🧬 biology

Querying Counterfactuals on Tissue Graphs with Supervised Disentanglement

This paper introduces Cellina, a supervised disentanglement framework that formalizes tissue graph counterfactuals by separating a cell's intrinsic state from its spatial context to accurately predict expression changes under rewired or modified neighbor interactions, outperforming existing methods on large-scale spatial transcriptomics datasets.

Original authors: Abdul Moeed, Stefan Schrod, Martin Rohbeck, Marc Jan Bonder, Pavlo Lutsik, Oliver Stegle, Daniel Dimitrov

Published 2026-06-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Abdul Moeed, Stefan Schrod, Martin Rohbeck, Marc Jan Bonder, Pavlo Lutsik, Oliver Stegle, Daniel Dimitrov

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Picture: The "Cell Neighborhood" Problem

Imagine a city where every person (a cell) has a unique personality and a specific job. In biology, we want to predict how a person will act if they move to a different neighborhood.

The Old Way: Previous computer models treated people like they were living in isolation. They assumed that if you gave everyone the same "stimulus" (like a new law or a virus), everyone would react the same way, regardless of who their neighbors were.
The Reality: In a living body, your behavior is heavily influenced by your neighbors. A person living next to a bakery might smell bread and feel hungry; a person next to a construction site might feel stressed. In biology, a cell's "mood" (gene expression) is shaped by the cells right next to it.

The paper argues that to predict how a cell will change, we can't just look at the cell itself; we have to simulate what happens if we swap its neighbors or change what its neighbors are saying.

The Solution: "Cellina" (The Neighborhood Simulator)

The authors built a tool called Cellina. Think of Cellina as a super-smart translator that can separate a person's "inner self" from their "environment."

The Two Parts of a Person:
- The Inner Self (Intrinsic): Who you are at your core (e.g., "I am a baker"). This doesn't change just because you move.
- The Environment (Extrinsic): How your surroundings affect you (e.g., "I smell like bread because I live next to a bakery").
The Magic Trick (Disentanglement):
Many AI models mix these two things up. If you move a baker to a construction site, a bad model might think the baker became a construction worker. Cellina is special because it uses "supervised disentanglement."
- Analogy: Imagine you have a photo of a person. Cellina uses a special filter to separate the person from the background. It learns to recognize the person even if the background changes, and it learns to recognize the background even if the person changes.
- How it learns: It uses a "teacher" (cell type labels) to say, "This part of the data is the person's identity," and a "trickster" (adversarial training) to say, "Make sure the 'person' part doesn't accidentally learn anything about the 'neighborhood'."

The Two Types of Experiments

The paper tests Cellina with two specific "What if?" scenarios, which they call Counterfactuals:

1. Edge Perturbation (The "Moving House" Test)

The Question: "What would this cell look like if we physically moved it to a different neighborhood and gave it a completely new set of neighbors?"
The Analogy: Imagine taking a baker and moving them to a construction site, replacing their old neighbors with construction workers. Cellina predicts how the baker's behavior changes just by being surrounded by these new people.
The Result: Cellina was much better at this than other tools. It successfully predicted that the baker would start acting differently (e.g., smelling like sawdust) without losing their identity as a baker.

2. Node Perturbation (The "Changing the Message" Test)

The Question: "What if the neighbors stay in the same house, but they start shouting different things?"
The Analogy: The baker stays next to the construction workers, but instead of shouting about drills, they start shouting about blueprints. Cellina predicts how the baker reacts to this change in conversation without moving.
The Result: Cellina could also simulate this, predicting how specific biological pathways (like a specific gene program) would turn on or off based on these new signals.

Why This Matters (According to the Paper)

The authors tested Cellina on two massive datasets:

Colorectal Cancer: Looking at cells in healthy colon tissue vs. cancerous tissue.
Mouse Brain: Looking at different regions of a mouse brain.

