acceptodds
Under review as a conference paper at ICLR 2027

Hypothesis Generation and Semantic Attribution in Inverse Perturbation Modeling

Abstract

Understanding gene regulatory relationships at scale can help identify intervention points and anticipate unintended effects. However, connecting a black-box model's learned computation to biological knowledge remains difficult, particularly for unfamiliar cellular states. We investigate whether a sparse, typed gene network can generate inspectable hypotheses for steering malignant transcriptional programs toward expression patterns associated with favorable prognosis. Using LINCS L1000 CRISPR/Cas9 profiles paired with matched plate controls, we train a graph neural network to recover the perturbed gene from baseline and endpoint expression. Its nodes combine ProTrek-derived protein representations with expression features; its edges encode signed regulatory relationships and other gene interactions. Our focus is how representations learned from perturbation responses use these identities and relationships to support individual nominations. We then query the model in a specified cancer-cell context with endpoints that upregulate favorable prognostic markers and downregulate unfavorable ones. At this fixed request, integrating from mean projected protein representations to actual representations allocates a nomination score difference through node states, individual layer-edge messages, and residual connections. In detailed A549-PYGL and A549-CCDC86 cases, supporting and opposing contributions coexist. Connected supporting structures suggest hypotheses involving glycogen-associated regulatory/metabolic propagation and physical interactions associated with ribosome biogenesis. Preserving gene and relation identities makes these learned computational dependencies explicit, allowing comparison with biological knowledge and framing mechanistic hypotheses for experimental investigation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
GNN, Graph, Proteins, Genes, Regulatory Networks, Deep Learning, Interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_rayGOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a graph neural network to identify perturbed gene based on expression profiles, then uses it to nominate perturbations in a cancer expression profile that's held out. Gradient attribution is used to analyse nominations.

Critical strengths
  1. The focus on hypothesis generation is interesting. Also I quite like the careful experimental and computational setup. the model also explicitly separate biological causal effect from message passing.

  2. The benchmarking is robust.

Critical weaknesses
  1. I am not too sure about why reverse prediction is required here. If we want to match perturbation responses, why not predict each candidate's response with a forward perturbation model then rank? I think comparison to these kind of the models would make the empirical results a lot stronger.

  2. The model depends on selecting valid endpoints. However the correctness of endpoints are not validated on held out data

Questions for authors
  1. Can you isolate the effect of classifier trivially predicting which gene are requested to decrease rather than predicting genes that are responsible for the perturbed profile?
1 comment

We thank the reviewer for the interest in hypothesis generation and the comments on our setup and benchmarking.

1. Why inverse rather than forward-then-rank? (W1) We start from a desired change in the cell and ask what supports a nomination for it. The value of this analysis is to move from a candidate name to a specific account of which parts of the desired program support it and which biological relationships provide supporting or opposing contributions. Because the desired program is the inverse model's input, attribution traces each nomination to individual requested changes and the relationships that combine them. For TTK in A549, decreases in checkpoint genes support the nomination, while E2F4, a regulator of TTK transcription, opposes it (Figure 2). The same analysis could serve requests for treatment sensitivity, lineage maturation or beta-cell function (Introduction).

Forward-then-rank is a valid route, and a matched comparison would be the right test of ranking performance. Our claims concern what supports a nomination rather than ranking, so we did not train that benchmark.

2. Does the model just pick the gene requested to decrease? (Q1) We tested this directly. For the original examples it largely does: removing the own decrease moves PYGL and CCDC86 from the top 10 to beyond rank 2,000 (Table 1). We then applied this control to every candidate, requiring its score to rise under the remaining request and fall under reversal: 333 nominations pass in nine held-out lines. TTK/A549 is not among the requested changes and passes in both models that held A549 out (Section 6.1).

3. Is the endpoint supported on held-out data? (W2) The A549 markers come from HPA's validated tier, whose prognostic associations replicate across cohorts (Appendices C.1 and K list every context's tier). In PRECOG, a separate patient-cohort compendium not used to select markers, directions agree for 85.8% of 2,051 matched entries. Measured TTK knockouts reach the 98.2nd percentile of requested-direction agreement among 5,115 A549 knockouts (10 wells). Two SD is our chosen amplitude; at a smaller size anchored to measured knockout displacement, TTK keeps its request and reversal signs, while CKS2 weakens (Appendix M).

