💻 bioinformatics

RAPTORGraph: graph-based pathway modeling for causal discovery in single-cell perturbations

RAPTORGraph is a novel graph-based variational autoencoder that integrates curated biological pathways with optimal-transport alignment to eliminate intervention spillover and enable identifiable causal discovery, thereby improving the prediction of unseen perturbations and non-additive interactions in single-cell experiments.

Original authors: Adhisantoso, Y. G., Schröder, S. K., Gress, M., Hernaez, M., Voges, J.

Published 2026-11-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Adhisantoso, Y. G., Schröder, S. K., Gress, M., Hernaez, M., Voges, J.

Original paper licensed under CC BY 4.0 (https://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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

Cells are the fundamental units of life, each carrying a complex set of instructions written in genes that dictate how an organism grows, functions, and responds to its environment. Scientists have long sought to understand how these genes interact, much like trying to figure out how turning a single switch in a vast, dark building affects the lights in other rooms. To do this, researchers use powerful tools to deliberately disrupt specific genes, a process known as perturbation, and then observe how the cell's behavior changes. This approach is central to modern biology and drug discovery, offering a way to map the hidden machinery of life. However, because there are thousands of genes and an almost infinite number of ways they can be combined, it is impossible to test every single scenario in a laboratory. This limitation has driven scientists to build computer models that can predict how cells will react to genetic changes they have never seen before. The challenge has been creating models that are not only accurate but also transparent, allowing researchers to understand the "why" behind a prediction, rather than just accepting a "black box" answer.

A new study introduces a sophisticated computer framework called RAPTORGraph, designed to solve two major problems that have plagued previous attempts at this task. The first issue is a phenomenon the researchers call intervention spillover. In many existing models, when a scientist tells the computer to change one specific gene, the model's internal logic gets confused, spreading that change across many different hidden variables instead of isolating it to a single cause. It is as if flipping a single light switch caused the entire building's wiring to flicker unpredictably, making it impossible to trace the effect back to its source. The second problem is known as mean collapse. Because it is impossible to observe the exact same cell before and after a genetic change, researchers often have to pair different cells together to train their models. If these pairings are done randomly, the model tends to learn only the average behavior of the entire group, smoothing out the unique and diverse responses of individual cells. This erases the very biological variety that scientists are trying to understand.

To fix these issues, the team built a model that forces a clear, one-to-one connection between a targeted gene and its effect within the computer's internal representation. They achieved this by structuring the model's learning process so that a change in a specific gene can only enter the system through a single, dedicated pathway. This design ensures that the model can isolate the impact of a single genetic change, keeping the signal clean and interpretable. Once this clean signal is established, the model uses a separate learning step to figure out how these isolated changes ripple through the network of biological processes. This allows the system to learn the complex, real-world relationships between genes without the confusion caused by the initial "spillover." Furthermore, to prevent the loss of cellular diversity, the researchers used a mathematical technique to carefully match control cells with perturbed cells based on their overall similarity, rather than pairing them at random. This ensures the model retains the rich variety of individual cell responses instead of collapsing into a generic average.

When tested on real data from experiments involving human leukemia cells, the new framework proved highly effective. It successfully predicted how cells would respond to genetic changes it had never encountered, including complex scenarios where two genes were altered at once. In these cases, the model could identify non-additive interactions, where the combined effect of two changes is different from simply adding their individual effects together. For instance, the model correctly identified that certain gene combinations could lead to a complete halt in cell division, a process known as cell cycle arrest, driven by specific biological programs. The researchers also demonstrated that the model could work in reverse: given a specific cellular outcome, it could suggest which genetic changes were likely responsible. This capability is crucial for generating new hypotheses and guiding future experiments.

The study confirms that by respecting the rules of causal logic—ensuring that causes are isolated and effects are traced accurately—computational models can become powerful tools for discovery. The researchers found that their approach not only matched the predictive accuracy of the best existing methods but also provided a clear window into the biological mechanisms driving those predictions. By avoiding the pitfalls of entangled signals and averaged data, RAPTORGraph offers a way to explore the vast landscape of genetic interactions with greater clarity. The work suggests that with the right architectural constraints, computers can help scientists move beyond simply predicting outcomes to truly understanding the causal chains that shape life, potentially accelerating the search for new treatments and deeper biological insights.

Technical Summary: RAPTORGraph

Problem Statement
Current computational models for predicting cellular responses to genetic perturbations face two primary limitations that hinder causal discovery and biological interpretability:

Intervention Spillover: Existing causal representation learning (CRL) models, such as DiscrepancyVAE (dVAE) and SENA, often rely on dense encoders (e.g., MLPs). In these architectures, a sparse intervention on a single gene in the high-dimensional observation space propagates through the network to create a dense, multi-node signal in the latent space. This "spillover" violates the theoretical requirement for atomic, single-node interventions, making it impossible to uniquely identify the latent target of an intervention and rendering causal identifiability ill-posed.
Mean Collapse: Single-cell RNA sequencing (scRNA-seq) is destructive, preventing paired observations of the same cell before and after perturbation. Existing methods often resort to randomly pairing control and perturbed cells for training. This forces models to minimize error across arbitrary pairings, leading them to learn the population-mean response rather than preserving cellular diversity, a phenomenon termed "mean collapse."
Confirmatory Bias: Models that incorporate curated biological pathways (e.g., SENA) to enhance interpretability often suffer from confirmatory bias. They are constrained to explain data through preexisting annotations, potentially discarding novel signals that lie outside known pathways and failing to generalize across different cell types where pathway specificity varies.

