acceptodds
Under review as a conference paper at ICLR 2027

Mechanisms Matter: Transportability of Cellular Perturbation Effects

Abstract

Predicting cellular responses to perturbations across biological contexts is central to drug development, yet deep learning models often fail to outperform simple baselines. Leveraging causal transportability theory, we show that cross-context generalization is governed by shared causal mechanisms, not merely distributional similarity. To enable controlled evaluation, we develop a causal simulator that generates realistic semi-synthetic Perturb-seq datasets with tunable mechanistic divergence, providing benchmarks with known ground-truth causal structure. Further, we adapt the Vendi score to the perturbation setting to quantify effective response diversity and diagnose mode collapse, a failure mode invisible to standard per-perturbation metrics. Extensive experiments across six deep learning models and six simple baselines on semi-synthetic and real Perturb-seq datasets reveal a cross-context generalization gap: performance under cross-context splits drops substantially, often to simple baseline levels. Notably, even on synthetic data with fully specified causal structure, no model generalizes across contexts with different causal mechanisms. Across real and semi-synthetic datasets, lower cross-context concordance of transcriptome-wide impact tracks larger transfer gaps, providing an empirical proxy for perturbation-relevant mechanistic divergence.

Primary area:
datasets and benchmarks
Keywords:
Cellular perturbation effect, transportability, Perturb-seq, causal mechanisms, single-cell genomics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_giTjOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

Summary

  • This paper investigates why existing perturbation prediction models often fail to outperform simple baselines, focusing on cross-context generalization and limitations of evaluation metrics.
  • Causal Transportability: The authors theoretically characterize when perturbation effects can be transferred across biological contexts based on shared causal mechanisms.
  • Causal DGP: They develop an OU process-based semi-synthetic simulator to systematically evaluate generalization under changes in causal regulatory networks (ACA_C) and basal expression (BCB_C).
  • Vendi Score: They adapt the Vendi Score with null calibration to detect diversity collapse, complementing PDS in evaluating perturbation-response diversity and discrimination.
  • Experiments on synthetic and real Perturb-seq datasets demonstrate substantial cross-context generalization gaps and investigate their relationship with transcriptome-wide impact (TI) concordance.
Critical strengths

Strengths

  • Purpose-driven causal simulation: Rather than attempting to model every biological factor, the authors construct a controlled benchmark specifically designed to investigate causal mechanism shifts. Extending OU-based dynamics to cross-context evaluation is an interesting and meaningful contribution.
  • Theoretical analysis: The decomposition into AC,BCA_C, B_C, and Γq\Gamma_q provides a clear framework for understanding perturbation transportability. The derivation of perturbation-specific transportability conditions is particularly insightful.
  • Noise aware evaluation: The null calibration of the Vendi Score using empirical control cells is a thoughtful approach to accounting for finite-sample variability.
Critical weaknesses

Weaknesses

  • Simplified causal structure: Although the Causal DGP enables controlled evaluation, it does not fully capture biological complexities such as nonlinear regulation, context-dependent perturbation efficiency, and context-specific measurement effects.
  • Limited novelty of the Vendi Score adaptation: Limitations of existing evaluation metrics and mode collapse in cellular perturbation prediction have already been investigated (Viñas Torné et al., 2025; Wu et al., 2025; Mejia et al., 2025). Moreover, the Vendi Score has previously been applied to diversity evaluation in single-cell transcriptomics (Xu et al., 2026; DenAdel et al., 2026). While the proposed null-calibrated pseudobulk Vendi Score is a useful adaptation, its additional benefits over existing diversity and discrimination metrics could be more clearly demonstrated.
  • Causal mechanism identifiability: It remains unclear whether the models fail because they cannot learn causal mechanisms or because the available observations are insufficient to identify the target-context mechanisms. This distinction is important for interpreting the generalization failures (Pratapa et al., 2020).
  • Limitations of TI concordance: TI summarizes the magnitude of perturbation effects but does not preserve gene-specific directions or regulatory relationships. Therefore, high TI concordance does not necessarily imply similar causal mechanisms, raising questions about its reliability as a proxy for mechanistic divergence.
Questions for authors

