acceptodds
Desk rejected at ICLR 2027

scPathCF: Intervention-Based Path Decomposition for Perturbation-Aware Protein Prediction

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

A central goal in functional genomics and drug discovery is to predict how protein levels change after a genetic perturbation, especially for surface proteins that serve as immune checkpoints and therapeutic targets. A protein change can follow a corresponding RNA change or arise from regulation that is not explained by the measured RNA. Existing methods typically predict protein directly from RNA or predict the perturbed transcriptional state, without separating these two sources of protein change. We present \method, a perturbation-aware predictor that decomposes protein abundance into a control baseline, an RNA-mediated increment, and a perturbation-dependent residual increment. The mediated head reads RNA-related representations but not perturbation identity, preventing a direct identity-to-protein shortcut. The residual head reads the perturbation embedding and an optional offline screening prior, but not factual RNA or cell context. Hard branch switches assemble a control, a mediator-blocked, and a factual state; their differences define the two increments as executable interventions. Across three public CRISPR–CITE-seq screens, the decomposition retains competitive predictive accuracy with a strict-zero prior on Papalexi, a full-data screen on Frangieh, and curated annotations on CaRPool, and the branch interventions reproduce contrasts consistent with known post-transcriptional and RNA-coupled regulation. The state differences provide an explicit account of how the fitted predictor allocates each response between its branches. Code is available at https://osf.io/xhzes/?view_only=d03a9e76234b425ead29d60f35586c85.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell perturbation prediction, CRISPR-CITE-seq, protein abundance prediction, transcript-to-protein modelling, attribution, counterfactual decomposition, foundation model adapter

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_gU9hOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes scPathCF, a perturbation-aware framework for predicting protein abundance from RNA measurements and perturbation identities. The method decomposes protein predictions into a control baseline, an RNA-mediated component, and a perturbation-dependent residual component. Hard branch switches provide operational interventions for examining these contributions. Experiments on three CRISPR–CITE-seq datasets evaluate predictive performance and the biological consistency of the resulting decomposition.

Critical strengths

The paper addresses an important problem: distinguishing RNA-associated and potentially post-transcriptional contributions to protein responses. The explicit separation of prediction pathways offers a potentially useful approach to model interpretability. Evaluation across three datasets, together with component ablations and screening-prior provenance analyses, provides encouraging preliminary evidence. The authors also acknowledge several important limitations of their decomposition.

Critical weaknesses

Weakness 1: The proposed path decomposition lacks sufficient validation, making its central interpretability claim questionable.
a. The decomposition is architectural rather than biologically identifiable. The so-called RNA-mediated branch also receives perturbation-specific information through identity-indexed RNA prototypes, while the residual branch uses perturbation embeddings. Consequently, perturbation-specific effects can potentially be redistributed between the two branches without changing the final prediction. The hard branch interventions demonstrate how the architecture allocates predictions, but do not independently establish genuine RNA-mediated and RNA-independent effects.
b. The biological validation is insufficiently independent. The reported agreement between residual contributions and biological annotations is partly influenced by screening-derived labels and supervision. The paper does not demonstrate that the same path assignments remain stable without these priors or under alternative decompositions. The presented examples establish biological plausibility rather than validating the reliability of the inferred paths.
This is consequential because interpretable path decomposition is the paper's primary contribution. Without stronger evidence that these assignments are meaningful and stable, the method risks providing architecture-dependent explanations rather than reliable biological insights.
Weakness 2: The predictive advantages are not sufficiently convincing, particularly under controlled comparisons and stronger generalization settings.
a. Marginal improvements and unequal prior information. The performance differences against the closest baselines are not statistically significant across all three datasets. More importantly, the Frangieh comparison uses a target-derived full-data screening prior unavailable to the baselines. Removing this prior increases MSE by 0.0195, substantially exceeding the method's original 0.0045 advantage over cTP-net. This raises serious concerns about whether the reported improvement originates from the proposed architecture.
b. Limited evidence of generalization to unseen perturbations. Although the method performs competitively under unseen-guide splits, the more challenging held-out CaRPool three-gene combination reveals a substantial weakness: scPathCF achieves MSE 0.6904, compared with 0.4846 for GEARS. Furthermore, some reported aggregate metrics are influenced by baseline protein-abundance differences rather than perturbation-specific responses.
The results do not yet convincingly establish that the proposed decomposition improves either predictive robustness or perturbation-specific generalization over competitive alternatives.

Questions for authors

Q1. Since perturbation identity enters both the mediated and residual pathways, how can the authors establish that the learned decomposition reflects meaningful RNA-mediated versus residual effects rather than an arbitrary allocation between model branches? Can the authors demonstrate the stability of these assignments under strict-zero priors, prototype ablations, and alternative training configurations?
Q2. Can the authors provide strictly comparable experiments without target-derived screening priors, particularly on Frangieh, and demonstrate that scPathCF retains a meaningful advantage over strong RNA-to-protein baselines? How much of the improvement remains after controlling for differences in prior information?
Q3. How does the method perform under genuinely unseen perturbation identities and combinations, using evaluation metrics focused on perturbation-induced protein changes rather than overall abundance? Can the authors explain the substantial performance gap against GEARS on the held-out CaRPool combination?

