactual querying.
Experimental Design and Benchmarks
Evaluation of counterfactual accuracy is structured around context transfer, wherein cell types and spatial domains are split (leave-one-cell-type-out within target domain), with the model tasked to predict held-out gene expression under novel spatial neighborhoods. Performance is quantified by Pearson correlation, signed precision, root mean squared error (logFC), and local E-distance metrics, emphasizing gene-level recovery and global distributional fidelity.
Baseline competitors include both non-spatial perturbation models (scGen, CPA), population-average methods (mean shift), and the most recent spatially-aware methods (MintFlow, SpatialProp). Notably, state-of-the-art models like SIMVI are omitted from primary comparison owing to infeasibility at scale and a lack of native counterfactual prediction support.
Cellina is benchmarked on a 2.4 million cell colorectal cancer spatial transcriptomic dataset as well as a whole-mouse-brain MERFISH cohort, supporting comparative cross-tissue and cross-species validation.
Empirical Results
Cellina and its GAT variant achieve robust state-of-the-art performance. The base Cellina outperforms all spatially-informed and spatially-unaware baselines in both Pearson correlation and signed precision—by +0.14 and +0.17 respectively over MintFlow on colorectal cancer, with Cellina-GAT yielding marginally superior correlation (0.85). Node perturbation tasks further demonstrate Cellina's superiority; its node-intervention predictions consistently dominate SpatialProp, with a +0.33 improvement in Pearson correlation and −5.21 in RMSE.
On the mouse brain cohort, Cellina variants retain top ranking for three out of four metrics, demonstrating the method's generalizability and data efficiency. The model also maintains computational tractability, training substantially faster than comparable GNN-based methods.
Disentanglement and Biological Validity
A central claim, empirically validated, is that Cellina's supervised disentanglement confers measurable gains in both counterfactual accuracy and interpretability of latent structure. Bio-conservation benchmarking confirms that z and s capture orthogonal biological axes (cell type and spatial domain), achieving superior clustering consistency relative to MintFlow and SIMVI.
Cellina's spatial latent s recovers discrete—yet annotation-agnostic—subdomains within tumor tissue, mapping to known signaling programs such as TGFβ and NFKB/MAPK activation. Simulated pathway-specific neighbor perturbations, informed solely by pathway-gene dictionaries, reproduce observed domain stratification of fibroblast gene expression. These results underscore the model's potential for hypothesis generation on regulatory programs governing microenvironmental reactivity.
Limitations
The approach is dependent on the quality and granularity of pre-defined cell-type and domain labels, constraining the fidelity of spatial disentanglement. The analysis presupposes accurate cell segmentation from imaging-based spatial transcriptomics, which remains an active technical limitation due to widespread misassignment of transcripts. Furthermore, Cellina provides operational (rather than causal) counterfactuals; the current implementation does not enforce the stronger identifiability constraints required by structural causal models, though the architecture is extendable in this direction.
Implications and Future Prospects
Cellina establishes a unified, mechanistically grounded framework for spatial counterfactual inference at single-cell resolution. Practically, this method enables in silico prediction of perturbational responses under arbitrary local tissue reforms—supporting both virtual knock-in/out studies and exploration of combinatorial neighborhood effects. The disentangled latent codes are immediately applicable for refining spatially-resolved cell atlases, guiding experimental design for spatial perturbation screens, and integrating with causal inference frameworks incorporating intervention-specific priors (e.g., sparse mechanism shift modeling).
Potential extensions include integration of raw imaging data for more robust spatial context learning, compositional modeling of higher-order cellular circuits, and explicit enforcement of identifiability and mechanism sparsity for enhanced interpretability and causal discovery.
Conclusion
This paper defines a comprehensive theory and practical implementation for querying counterfactuals on tissue graphs, overcoming the limitations of prior i.i.d. and label-swap protocols. Supervised disentanglement via Cellina provides biologically meaningful and highly accurate counterfactual predictions, enabling not only methodological advances for spatial omics but also translational applications in virtual tissue engineering and mechanistic cell biology. The approach forms a foundation for increasingly data-efficient and causally-grounded models of cellular microenvironmental response.