Querying Counterfactuals on Tissue Graphs with Supervised Disentanglement
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: We're looking at the paper: "Querying Counterfactuals on Tissue Graphs with Supervised Disentanglement".
Marcus: Here's the gist of the paper. Tissue graph counterfactuals ask how a cell’s expression would change under altered spatial neighbor contexts, which is central to predicting cell behavior in tissues The paper introduces Cellina, a framework that uses supervised disentanglement to decompose a cell’s intrinsic state from its spatial context, using the latter as a conditioning input for counterfactual predictions.
Ines: Formalizing Tissue Graph Interventions The authors first formalize tissue graph counterfactuals as a class of spatial interventions that either rewire connections between cells (edge perturbation) or modify the expression of their neighbors (node perturbation) These two types of interventions correspond to the two mutable components of the tissue graph, namely its edges and the expression of its neighbors.
Marcus: Cellina Architecture and Disentanglement Cellina is a (graph) variational autoencoder (VAE) that separates each cell’s gene expression into two latent components: an intrinsic representation z encoding cell identity, and an extrinsic (spatial) representation s encoding the effect of its microenvironment The model uses supervised disentanglement to anchor z to cell-type identity and adversarially remove spatial-domain information from z, routing microenvironmental variation to s by removing it from z.
Ines: Performance on Benchmarks Cellina outperforms spatially-informed and non-spatial competitors in insilico graph perturbations, disentanglement, and scalability across benchmarks spanning over 2.5 million spatially-resolved cells in colorectal cancer and mouse brain Specifically, Cellina’s best variant leads the strongest baseline on colorectal cancer data by +0.17 on both Pearson and Signed Precision.
Title and authors: Ines: So we're looking at this paper, "Querying Counterfactuals on Tissue Graphs with Supervised Disentanglement," and what it's doing is taking those questions about how a cell changes when its neighbors change, which is super important for predicting tissue behavior.
Marcus: Yeah, the core idea here is that we can actually do that in silico using a framework called Cellina to separate what makes a cell itself from what its surroundings are doing.
Ines: Exactly, it introduces this dual-encoder graph variational autoencoder, Cellina, which splits each cell's gene expression into an intrinsic representation z and an extrinsic spatial representation s.
Marcus: The key part is that they use supervised disentanglement to anchor the intrinsic representation z to cell type identity while adversarially removing any spatial domain information from it.
Yuki: I wonder what this means for us in terms of understanding how cells interact in a tissue, because if we can separate the intrinsic state from the spatial context, we might be able to isolate specific neighborhood effects.
Ines: Well, they formalize tissue graph counterfactuals as two types of spatial interventions: rewiring the connections between cells and modifying the expression of their neighbors.
Marcus: That means they’re treating those two things—the edges and the neighbor expression—as the mutable parts of the tissue graph that we can actually perturb.
Yuki: It sounds like a lot to unpack, because it moves us away from just looking at static structures and toward simulating dynamic neighborhood changes.
Ines: They show that Cellina outperforms both spatially-informed and non-spatial competitors on benchmarks covering over two point five million cells in colorectal cancer and mouse brain data.
Marcus: The paper claims that their best variant leads the strongest baseline on colorectal cancer data by plus zero point one seven on both Pearson correlation and signed precision metrics.
Yuki: That's a solid number, but what does that actually tell us about the biological reality of these subdomains they find?
Ines: It enables them to discover biologically distinct cancer subdomains in an unsupervised way just by clustering the latent s over colorectal cancer cells.
Title and authors: Marcus: And they then score those clusters against PROGENy pathway gene sets to assign specific biological identities to each microenvironment, which is pretty powerful for annotation.
Yuki: So it’s not just finding patterns; it’s giving those patterns a functional name based on known pathways.
Ines: Right, and they also use this disentangled spatial representation s to enable targeted neighbor perturbation simulations using existing priors, which we saw in experiment four point four where node perturbations with PROGENy pathway weights recover subdomain effects quite well.
Marcus: That’s a big step because it shows you can actually test hypotheses about how changing the neighbors' expression affects the cell without needing massive amounts of new experimental data for every single perturbation.
