🤖 machine learning

scCBGM: Interpretable Single-Cell Counterfactual Editing

The paper introduces scCBGM, an interpretable single-cell framework that adapts concept bottleneck architectures and flow matching to enable precise counterfactual editing and combinatorial generalization of cellular phenotypes, addressing the infeasibility of exhaustive experimental mapping.

Original authors: Alma Andersson, Aya Abdelsalam Ismail, Edward De Brouwer, Doron Haviv, Tommaso Biancalani, Kyunghyun Cho, Gabriele Scalia, Aïcha BenTaieb, Hector Corrada Bravo

Published 2026-06-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Alma Andersson, Aya Abdelsalam Ismail, Edward De Brouwer, Doron Haviv, Tommaso Biancalani, Kyunghyun Cho, Gabriele Scalia, Aïcha BenTaieb, Hector Corrada Bravo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a master chef trying to understand how a specific dish changes when you tweak the ingredients. You have a recipe for a "Spicy Tomato Soup" (a cell in its natural state). You want to know: "If I add extra chili, what will this exact same bowl of soup taste like?"

In the real world, you can't just magically add chili to the soup you're already eating and see the result without making a whole new batch. In biology, scientists face the same problem with cells. They can't take one specific cell, turn off a gene, turn on a drug, and watch it happen in real-time to see the result. They can only observe the cell before and the cell after in different experiments, but they can never see the "what if" scenario for that single individual cell.

This paper introduces a new tool called scCBGM (Single-Cell Concept Bottleneck Generative Model). Think of it as a "Time-Traveling Recipe Simulator" for cells.

Here is how it works, broken down into simple concepts:

1. The Problem: The "Black Box" of Biology

Previously, computer models could tell you what a group of cells might look like after a treatment (like an average soup recipe). But they couldn't tell you what your specific cell would look like if you changed one thing. Also, these models were often "black boxes"—they gave an answer, but they couldn't explain why or let you control specific parts of the recipe (like "turn up the heat" or "add more salt").

2. The Solution: The "Concept Bottleneck"

The authors built a model that doesn't just guess; it understands the concepts (the ingredients) of the cell.

The Ingredients: Instead of looking at millions of gene numbers at once, the model breaks the cell down into understandable "concepts," like "Is this a T-cell?" or "Is it currently being attacked by a virus?" or "How active is its immune system?"
The Bottleneck: Imagine a funnel. All the complex data about the cell has to pass through this funnel to become a list of these simple concepts. This forces the computer to organize the messy data into clear, understandable categories.

3. The Magic Trick: "Counterfactual Editing"

This is the core superpower of the paper.

The Setup: You give the model a photo of a specific cell (the "Factual" cell).
The Edit: You tell the model, "Okay, keep everything about this cell the same, but pretend its 'Immune System' concept is turned OFF."
The Result: The model generates a brand new image (a new gene profile) of what that exact same cell would look like if it had no immune system.

This is called a counterfactual. It's answering the question: "What would have happened if...?"

4. Why This Model is Special (The "Secret Sauce")

The paper explains that previous models had trouble with two things:

Noise: Real biological data is messy (like a recipe with typos).
Mixing: The model often confused the "ingredients" (concepts) with the "cooking style" (random noise).

The authors added two special features to fix this:

The "Skip Connection" (The Direct Line): Imagine you are painting a picture. Usually, you might forget the original colors as you paint over them. This model keeps a direct line from the "ingredients" (concepts) all the way to the final picture. This ensures that if you say "Red," the final picture stays Red, even if the rest of the painting gets messy.
The "Cross-Covariance Penalty" (The Separation Wall): This is a mathematical rule that forces the model to keep the "ingredients" separate from the "random noise." It's like having a strict chef who refuses to let the salt shaker fall into the pepper grinder. This ensures that when you change a concept, you aren't accidentally changing random noise.

5. How They Tested It

Since you can't actually time-travel to see if the model is right, they created a fake universe (synthetic data) where they knew the exact answer.

They built a digital world where they knew exactly how a cell should change if you tweaked a concept.
They tested their model against this fake world.
The Result: The model was much better at predicting the "what if" scenarios than other existing models, even when the data was noisy or missing pieces.

