acceptodds
Under review as a conference paper at ICLR 2027

Beyond Perturbation Prediction: Learning Directed Gene Structure with RISCAL

Abstract

Predicting genetic perturbation responses and recovering their regulatory mechanisms are central to understanding gene function and therapeutic target prioritization. Existing methods often emphasize response prediction or structure recovery, while assuming perfect perturbations or modeling imperfect ones as unrestricted mechanism changes. We introduce RISCAL (Residual-activity Interventions for Structural Causal Learning), a nonlinear gene-level structural causal model that jointly learns directed dependencies and predicts responses under residual-activity interventions. Its intervention model captures imperfect loss-of-function by separating gene-specific residual functional activity from measured transcript abundance. RISCAL accurately recovers synthetic ground-truth DAGs and matches the strongest predictive baselines on LINCS L1000 profiles from ten cell lines. Its predicted graph cores are enriched for CollecTRI/SIGNOR interactions and form coherent, partially conserved functional programs. RISCAL can also incorporate unseen regulators and predict responses consistent with experimental evidence.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Causal discovery, Structural Causal Models, Interventional Learning, Interventional Causal discovery, Perturbation Response Prediction, Gene Regulatory Networks, Computational Biology

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_eT2TOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RISCAL learns a directed gene graph from paired control and CRISPR perturbation profiles and uses the same model to predict perturbation responses. Each gene has its own nonlinear mechanism determined by its inferred parents. A perturbation is modeled as only partially removing the target gene's effective activity, while the target's measured expression is handled by a separate readout. Evaluation covers synthetic data with known ground-truth graphs, ten LINCS cell lines, and an experiment involving an unseen regulator.

Critical strengths

Separating functional activity from measured RNA abundance is a useful modeling idea, since incomplete loss of function does not necessarily correspond to the measured target transcript going to zero. The synthetic generator is also unusually detailed and introduces biological complications incrementally. The real-data splitting, control pairing, and batch handling appear more careful than in many perturbation-prediction evaluations.

Critical weaknesses

My main concern with the synthetic evaluation is that the later generator stages closely mirror RISCAL's modeling assumptions, including residual activity, nonlinear child-specific mechanisms, sparse directed structure, and reused controls. RISCAL therefore has a substantial specification advantage on these benchmarks. The results still show that the implementation can recover the structure it is designed for, but they tell me less about performance under misspecification. SDCD being essentially at the Random-DAG level on the 3,710-gene benchmark also makes me wonder whether this is a meaningful comparison at that scale. Several closely related joint structure-and-prediction methods discussed in the paper, including DCD-FG, GIM, and Bicycle, are not included in the empirical comparison.

On real data, I am not yet convinced what predictive benefit comes specifically from RISCAL's learned causal structure. RISCAL, GEARS, and the axial transformer operate in almost the same predictive regime, even though GEARS does not receive each test profile's paired control as an input. That makes it difficult to isolate the value of the paired-control information and learned DAG. A very simple reference such as a mean-response predictor would help. The linear-regression and MLP baselines also have macro intervention MSE more than an order of magnitude worse than RISCAL, GEARS, and the axial transformer, which is surprising enough that I would like a sanity check or explanation.

The learned residual-activity distributions on LINCS are also extremely concentrated: the global means are very similar across cell lines and the across-gene standard deviations are small. This makes it unclear how much the gene-specific part of the residual-activity model contributes on real data. Finally, the curated-interaction enrichment is statistically significant against the degree-preserving nulls in many cell lines, but the absolute overlap is still only a small number of curated edges among thousands of learned edges. I would find the evidence more convincing with the same enrichment analysis applied to alternative learned graphs.

The paper is also difficult to read. Much of the information needed to interpret the main claims is in a very long appendix, and I had to move back and forth repeatedly to understand the evaluation.

Questions for authors
  1. What does a very simple mean-response baseline achieve on the LINCS prediction task?

  2. On LINCS, what happens if the gene-specific residual-activity parameters are replaced by a shared scalar residual activity? Does the gene-specific parameterization materially improve prediction or graph stability?

  3. What does the same CollecTRI/SIGNOR enrichment analysis give for alternative inferred graphs such as GENIE3 or a correlation-network baseline?

  4. For the unseen-regulator experiment, how well does simply averaging two measured replicates predict the third? I would like a simple experimental-variability reference before interpreting agreement with the external experiment as evidence for generalization.

