acceptodds
Under review as a conference paper at ICLR 2027

Hypothesis-Evidence Feedback Reasoning for Cellular Perturbation Prediction

Abstract

Predicting cellular responses to unseen perturbations requires generalizing from sparse experimental knowledge across interventions and cellular contexts. Which experimental evidence is most useful can depend not only on the perturbation query but also on the current response prediction, motivating feedback between prediction and evidence selection. We therefore introduce Hypothesis-Evidence Feedback Reasoning (**HEFR**), a unified *Propose–Reason–Realize* framework that turns response prediction into an adaptive process of hypothesis initialization, hypothesis-evidence interaction, and cell-population realization. Starting from an initial latent hypothesis of the perturbation response, HEFR iteratively uses the current hypothesis to retrieve evidence that grounds its revision into the next hypothesis. This loop terminates once further reasoning is no longer expected to yield substantive refinement. The final hypothesis is decoded into an aggregate response and realized as a heterogeneous population of perturbed cells consistent with that response. Across chemical and genetic perturbation benchmarks, HEFR substantially outperforms existing methods in key metrics of aggregate accuracy and distributional fidelity. Theoretical and biological interpretability analyses further demonstrate the validity of HEFR. Ultimately, HEFR shifts virtual-cell modeling from passive pattern mapping toward adaptive, feedback-driven scientific reasoning. Code will be released publicly upon publication.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Cellular Perturbation Prediction, Hypothesis-Evidence Feedback, Latent Reasoning, Biological Analysis

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_iPcPOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

A notable strength of this paper is that it introduces a reasoning-like mechanism for single-cell perturbation prediction, where information retrieved from the training data is used to assist model updates during inference. I find this formulation interesting and potentially useful. However, I have some concerns regarding the model design. In addition, the general idea of querying or retrieving information from the training data has been explored in prior work. The current manuscript may therefore benefit from a more comprehensive discussion of the relevant literature, together with appropriate comparisons and citations to clarify the methodological distinction and novelty of the proposed approach.

Critical strengths

This paper proposes a reasoning-like mechanism for single-cell perturbation prediction, in which information retrieved from the training data is explicitly used to guide model updates during inference. I find this design particularly interesting, as explicit reasoning mechanisms of this kind remain relatively underexplored in the current single-cell perturbation prediction literature. This direction could be valuable for improving the flexibility and interpretability of perturbation prediction models.

Critical weaknesses

Although the implementation of the proposed reasoning-like mechanism is interesting, the related-work discussion appears incomplete and lacks important citations. Prior single-cell perturbation work [1] also leverages information from observed training data to guide perturbation-specific prediction or generation. While HEFR introduces a different mechanism based on iterative, hypothesis-dependent evidence retrieval and adaptive refinement, the conceptual connection to prior training-data-guided inference should be made explicit. The authors should therefore discuss [1] and clarify the methodological distinction between the two approaches, particularly which aspects of HEFR are genuinely new beyond this existing line of work.

[1] Doloris: Dual Conditional Diffusion Implicit Bridges with Sparsity Masking Strategy for Unpaired Single-Cell Perturbation Estimation

Questions for authors
  1. [1] also leverages information from observed training data to guide perturbation-specific prediction or generation. While HEFR introduces iterative, hypothesis-dependent evidence retrieval and adaptive refinement, the conceptual relationship to this prior line of work is currently unclear. Could the authors discuss and clarify the methodological distinction between the two approaches?

  2. Why is cosine improvement used as the sole objective for the value-guided critic? The critic accepts or rejects a revision based on the predicted improvement in cosine similarity between the decoded aggregate response and the ground truth. While cosine similarity captures response direction, it is insensitive to response magnitude, which is also evaluated through RMSE and directly affects the target mean used by the REALIZE stage. Could the authors justify this design choice and evaluate alternative critic objectives, such as RMSE improvement or a combination of directional and magnitude-based criteria?

  3. Since ϵ>0\epsilon>0, any gene with μg>0\mu_g>0 appears to have strictly positive expression in every predicted cell after projection. How can the model represent the large number of exact zeros typically present in single-cell data?

  4. Eq.~(13) defines REALIZE without an explicit stochastic latent variable or noise input. Could the authors clarify whether REALIZE is deterministic conditional on the input control cell, and, if so, how perturbation-induced variability beyond the heterogeneity already present in the control population is represented?

Overall, I would be open to increasing my score if the rebuttal can satisfactorily address the concerns above and clarify the key methodological choices and novelty positioning.