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

This paper presents a method, called scPathCF, a perturbation-aware predictor that converts protein abundance into a control baseline, an RNA-mediated increment and a perturbation-dependent residual. Understanding how protein levels change following to a genetic perturbation is an important task in drug discovery. The method is tested on three datasets and achieves competitive predictive accuracy, albeit with a small difference in several experiments. However, while the method predicts the overall protein abundance accurately, there is a gap in that it’s unclear if the intermediate representation (mediated head, residual head) outputs correspond to the true biological processes they are intended to represent.

Critical strengths

The paper addresses an important problem in predicting protein responses to genetic perturbations. The argument about how protein levels are related to, but separate from RNA changes helps motivate the approach taken in the paper. It also gives the paper a clear biological purpose.

The architecture is logical and the explicit branch switching enables inspection of how the model allocates its predictions across the mediated and residual heads.

Evaluation across three datasets is appreciated by this reviewer, and the method achieves competitive performance.

Critical weaknesses

While the paper shows competitive performance across different datasets, accurate prediction of protein abundance does not establish whether the mediated and residual heads cleanly separate the biological contributions they are intended to represent. They may sum up to the predicted protein abundance, but it wasn’t clear to this reviewer what prevents the mediated and residual heads from capturing overlapping, confounded biological effects. Independent validation, or controlled experiments with known mediated and residual effects, is needed to establish the decomposition’s meaning. This concern relates to the main contribution of the paper.

Another weakness is the mathematical notation and description of the method. For example, Equation 1 introduces some symbols that aren’t defined when introduced, like \mathbf{M} and in Equation 1, \mathbf{z}_C is a vector but in the Figure 1 caption it is a scalar. The indexing 0M0, 1M1 and 1M0 isn’t clearly described so the reader must try to infer what it means. Architectural components and training objectives are introduced densely without a clear explanation of their roles and relationships. Unfortunately the text feels rather AI-written and hard to follow as a human reader in places. These issues make the exposition difficult to understand, particularly what is being intervened on and how.

Questions for authors
  1. What evidence shows the mediated and residual heads separate their intended biological contributions, rather than capture overlapping, confounded biological effects?

  2. Please clearly define subscripts 0M0, 1M0 and 1M1, and clarify how the head functions in Equation 1 relate to m_0, m_1 and r in Equations 3–4.

  3. Are the results that show the best performance using the proposed method statistically significant?

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

scPathCF predicts surface protein abundance from post-perturbation RNA and perturbation identity, decomposing it into a control baseline, an RNA-mediated increment, and a residual increment. Hard branch interventions quantify these contributions, evaluated for predictive accuracy and biological consistency across three CRISPR–CITE-seq datasets.

Critical strengths
  • S1. Separating RNA-coupled and RNA-independent protein responses is biologically motivated, and explicit path contributions support interpretability.
  • S2. The CMTM6–PD-L1 and IFN signaling examples show path assignments consistent with known biology.
  • S3. The model achieves competitive prediction across distinct datasets. The screening-prior provenance analysis, including prior-free evaluation on Papalexi, is also valuable.
Critical weaknesses
  • W1. The meaning and stability of path assignments remain unclear because both branches receive perturbation identity.
    • Branch interventions define path contributions by construction, but both heads access perturbation identity through embeddings or RNA prototypes. Moreover, removing the residual head barely affects prediction in the archived screen-enabled Papalexi ablation, even though Papalexi provides the representative residual-regulation case. Thus, the decomposition's added interpretive value remains unclear.
    • The key CMTM6→residual and IFN→mediated assignments have not been verified without screening priors. Target-derived RH annotations and associated losses can influence these assignments, yet the prior setting for the biological-consistency analysis is unspecified. The existing strict-zero checkpoints could directly test their stability.
  • W2. The predictive comparisons do not establish the added value of perturbation-aware protein modeling.
    • The main Frangieh comparison uses a target-derived screening prior unavailable to the baselines. Removing the prior reduces performance by more than the reported margins over the closest competitors, leaving prior-independent superiority uncertain. A matched strict-zero or train-only comparison is needed.
    • scPathCF receives factual post-perturbation RNA, unlike GEARS and N-gram ridge. Against cTP-net, which also uses factual RNA, its performance differences on Papalexi and Frangieh have confidence intervals including zero. On the archived Papalexi predictions, differential correlation is no better than RNA-only models. The benefit of additionally modeling perturbation identity therefore remains unproven.
  • W3. The generalization results depend on perturbation-specific information available under each split.
    • The main unseen-guide split allows other guides for the same gene in training, making learned perturbation embeddings and identity-indexed RNA prototypes available. This evaluates new-guide prediction rather than generalization to unseen perturbation identities.
    • The identity-holdout evidence is also limited: Frangieh reuses baseline results from the unseen-guide evaluation, while scPathCF underperforms GEARS and n-gram ridge on the single held-out CaRPool combination. How the residual head represents genes absent from training is also unclear.
Questions for authors
  • Q1. How are the learned prototype gates distributed, and how do path assignments change when prototypes are removed or restricted?
  • Q2. Do the CMTM6 and IFN assignments persist in strict-zero checkpoints? Which prior setting was used for the biological-consistency analysis?
  • Q3. How does Frangieh perform with strict-zero or train-only priors under matched information access?
  • Q4. Do macro per-protein and differential-response metrics support the results on strict-zero Papalexi and the other datasets?
  • Q5. How are embeddings defined for unseen perturbation genes? Can Frangieh baselines be rerun under identity holdout?

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

Related papers

Discussion (0)

Sign in to comment.