Does this address the forward-model and endpoint questions, or is there a remaining point we can clarify?

Reviewer_fuM7OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper trains a graph-based inverse model to identify gene knockouts from control and perturbed expression. It then uses a desired expression program based on prognosis markers to nominate interventions. Protein representations, gene identities, and typed interactions are kept explicit. Conductance attributes each nomination's score difference to nodes, messages, and residual connections relative to a protein reference. Two A549 examples illustrate different supporting networks, and message suppression checks whether the attributions track changes in the model's scores.

Critical strengths
  • Making inverse perturbation predictions inspectable is a useful problem. Keeping gene and interaction identities in the computation makes the proposed hypotheses easier to examine.
  • The attribution accounting is careful. The paper separates positive and negative contributions, tracks residual paths, and checks finite message suppression. This gives more evidence about model behavior than a plausible-looking network alone.
Critical weaknesses
  • Both case studies explicitly request lower expression of the nominated gene itself. The paper does not show whether the broader prognosis program drives these nominations beyond that direct cue. The attribution varies protein representations while keeping expression fixed, so it does not answer this central question.
  • The biological value beyond the supplied interaction graph is not established. The method combines existing graph modeling and conductance ideas, but the predictive comparison mainly changes graph depth. The selected pathways and message suppression show how the fitted model works, without showing that its learned structure produces better intervention hypotheses than a simpler inverse model. No measured response is used to check whether either nomination moves the other requested markers in the intended direction.
Questions for authors
  1. If each nominee's own expression edit is removed, do its ranking and supporting network still respond to the remaining prognosis program? What happens when that program is reversed?
  2. Do the available perturbation responses support the requested changes in the other markers for these nominees, and does a graph-free inverse predictor give similar nominations?
3 comments

We thank the reviewer for questions we could test directly. We tested each one.

1. Own edit removed, program reversed (W1, Q1). For PYGL and CCDC86, the test confirms the concern: removing the own decrease moves both beyond rank 2,000, and both rank first when only that decrease is kept and all other changes are reversed (Table 1; Table 8). Used as a screen over every candidate (own change removed, rank ≤200, score rising under the request and falling under reversal in every eligible model), the test passes 333 nominations in nine held-out lines. TTK/A549 is not among the requested changes; its score rises by +1.52/+1.33 under the request and falls by −1.83/−1.22 under reversal in H1/H3, both of which held A549 out (Section 6.1).

2. Beyond the supplied graph (W1, W2). We added attribution from no expression change to the request, with proteins fixed (Section 5.1). The graph supplies possible relationships; the fitted model decides how each one counts. For TTK, decreases in MAD2L1, BUB1B and PLK1 support the nomination, while the BRCA1 decrease opposes it in H1. Unedited MAD1L1 supplies the strongest supporting final message, and the message from E2F4, a regulator of TTK transcription, opposes it (Figure 2; Table 10). Of the input attribution, 30.2%/26.3% comes from outside TTK's 92 immediate graph neighbors. On the original protein-reference attribution, resetting the trained weights while keeping the graph gives allocations uncorrelated with the trained ones (|ρ| ≤ 0.045; Appendix R).

3. Measured responses of the other markers (Q2). Excluding each target's own marker, CCDC86 knockouts reach the 88.5th percentile of requested-direction agreement (9 wells), meeting our ≥75th-percentile criterion; PYGL reaches the 70.8th (17 wells). TTK reaches the 98.2nd (10 wells), and its input-attribution profile correlates with the measured marker displacement (ρ = 0.492; 98.8th percentile among 4,654 other A549 knockouts; Appendix O). Table 15 reports all 19 screen cases, including those the measurements do not support.