[1] Doloris: Dual Conditional Diffusion Implicit Bridges with Sparsity Masking Strategy for Unpaired Single-Cell Perturbation Estimation

5 comments
AuthorsOpenReview ↗

Part 1/3 of the responses

We sincerely thank you for your time and thoughtful comments, which have helped us clarify and strengthen the manuscript. We respond point by point below.

W1&Q1: Incomplete related work and conceptual distinctions.

Response: We have added Doloris [1] to Related Work and clarified their distinctions (Line 111 in the revised paper). Both methods utilize observed training information, but differ fundamentally at both conceptual and mechanistic levels.

  • Conceptually, the methods use different training information for different purposes. Doloris’s reference mechanism uses observed training-cell activation patterns as templates for coherent sparsity in generated cells. HEFR uses observed perturbation-response records as evidence for correcting the aggregate-response hypothesis. The former supports cell-level sparsity generation, while the latter supports perturbation-response reasoning.
  • Mechanistically, one key distinction lies in how the two methods select and incorporate observed training information. Doloris selects reference subsets by matching their empirical gene-activation probabilities to model-predicted probabilities using L2 distance. It then samples a reference cell’s binary activation mask and adjusts it according to the predicted probabilities. HEFR instead selects response records (observed training data) by their estimated revision utility, conditioned on the query, current hypothesis, and candidate evidence. The retrieved evidence informs a latent hypothesis update, and the updated hypothesis guides the next retrieval. HEFR therefore introduces a distinct feedback-reasoning mechanism that uses observed perturbation responses to revise a prediction hypothesis, with hypothesis-dependent evidence selection and gain-guided stopping.

Overall, Doloris centers on modeling unpaired control and perturbed-cell distributions through dual conditional diffusion bridges, with sparsity masking to preserve expression diversity, whereas HEFR centers on hypothesis-evidence feedback reasoning to refine aggregate perturbation-response hypotheses and adapt reasoning depth.

[1] Doloris: Dual Conditional Diffusion Implicit Bridges with Sparsity Masking Strategy for Unpaired Single-Cell Perturbation Estimation.

Q2: Justification for the cosine-only critic objective and comparison with magnitude-sensitive alternatives.

Response: We chose cosine gain to prioritize response-direction refinement among candidates produced by magnitude-supervised prediction modules. The regression objective for f0f_0 and latent residual losses for Gθ/HνG_\theta/H_\nu provide magnitude-sensitive supervision. Cosine is therefore the critic’s selection criterion, rather than the model’s sole training signal.
Following your suggestion, we directly compare cosine gain, RMSE gain, and their normalized combination on sciPlex3. RMSE gains are computed from the target mean-expression profiles passed to REALIZE. From the table below we can see that Cosine gain yields the highest Cos-FC. The RMSE gain slightly improves mean-expression accuracy but lowers directional accuracy. The combined objective also yields a lower Cos-FC, with small improvements in RMSE and MMD-G. These results support our direction-focused critic design: it provides the strongest directional prediction in this comparison, with only small differences in mean-expression error and population discrepancy. We have included the results in Appendix D.5 in the revised paper (Table 16).

Critic objectiveCos-FC ↑RMSE ↓MMD-G ↓
Cosine gain0.69860.02310.1046
RMSE gain0.66140.02280.1052
Normalized combination0.67380.02300.1044
AuthorsOpenReview ↗

Part 2/3 of the responses

Q3: Exact-zero representation under the smoothed mean projection.

Response: Thank you for your insightful comment. For genes with a positive target mean, the projection does not explicitly represent exact zeros. Eq. (15) is designed to enforce the predicted cell population to match the reasoning-predicted response means through a smooth, differentiable (important for our distributional optimization) normalization. Adding ϵ\epsilon to both the preliminary expression values and their mean keeps the normalization well-defined while preserving the target mean exactly. Importantly, the projection preserves key aspects of continuous cell-to-cell heterogeneity, and exact zeros can be introduced through mean-preserving post-processing.
What the projection preserves: Let mgm_g and sg>0s_g>0 be the preliminary expression mean and standard deviation of gene gg. For μg>0\mu_g>0, Eq. (15) gives x^i,g−μg=ag(x~i,g−mg)\hat{x}_{i,g}-\mu_g=a_g(\tilde{x}_{i,g}-m_g), where ag=μgmg+ϵ>0a_g=\frac{\mu_g}{m_g+\epsilon} > 0. Since SD(x^g)=agsg\mathrm{SD}(\hat{x}_g)=a_g s_g, we have x^i,g−μgSD(x^g)=x~i,g−mgsg\frac{\hat{x}_{i,g}-\mu_g}{\mathrm{SD}(\hat{x}_g)}=\frac{\tilde{x}_{i,g}-m_g}{s_g}. Thus, projection preserves key properties such as standardized expression patterns, within-gene cell rankings, and Pearson correlations between nonconstant positive-mean genes. These properties establish preservation of relative expression organization. Moreover, REALIZE’s distributional losses act directly on the projected populations, so training optimizes population fidelity under the mean constraint.