Questions

  1. Identifiability: Could the authors distinguish failures caused by insufficient causal learning from those caused by fundamental identifiability limitations, potentially using an oracle or mechanism-aware baseline?
  2. TI Concordance: Since TI does not preserve gene-specific effect directions, how reliable is it as a proxy for mechanistic divergence? Could it be directly compared with ground-truth mechanism divergence in the synthetic data?
  3. Theory vs. Observation: Proposition 1 establishes transportability conditions for latent mean perturbation effects, whereas the experiments evaluate observed expression after nonlinear emission. How does this difference affect the connection between the theoretical guarantees and empirical findings?
4 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer for the insightful and detailed comments, and for recognizing our purpose-driven simulator, perturbation-specific transportability analysis, and null-calibrated Vendi score. Changes in the revised paper are in blue; we address weaknesses [W] and questions [Q] below.


[W1] Simplified causal structure

We agree. This formulation provides a tractable setting in which the effects of mechanism shifts can be precisely characterized. Context-dependent efficiency could be added by letting Γq\Gamma_q vary across contexts, measurement effects by varying the emission model, and nonlinear regulation by using a nonlinear drift. Each adds a context-specific mechanism, making transport harder, yet models already fail in the simplified setting (Figure 5). The real datasets likely include these complexities, and TI concordance tracks the gap in synthetic and real data (Table 9). We have added these extensions to Appendix B.2.1 (Lines 1104-1113); measurement-level shifts are also noted in Section 2.2 (Lines 160-191).


[W2] Novelty of the Vendi adaptation

We agree that metric limitations and mode collapse have been studied (Lines 45-72), and now cite prior single-cell uses of Vendi, which measure cell-level diversity without null calibration (DenAdel et al., 2026; Xu et al., 2026; Lines 262-264). Our pseudobulk Vendi instead treats each perturbation response as a unit with a null-calibrated RBF kernel; its predicted-to-observed ratio distinguishes collapse from spurious diversity. Unlike uncalibrated effective rank, which stays high under weak effects, the calibrated Vendi score tracks effect strength from the null (Figure 3a). Paired with PDS-ℓ1\ell_1, Vendi separates mode collapse from discrimination (Figures 3b and 11): across contexts, deep models lose discrimination but stay more diverse than simple baselines, a pattern neither metric reveals alone. At the dataset level, Vendi and split-half PDS-ℓ1\ell_1 indicate response diversity and reproducibility, a proxy for noise (Table 5).


[W3/Q1] Causal mechanism identifiability

We agree this distinction matters. In causalDGP, it can be largely separated; in real data, only partially. Matched splits differ only in target data for held-out perturbations (Lines 357-360), so the gap isolates the cost of withholding them; with Proposition 1, they serve as a mechanism-aware reference. Under baseline-only shifts, target effects equal source effects, so remaining gaps reflect insufficient learning (Figure 5). Under regulatory shifts, transport requires an identifiable transport map AC′−1ACA_{C'}^{-1}A_C, whose recovery remains challenging (Pratapa et al., 2020; Section 2.3), so part of the gap may reflect a fundamental identifiability limit. In real data, failures are consistent across models and datasets, and low TI concordance indicates where responses diverge (Table 9). Notably, Zhu25 has more control cells per context than Replogle22/Nadig25 (Table 1), so its failures (Figures 4 and 5) are unlikely to reflect limited sample size.


[W4/Q2] TI concordance

We agree that TI summarizes effect magnitudes, so high concordance does not guarantee similar mechanisms. Its reliability is nonetheless supported where ground truth is known. In causalDGP, where rewiring is known by construction, TI concordance declines monotonically with rewiring (0.91, 0.48, 0.25, 0.18 at 0, 25, 50, and 75%; Table 11) and tracks the transfer failures (Figure 5). In real data, it is consistent with known cell-line differences (Replogle22/Nadig25; Figure 14) and coincides with cross-donor failures (Zhu25; Figure 5). Although magnitude-preserving direction changes are possible in principle, TI concordance distinguished every known mechanism shift in causalDGP. We now present low TI concordance as an empirical proxy for mechanistic divergence, not a guarantee of similarity (Section 4.3.3, Lines 462-464).