Yuki: If we can simulate those targeted perturbations, it lets us ask 'what if' questions about specific signaling programs like TGF beta dominance in colorectal cancer.
Ines: Exactly, and they also show that their latent z and s encode the desired source of variation while adequately removing nuisance sources of variation, which suggests robustness in both latent spaces.
Marcus: And on the speed front, Cellina is one of the fastest to train because they use an efficient pseudobulk based phi v computed a priori, which skips the expensive message passing from graph training.
Yuki: That efficiency is interesting when you consider how much data we’re dealing with in real biology; skipping that message passing saves a lot of time and computational resources.
Ines: They also found that counterfactual predictions are better when conditioned on neighborhoods sampled from the matched subdomain than those from the whole colorectal cancer domain.
Marcus: That suggests that focusing your neighborhood perturbations on a specific biological context yields more accurate results than doing a global simulation.
Yuki: So, this moves us toward using these simulations to refine pathology annotations and understand how different tissue contexts influence cell fate decisions.
Ines: This paper, "Querying Counterfactuals on Tissue Graphs with Supervised Disentanglement," provides a unified framework for studying neighborhood-driven cell responses.
Marcus: It's a principled test of whether the representations we learn truly separate intrinsic cell identity from the microenvironmental effects it experiences.
Title and authors: Yuki: I think what this really changes for us is providing a principled way to study how spatial context drives cellular behavior across different tissues and even species.
Ines: We’re looking at how this model allows us to query counterfactuals, meaning we can ask 'what if' questions about expression changes under altered spatial neighbor contexts, which is central to predicting cell behavior in tissues.
Marcus: And the way they achieve that by decomposing the state into z and s means we have a clear way to interpret those predictions as simulated outcomes under different inputs and contexts.
Yuki: So, looking forward, this framework suggests that encoding neighborhoods continuously yields a strong prior even without any supervision, which is a key factor in Cellina’s performance.
Ines: It’s biologically motivated design choices that are really paying off here across all the metrics they evaluated, including Pearson r and Signed Precision.
Marcus: The limitation they flag is that their supervised disentanglement objectives rely on cell-type and spatial domain annotations, which are simplifications of continuous biology, bounding how well the spatial latent s can reflect neighborhood variation.
Yuki: That’s a fair point; the method stops working where those labels don't capture all the complexity of continuous biological systems.
Ines: They suggest that future methods that integrate raw imaging signal alongside segmented counts could denoise both the expression signal and those counterfactuals derived from it, which they say is compatible with their latent decomposition because it doesn't yet impose the structural assumptions required for identifiability.
Marcus: So, to wrap up, Cellina accurately predicts in silico perturbations in colorectal cancer data, and its performance across various metrics suggests that its design choices are biologically motivated rather than architecturally complex.
Yuki: I think the main implication is that we now have a principled test of whether the representations separate intrinsic from microenvironmental variation.
Ines: This work provides a unified framework for studying neighborhood-driven cell responses, and it gives us a principled test of whether the representations separate intrinsic from microenvironmental variation.
Marcus: It's a powerful tool for simulating how cells respond to altered inputs and contexts across different scales of tissue structure.
The paper's summary: Ines: So, this paper is about taking those questions we talked about—how does a cell change if its neighbors change? It formalizes that as tissue graph counterfactuals, which is just a way of saying we can test how cell behavior shifts based on its spatial context.
Marcus: Right, and they use this framework to build Cellina. It’s a dual-encoder graph variational autoencoder that splits the expression of each cell into two parts: an intrinsic representation called z and an extrinsic one called s.
Yuki: That separation is key, because they use supervised disentanglement to make sure the intrinsic part z actually captures the cell type identity, and then they push all the spatial information into that extrinsic part s.
Ines: Exactly. They use a classifier to anchor z to cell type and an adversarial discriminator to pull spatial stuff out of z, routing it entirely into s. So, when you want to predict what happens under a different neighborhood context, you condition the prediction on that spatial representation s.
Marcus: They’re using this setup for counterfactual predictions. The core idea is that you can use the cell's intrinsic state and then feed in a different spatial state to see how the expression changes.
Yuki: And they show how robust this works by demonstrating it across massive datasets, over two point five million cells from colorectal cancer and mouse brain samples.