4. Graph-free predictors (W2, Q2). A node-local model, which cannot use other genes' edits, reproduces PYGL and CCDC86, as their own-edit dependence predicts. A global MLP shares only 2–7 of its top 50 nominations with the GNN. In observed-knockout recovery, the GNN leads on held replicates and the node-local model on unseen cells (Appendix R). We use the GNN for inspectability rather than ranking: its cross-gene computation runs through named, typed relationships that can be examined, as in point 2.

Does this address your questions, or is there a remaining point we can clarify?

I thank the authors' response. In general, I think the answers resolve my previous concerns. However, I have a further quesion about the method part. I think PDGrapher seems applicable to the tasks of recovering knockout targets from control–perturbed expression pairs and nominating interventions for a desired expression program. And this paper is cited but not included in the benchmark comparisons. I think it is a good method to compare to see the method contribution of the paper. Is there a specific reason it was not included as a baseline?

We thank the reviewer for the follow-up. We agree that PDGrapher is relevant to the shared prediction task and cited it as a precedent for expression-based intervention inference. Our study prioritized the question of how the model learns to combine expression evidence and specific biological relationships to support an individual nomination. We attempted a preliminary PDGrapher comparison during development. Its published formulation differs from our task in how interventions and expression are represented and in what training objective it optimizes; adapting it to our benchmark produced numbers that were difficult to interpret as a test of the method rather than of our adaptation. At that point, we judged that targeted baselines constructed for our question would be more informative than a partially aligned external comparison.

In the revision, we report two fitted controls to examine this question under different information and architectural constraints. The expression-only MLP receives the full control and endpoint expression vectors, without protein features or graph connectivity. The separately trained protein-informed node-local model retains the protein projector, expression encoders, and shared decoder, but excludes cross-gene message passing: each candidate’s raw score depends only on its own expression and protein representation. These controls examine recovery without explicit protein or graph features and without cross-gene communication, respectively. Attribution and message-deletion analyses then examine how the full model uses the supplied biological relationships to combine evidence across genes for individual nominations. These analyses address how nominations depend on expression evidence and the supplied biological relationships; they do not establish comparative performance against PDGrapher.

Reviewer_fx3COpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an interpretable framework for inverse gene perturbation modeling that identifies candidate perturbation targets from baseline and desired gene expression profiles. The authors train a relation-aware graph neural network on LINCS L1000 CRISPR/Cas9 perturbation data, integrating protein representations, gene expression, and biological interaction networks. To interpret individual predictions, they employ conductance-based attribution to decompose prediction scores into contributions from gene representations, relation-specific messages, and residual connections. The approach is evaluated through perturbation target recovery across cell lines and detailed case studies involving PYGL and CCDC86 in A549 cells. These analyses demonstrate how model computations can be connected to biologically interpretable hypotheses.

Critical strengths
  1. Fine-grained attribution of gene-level and interaction-level contributions.

A key strength is the attribution framework that explicitly traces prediction scores through gene representations, individual graph messages, and residual connections while preserving biological identities and relation types. The authors also examine signed contributions and conductance completeness, providing a more transparent account of the model's internal computations than conventional feature-level explanations. This is valuable for interpreting graph-based perturbation models.

  1. Integration of biological knowledge with computational hypothesis generation.

The framework combines pretrained protein representations with regulatory and functional interaction networks, enabling model predictions to be examined in terms of recognizable biological entities and relationships. The PYGL and CCDC86 case studies illustrate how distinct computational dependencies can motivate biologically meaningful hypotheses. The additional message-suppression analyses provide evidence that attributed interactions influence the fitted model's predictions.

Critical weaknesses
  1. Lack of direct validation of the proposed inverse perturbation recommendations.

The quantitative evaluation primarily measures whether the model can recover the identity of an observed perturbation from baseline and post-perturbation expression profiles. However, the intended application is to recommend perturbations that produce a desired transcriptional state, which is a substantially different objective. The paper does not quantitatively establish whether its recommended interventions actually induce the desired expression changes. Moreover, the desired profiles are constructed using prognostic associations rather than experimentally validated intervention outcomes. This gap substantially limits the demonstrated effectiveness of the proposed inverse modeling framework.