The Findings:

Better Predictions: Cellina beat all other existing methods (like scGen, CPA, and MintFlow) in predicting how cells would change in these "virtual experiments."
Finding Hidden Patterns: Because Cellina separates the "neighborhood" from the "cell," it found hidden sub-groups within the cancer tissue that other methods missed. It realized that not all cancer cells are in the same "environment," even if they are in the same tumor.
Scalability: It can handle millions of cells without crashing, which is a big deal for modern biology.

What Cellina Does Not Do (Based strictly on the text)

It is not a crystal ball for cures: The paper explicitly states these are "in-silico" (computer simulation) predictions. They are "generative hypotheses" that need to be tested in a real lab (wet-lab validation) before we know if they are true.
It doesn't replace doctors: The authors do not claim this tool can diagnose patients or prescribe medicine yet. It is a research tool to help scientists understand how cells interact.
It relies on labels: To work its magic, Cellina needs to be told what "type" of cell it is looking at (e.g., "This is a T-cell") and what "region" it is in (e.g., "This is the tumor"). It learns from these labels to separate the identity from the environment.

Summary

Cellina is a new computer program that treats tissue like a graph of connected neighbors. It learns to separate a cell's identity from its surroundings. By doing this, it can answer "What if?" questions: "What if this cell had different neighbors?" or "What if its neighbors said something else?" It does this better than any previous tool, helping scientists understand the complex social life of cells in tissues like tumors and brains.

Technical Summary: Querying Counterfactuals on Tissue Graphs with Supervised Disentanglement

Problem Definition

Single-cell biology aims to predict cellular responses to perturbations. However, existing methods typically rely on two assumptions that tissues violate: (i) perturbations act as shared stimuli applied uniformly, and (ii) cells are conditionally independent (i.i.d.). In living tissues, a cell's transcriptional state is shaped by its local neighborhood, which varies uniquely for every cell.

The authors formalize tissue graph counterfactuals as a class of spatial interventions asking: What would a cell express if its neighborhood context were altered while its intrinsic identity remained fixed? These interventions target two mutable components of the tissue graph:

Edge Perturbation: Rewiring a cell's neighborhood connections (replacing neighbors).
Node Perturbation: Modifying the gene expression features of existing neighbors while preserving graph topology.

Current methods lack a unified framework to handle these queries, often treating cells as i.i.d. or targeting specific intervention types without separating intrinsic cell identity from extrinsic microenvironmental influence.

Methodology: Cellina

The paper introduces Cellina, a graph Variational Autoencoder (VAE) designed to render tissue graph counterfactuals tractable through supervised disentanglement.

Model Architecture

Cellina decomposes a cell's gene expression $x$ into two latent components:

Intrinsic Representation ( $z$ ): Encodes cell identity (cell type), independent of spatial context.
Extrinsic Representation ( $s$ ): Encodes the effect of the spatial microenvironment (neighborhood).

The generative model assumes standard normal priors for $z$ and $s$ , with the likelihood $p(x|z, s)$ modeled as a Negative Binomial distribution parameterized by a decoder $Dec([z; s])$.

Inference Variants

Two variants are proposed to encode the spatial representation $s$ :

Cellina (Base): Uses a degree-normalized aggregation of log-normalized neighbor expression $\phi(v)$ , encoded via an MLP.
Cellina-GAT: Replaces the fixed aggregator with a multi-layer Graph Attention Network (GATv2) operating on the local subgraph. This variant excludes self-loops to ensure the focal cell's own expression is captured solely by $z$ .

Supervised Disentanglement

Purely unsupervised factorization is not identifiable without inductive biases. Cellina injects biological supervision to anchor $z$ to cell-type identity and route spatial variation to $s$ :

Cell-Type Classifier ( $L_{clf}$ ): Anchors $z$ to known cell-type labels.
Adversarial Domain Discriminator ( $L_{adv}$ ): A two-step alternating optimization where a discriminator attempts to predict the spatial domain from $z$ . The encoder is trained to minimize this prediction (maximize entropy), effectively stripping spatial-domain information from $z$ and forcing it into $s$ .
Contrastive Loss ( $L_{spatial}$ ): Applied only in Cellina-GAT, this loss encourages similarity in $s$ for spatial neighbors while separating distinct domain regions, acting as a biologically grounded inductive bias.