Ines: The performance numbers are pretty strong. Their best variant on colorectal cancer data beats competitors on both Pearson correlation and signed precision metrics by a noticeable margin, about zero point one seven points better.
Marcus: That’s significant because they’re talking about simulating these perturbations in silico, which is usually a huge computational hurdle. They show this model is one of the fastest to train because they use some pre-computed values for something called phi v, which skips the expensive message passing step in graph training.
Yuki: It's interesting that they also found that the latent s space actually reveals biologically distinct cancer subdomains without them needing to supervise it with those annotations beforehand.
Ines: They cluster the latent s over colorectal cancer cells and then score those clusters against known pathway gene sets, which allows them to assign biological identities to each microenvironment based on inferred signaling programs, like TGF beta dominance.
Marcus: So they aren't just finding random patterns in the data; they’re giving those patterns a functional meaning by linking them back to known biology.
Yuki: And that leads into their ability to do targeted neighbor perturbation simulations. They use that disentangled s representation to simulate how changing a specific pathway, like PROGENy, affects the cell expression without needing new wet lab experiments for every single change.
Ines: That’s what they’re showing with experiment four point four where node perturbations weighted by PROGENy pathway effects recover the subdomain structure pretty well. It lets you test hypotheses about specific neighborhood effects directly in the model.
Marcus: The limitation they mention is that their supervised disentanglement relies on those cell type and spatial domain annotations, which are just simplifications of continuous biology, so it bounds how well the spatial representation s can actually reflect every nuance of neighborhood variation.
Yuki: That makes sense. It’s a great model for testing the hypothesis that encoding neighborhoods continuously yields a strong prior even when you don't have perfect supervision.
Ines: The paper concludes that this framework gives us a unified way to study how cells respond to neighborhood changes, and it provides a principled test of whether the representations they learned truly separate what’s intrinsic from what’s coming from the microenvironment.
Marcus: It really shows how these latent decompositions are compatible with interpreting those counterfactual predictions as actual simulated outcomes under altered inputs and contexts.
Yuki: This gives us a solid tool for asking those 'what if' questions about cellular behavior across different tissue structures and species, which is where we’ll be headed next.
The paper's improvements: Ines: So, we’re looking at what this paper suggests for future work—how do they plan to make this system even better? It focuses on how to improve the way they learn those intrinsic and extrinsic representations.
Marcus: They point out that their current supervised disentanglement relies on those cell type and spatial domain labels, which are just simplifications of continuous biology, so it bounds how well the spatial representation s can actually reflect every nuance of neighborhood variation.
Yuki: That means they know the current method has limits because those annotations aren't perfectly capturing the messy reality of biological systems.
Ines: Exactly. They suggest that a future direction involves integrating raw imaging signals alongside segmented counts to denoise both the expression signal and those counterfactuals derived from it.
Marcus: That makes sense for improving accuracy, but they say this is compatible with their current latent decomposition because it doesn't yet impose the structural assumptions required for identifiability.
Yuki: So, they’re saying you can clean up the data input without breaking the core mathematical structure of how z and s are learned. That’s a clever way to address that limitation.
Ines: It suggests that if we can get better signal quality from imaging, we might get better predictions for those neighbor perturbations too.
Marcus: They also mention that the latent decomposition itself is compatible with this interpretation because it doesn't yet force the model to assume a specific structural setup for how those two components interact.
Yuki: It’s about improving the input quality so the learning process has more data to work with without fundamentally changing the way we understand intrinsic versus extrinsic variation.
Ines: So, they are suggesting a path where better data cleaning can enhance their counterfactual simulations while keeping the latent structure intact for interpretation.
Marcus: And it’s a practical suggestion because it moves away from just relying on perfect pre-segmented counts, which is often not possible in real-world sequencing data.
Yuki: It’s interesting that they keep the focus on the compatibility with their latent decomposition; they aren't trying to reinvent how z and s work, just how clean we can make the picture we feed into it.
Ines: So, a path forward is using better imaging data to refine those counterfactual predictions while maintaining their current framework for understanding cell behavior.
Marcus: It gives them a clear direction on where the model needs to evolve to handle more complex biological noise and data structures.