5 comments
AuthorsOpenReview ↗

Weaknesses — paragraph 1: synthetic evaluation and related causal methods

We sincerely thank the reviewer for the careful assessment and constructive feedback. Their comments have prompted additional experiments and important clarifications, substantially improving the scientific rigor of our evaluation and the support for our claims. Below, we address each concern and question in detail. All changes are highlighted in blue in the revised manuscript.

Weaknesses — paragraph 1: synthetic evaluation and related causal methods

Thank you for highlighting these concerns. To assess RISCAL under alternative data generating mechanisms and against closely related causal learning methods, we expanded the synthetic evaluation and clarified the comparisons

Recovery with different generating mechanisms: we generated two new families of synthetic data at PC3 scale, retaining the same DAG for each pair of seeds. RISCAL retains its architecture, tanh mechanisms, and official training recipe. We did not adapt the model to the new families or use the ground-truth DAG to select checkpoints.

The first family uses additive softplus mechanisms. The second introduces multiplicative interactions between parents, so that the influence of a gene depends on the state of other parents:

Generating familyAUPRCF1Acyclic graphs
Original0.633 ± 0.0160.649 ± 0.0185/5
Additive softplus0.663 ± 0.0180.695 ± 0.0185/5
Multiplicative interactions0.510 ± 0.0340.587 ± 0.0335/5

RISCAL maintains strong structural recovery under additive softplus mechanisms and recovers a more conservative subset of relationships under multiplicative interactions. In the latter setting, it achieves precision of 0.829 ± 0.042 and recall of 0.455 ± 0.039, selecting approximately 7,566 edges out of 13,732 true edges. These results extend the evaluation beyond the original generator, demonstrating recovery under different mechanisms while characterizing its reduced recall with multiplicative interactions.

See Section 7.1 and Appendix B.4, Tables 11–12.

Synthetic benchmark at PC3 scale: to provide more informative comparisons at this scale, we expanded the evaluation with InterSort and CASCADE, reporting native F1, oracle F1, and threshold-independent AUPRC:

Method, PC3 scaleNative F1Oracle F1AUPRC
RISCAL0.6490.6720.633
InterSort0.0050.5780.556
CASCADE0.460≈0.5780.430
SDCD<0.001<0.0010.001

InterSort and CASCADE recover structural information clearly above chance. Oracle threshold selection narrows the F1 gap, but RISCAL retains the highest F1 and AUPRC.

See Section 7.1, Table 1.

Comparison with closely related methods: given the limited rebuttal period, we prioritized CASCADE as a particularly close reference: it jointly learns a gene-level DAG, nonlinear mechanisms, and intervention coefficients to model perturbation responses. We added its evaluation on the two synthetic benchmarks:

Synthetic scaleRISCAL: F1 / AUPRCCASCADE: F1 / AUPRC
100 genes0.812 / 0.8500.418 / 0.365
PC3 scale0.649 / 0.6330.460 / 0.430

RISCAL achieves higher F1 and AUPRC at both scales. We also added its enrichment evaluation across all ten real cell lines, where RISCAL continues to show higher enrichment against CollecTRI and SIGNOR. We have also revised the related work to give DCD-FG, GIM, and Bicycle appropriate recognition and explain their contributions.

See Sections 2 and 7.1, and Tables 1–2.

AuthorsOpenReview ↗

Weaknesses — paragraph 2: graph utility, controls, and predictive baselines

Weaknesses — paragraph 2: graph utility, controls, and predictive baselines

Thank you for these observations. To assess the contributions of the learned graph and individual controls, we conducted additional ablations and evaluated simple predictive baselines. We also verified the implementations and results of linear regression and the MLP.

Contribution of the graph to prediction and interpretive value: we evaluated the RISCAL trained without the SCM term on PC3. The means across five fits are:

PC3Full RISCALWithout SCM
Edges per fit10,357.60
Response R²0.4250.427
Macro MSE × 10³7.2377.216
Response Pearson0.6460.649

Without the SCM term, the model predicts each non-target gene's learned baseline μ, while retaining the perturbed gene's specific readout. Both variants achieve comparable predictive performance, while the full RISCAL additionally learns interpretable regulatory relationships. With the same edge budget, RISCAL recovers more curated undirected pairs than coexpression:

ResourcePairs per cell line, RISCALPairs per cell line, coexpression
CollecTRI26.311.7
SIGNOR16.211.3

Furthermore, we identified 54 candidate relationships present in all five fits across all ten cell lines and absent from both resources. Our literature review found published biological support for 33 of them.

These results illustrate the graph's interpretive utility in recovering documented relationships and prioritizing candidates for further investigation. Additional biological evidence is presented in our response to Reviewer 2, Weaknesses 3 and Question 4.