[Q3] Theory vs. observation

Proposition 1 concerns latent transport, and nonlinear emission can alter measured effects across baselines even when latent effects are equal (Section 2.3, Lines 203-209). The link to our experiments is identifiability: since the latent-to-mean map is monotone and shared across contexts (Section 2.2, Lines 160-191), target effects remain identifiable in principle, though information is attenuated for weakly expressed genes. Baseline-only shifts can still produce small observed gaps unless models account for the emission. Empirically, the gap follows the predicted ordering: small under no-shift and baseline-only settings, larger under regulatory and combined shifts (Figure 5).

AuthorsOpenReview ↗

Rebuttal Response by Authors

Dear Reviewer,

Thank you for your thoughtful review. As the discussion period closes on Wednesday, we wanted to check whether you have had a chance to read our rebuttal.

We have made a careful effort to address each of your concerns, and we would be grateful if you would consider updating your score if you feel they have been resolved. We are happy to clarify any remaining points before then.

Thank you for the detailed responses and additional analyses. The explanations on TI concordance and the Vendi score address some of my concerns. However, I think the distinction between learning failures and identifiability limitations, as well as the gap between theory and observed outcomes, remains somewhat unclear. I will therefore maintain my original rating.

AuthorsOpenReview ↗

Rebuttal Response by Authors

Thank you for engaging and for noting that our responses on TI concordance and the Vendi score addressed some of your concerns. We clarify the two remaining points below.


[W3/Q1] Learning failures vs. identifiability limits

causalDGP separates the two because mechanism shifts are known by construction. The matched in-context split serves as a model-specific oracle: test cells are identical, but in-context training also includes target-context cells of the held-out perturbations (Lines 357-360). The transfer gap (in-context minus cross-context performance) thus isolates each model's cost of transfer. Proposition 1 determines identifiability:

  • BCB_C shifts: target effects equal source effects and are identifiable from source data, so any transfer gap is a learning failure; scLDM shows no transfer gap here, confirming the information is available (Figure 5).
  • ACA_C shifts: target effects are exactly recoverable given AC′−1ACA_{C'}^{-1}A_C, so the transfer gap reflects missing information about this map, not model capacity. Target-context controls and seen perturbations identify it only partially (Section 2.3; Pratapa et al., 2020), and no evaluated model closes this gap, which grows with rewiring (Figure 5; Table 11).

In real data, the type of mechanism shift is unknown, so learning failures and identifiability limits cannot be fully separated. Instead, low TI concordance flags mechanistic divergence, where identifiability limits are likely: Zhu25 has the lowest concordance and largest failures, Replogle22/Nadig25 moderate concordance and partial transport (Table 9; Figures 4 and 5). Zhu25 also has more control cells per context than Replogle22/Nadig25 (Table 1), so limited data is an unlikely cause.


[Q3] Theory vs. observation

The nonlinear emission changes observed effect magnitudes but does not alter the underlying latent-space identifiability conditions established by Proposition 1. Hence, the theoretical conditions and empirical findings remain consistent. The emission (softplus mean with negative binomial sampling) maps latent states to observed counts and is monotone and shared across contexts (Section 2.2, Lines 160-191), so latent effects remain identifiable from observed expression. Under BCB_C shifts, the emission can alter observed effect sizes, especially for weakly expressed genes, but accounting for it can close this gap: scLDM shows no transfer gap under no shift or BCB_C shifts (Figure 5). ACA_C shifts change the latent effects themselves and require AC′−1ACA_{C'}^{-1}A_C. Consistent with Proposition 1, transfer gaps are small under no shift and BCB_C, larger under ACA_C and combined shifts (Figure 5), and increase with rewiring (Table 11).


We hope this addresses the reviewer's remaining concerns, and we would be grateful if they would consider these points in their final assessment. We are happy to provide further clarification.

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

This submission takes a carefuly modelling approach to single cell perturbation prediction lending from ideas in causal transportability. It builds a corresponding synthetic data generating process and then carefully develops evaluation procedures including an instantiation of the Vendi diversity score to empirically assess implied transportability claims across methods and datasets.

Critical strengths

I will prefix this assessment by a disclaimer that I am not an expert in scRNA modeling or perturbation effect estimation in single cell transcriptomics.

  • I found the paper very well written and motivated. Limitations are always labeled clearly at the relevant locations. The modelling approach is well motivated. Claims are not overstated, the authors proceed carefully and honestly throughout most parts of the paper. The modeling approach is laid out cleanly and the authors do a good job in my view connecting theoretical implications for the chosen modeling approach to what one can/should expect in empirical evaluation. So the bridge between causal/modeling assumptions, the simulation process, and the final evaluation (with intermediate checks of whether all parts roughly make sense) is well built.