Conclusion: Ines: So to wrap up, this paper, "Querying Counterfactuals on Tissue Graphs with Supervised Disentanglement," gives us a unified framework for studying how cells respond when their neighbors change.
Marcus: It’s essentially showing that you can separate the intrinsic cell identity from the spatial context using Cellina’s dual-encoder graph variational autoencoder.
Yuki: And they proved that this separation works well by uncovering biologically meaningful cancer subdomains without needing perfect supervision for every detail.
Ines: They showed that by using this latent space, we can do targeted neighbor perturbation simulations based on known pathways like PROGENy, which is really useful for testing hypotheses.
Marcus: The big picture here is that the model is fast enough to train and it performs better than competitors on large cohorts, like over two point five million cells from colorectal cancer.
Yuki: So it supports the idea that we can use these simulations to refine pathology annotations by looking at how specific signaling programs affect cell fate decisions in a tissue context.
Ines: They also showed that encoding neighborhoods continuously gives a strong prior even without supervision, which is a big factor in why this approach performs so well.
Marcus: But they do admit that the supervised disentanglement objectives rely on those cell type and spatial domain labels, which are just simplifications of continuous biology.
Yuki: That means the method has a bound on how accurately that spatial representation s can capture every single nuance of neighborhood variation in reality.
Ines: Still, they suggested a path forward using better raw imaging data to denoise the expression signal and those counterfactual predictions, which is compatible with their latent decomposition.
Marcus: That’s smart because it gives them a clear way to improve the input quality while keeping the core mathematical structure intact for interpretation.
Yuki: This work provides a principled test of whether these learned representations actually separate what's intrinsic from what's coming from the microenvironment across different species and tissues.
Division of Computational Genomics and Systems Genetics, German Cancer Research Center (DKFZ) · Helmholtz Information & Data Science School for Health for Health, Germany · Genome Biology Unit, European Molecular Biology Laboratory, Heidelberg, Germany · Department of Genetics at University Medical Center Groningen, University of Groningen · Oncode Institute at Utrecht · KU Leuven in Leuven, Belgium · Wellcome Sanger Institute in Hinxton, UK
q-bio.GN, cs.LG, stat.ML
Submitted: 2026-06-07
Updated: 2026-11-05
Code: https://github.com/PMBio/cellina
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 83/100
The gist: Tissue graph counterfactuals ask how a cell’s expression would change under altered spatial neighbor contexts…
Key concepts
- Tissue Graph Counterfactuals
- These are questions about how a cell's gene expression would change if its surrounding neighbors or connections were different. This framework formalizes these as two types of spatial interventions: changing the connections (edges) or changing the expression of neighboring cells (nodes). It provides a unified way to study how local neighborhood changes drive cell behavior.
- Cellina Architecture
- Cellina is a graph variational autoencoder that decomposes gene expression into two latent parts: z, which represents the cell's inherent identity or type, and s, which captures the influence of its microenvironment. The model uses supervised disentanglement to ensure z is tied to cell type while s captures spatial variation. This separation allows predictions to be conditioned on specific spatial contexts.
- Supervised Disentanglement
- This technique uses two training objectives: one forces the intrinsic representation (z) to predict the known cell-type label, and an adversarial discriminator tries to predict the spatial domain from z. This process pushes spatial information out of z and into s, ensuring that z reliably encodes identity while s reliably encodes environment.
- Latent Spatial Representation (s)
- The extrinsic representation 's' captures the effect of a cell's microenvironment, such as its local neighborhood structure or surrounding cell types. By clustering this latent space across cancer cells, researchers can uncover distinct subdomains that are spatially autocorrelated and assign biological identities to these microenvironments based on pathway gene sets.
