acceptodds
Withdrawn from ICLR 2027

Interventional World Models with Conditional Gene Graphs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Predicting cellular responses to unseen intervention combinations requires modeling how interventions transform cellular states and interact. We introduce the Interventional Graph World Model (IGWM), which represents interventions as reusable state transition operators rather than endpoint labels. IGWM learns a state and intervention conditioned graph editor that selects local gene interactions from a candidate graph. Message passing on the resulting sparse graph defines a conditional velocity field for response prediction. Flow matching enables learning from unpaired endpoint single cell data without longitudinal trajectories. This design supports direct prediction and sequential application of interventions. Across two datasets, IGWM-sparse improves all reported mean metrics over prior baselines for unseen combinations on Norman19 and held-out cells under observed interventions on VCC-H1. IGWM provides a structured framework for modeling context dependent intervention dynamics.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
world models, single-cell perturbation prediction, compositional generalization, conditional gene graphs, flow matching

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_WGEEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes the Interventional Graph World Model (IGWM), which represents interventions as reusable state transition operators. IGWM constructs a candidate gene graph from control cells and employs a state- and intervention-conditioned graph editor to select Top-K edges. Message passing on the resulting sparse graph parameterizes a conditional velocity field, while conditional flow matching learns state transitions from unpaired control and treated endpoint populations.

Critical strengths

1. Predicting unseen intervention combinations is an important challenge in virtual-cell modeling. Combining state- and intervention-conditioned graph editing with continuous transition modeling provides a potentially useful approach to capturing context-dependent gene interactions.

2. Promising empirical performance on unseen intervention combinations. IGWM achieves meaningful improvements in perturbation-response prediction on Norman19, particularly in DEG F1 and Delta Pearson. These results suggest that the proposed architecture can benefit direct prediction of unseen combinations of observed single-gene interventions.

Critical weaknesses

1. Limited conceptual novelty. The central idea of modeling perturbation responses through conditional flow matching has already been explored. For example, CellFlow [1] uses flow matching to generate perturbed cell populations conditioned on complex interventions. More closely related, scDFM [2] combines conditional flow matching with gene interaction graphs through its Perturbation-Aware Differential Transformer. Against these closely related approaches, the methodological novelty of IGWM appears incremental.

2. Insufficient evaluation of generalization to unseen interventions. The evaluation is limited to unseen double-gene combinations whose constituent single-gene interventions are observed during training (Norman19), and held-out cells under observed interventions (VCC-H1). Neither setting evaluates generalization to genuinely unseen intervention targets. Moreover, intervention representations are learned from gene identities, while the candidate graph is constructed solely from control-cell expression correlations, without incorporating external biological knowledge. It remains unclear how IGWM would generalize to perturbation targets absent from training, which substantially limits the scope of its demonstrated generalization capabilities.

3. Insufficient comparisons with competitive baselines. The experiments omit closely related flow-matching approaches such as CellFlow and scDFM, as well as more recent perturbation prediction models such as TxPert.


[1] CellFlow enables generative single-cell phenotype modeling with flow matching

[2] scDFM: Distributional Flow Matching for Robust Single-Cell Perturbation Prediction

Questions for authors

See weaknesses.

Reviewer_7vM4OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces a world model for unseen combinatorial perturbation prediction based on graph conditioned vector field learned from flow matching. The model leverages the graph information to represent perturbation effect as transitions rather than endpoints to generate a response vector field based on sparse graphs. On two independent datasets the method yields superior performance against baselines like CPA and GEARS.

Critical strengths
  1. The idea of basing intervention effect on top of graph propagation is interesting and promising in terms of representing information such as regulatory dynamics.

  2. The attempt of modelling interventional models for perturbation is towards the correct direction of analysing perturbation screens

Critical weaknesses
  1. I don't understand the point of the flow matching model. Maybe i am missing something here. Why does a conditional vector field help perturbation prediction on unseen combinations? Or rather, why does traditional foundation models fail at this task? A lot of further justification including ablation studies is required.

  2. It's not clear to me why the usage of candidate graph provide additional information for predicting perturbation response. Two genes need not to be co-regulated to have similar perturbation response and vice versa. So what does a correlation based graph add here?