  • The evaluation is refreshingly thoughtful compared to what large parts of the sc perturbation literature does. I think the community has very well understood that we have perhaps been a bit too fast in developing more and more methods that claim superior performance, but with too little attention on the metrics and evaluation protocols that should measure this progress. However, still few act on this understanding. It's great to see meaningful simple baselines included and a paper focusing more on evaluation and careful modeling rather than yet the next perturbation prediction method.

Critical weaknesses
  • The theoretical contribution is quite narrow in its scope and novelty. This is perhaps the one claim that I would tone down in the abstract/intro. Transportability in this specific model under stationary linear processes with shared perturbations is perhaps from a theoretical perspective not jaw-dropping, but I still find it valuable to be stated cleanly in this specific context.
  • Absence of guaranteed transfer does not yet imply that there couldn't be some success for context-aware prediction. So here it would be interesting to probe a bit more when/whether the absence of transfer really implies some sort of empirical impossibility of predicting perturbation effects (or under which assumptions this may still be empirically possible).
  • The model is of course still quite limited in many (stated) ways and very likely still an overly simplistic description of the complex processes governing single cell transcriptomics under perturbations. In particular, the justification for the simulator relies on matching broad expressed statistics, which does not yet mean that the simulator is biologically plausible.
Questions for authors

See weaknesses

4 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer for the encouraging and insightful comments, and for recognizing the care taken in our modeling and evaluation. Changes in the revised paper are in blue; we address each weakness below.


[W1] Scope of the theoretical contribution

We agree. Proposition 1 follows directly from stationary linear dynamics with shared perturbations. We present it as a formalization that frames cross-context prediction as an identifiability problem and guides our analyses. We have toned down its framing in the abstract and introduction (Lines 13-15, 87-88, 105-107).


[W2] Non-identifiability and cross-context prediction

We agree: non-identifiability means the data cannot determine target effects in general, not that prediction is impossible. Unequal effects remain transportable whenever the transport map AC′−1ACA_{C'}^{-1}A_C can be recovered, e.g., from target perturbation data or structural assumptions (Section 2.3). Typical screens (P≪GP \ll G) recover it only partially, so transportability is perturbation-specific (Proposition 1), as in real data (Table 10). TI concordance is associated with transfer performance in the evaluated settings: models fail under extremely low concordance (Zhu25) but partially transport under moderate concordance (Replogle22/Nadig25), varying by source cell line (Table 9; Figures 4, 5 and 14).


[W3] Biological plausibility of the simulator

We agree that matching expression statistics does not establish biological plausibility. The causalDGP simulator is purpose-driven, designed for tractable theory (Proposition 1) and controlled validation, and statistically calibrated to real data beyond marginal statistics, including gene-pair correlations, effect sizes (TI\sqrt{\mathrm{TI}}) and perturbation geometry (Vendi, PDS) (Table 5). Even under these simplified dynamics, models fail to transport, and TI concordance tracks the known mechanism shifts and real-data gaps. The simulator readily extends to context-dependent efficiency, measurement effects and nonlinear regulation (Appendix B.2.1, Lines 1104-1113). We now describe causalDGP as statistically calibrated rather than realistic (Lines 16 and 90).

AuthorsOpenReview ↗

Rebuttal Response by Authors

Dear Reviewer,

Thank you for your thoughtful review. As the discussion period closes on Wednesday, we wanted to check whether you have had a chance to read our rebuttal.

We have made a careful effort to address each of your concerns, and we would be grateful if you would consider updating your score if you feel they have been resolved. We are happy to clarify any remaining points before then.

Thanks a lot for the responses and the revised version. The more contextualized claims now fit better. I remain carefully positive on the submission held back to increase my score even further primarily by the same reasons already mentioned and acknowledged by the authors as well.

I have selected 2 as my confidence, because I am not super active in this field, but would fairly confidently disagree with the clear rejection by reviewer JxVD as I see the value of the submission along a different dimension than the weaknesses mentioned there.

AuthorsOpenReview ↗

Rebuttal Response by Authors

We thank the reviewer for engaging, and for recognizing the value of connecting causal assumptions, simulation, and evaluation.