See Appendix G.1, Table 38, and Section 7.3, Table 2.

Contribution of the individual control: prompted by the reviewer's comment, we replaced the individual controls with the same mean training control profile and retrained the model. Aggregate performance remains practically unchanged: on PC3, response R² is 0.4252 with the individual control and 0.4253 with the mean control. The recovered graphs are exactly identical with individual and mean controls across all five fits, both on real PC3 data and on the synthetic benchmark at PC3 scale.

While aggregate performance remains comparable, individual controls allow sample-specific predictions. Using the same mean control produces identical predictions for samples of a given intervention, whereas individual controls introduce within-intervention variation in the full predicted profile, increasing from 0 to 0.001324 ± 0.000052.

Thus, individual controls enable sample-specific variation, although this does not translate into improved aggregate performance in the evaluated setting.

See Appendix G.7, Table 43.

Absolute performance against simple baselines: we evaluated three simple baselines, computed using training data and evaluated on the same splits. RISCAL outperforms all three on average across the ten cell lines; the results by cell line are included in the paper.

Method, mean across ten cell linesResponse R²Macro MSE × 10³
RISCAL0.4287.473
Mean treated profile0.2729.362
Control + mean response−0.33217.145
Copy control−0.00513.129

See Section 7.2 and Appendix E.2, Tables 34–35.

Checks of linear regression and the MLP: we reviewed the code, configurations, and saved predictions for PC3. The models use the same splits; recalculating macro MSE reproduces the reported values, and loading the MLP checkpoint reproduces its predictions. We found no implementation or evaluation errors.

The MLP systematically underpredicts across all five fits. In fold 0, the mean bias per gene accounts for 97.5% of its MSE. The regression model, fitted without regularization, has large coefficients and considerably higher validation errors than training errors.

The numbers are therefore reproducible for the training recipes used, but we acknowledge that these baselines could be improved. We do not interpret their results as a general limitation of linear regression or MLPs, and we base the predictive comparison on the strong baselines and simple controls described above.

AuthorsOpenReview ↗

Weaknesses — paragraph 3: residual activity and biological enrichment

Weaknesses — paragraph 3: residual activity and biological enrichment

We thank the reviewer for these observations and separately address the contribution of gene-specific residual activity and the comparison of biological enrichment with other methods.

Contribution and interpretation of gene-specific residual activity: to explore this limitation, we generated a new synthetic dataset with more heterogeneous ρ: [0.05, 0.95] versus [0.30, 0.50] in the PC3-scale benchmark. We also increased the learning rate from 1e-5 to 3e-4. The Spearman correlation with true ρ is 0.425 for genes with children and 0.557 for genes with at least three children. These results support the parameter's ability to recover differences between genes.

To examine its consistency with real data, we froze the RISCAL models for all ten cell lines and re-estimated only residual activity. Its gene-specific component η has an adjusted Spearman correlation of 0.359 with the change in RNA expression of the perturbed gene, with positive correlations and p<0.05 in all ten cell lines. Although we distinguish RNA expression from functional activity in the paper, this agreement provides experimental support for the estimated differences.

See Appendix G.6 and Appendix D.6, Table 32.

We also evaluated a variant with a learned global ρ in response to the reviewer's comment. The means across five fits are:

EvaluationGene-specific ρGlobal ρ
Real PC3: response R²0.4250.425
Real PC3: macro MSE0.0070.007
PC3-scale synthetic: response R²0.4950.495
PC3-scale synthetic: macro MSE0.0070.007
PC3-scale synthetic: graph F10.6490.653
PC3-scale synthetic: graph AUPRC0.6330.637

Both variants achieve comparable predictive and structural performance. The contribution of gene-specific ρ is its ability to represent and investigate differences between perturbations, supported by the recovery and biological association analyses above.

One example is the comparison with gene essentiality in DepMap. This dataset measures how much cell survival depends on each gene. We compared this measure with η from the RISCAL models and obtained a Pearson correlation of 0.241, with positive correlations and p<0.05 in all ten cell lines. This association is consistent with essential-gene perturbations producing stronger downstream transcriptional responses, represented through lower residual activity. Although essentiality and activity reduction are distinct concepts, this relationship further supports the parameter's interpretive potential.

See Appendix G.5, Table 42, and Appendix D.6, Table 31.