The total training objective is:
$\mathcal{L} = \mathcal{L}_{VAE} + \alpha_{clf} \mathcal{L}_{clf} + \alpha_{spatial} \mathcal{L}_{spatial} - \alpha_{adv} \mathcal{L}_{adv}$

Counterfactual Simulation

Edge Perturbation: Neighbors of a focal cell are swapped with cells from a target domain (e.g., replacing healthy neighbors with cancer neighbors), excluding the focal cell's own type to ensure a conservative evaluation.
Node Perturbation: The graph topology remains fixed, but neighbor gene expression vectors are modified by a gene-specific transformation (e.g., multiplicative scaling based on log-fold changes between domains).

Key Contributions

Formalization: The authors define tissue graph counterfactuals as a unified class of spatial interventions encompassing both edge (topology) and node (feature) perturbations.
Cellina Framework: They introduce a dual-encoder graph VAE with supervised disentanglement that separates intrinsic identity from spatial context.
Performance: Cellina outperforms spatially-informed and non-spatial baselines (including scGen, CPA, MintFlow, and SpatialProp) in counterfactual prediction, disentanglement, and scalability across datasets containing over 2.5 million cells.
Biological Discovery: The framework enables unsupervised identification of biologically distinct cancer subdomains and facilitates targeted pathway-specific neighbor perturbation simulations.

Experimental Results

The model was evaluated on two large-scale spatial transcriptomics datasets:

Colorectal Cancer (CRC): ~2.4 million cells across six slides. The task involved predicting the expression of healthy colon cells (REF) if placed in a tumor region (CRC).
- Results: Cellina variants significantly outperformed all baselines. The base Cellina improved Pearson correlation by +0.14 and Signed Precision by +0.17 over the strongest non-Cellina baseline. Cellina-GAT achieved the highest Pearson score (0.85) and lowest RMSE (1.14).
- Ablation: Even the "Cellina (ablated)" variant (without supervision) outperformed non-spatial baselines, but supervised disentanglement provided further gains (+0.12 Pearson), confirming the efficacy of the separation strategy.
Whole Mouse Brain: ~146K cells across three slides (MERFISH). The task involved transferring predictions between spatial domains (Thalamus to Isocortex/Fiber-tracts).
- Results: Cellina variants again led across three of four metrics, demonstrating generalization across tissues and species.

Biological Insights

Subdomain Discovery: Clustering the latent spatial representation $s$ revealed distinct CRC subdomains (CRC1 and CRC2) with different signaling profiles (TGF $\beta$ -dominant vs. NF $\kappa$ B/MAPK-dominant) without direct supervision.
Pathway Perturbation: Simulating node perturbations using PROGENy pathway weights successfully recapitulated the biological responses of these subdomains, identifying key genes (e.g., FN1, MMP3) consistent with known fibroblast programs.

Significance and Claims

The paper claims that Cellina provides a principled framework for studying neighborhood-driven cell responses, moving beyond the assumption of independent cells. By explicitly disentangling intrinsic state from extrinsic influence, the model enables:

Contextual Prediction: Predicting cell behavior not in isolation but within specific spatial contexts.
Reduced Experimental Search Space: Generating "virtual" cell responses to altered neighborhoods to guide wet-lab validation.
Scalability: The base Cellina model scales efficiently (training time comparable to non-spatial scVI) by decoupling neighborhood construction from training, unlike graph-attention-based competitors that require expensive message passing.

The authors note that while counterfactuals are currently generative hypotheses requiring experimental validation, the operational interpretation of these simulations offers a promising avenue for querying virtual cell responses in complex tissue environments. They acknowledge limitations regarding the reliance on pre-defined cell-type and domain annotations and the potential for segmentation errors in spatial data, suggesting future work could integrate raw imaging signals to denoise these inputs.