They also tested it on real data (like immune cells reacting to viruses). While they couldn't know the "true" answer for a single cell, they checked if the groups of edited cells looked like real cells that had actually been treated. The model's predictions matched the real-world groups very well.

Summary

Think of scCBGM as a highly intelligent, transparent simulator for biology.

It takes a messy cell.
It breaks it down into clear, understandable "concepts" (like cell type or drug response).
It lets you "edit" those concepts (e.g., "Turn off the drug response").
It predicts exactly what that specific cell would look like with that change, without needing to run a new, expensive experiment.

The paper claims this helps scientists understand disease mechanisms and design better treatments by allowing them to run "virtual experiments" on individual cells to see how they might react to different interventions.

Technical Summary: Single-Cell Concept Bottleneck Generative Models for Interpretable and Controllable Cellular Editing

Problem Statement

Understanding cellular phenotypes and their responses to perturbations (e.g., drugs, cytokines) is critical for disease biology and therapeutic design. While single-cell RNA sequencing (scRNA-seq) enables characterization at cellular resolution, the combinatorial space of conditions (treatments, doses, cell types) makes exhaustive experimental mapping infeasible. Computational models are required to predict cellular responses under unseen conditions.

A specific capability required of these models is precise cellular editing: the ability to start with an observed cell and systematically modify specific biological properties (concepts) while preserving others. This differs from conditional generation, which produces any cell under given conditions, whereas editing targets a specific cell under altered conditions to enable counterfactual reasoning (e.g., "What would this specific T-cell look like if its NF-kB pathway were turned off?").

Existing methods face two primary limitations:

Lack of Cell-Level Counterfactuals: Many models capture population-level effects or conditional distributions but fail to generate counterfactuals for individual cells.
Lack of Interpretable Control: While some methods offer interpretability, they are often descriptive (explaining correlations) rather than actionable (simulating or editing responses). Conversely, generative models with fine-grained control often rely on opaque latent variables rather than biologically meaningful concepts.

Methodology: scCBGM

The authors introduce single-cell Concept Bottleneck Generative Models (scCBGM), a framework designed for interpretable and controllable cellular editing. The approach adapts Concept Bottleneck Models (CBMs) and Concept Bottleneck Generative Models (CBGMs) to the unique challenges of scRNA-seq data, including high heterogeneity, technical noise, and unreliable concept annotations.

1. Problem Formalization

The framework adopts Pearl's Structural Causal Model (SCM). Gene expression $X$ is influenced by observed concepts $C$ (e.g., cell type, stimulation status) and unobserved residual factors $U$ (e.g., batch effects, cell identity). The goal is to predict the counterfactual outcome $X'$ for a specific cell $x$ if its concepts were changed from $c$ to $c'$ , while keeping the unobserved factors $U$ constant. This is formalized as learning a predictor $f_\theta$ such that $\hat{x}' = f_\theta(x, c') \approx \mathbb{E}[f_X(c', U) | X=x]$ .

2. Architecture

The model follows an encoder-decoder design with a concept bottleneck:

Encoder: Maps input gene expression $x$ to a latent distribution $z$ .
Concept Bottleneck: Decomposes $z$ $z$ into:
- Known factors ( $\hat{c}$ ): Predictions of the $K$ biological concepts.
- Unknown factors ( $u$ ): A representation capturing variation not explained by $\hat{c}$ .
Decoder with Skip Connections: Unlike standard CBGMs, the decoder receives the concatenation of $u$ and $\hat{c}$ at the input. Crucially, skip connections feed the known concepts $\hat{c}$ into every layer of the decoder. This enforces systematic conditioning on concepts throughout the generation process, improving robustness to noisy annotations.