Biological enrichment compared with other graphs: we agree that enrichment should be interpreted alongside absolute counts and comparisons with other methods. We extended the same analysis to InterSort, SDCD, CASCADE, and KO→response. For directed relationships with native support selection, the means across the ten cell lines are:

MethodEdgesCollecTRI / SIGNOR pairsCollecTRI / SIGNOR enrichment
RISCAL8,592.49.6 / 10.13.376× / 3.587×
InterSort4,682,003.81,292.0 / 1,513.61.008× / 1.012×
CASCADE120,786.7573.9 / 346.51.054× / 1.057×
KO→response8,592.47.5 / 9.02.290× / 1.832×
SDCD17.10 / 0Undefined

RISCAL achieves the highest mean enrichment in both resources, while selecting considerably smaller graphs than InterSort and CASCADE. It also achieves higher enrichment than KO→response with the same edge budget. These results support the biological relevance of RISCAL's learned relationships compared with alternative inferred graphs.

See Section 7.3, Table 2, and Appendix D.3.2, Table 17.

Weaknesses — paragraph 4: readability

We thank the reviewer for this feedback. Following this comment and Reviewer 1's suggestion, we moved the compact training and prediction algorithm to the main text (Section 4.6, Algorithm 1) and shortened the synthetic-generator presentation (Appendix B). The algorithm summarizes the full procedure and points directly to the defining equations and parameter and hyperparameter tables (Tables 4 and 5). These changes aim to reduce back-and-forth between the main text and the appendix.

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

RISCAL jointly performs perturbation-response prediction and directed gene-structure learning while distinguishing residual functional activity from measured transcript abundance under incomplete CRISPR loss-of-function. It evaluates DAG recovery on a synthetic benchmark and assesses predictive performance and the biological plausibility of the learned graphs on LINCS data.

Critical strengths
  • S1. Separating functional activity from measured transcript abundance is a biologically well-motivated formulation, and it is carried consistently through both the structural model and response prediction.
  • S2. The synthetic benchmark adds biological complexity in controlled stages, and structure recovery remains informative at realistic network scale. The component ablations and oracle substitutions are also useful for understanding the roles and limitations of the design choices.
  • S3. Across ten LINCS cell lines, the model learns directed graphs while matching strong predictive baselines. The stability and enrichment analyses are appropriately not overinterpreted as evidence of causal correctness.
Critical weaknesses
  • W1. The added value of gene-specific residual-activity modeling is not isolated.
    • The learned residual activities show substantial deviations from the generating values and markedly reduced gene-specific variation at realistic scale. Real-data estimates are also nearly identical across cell lines. While the authors interpret this as numerical stability, consistency across fits does not establish that meaningful gene-specific differences have been learned.
    • Removing residual attenuation demonstrates its importance but does not isolate the contribution of gene-specific parameterization. A direct comparison with global attenuation is needed.
  • W2. The added predictive value of the learned structure remains unclear.
    • Although the paper describes structure learning and prediction as complementary objectives, removing the SCM term eliminates all selected edges while leaving predictive performance nearly unchanged. This supports their compatibility but provides little evidence that the learned structure materially contributes to prediction.
    • No corresponding no-graph baseline is evaluated on real data, and the absolute perturbation-response R² is not reported, making it difficult to assess improvements over simple baselines.
  • W3. The methodological distinction from existing causal-learning methods and the biological utility of the learned graphs are insufficiently established.
    • The child-specific mechanisms build on existing approaches such as DCDI and GIM, yet closely related causal-learning methods, including CASCADE, are not directly compared. This makes the added contribution of the residual-activity formulation difficult to assess.
    • Post-hoc threshold adjustment substantially narrows the structural recovery gap with competing methods. Although RISCAL retains an advantage in edge ranking, part of its reported advantage reflects differences in support-selection rules.
    • Graph stability and biological enrichment support the plausibility of the inferred structure, but do not show whether the learned directed relationships provide useful information beyond existing biological interaction resources.
Questions for authors
  • Q1. How does a global-attenuation variant compare in structure recovery and prediction? Could the authors also report the initialization and learned changes in residual-activity parameters?
  • Q2. How does removing the SCM term affect predictive performance on real LINCS data? Could the authors report the absolute perturbation-response R² alongside simple baselines?
  • Q3. What does RISCAL add relative to closely related causal-learning methods under matched experimental conditions?
  • Q4. Can independent perturbation evidence validate whether the inferred edge directions provide information beyond graph-level enrichment and functional coherence?
6 comments
AuthorsOpenReview ↗

W1. Gene-specific residual activity

We sincerely thank the reviewer for the thoughtful feedback and constructive suggestions, which have substantially strengthened the scientific quality of the manuscript, the robustness of our claims, and the rigor of our evaluation. Below, we address each weakness and question in detail. All changes are highlighted in blue in the revised manuscript.