Questions for authors
  1. The benchmarking experiments are very limited. Comparisons to other flow matching based/generative models such as CellFlow or [1]. Maybe even a simple conditional flow matching model needs to be added. Additionally, like i mentioned in the weakness section comparisons to single cell foundation models are crucial to justify adding the graph component and actions on top of the graph.

  2. Theoretical results needs a lot more work. The authors should either establish more significant connection to support their modeling choices or just remove them. For example proposition 1 is just a prediction invariance from a score margin divided by a Lipschitz bound and proposition 3 reduces to standard flow matching conditional expectation result established in [2] if the noise is not time dependent.

Reference [1] Schneider, Nora, Lars Lorch, Niki Kilbertus, Bernhard Schölkopf, and Andreas Krause. 2024. “Generative Intervention Models for Causal Perturbation Modeling.” In arXiv [Cs.LG]. November 21. arXiv. https://doi.org/10.48550/arXiv.2411.14003. [2] Lipman, Yaron, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. 2022. “Flow Matching for Generative Modeling.” In arXiv [Cs.LG]. October 6. arXiv. http://arxiv.org/abs/2210.02747.

Reviewer_rJhsOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces the Interventional Graph World Model (IGWM) to predict single-cell transcriptomic responses to unseen genetic intervention combinations. The model represents interventions as continuous state transition operators using conditional flow matching. It constructs a base candidate graph using Pearson correlations from untreated control cells, then dynamically sparsifies this graph using a state- and intervention-conditioned editor to select K=512K=512 active edges. The active graph drives message passing to parameterize a velocity field, which is integrated via explicit Euler steps to predict the treated cell state. The model is evaluated on Norman19 and VCC-H1, showing improvements over CPA and GEARS.

Critical strengths

Formulating the cellular transition via conditional flow matching without relying on paired trajectories is a modern upgrade to GEARS in terms of population-level distribution matching.

Critical weaknesses
  1. I think this paper has fundamentally unsound causal foundations. The paper claims to present an "interventional world model," but there are two big issues. First, the underlying candidate graph is constructed entirely using Pearson correlations of steady-state control cells. Correlation captures co-expression, which is heavily confounded by shared upstream regulators, rather than direct causal gene-to-gene interactions. The bigger issue: as pointed out by the graphVCI paper, older matching methods like GEARS often lack rigorous supervision for counterfactual outcomes and struggle with true causal Individual Treatment Effect (ITE) estimation. GEARS' random pairing of samples in treatment and control groups fundamentally does not respect the third rung on the ladder of causality, making it a very questionable method for individual-level treatment effect estimation and more suitable for estimating shared effect across population, and this work inherited the same problem. The author even stated "Recent benchmarks also show that apparent gains may reflect shared treatment effects rather than perturbation specific biology" in Introduction, yet inherited from GEARS the exact training mechanism that is one of the most blatant culprits of this issue. If positioned as a empirical graph paper focusing on biological analysis like GEARS, I can see certain venues like the one GEARS was published in can let the theoretical flaws slide. I don't think I can let it slide in this venue, especially when the authors specifically framed it as a causal method. I think treating a correlation-based heuristic with random pairing as an interventional mechanism demonstrates a fundamental misunderstanding of causal inference requirements.

  2. The baselines in experiments are very outdated. The experimental validation relies exclusively on CPA and GEARS. While these are popular historical baselines, they are no longer state-of-the-art. The authors explicitly cite much newer 2025 and 2026 models such as PerturbNet and STAMP in their Related Work, yet excluded them from the experiments. And as stated in the previous weakness, the authors cited recent benchmark studies (e.g. Ahlmann-Eltze et al., 2025) which warn that deep learning perturbation models frequently fail to outperform simple linear or additive baselines. Despite acknowledging this literature, the authors did not include a naive baseline to demonstrate that their highly complex flow-matching and graph-editing architecture is actually necessary or biologically meaningful.

Questions for authors

Questions corresponds to the weaknesses above.

  1. How do you theoretically justify characterizing IGWM as an "interventional" model when the foundational transition graph is derived from non-causal Pearson correlation, and with a random pairing mechanism that fundamentally does not respect the basics of individualized treatment effect?

  2. Why were the 2025/2026 state-of-the-art models cited in Related Work excluded from empirical evaluations?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.