Adding exact zeros by mean-preserving post-processing: We threshold low expression values as qi,g=x^i,g1x^i,g>τgq_{i,g}=\hat{x}_{i,g}\mathbf{1}\\{\hat{x}_{i,g} > \tau_g\\}, compute qˉg=1N∑iqi,g\bar{q}_g=\frac{1}{N}\sum_i q_{i,g}, and rescale the retained values as xi,g′=μgqi,gqˉgx'_{i,g}=\mu_g\frac{q_{i,g}}{\bar{q}_g}. Thresholds τg\tau_g are determined exclusively from training/calibration data. If all entries are removed for a positive-mean gene, its largest entry is retained before rescaling. Zero-mean genes remain entirely zero. This post-processing preserves gene-wise means exactly and requires no retraining.
We further evaluate distributional performance before and after post-processing across the settings in Tables 1–3. All relative changes are below 0.5%, preserving the reported advantages, as shown below. When mg≫ϵm_g \gg \epsilon, preliminary zeros are mapped to a small positive floor, μgϵ/(mg+ϵ)\mu_g\epsilon/(m_g+\epsilon), helping explain why removing these low-level values and rescaling to restore the means has little effect on distributional performance.

Distributional metrics changes with mean-preserving post-processing:

SettingMMD-G: before → after ↓MMD-P: before → after ↓DEG-R: before → after ↑
Unseen perturbations — drug0.1046 → 0.10510.1095 → 0.10910.1019 → 0.1023
Unseen perturbations — gene0.3347 → 0.33380.3981 → 0.39930.1402 → 0.1397
Unseen combinations — drug1.1095 → 1.11282.2406 → 2.23390.1128 → 0.1132
Unseen combinations — gene0.3712 → 0.37050.2438 → 0.24460.2941 → 0.2934
Unseen cellular contexts0.4216 → 0.42290.5008 → 0.49940.1861 → 0.1866

We also compare zero-frequency agreement against observed treated populations on the sciPlex3 unseen-perturbation benchmark, results are shown below. The output achieves competitive agreement in both per-gene zero frequencies and per-cell detection fractions. Thus, the post-processing addresses exact-zero representation while retaining the predicted means and distributional performance. We have clarified the original smooth output’s scope and documented the post-processing in Appendix D.6 of the revised paper.

Zero-frequency agreement:

MethodGene-wise zero-fraction MAE ↓Detected-gene-fraction Wasserstein distance ↓
CellFlow0.0560.075
State0.0440.030
HEFR + sparsification0.0410.032
AuthorsOpenReview ↗

Part 3/3 of the responses

Q4: Determinism of REALIZE and its representation of perturbation-induced variability.

Response: REALIZE is deterministic given the input control cell, perturbation descriptor, and terminal hypothesis. However, the deterministic transformation in REALIZE can change population variability rather than simply preserve the variability of the controls. For our population-distribution prediction task, control cells provide diverse initial states, and REALIZE learns how each state responds to the perturbation. Because this transformation is nonlinear and cell-state-dependent, different cells can undergo changes of different magnitudes and directions. It can therefore amplify, suppress, or reshape variability, altering population variance, gene correlations, and state composition without an explicit noise input.
These distributional changes are supervised directly using treated populations. The objective includes gene- and PCA-space distributional discrepancies and a PCA-space variance-matching term, all evaluated after mean projection (Appendix A.3 of the initial manuscript). The mean-matched control supports this design: rescaling control cells to the same terminal predicted means yields MMD-G 0.2903, compared with 0.1046 for REALIZE (Table 4 of the initial manuscript), demonstrating improved population fidelity beyond mean adjustment alone. This represents perturbation-induced changes in the population distribution through a learned transformation of control-cell states.

We hope our responses have addressed your concerns, and we welcome further discussion if any clarification is needed, thank you!

Reviewer_WsEgOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces HEFR for predicting single-cell responses to chemical and genetic perturbations. A kernel-based predictor initializes a latent response hypothesis, which is iteratively revised using dynamically retrieved experimental evidence and a learned stopping critic. A conditional population model then generates cell profiles constrained to match the predicted mean response. The framework is evaluated on unseen perturbations, combinations, and cellular contexts using aggregate and distributional metrics. Component ablations, policy comparisons, and evidence-replacement experiments assess the proposed mechanisms. Results indicate competitive predictive performance, while theoretical analyses characterize potential benefits of adaptive retrieval and stopping under stated assumptions.