W1. Gene-specific residual activity

Thank you for raising this point. To assess the contribution of gene-specific residual activity, we conducted additional analyses of parameter recovery, stability, biological relevance, and a direct comparison with global attenuation.

Recovery of gene-specific differences: we generated a new PC3-scale synthetic dataset with more heterogeneous ρ ([0.05, 0.95] vs. [0.30, 0.50] in the original benchmark) and increased the learning rate from 1e-5 to 3e-4. The learned values correlate with ground-truth ρ, achieving Spearman correlations of 0.425 for genes with children and 0.557 for genes with at least three children, supporting the model's ability to recover gene-specific differences.

To further assess biological relevance, we froze RISCAL models for all ten cell lines and re-estimated only residual activity. Its gene-specific component η correlates with the observed change in target-gene RNA expression (adjusted Spearman = 0.359), with positive correlations and p<0.05 in all ten cell lines. Although RNA expression and functional activity are distinct, this association provides experimental support for the learned differences.

See Appendix G.6 and Appendix D.6, Table 32.

Stability and meaning of the learned differences: in the RISCAL models, η has a Spearman correlation of 0.721 across the five fits of the same cell line, with p<0.05 in all 100 fit-pair comparisons across the ten cell lines, indicating stability in the relative ordering of genes.

After averaging the five fits for each gene, the mean Spearman correlation across the 10 cell lines is 0.492. This similarity is consistent with the experimental design, in which 96.1% of genes use the same CRISPR reagent identifiers across cell lines. Together, this consistency and the preceding recovery and agreement analyses support the parameter's ability to represent relative differences between perturbations.

See Appendix D.6 and Supplementary Table 30.

Direct comparison with global attenuation: in response to the reviewer's comment, we evaluated a variant with a learned global ρ. The means across five fits are:

EvaluationGene-specific ρGlobal ρ
PC3: response R²0.4250.425
PC3: macro MSE0.0070.007
Stage 13: response R²0.4950.495
Stage 13: macro MSE0.0070.007
Stage 13: graph F10.6490.653
Stage 13: graph AUPRC0.6330.637

Both variants achieve comparable predictive and structural performance. The contribution of gene-specific ρ is therefore its ability to represent and investigate differences between perturbations, supported by the recovery, stability, and biological association analyses above.

An additional example is the comparison with gene essentiality in DepMap, which measures how strongly cell survival depends on each gene. We compared this measure with η from the RISCAL models and obtained a Pearson correlation of 0.241, with positive correlations and p<0.05 in all ten cell lines. This association is consistent with essential-gene perturbations producing stronger downstream transcriptional responses, represented through lower residual activity. Although essentiality and activity reduction are distinct concepts, this relationship further supports the parameter's interpretive potential.

See Appendix G.5, Table 42, and Appendix D.6, Table 31.

AuthorsOpenReview ↗

W2. Predictive value of the learned structure & W3. Methodological distinction, threshold selection, and biological utility (I)

W2. Predictive value of the learned structure

Thank you for highlighting this point. To assess the contribution of the learned structure and contextualize predictive performance, we conducted additional experiments addressing both aspects.

Contribution of the graph to prediction: we evaluated the model trained without the SCM term on PC3. The means across five fits are:

PC3Full RISCALWithout SCM
Edges per fit10,357.60
Response R²0.4250.427
Macro MSE × 10³7.2377.216
Response Pearson0.6460.649

Without the SCM term, the model predicts each non-target gene's learned baseline μ, while retaining the perturbed gene's specific readout.

Both variants achieve comparable predictive performance, while the full RISCAL additionally learns interpretable regulatory relationships. With the same edge budget, RISCAL recovers more curated undirected pairs than coexpression: 26.3 vs. 11.7 in CollecTRI and 16.2 vs. 11.3 in SIGNOR, averaged across cell lines. Furthermore, we identified 54 candidate relationships absent from both resources, with published biological support for 33 of them. Further evidence of the graph's interpretive utility is presented in Weaknesses 3.

See Appendix G.1, Table 38, and Section 7.3, Table 2.

Absolute performance against simple baselines: we evaluated three simple baselines, computed using training data and evaluated on the same splits. RISCAL outperforms all three on average across the ten cell lines; the values by cell line are included in the paper.

Method, mean across ten cell linesResponse R²Macro MSE × 10³
RISCAL0.4287.473
Mean treated profile0.2729.362
Control + mean response−0.33217.145
Copy control−0.00513.129

See Section 7.2 and Appendix E.2, Tables 34–35.