Terminology
Summary
Tissue graph counterfactuals ask how a cell’s expression would change under altered spatial neighbor contexts, which is central to predicting cell behavior in tissues The paper introduces Cellina, a framework that uses supervised disentanglement to decompose a cell’s intrinsic state from its spatial context, using the latter as a conditioning input for counterfactual predictions. p. 0 In this work, we first formalize tissue
graph counterfactuals as a class of spatial interventions that either rewire connections between cells (edge perturbation) or modify the expression of their neighbors
(node perturbation). We then introduce Cellina ‡
, a framework that uses supervised
disentanglement to decompose a cell’s intrinsic state from its spatial context, using
the latter as a conditioning input for counterfactual predictions.Where the paper says it
Formalizing Tissue Graph Interventions
The authors first formalize tissue graph counterfactuals as a class of spatial interventions that either rewire connections between cells (edge perturbation) or modify the expression of their neighbors (node perturbation) These two types of interventions correspond to the two mutable components of the tissue graph, namely its edges and the expression of its neighbors.
Cellina Architecture and Disentanglement
Cellina is a (graph) variational autoencoder (VAE) that separates each cell’s gene expression into two latent components: an intrinsic representation z encoding cell identity, and an extrinsic (spatial) representation s encoding the effect of its microenvironment The model uses supervised disentanglement to anchor z to cell-type identity and adversarially remove spatial-domain information from z, routing microenvironmental variation to s by removing it from z.
Performance on Benchmarks
Cellina outperforms spatially-informed and non-spatial competitors in insilico graph perturbations, disentanglement, and scalability across benchmarks spanning over 2.5 million spatially-resolved cells in colorectal cancer and mouse brain Specifically, Cellina’s best variant leads the strongest baseline on colorectal cancer data by +0. p. 1 On colorectal cancer data, our best Cellina variant leads the strongest baseline by
+0. 17 on both Pearson and Signed Precision; on the whole-mouse-brain cohort, it remains
top-ranked across two held-out spatial domains. On colorectal cancer data, our best Cellina variant leads the strongest baseline by
+0. 17 on both Pearson and Signed Precision; on the whole-mouse-brain cohort, it remains
top-ranked across two held-out spatial domains.Where the paper says it
Where the paper says it
Biological Insights from Latent Space
Cellina reveals biologically distinct cancer subdomains in an unsupervised manner and enables targeted neighbor perturbation simulations using its disentangled spatial representation This is demonstrated by clustering the latent s over CRC cells to discover spatially autocorrelated subdomains, which are then scored against PROGENy pathway gene sets to assign biological identities to each microenvironment.
Scalability and Robustness
Cellina comes out as one of the fastest-to-train models in the benchmark suite owing to the efficient pseudobulk-based φ(v) computed a priori, which omits expensive message passing from model training over graphs The authors also show that Cellina’s latent z and s encode the desired source of variation while adequately removing nuisance sources of variation, suggesting robustness in the latent spaces.
Conclusion and Future Directions
The formalization of tissue graph counterfactuals provides a unified framework for studying neighborhood-driven cell responses Cellina’s design choices are biologically motivated, and its latent decomposition is compatible with the interpretation of counterfactuals as simulated outcomes under altered inputs and contexts. The work suggests that encoding neighborhoods continuously yields a strong prior even without any supervision, which is a key factor in Cellina’s performance. p. 7 Two design choices follow directly from this
observation and underpin Cellina’s performance. First, encoding neighborhoods continuously yields
a strong prior even without any supervision: Cellina (ablated), trained without cell-type or domain
labels, already outperforms or matches all non-Cellina baselines.Where the paper says it
Pruning redundant reasoning tokens
Tissue graph counterfactuals ask how a cell’s expression would change under altered spatial neighbor contexts, which is central to predicting cell behavior in tissues The paper introduces Cellina, a framework that uses supervised disentanglement to decompose a cell’s intrinsic state from its spatial context, using the latter as a conditioning input for counterfactual predictions.
How the training signal is built
The joint generative model is defined by p(x, z, s) = p(x z, s) p(z) p(s), where the likelihood used in the evaluations throughout this study is a Negative Binomial (NB) distribution over counts. The dual supervised disentanglement on z is enforced by training a cell-type classifier fclf to predict the cell-type label y from z and an adversarial domain discriminator fdisc to predict the spatial domain label d from sg(z). p. 3 Specifically, (i) we anchor z to cell-type identity using a cell
type classifier with loss Lclf, and (ii) adversarially strip spatial-domain information from z through
a two-part adversarial objective. First, Ldisc is optimized to train the discriminator to predict the
domain label from z; second, Ladv is optimized to encourage the encoder to render z domaininvariant.Where the paper says it
Where the gains come from
The authors show that Cellina’s explicit disentanglement additionally enables the discovery of the subdomain structure that can be used to refine pathology annotations and enables pathway-specific neighbor perturbations The combination of these components allows Cellina to consistently surpass all competitors on the most biologically-relevant metrics across both cohorts.