3. Training Objectives

The model is trained by minimizing a loss derived from the $\beta$ -VAE Evidence Lower Bound (ELBO), augmented with specific penalties:

Reconstruction Loss: Minimizes the difference between predicted and observed gene expression.
Concept Supervision ( $L_{concept}$ ): Predicts concepts from the latent space, handling missing annotations via a binary mask.
Cross-Covariance Penalty ( $L_{cc}$ ): To enforce independence between the known concepts $\hat{c}$ and unknown factors $u$ , the model penalizes the squared Frobenius norm of their empirical cross-covariance. Unlike cosine-similarity losses, this approach does not impose dimensional constraints ( $K \neq d_u$ ), allowing for flexible embedding sizes.

4. Extension to Flow Matching

The framework is extended beyond Variational Autoencoders (VAEs) to Flow Matching (FM) models. By conditioning the flow matching vector field on the scCBGM embeddings ( $[u, \hat{c}]$ ), the authors enable concept-guided editing in two regimes:

Decoding-only: Sampling from noise conditioned on edited concepts (useful for generating diverse examples).
Encoding-Decoding: Mapping a specific cell to noise (abduction), editing the concepts, and decoding back (action/prediction). This strictly implements the abduction-action-prediction framework for counterfactual generation.

Key Contributions

Architectural Modifications: Introduction of decoder skip connections for persistent concept conditioning and a cross-covariance penalty for decoupled embeddings without dimensional constraints.
Framework Extension: Extending CBGMs to Flow Matching models, enabling high-fidelity generation alongside interpretable control.
Synthetic Benchmark: Development of a synthetic data generation process based on a hierarchical overdispersed Poisson model that separates exogenous noise from conditions, providing access to ground-truth counterfactuals for rigorous evaluation.
Empirical Validation: Demonstration of superior performance in combinatorial generalization and counterfactual prediction across multiple real-world datasets and the synthetic benchmark.

Results

The paper evaluates scCBGM on three real-world datasets (Kang et al., 2017; Cui et al., 2024; Nault et al., 2023) and synthetic data.

Synthetic Data Evaluation: scCBGM outperforms standard CBGMs across all noise levels and scenarios. The ablation study confirms that the combination of the concept bottleneck architecture, cross-covariance loss, and skip connections yields the best performance.
Real Data Benchmarking: Using population-level proxy metrics (rMMD, rFID, rSD) where ground-truth cell-level counterfactuals are unavailable, scCBGM and its Flow Matching variant (scCBGM-FM) consistently outperform state-of-the-art baselines (including scGen, CINEMA-OT, biolord, and Vanilla-FM) in 5 out of 7 experiments on the Kang dataset and 4 out of 7 on the Cui dataset.
Zero-Shot Generalization: The model successfully predicts responses for held-out cell type–stimulation combinations (e.g., stimulated Naive CD4 T cells not seen during training), demonstrating strong compositional generalization.
Continuous Dose Response: In the Nault et al. dataset, scCBGM effectively models continuous dosage responses, predicting gene expression for intermediate TCDD dosages not present in the training set.
Case Study (Drug Response Enhancement): The authors demonstrate a use case where scCBGM identifies key pathways (TGF $\beta$ , PI3K, p53) mediating treatment response in hepatic stellate cells. By in silico editing control cells to activate these pathways, the model successfully shifts the population's phenotype to resemble high-dosage responders, suggesting a mechanistic route to overcome cellular resistance.

Significance and Claims

The paper claims that scCBGM addresses the open challenge of enabling precise counterfactual editing at the level of individual cells with biologically interpretable control.

Interpretability and Actionability: Unlike descriptive interpretability methods, scCBGM allows researchers to intervene on specific biological concepts (e.g., pathway activity, drug dosage) to simulate cellular responses.
Robustness: The architecture is specifically designed to handle the noisy and incomplete concept annotations typical of scRNA-seq data.
Therapeutic Design: By navigating combinatorial spaces of cellular responses that are experimentally infeasible to map, the framework aims to accelerate therapeutic design and mechanistic hypothesis generation.

The authors remain modest regarding limitations, acknowledging that true cell-specific counterfactuals cannot be observed in real data (requiring population-level proxies for validation) and that the model generalizes to unseen combinations of known concepts but not to entirely new concepts absent from training. They also note the ethical risks associated with high-fidelity generative models, such as the potential for fabricating biological data or generating false therapeutic leads without rigorous experimental validation.