W3. Methodological distinction, threshold selection, and biological utility

Thank you for these comments. To address the concerns regarding methodological novelty, structural recovery, and biological utility, we clarified the contributions of RISCAL and expanded the experimental comparisons.

Methodological distinction and comparison with closely related methods: we have clarified in the main text that our contribution combines gene-specific mechanisms, incomplete perturbations through ρ, separation of measured RNA from the target's functional activity, and joint learning of structure and prediction. We also now acknowledge more explicitly the components with precedents in DCDI and GIM.

We additionally evaluated CASCADE on both synthetic benchmarks. With 100 genes, RISCAL/CASCADE achieves F1 scores of 0.812/0.418 and AUPRC values of 0.850/0.365. Results at PC3 scale are shown below. The specific contribution of ρ is addressed in W1.

See Sections 2 and 4.3, and Section 7.1, Table 1.

Effect of the threshold on structural recovery: to disentangle the effect of support selection from structural recovery, we highlight oracle F1 and threshold-independent AUPRC in the main text. On the synthetic benchmark at PC3 scale:

MethodNative F1Oracle F1AUPRC
RISCAL0.6490.6720.633
InterSort0.0050.5780.556
CASCADE0.4600.5780.430
SDCD<0.001<0.0010.001

RISCAL achieves the highest oracle F1 and AUPRC, and its native F1 exceeds the oracle F1 of the competing methods. This supports its advantage beyond differences in support-selection rules.

We also transfer the mean of the five thresholds selected on the synthetic benchmark to all ten real cell lines. The table reports means across the ten lines:

MethodTransferred thresholdMean edgesCollecTRI pairsCollecTRI enrich.SIGNOR pairsSIGNOR enrich.
RISCAL0.1294611,245.73.17.181×2.75.660×
InterSort3.46344710,638.75.21.865×8.71.357×
SDCD1.1795432.60.0Undefined0.0Undefined
CASCADE0.9786675,655.737.22.060×29.01.639×

RISCAL retains a higher enrichment factor than both InterSort and CASCADE in all 20/20 combinations of cell line and resource. SDCD recovers no curated matches with the transferred threshold, so its enrichment factor is undefined.

See Section 7.1, Table 1, and Appendix D.3.3, Tables 18–20.

AuthorsOpenReview ↗

W3. Biological utility beyond existing resources (II)

W3. Biological utility beyond existing resources (continued)

Biological utility beyond existing resources: we extend the comparison on real data to InterSort, SDCD, CASCADE, coexpression, and KO→response. The following table shows the means across the ten cell lines:

Directed relationships

MethodEdgesCollecTRI pairsCollecTRI enrich.SIGNOR pairsSIGNOR enrich.
RISCAL8,592.49.63.376×10.13.587×
InterSort4,682,003.81,292.01.008×1,513.61.012×
SDCD17.10.0Undefined0.0Undefined
CASCADE120,786.7573.91.054×346.51.057×
KO→response8,592.47.52.290×9.01.832×

Undirected pairs

MethodEdgesCollecTRI pairsCollecTRI enrich.SIGNOR pairsSIGNOR enrich.
RISCAL8,592.426.33.605×16.22.845×
Coexpression8,592.411.75.537×11.32.682×

RISCAL shows the highest directed enrichment among the compared methods relative to their respective null models. Furthermore, when evaluating undirected pairs, it recovers more curated relationships than coexpression: 26.3 versus 11.7 in CollecTRI and 16.2 versus 11.3 in SIGNOR, with the same edge budget. These results support the interpretive value of the learned graph compared with other methods.

Similarly, stronger edges allow candidates with greater support to be prioritized. From the weakest to the strongest quintile, enrichment increases from 2.511× to 5.989× in CollecTRI and from 1.824× to 5.522× in SIGNOR.

To examine the contribution beyond these resources, imagine a researcher with access to CollecTRI, SIGNOR, and the RISCAL models trained on the ten cell lines. By selecting connections present in all five fits across all ten cell lines and absent from both resources, the researcher obtains 54 candidates. The literature review finds:

Evidence foundCandidates
Compatible direction and sign of expression2
Other biological support, without full validation of direction and sign31
Published regulation with the opposite sign2
No sufficient specific experimental support found19

Thus, the researcher would recover relationships with published support that were absent from the resources used and would have additional candidates to investigate. Of the 19 without sufficient experimental support, one has a prior computational prediction, while the other 18 remain hypotheses for future validation; experimentally testing them exceeds the time and resources available during the rebuttal.