Limitations and Future Directions
Cellina’s supervised disentanglement objectives rely on cell-type and spatial-domain annotations, which are simplifications of continuous biology, bounding how well the spatial latent s can reflect neighborhood variation Future methods that integrate raw imaging signal alongside segmented counts could denoise both the expression signal and the counterfactuals derived from it. The latent decomposition is compatible with this interpretation, as it does not yet impose the structural assumptions required for identifiability. p. 9 Cellina’s latent decomposition is compatible with this interpretation. It does not yet
impose the structural assumptions [Khemakhem et al.Where the paper says it
Evaluation Metrics
All metrics are computed using library-size–normalized gene expression, with fixed library size l0 = 10 4 The evaluation includes Pearson r, Spearman ρ, Signed Precision, RMSE (counts), RMSE (log-fold change), and E-distance (local) to assess different aspects of counterfactual predictions.
Data Availability and Pre-processing
The CRC data was downloaded as AnnData files from Zenodo, where a standard feature selection procedure was applied to subset each slide to the 2000 highly variable genes implemented in scanpy For the whole-brain mouse cohort, analysis was restricted to three anatomical domains selected based on abundance, consistent representation across sections, and cell type diversity.
Empirical Convergence of Perturbations
In convergence analyses, it is observed that all four metrics improve with k for both logFC variants and saturate by k ≈ 100–200, collapsing shortly after At k = 200 (in-distribution), the cell-type-specific variant consistently outperforms the global one and reaches Pearson r ≈ 0.89, Signed Precision ≈ 0. p. 1 On colorectal cancer data, our best Cellina variant leads the strongest baseline by
+0. 17 on both Pearson and Signed Precision; on the whole-mouse-brain cohort, it remains
top-ranked across two held-out spatial domains. shifts agree in absolute magnitude (RMSE of log fold-change vectors), and (iv) overall distributional
fit (E-distance). All metrics are assessed for each held-out cell type; formal definitions and an
expanded metrics set with per-cell-type breakdown appear in Appendices A and G. All three Cellina variants outperform every baseline method with clear margins that hold across
individual patient slides and cell types; Cellina-GAT edges ahead on Pearson (0. 85) and RMSELFC
(1.Where the paper says it
Where the paper says it
Where the paper says it
Extended Results
Cellina generalizes to spatial transcriptomics across tissues and species by maintaining the highest scores in correlation metrics of log-fold changes as well as Signed Precision and RMSELFC on the whole mouse brain data. The analysis also shows that counterfactual predictions are better when conditioned on neighborhoods sampled from the matched subdomain than those from the global CRC domain. p. 7 43
SpatialPropnode-pert 0. 56
conditioned on neighborhoods sampled from the matched subdomain outperform those conditioned
on neighborhoods from the global CRC domain, across all five cell types (Figure 3(b)).Where the paper says it
Scalability Benchmark
Cellina comes out as one of the fastest-to-train models in the benchmark suite owing to the efficient pseudobulk-based φ(v) computed a priori, omitting the expensive message passing from model training over graphs. This contrasts with GAT-based methods that require subgraph sampling during training, such as Cellina-GAT. p. 3 For this variant, we additionally add a
modified contrastive loss Lspatial (Appendix D). These two variants trade off efficiency and expressivity: Cellina’s linear aggregator decouples
neighborhood construction from training and exhibits training-time scaling similar to non-spatial
baselines (Figure A6), whereas Cellina-GAT learns attention over each subgraph at additional
computational cost per step (Appendix D. 3 Results
Cellina accurately predicts in-silico perturbations in colorectal cancer data. To evaluate counterfactual predictions in a clinically relevant setting, we consider a spatial transcriptomics dataset of approximately 2.Where the paper says it
Where the paper says it
Final Summary
Cellina accurately predicts in-silico perturbations in colorectal cancer data, and its performance across various metrics suggests that its design choices are biologically motivated rather than architecturally complex. Cellina’s explicit disentanglement additionally enables the discovery of the subdomain structure that can be used to refine pathology annotations and enables pathway-specific neighbor perturbations. p. 8 adds substantial further gains. Cellina’s explicit disentanglement additionally enables the discovery