The linear theory and statistically calibrated simulator are deliberate scope choices that enable controlled analysis, and models already fail to transport in this simplified setting. We state these choices in the revised paper (Lines 13-15, 16, 87-90, 105-107; Appendix B.2.1, Lines 1104-1113) and see their extensions as directions for future work.

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

The paper studies when cellular perturbation effects can be transferred across biological contexts. It introduces CausalDGP, a causal simulator with controlled changes in regulatory interactions and basal activity. Also, it adapts the Vendi score to assess response diversity, and evaluates six deep learning models and five baselines on both simulated and real data. The study provides useful empirical evidence of transfer failures. However, the qualitative result is largely expected, and I have some concerns about the conceptual contribution, the robustness of the simulator findings, and the scope of the mechanistic conclusions.

Critical strengths

The paper addresses a relevant problem and provides a controlled framework with known causal mechanisms, separating changes in regulatory interactions from changes in basal activity.

The joint assessment of accuracy, diversity, and discrimination provides complementary information about prediction failures that individual metrics may miss.

Critical weaknesses

The drop in predictive performance after changes in relevant regulatory interactions is expected, and the direct transportability condition follows straightforwardly from the linear formulation. The main value of the paper seems to be showing how much current models lose accuracy when predicting perturbation responses in a different cellular context. However, these are already addressed in previous benchmarking studies.

The contrast between changes in (A_C) and (B_C) depends on the linear latent dynamics. The authors acknowledge this limitation, but do not establish whether the findings remain under non-linear regulation. Sensitivity to network topology, perturbation strength, and the degree of rewiring is also insufficiently explored, and thus further sensitivity studies are needed.

STATE-Geneformer provides a useful comparison of input representations, but its contribution to methodological diversity is limited because STATE is already included. An approach that explicitly represents gene regulatory programs, such as MORPH, would broaden the evaluation and would be relevant to the simulator’s focus on regulatory mechanisms.

Questions for authors

Would the contrast between changes in (A_C) and (B_C) remain under nonlinear regulatory dynamics? Given that 50% of connections per row are rewired, how do the findings change with smaller modifications or weaker perturbations?

How can we distinguish failures caused by insufficient information from failures to use the available information effectively? Do the experiments allow these explanations to be separated?

Do the findings suggest prioritizing models that explicitly represent gene regulation and its dependence on context, or collecting additional perturbation measurements in the target context?

3 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer for their feedback and for recognizing our controlled framework and joint metric assessment. Changes in the revised paper are in blue; we address weaknesses [W] and questions [Q] below.


[W1] Novelty relative to prior work

We agree the performance drop is expected and Proposition 1 follows directly from linearity. Prior cross-context benchmarks report unseen-context performance and examine metric distortion, misspecification and mode collapse (Lines 45-72), but do not link the generalization gap to mechanism shifts. We frame cross-context prediction as a transportability problem analyzed through identifiability: matched splits isolate the gap (Lines 357-360), and causalDGP, with known mechanisms, supports theoretical (Proposition 1) and empirical analysis. This attributes gaps to mechanism changes, separating avoidable from non-identifiable failures (Figure 5), and reveals (i) graded degradation, substantial already at 25% rewiring (Table 11); (ii) diverse rather than mean-collapsed deep-model predictions (Vendi; Figures 3b, 11); and (iii) TI concordance predicting transfer failures (Table 9). Notably, our framework distinguishes gaps needing target-context experiments from those needing better models.


[W3] Baseline model diversity

Thanks for suggesting MORPH, which we cite (Line 78). We evaluate transport-based (STATE, scLDM), graph-based (GEARS), VAE-based (CPA, scVI) and foundation-model (STATE-Geneformer) approaches (Section 4.1); STATE-Geneformer isolates pretraining effects. MORPH, also VAE-based, reuses its source-learned perturbation encoder and gene-program cross-attention and adapts its decoder to target controls, assuming the mechanism invariance we formalize (Proposition 1). Under regulatory shifts, transport requires an identifiable AC′−1ACA_{C'}^{-1}A_C, whose recovery remains challenging (Pratapa et al., 2020; Section 2.3). Consistently, MORPH's transfer error correlates with cross-cell-line divergence (r = 0.52; He et al., 2025, Fig. 3b), and all evaluated models show this limit (Figure 5). As with GEARS, representing gene programs does not guarantee transportability unless it is context-specific and identifiable. We now cite MORPH's results (Lines 469-471).