See Section 7.3, Table 2, and Appendix D.5, Table 28.

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

The paper studies gene perturbation response prediction, grounded by a differentiable method for interventional causal discovery. Specifically, their RISCAL method (Residual-activity Intervention Structural CAusal Learning) assumes soft (i.e., imperfect) single-node interventions and paired data, with a measurement model to differentiate between latent functional activity of each gene, and the measured transcript abundance. More specifically, the functional form of the soft interventions is motivated by Loss of Function perturbations, where an intervention on kk is instantiated as multiplication of the latent activity xk=μk+zkx_k = \mu_k + z_k by a latent residual-activity variable rkr_k; meanwhile, the expression readout x^k\hat{x}_k is an affine function of xkx_k. The training loss combines terms for reconstruction (via negative log-likelihood), sparsity, and acyclicity, where the learned parameters include an adjacency matrix AA, NN parameters for each causal mechanism, parameters for the distribution of each rkr_k, and parameters for the affine readout x^k\hat{x}_k.

The experiments include synthetic benchmarks and an application to the LINCS L1000 dataset, comparing their method against causal baselines (such as SDCD and InterSort) and perturbation prediction baselines (such as linear regression, MLP, a novel transformer, and GEARS). On the synthetic benchmarks, the paper evaluates causal discovery metrics such as AUPRC, precision, recall, and F1 score, finding that their method performs better on all metrics except recall. For the L1000 dataset, the paper evaluates perturbation-response metrics such as R2R^2 and intervention MSE. While two of the predictive baselines (GEARS and the axial transformer) outperform RISCAL, they find that RISCAL remains in a similar predictive regime and outperforms other structure learning methods. Finally, the authors study the stability and biological plausibility of their learned graphs. In terms of stability, the consensus cores over 5 runs contain, on average, 46.5% of the edges from individual runs, indicating decent reproducibility. Other analyses - e.g. gene ontology enrichment, functional coherence of the graph communities, and extrapolation to an unseen regulator - also support the biological plausibility of the learned graphs.

Critical strengths

The paper takes the scientific problem and the scientific evaluation very seriously. The domain-specific modeling is thoughtful and well-informed, both in terms of modeling the interventions and modeling the measurement process. For scientific evaluation, the proposed generator is well-aligned to the biological application, with a thorough incremental construction supporting a controlled assessment of how each biologically-motivated aspect (e.g., the measurement model and graph topology) affects recovery, and the PC3-scale regime uses a realistic gene number GG and number of samples. Moreover, the biological plausibility evaluation is fairly thorough, especially relative to other papers in the space. While some other papers report one or two results on stability, gene ontology enrichment, and/or extrapolation to an unseen perturbation, this paper reports all of these, as well as functional coherence of the graph communities; I found the framing in terms of "converging evidence" especially nice.

Critical weaknesses

The main weakness is the assumed setting of paired data and the related fact that the real-data analysis uses a bulk RNA sequencing dataset. Most interventional causal structure learning algorithms do not assume paired data, and such data is usually unavailable at the single-cell level: since the assay destroys the cell, it can only measure it once. To get around this, the authors use L1000, where each sample is an entire well (containing many cells), rather than a cell. However, this aggregation changes the data-generating structure: if x1,…,xNx_1, \ldots, x_N are sampled i.i.d. from a distribution that factorizes according to a DAG GG, the distribution of their average xˉ=1N∑n=1Nxn\bar{x} = \frac{1}{N} \sum_{n=1}^N x_n does not generally factorize according to GG. Since the paper is not theoretical, this issue is slightly less problematic, but it does introduce a notable mismatch between the method's implicit assumptions and the dataset on which it is applied.

Another limitation of the setting is the focus on known-target interventions, whereas gene perturbation experiments are thought to produce off-target effects, which has been addressed by other methods in causal structure learning (e.g., Mooij et al. (2020) and Squires et al. (2020)). It would be useful to emphasize this limitation more in the abstract, introduction, and conclusion.

References

Mooij et al. (2020), Joint causal inference from multiple contexts
Squires et al. (2020), Permutation-based causal structure learning with unknown intervention targets

Questions for authors
  1. Other related work: The distinction between the value of a variable (here, "remaining activity") and the measured value (here, "measure target transcript abundance") has a long history in causal discovery, see e.g., Silva et al. (2006) and Saeed et al. (2020). Could you include this context (and other important context, e.g. on off-target effects) in the Related Work?