of the subdomain structure that can be used to refine pathology annotations and enables pathwayspecific neighbor perturbations. We validate this separation
under out-of-distribution regimes via in silico neighborhood alterations. A model that conflates
intrinsic and extrinsic variation cannot succeed at this task, making it a principled test of whether the
representations separate intrinsic from microenvironmental variation [Schölkopf et al. Moreover, we interpret counterfactuals operationally, as simulated outcomes under altered inputs
and contexts, rather than in the strict Pearlian sense, which presupposes a fully specified structural
causal model. Cellina’s latent decomposition is compatible with this interpretation. Contributions:
1. We formalize tissue graph counterfactuals as a class of spatial interventions encompassing edge and node perturbations; thereby we provide a unified framework for studying
neighborhood-driven cell responses.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improve prediction of cell behavior under altered spatial contexts by formalizing tissue graph counterfactuals as a class of spatial interventions encompassing edge and node perturbations, providing a unified framework for studying neighborhood-driven cell responses.
-
Introduce Cellina, a dual-encoder graph VAE with supervised disentanglement, which separates each cell’s gene expression into an intrinsic representation z encoding cell identity and an extrinsic (spatial) representation s encoding the effect of its microenvironment. This allows counterfactual predictions to be conditioned on the spatial context s.
-
Enable targeted neighbor perturbation simulations by using Cellina’s disentangled spatial representation to simulate pathway-targeted neighbor perturbations using existing priors, as shown in Experiment 4.4 where node perturbations with PROGENy pathway weights recover subdomain effects to a large extent.
-
Enhance biological interpretation of spatial heterogeneity by allowing the latent factor s to reveal biologically distinct cancer subdomains without supervision, which can be used to refine pathology annotations and assign biological identities to microenvironments based on inferred signaling programs like TGF beta dominance in CRC1.
Abstract
Tissue graph counterfactuals ask what a cell would express if its spatial neighborhood were altered, a question central to predicting cell behavior in tissues. Existing methods, however, present only specific instances of such queries. In this work, we first formalize tissue graph counterfactuals as a class of spatial interventions that either rewire connections between cells (edge perturbation) or modify the expression of their neighbors (node perturbation). We then introduce Cellina (https://cellina.readthedocs.io), a framework that uses supervised disentanglement to decompose a cell's intrinsic state from its spatial context, using the latter as a conditioning input for in-silico perturbation predictions. Across benchmarks spanning over 2.5 million spatially-resolved cells in observational and interventional screens, Cellina outperforms both spatially-informed and non-spatial competitors in in-silico perturbations, while training as fast as non-spatial methods. Moreover, we show that Cellina reveals cancer subdomains with distinct signaling programs in an unsupervised manner and enables targeted neighbor perturbation simulations.
Sources
- How Attentive are Graph Attention Networks?
- Fast Graph Representation Learning with PyTorch Geometric
- STRAND: Sequence-Conditioned Transport for Single-Cell Perturbations
- Demystifying Inter-Class Disentanglement
- Supervised Contrastive Learning
- Auto-Encoding Variational Bayes
- Variational Graph Auto-Encoders
- PerturBench: Benchmarking Machine Learning Models for Cellular Perturbation Analysis
Related papers
- SeekRBP: Leveraging Sequence-Structure Integration with Reinforcement Learning for Receptor-Binding Protein Identification
- Genome-Factory: A Library for Tuning, Deploying, and Interpreting Genomic Foundation Models
- AVA: Agentic Variant Adjudication for Automated, Auditable Read-Level Review in Cancer Diagnostic Sequencing
- C3P: Contrastive promoter-protein pretraining yields representations capturing bacterial gene regulation
- Leveraging genomic deep learning models for the prediction of non-coding variant effects
- Adaptive Allele-Haplotype Incidence Indexing for Pangenome Queries: Representation, Reconstruction, and Benchmarking