  1. Limited validation of attribution reliability and biological generalizability.

The mechanistic interpretation is demonstrated primarily through two selected perturbation candidates in a single cancer-cell context. Although conductance completeness and message-suppression analyses establish computational consistency, they do not demonstrate that the identified pathways represent genuine biological mechanisms. Furthermore, the analysis focuses on attribution relative to a protein representation reference rather than directly explaining which desired expression changes drive each recommendation. Broader quantitative evaluation across perturbation targets, cellular contexts, and independent model fits is necessary to establish the reliability and generalizability of the proposed explanations.

Questions for authors

See the Weakness

1 comment

We thank the reviewer for the assessment of the attribution framework, including its signed contributions and completeness checks. The revision adds measured knockout comparisons and attributes the desired expression changes directly, across targets, cell lines and fitted models.

1. Do nominated knockouts produce the requested changes? (W1) Excluding the target's own marker, existing TTK/A549 knockouts reach the 98.2nd percentile of agreement with the requested directions among 5,115 A549 knockouts (10 wells). For TTK, the attribution also points to the markers that move: its input-attribution profile correlates with the measured marker displacement (ρ = 0.492; 98.8th percentile among 4,654 other knockouts; Section 7). Table 15 reports all 19 screen cases, including those the measurements do not support.

2. Which desired changes drive the nomination? (W2) We added attribution from no expression change to the desired program, with proteins fixed; input attribution and network conductance account for the same score change to within 1.6 × 10⁻⁸ (Section 5.1). For TTK, decreases in MAD2L1, BUB1B and PLK1 support the nomination, the unedited checkpoint gene MAD1L1 supplies the strongest supporting final message, and the message from E2F4, a regulator of TTK transcription, opposes it. One prediction thus separates checkpoint support from transcriptional opposition (Figure 2). Appendix S adds six cases linked to experimental literature.

3. Reliability beyond the examples (W2). In a pre-declared study of 333 nominations in nine held-out lines, we deleted individual messages from panels containing the strongest positive and negative messages plus low-attribution and random controls. The mean correlation between attribution and deletion effect is positive in all nine lines at every transition (0.86, 0.74, 0.57, 0.43; p = 0.001), while message size alone gives 0.00–0.11 (Section 6.2; Figure 3). Agreement weakens with depth, and TTK's final-transition panels are negative (Table 13).

4. Separately fitted models (W2). H1 and H3 are trained on different cell-line partitions and both held A549 out; TTK's main inputs, MAD1L1 support and E2F4 opposition recur in both. They share a seed, so this tests partition variation. Independently seeded refits of the protein-reference attribution show positive within-fit deletion correspondence and weak early-layer agreement across seeds (Appendices Q–R).

We hope this addresses the request for broader evaluation, and we are glad to clarify any remaining point.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revision summary and response to the reviews

We thank the reviewers for specific questions that we could test directly. The revision keeps the submitted model, data, training and attribution method, and adds those tests.

Own requested change (fuM7, rayG). Removing each nominee's own requested change and reversing the rest confirmed that the original PYGL and CCDC86 nominations rest on their own decreases (Table 1). Applied to every candidate, the same control finds 333 nominations in nine held-out lines whose scores rise under the remaining request and fall under its reversal. Our worked example, TTK in A549, is not among the requested changes.

Attribution of the request (fuM7, fx3C). We now attribute along the path from no expression change to the request. The aim is to go beyond a target name: to show which requested changes support a nomination and which biological relationships support or oppose it. For TTK, decreases in checkpoint genes support it, the unedited checkpoint gene MAD1L1 carries the strongest final support, and E2F4, a regulator of TTK transcription, opposes it (Figure 2). Across the 333 nominations, message attributions correlate positively, on average, with the effects of deleting those messages in every line at every transition (Figure 3).

Measured knockouts (fuM7, rayG, fx3C). Existing TTK knockouts move the other requested markers in the requested directions at the 98.2nd percentile of 5,115 A549 knockouts (10 wells; Section 7).

Section 6 and Figure 2 are the clearest entry points to the revision. We welcome further questions before the discussion closes.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.