  2. Fixed-point iteration: How does the fixed-point iteration (FPI) discussed in Section 4.4 enter into training? From what's written, I assume that you differentiate through the entire FPI procedure, which could be expensive. However, the reference to deep equilibrium models suggests that the gradient of that "layer" could instead be computed using the implicit function theorem. Can you provide more detail in the main body on how the gradients are handled?

  3. Method presentation: Currently, the method (architecture + objective) is introduced over the course of ~3 pages. I appreciate the need to explain each choice, but I would find it useful to have a figure or algorithm environment that summarizes the entire method (e.g. all trainable parameters, how they influence predictions, and how they enter into the objective). Could you provide (or describe) such a high-level summary figure/algorithm/etc. in the rebuttal?

References

Silva et al. (2006), Learning the structure of linear latent variable models
Saeed et al. (2020), Anchored causal inference in the presence of measurement error

1 comment

We sincerely thank the reviewer for the thoughtful feedback and constructive suggestions, which have made the manuscript clearer and easier to follow. Below we address each of the weaknesses and questions in detail; all changes are highlighted in blue in the updated manuscript.

Weaknesses: paired data and population-level interpretation

We agree that averaging does not generally preserve a single-cell causal DAG. Our LINCS samples are well-level profiles (L1000, a bead-based assay that profiles each well as a cell population, not RNA-seq), and our “pairs” match each treated well to a distinct same-plate control, i.e., a batch-matched baseline rather than a repeated measurement of the same cells; our synthetic Stage 12 shows recovery is retained with such reused, independent controls. For linear mechanisms, the average of i.i.d. cells follows the same graph, but under nonlinear mechanisms, as in RISCAL, this is not necessarily the case. We therefore interpret the learned LINCS graphs as population-level effective regulatory structure, and have clarified in Sec. 3, Sec. 7.3 and Appendix C.1 that our real-data results support prediction and biological plausibility, not identification of an intracellular causal DAG.

Weaknesses: known intervention targets

We agree. RISCAL assumes a known intervention target; residual activity models incomplete inhibition of that target, not unknown off-target effects. We have made this assumption explicit in the introduction (Section 1), problem formulation (Section 3) and conclusion (Section 8), and now discuss methods that relax known targets, including Mooij et al. (2020) and Squires et al. (2020). Given the abstract’s length limit, we kept this scope statement in the main text.

Q1. Related work

We agree. We have expanded Related Work to acknowledge latent-variable measurement models and measurement-error-aware causal discovery (Silva et al., 2006; Saeed et al., 2020; Dai et al., 2022), alongside methods for unknown intervention targets (Mooij et al., 2020; Squires et al., 2020; Hägele et al., 2023). Our contribution is not the general distinction between causal state and measurement, but its application to imperfect loss-of-function: only the intervened gene’s residual activity is separated from its measured transcript, within a model that jointly predicts perturbation responses and learns structure.

See Section 2.

Q2. Fixed-point iteration and training gradients

We thank the reviewer for raising this concern. We do not differentiate through the fixed-point iteration, neither by unrolling nor by implicit differentiation. The iteration (Q = 25 damped steps at PC3 scale and on real data) is used only for response prediction. Training follows interventional likelihood-based structure learning (Brouillard et al., 2020): the SCM loss evaluates each non-target mechanism on teacher-forced profiles (observed non-target expression, with the target replaced by its effective activity), and the target loss trains a separate readout that does not depend on the rollout.

In the full model, the training objective therefore has exactly zero gradient with respect to the iteration, even though the implementation records the unrolled operations in autograd (we verified this numerically). The tradeoff is that multistep propagation error is not optimized directly; in the no-readout ablation the target loss does reach the rollout.

We have clarified this in Section 4.4 and qualified the reference to deep equilibrium models.

Q3. High-level method summary

We thank the reviewer for this suggestion and agree that a compact parameter-to-objective summary improves readability. In the revised manuscript, Section 4.6 (Algorithm 1) now includes a concise training algorithm. It summarizes the procedure step by step and points to the equations that define each part: path strengths and the soft gate, the SCM and target-readout losses, the sparsity and acyclicity penalties, and DAG checkpoint selection. The fixed-point rollout is shown as a separate prediction step, since no training gradient flows through it.

The algorithm also references Table 4 in the appendix, which lists every trainable parameter (U, v, μ, a, η, κ, s, b, and the two log-variances) with its shape, its role, how it enters the predictions, and which loss terms it appears in. It also references Table 5, which lists all hyperparameters and their values.

Together with Figure 1, which gives the architectural overview, we believe this answers the reviewer’s request: what is learned, how each parameter influences the prediction, and how it enters the objective.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.