[W2/Q1] Sensitivity and nonlinearity of causalDGP

Linear dynamics make it tractable to separate regulatory (ACA_C) from baseline (BCB_C) changes, and models already fail here, so nonlinearity is not needed for failure. By Proposition 1, the cross-context difference AC′−1(AC−AC′)τC(q)A_{C'}^{-1}(A_C - A_{C'})\boldsymbol{\tau}_C^{(q)} vanishes without rewiring and is bounded by the network change for any stable topology. Consistently, all models degrade from 0% to 50% rewiring (Figure 5), and an added sweep shows scLDM's Pearson δ\boldsymbol{\delta} falling from 0.78 (0%) to 0.54, 0.38, and 0.31 at 25, 50, and 75% rewiring, with PDS and TI concordance declining in step (Table 11). Scaling Γq\Gamma_q scales effects and their difference proportionally, so weaker perturbations lower the signal-to-noise ratio, not the relative gap. The ACA_C/BCB_C contrast is specific to the linear case (Section 2), and nonlinear regulation may let baseline shifts alter responses; the observed failures do not depend on it, as TI concordance tracks the gap in synthetic and real data (Table 9). Analyses and extensions are added to Appendix B.2.1 (Lines 1104-1113).


[Q2] Information versus learning failures

In causalDGP, largely; in real data, only partially. Matched splits differ only in target data for held-out perturbations (Lines 357-360), so the gap isolates the cost of withholding them, and Proposition 1 indicates when direct reuse of latent effects is valid. Under baseline-only shifts, latent target effects equal source effects, so remaining gaps reflect a failure to use available information (Figure 5). Under regulatory shifts, transport requires a context-specific adjustment the data may not identify (Section 2.3), so part of the gap may reflect insufficient information. In real data, the two cannot be fully separated, but low TI concordance indicates where responses diverge (Table 9), and Zhu25 has more control cells per context than Replogle22/Nadig25 (Table 1), so its failures (Figures 4 and 5) are unlikely to reflect limited sample size.


[Q3] Better models or more target data

This depends on identifiability (Proposition 1). Under baseline-only shifts, target effects are identifiable, so better models should close the gap; scLDM transports best (Figure 5). Under regulatory shifts, effects are not identifiable in general and context-specific GRN recovery remains challenging (Pratapa et al., 2020), so targeted target-context measurements are the more reliable near-term route. Representing gene regulation helps only if it is context-specific and identifiable. Context-aware models could reduce the measurements needed, and low TI concordance can show where (Lines 487-490).

AuthorsOpenReview ↗

Rebuttal Response by Authors

Dear Reviewer,

Thank you for your thoughtful review. As the discussion period closes on Wednesday, we wanted to check whether you have had a chance to read our rebuttal.

We have made a careful effort to address each of your concerns, and we would be grateful if you would consider updating your score if you feel they have been resolved. We are happy to clarify any remaining points before then.

AuthorsOpenReview ↗

Rebuttal Response by Authors

Dear Reviewer JxVD,

As the discussion closes today, we briefly highlight the main changes made in response to your review: a rewiring sensitivity analysis (Table 11), a discussion of MORPH (Lines 469-471), and a clarification of when gaps require target-context experiments versus better models (Lines 487-490). We would be happy to clarify any remaining concerns before the deadline. Thank you again for your time.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overall Rebuttal by Authors

We thank all reviewers for their constructive feedback and are encouraged that they found the paper well written and carefully argued, with a thoughtful evaluation (zVTP), the controlled framework and joint metric assessment valuable (JxVD, giTj), and the perturbation-specific transportability analysis and null-calibrated Vendi score insightful (giTj).

Our central message is unchanged: cross-context perturbation prediction is a transportability problem. Changes in perturbation-relevant mechanisms can invalidate direct transfer of latent effects, and target effects can be exactly recovered only if the transport map is identifiable.

The revised paper (changes in blue) (i) states the scope of the theory more precisely and notes its implications in the conclusion; (ii) describes the simulator as statistically calibrated, adds an analytical and empirical sensitivity analysis, and expands the discussion of its extensions; (iii) clarifies the interpretation of our metrics; and (iv) adds citations to related work. Below are our point-by-point responses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.