Methodology
The authors propose RAPTORGraph (Response Analysis of Perturbed TranscriptOmes using InteRpretable Graph), a $\beta$ -variational autoencoder framework designed to satisfy the theoretical requirements for causal identifiability while maintaining predictive accuracy. The architecture consists of four key components:

GraphPathway Encoder with Preconditioning:
To solve intervention spillover, the authors introduce a novel encoder architecture that enforces a sparse, one-to-one mapping between perturbed genes and latent variables.
- Preconditioning: The input gene vector is sorted such that perturbed genes appear first. The encoder's primary weight matrix is structured as a block matrix where the top-left submatrix ( $\Lambda_{11}$ ) is strictly diagonal. This ensures each of the $k$ perturbed genes influences exactly one of the first $k$ learned meta-pathways.
- Zero-Block Constraint: A zero block explicitly prevents perturbed genes from influencing the remaining $(d-k)$ latent variables, isolating the intervention to a single node.
- Flexibility: The remaining submatrices are dense, allowing the model to learn complex relationships between unperturbed genes and all latent pathways.
DAGMA Causal Graph Learning:
Instead of relying on predefined pathway annotations, RAPTORGraph learns the causal structure of the latent meta-pathways using the DAGMA (Differentiable Acyclicity via Graphical Matrix Algebra) framework.
- This module learns a Directed Acyclic Graph (DAG) representing the causal relationships between the sparse pathway activities.
- It utilizes a continuous acyclicity constraint based on a log-determinant function, allowing for end-to-end optimization of the graph structure without enforcing arbitrary hierarchical orderings.
Optimal Transport (OT) Preprocessing:
To address mean collapse, the model employs an offline optimal transport step to pair control and perturbed cells.
- Unlike random pairing, OT globally matches cells based on their distributional similarity, preserving the underlying population structure and cellular heterogeneity.
- This provides a well-defined mapping between control and perturbed samples, stabilizing the learned latent representations.
Training Objective:
The model is trained end-to-end using a composite loss function ( $L_{total}$ ) that balances four components:
- Reconstruction Loss ( $L_{MSE}$ ): Faithfully reconstructs basal (unperturbed) cell states.
- KL Divergence ( $L_{KL}$ ): Regularizes the latent space to match a prior distribution.
- Interventional Prediction Loss ( $L_{MMD}$ ): Uses Maximum Mean Discrepancy to minimize the distance between predicted and true distributions of perturbed cells.
- Causal Graph Loss ( $L_{DAGMA}$ ): Enforces acyclicity and sparsity in the learned causal graph.

Key Contributions

Architectural Innovation: The introduction of the preconditioned GraphPathway encoder, which mathematically enforces the sparse, single-node intervention requirement necessary for causal identifiability, resolving the intervention spillover problem.
Bias Mitigation: A framework that learns causal meta-pathways directly from data rather than relying on curated biological pathways, thereby avoiding confirmatory bias and allowing for the discovery of novel biological signals.
Diversity Preservation: The integration of Optimal Transport preprocessing to prevent mean collapse, ensuring the model retains cellular diversity rather than regressing to population averages.

Results
The authors benchmarked RAPTORGraph against state-of-the-art models (SENA, scGPT, dVAE, and GEARS) on the Norman et al. Perturb-seq dataset and the Replogle et al. CRISPRi dataset.

Generative Performance: RAPTORGraph achieved a strong balance between reconstruction fidelity (MSE) and distributional accuracy (MMD). On the Norman dataset, it achieved an MMD of $0.112 \pm 0.001$ and MSE of $0.045 \pm 0.000$ , outperforming SENA and scGPT in MMD while maintaining competitive MSE.
Non-Additive Interaction Prediction: In predicting complex genetic interactions (synergy, suppression, neomorphism, redundancy, epistasis), RAPTORGraph demonstrated superior performance in identifying Synergy ($0.367$ Precision@10) and Neomorphism ($0.433$ Precision@10) compared to most baselines. The authors note that OT preprocessing was a key driver for predicting redundancy and epistasis.
Reverse Perturbation Analysis: When tasked with identifying the genetic cause of an observed phenotype (reverse engineering), RAPTORGraph showed competitive performance in identifying relevant genes (Gene Overlap Hit Rate), consistently ranking among the top models, though exact-match hit rates remained modest for all models.
Biological Interpretability: Gene Set Enrichment Analysis (GSEA) on the learned latent factors revealed biologically accurate programs. For instance, the latent factor linked to EGR1 correctly identified differentiation-induced cell cycle arrest (negative enrichment for "E2F Targets" and "G2-M Checkpoint"), and the factor linked to JUN identified stress-induced growth arrest.

Significance and Claims
The paper claims that RAPTORGraph bridges the gap between predictive power and mechanistic insight in single-cell perturbation modeling. By explicitly resolving the intervention spillover and mean collapse problems, the framework enables identifiable causal discovery without relying on potentially biased, preexisting pathway annotations.

The authors emphasize that while the model enforces a single-node intervention architecture, this is a modeling requirement to satisfy identifiability theory, not an assertion that biological perturbations only affect one pathway. However, the results suggest that this architectural constraint allows the model to uncover distinct, biologically accurate causal programs. The work positions RAPTORGraph as a tool for generating testable hypotheses regarding how genes jointly shape cell behavior and for guiding the selection of informative experiments, particularly in the context of non-additive genetic interactions and drug target discovery. The authors acknowledge limitations, noting that generalizability evidence is currently strongest on the Norman dataset and that pinpointing exact higher-order combinations remains a challenge for all current models.

Technical Summary: RAPTORGraph

More like this