Critical strengths
  1. HEFR integrates response initialization, adaptive evidence-guided revision, and population generation in a clear architecture. It reports competitive results across chemical and genetic generalization tasks, including leading all seven reported metrics on unseen genetic combinations.
  2. Fixed-module policy comparisons, mean-matched realization ablations, and matched evidence replacements provide substantive support for the contributions of feedback and population modeling. The critic diagnostics also expose imperfect stopping decisions rather than reporting only favorable aggregate scores.
Critical weaknesses
  1. Critic training requires validation responses, but Table 9 lists no validation partition for unseen contexts. Appendix C.3 also discloses degenerate baseline encodings and unequal holdout difficulty, while the stated single-split ablations use a Full HEFR score matching the pooled result. These issues weaken confidence in the reported comparisons.
  2. Controlled mechanism analyses concentrate on chemical perturbations, and the theoretical results concern idealized policies under conditions not guaranteed by HEFR. The biological examples support useful prediction refinement, but the link between retrieved evidence and independently established biological mechanisms remains insufficiently developed.
Questions for authors
  1. Appendix D.1 describes a single sciPlex3 ablation split, but Table 4 reports a Full HEFR Cos-FC of 0.6986. Table 15 assigns that value to the pooled result, with different scores for every individual split. How was the ablation evaluation constructed to produce the reported Full HEFR score of 0.6986?
  2. Appendix C.3 reports that several comparators use degenerate encodings for held-out entities and that GEARS faces a stricter holdout in genetic context transfer. These differences complicate attribution of the performance gap to HEFR itself. How much of the reported advantage remains under comparisons with matched training access and valid representations of the held-out entities?
  3. Tables 4 and 12 provide useful mean-matched realization controls on sciPlex3. However, the main distributional comparisons across methods remain sensitive to differences in mean-response accuracy. Does HEFR retain its distributional advantage when comparator populations are adjusted to have the same gene-wise means as HEFR?
  4. Figure 3 includes a CYC116-MCF7 example with improved cell-state composition. The presentation does not trace how the selected experimental evidence explains the specific biological response. What interpretation grounded in independent biological knowledge connects the retrieved evidence to the predicted state shift in this example?
  5. Appendix A.3 uses validation responses to train the critic and select its stopping threshold, but Table 9 lists no validation partition for unseen cellular contexts. The source of calibration supervision is therefore unclear. How was the calibration partition constructed in these experiments while excluding treated responses from the held-out cell line?
6 comments
AuthorsOpenReview ↗

Part 1/4 of the responses

We sincerely thank you for your time and thoughtful comments, which have helped us clarify and strengthen the manuscript. We respond point by point below.

W1.1&Q5: Critic training for unseen contexts.

Response: Appendix A.3 described the default procedure with released validation data, omitting this special case. For unseen contexts, critic calibration uses an internal split drawn exclusively from source-context training data. In each outer fold, four of 18 drug–context groups (22.2%) are reserved for calibration, keeping every drug represented in support. Upstream modules and memory use support data only and are frozen for critic training. We then refit the upstream modules and rebuild memory on all source training data, keeping the critic fixed. Held-out-line treated responses are excluded from training, calibration, memory construction and model selection. This procedure is now documented in the revised paper (Line 744).

W1.2&Q2: Comparison fairness under encoding and holdout limitations.

Response: Our evaluation aimed to include as many methods as possible under the standard splits, and numerous comparisons with strictly matched protocols already support HEFR’s advantages. Following your suggestion, we additionally align query coverage, training access, and representation validity.

  • Existing strictly matched comparisons: The methods below already follow matched protocols under the official benchmark settings.
SettingComparatorsMetrics HEFR (ours) ranks first
Unseen drug perturbationsCellFlow, PerturbNet, State, cycleCDR, chemCPA, identity floor6/7
Unseen gene perturbationsscGPT, PerturbNet, State, PT-RAG, CellFlow, scDFM, identity floor5/7
Unseen drug combinationsCellFlow, chemCPA, identity floor,6/7
Unseen gene combinationsLatentAdditive, CPA, SAMS-VAE, biolord, scGPT, scDFM, CellFlow, PerturbNet, GenePert, Decoder-only, identity floor7/7
Unseen cellular contextsLatentAdditive, State, CellFlow, scGen, CellOT, cycleCDR, identity floor5/7
Unseen drug–context pairsLatentAdditive, CPA, chemCPA, SAMS-VAE, biolord, State, scDFM, CellFlow, scGen, PerturbNet, cycleCDR, CellOT, Decoder-only, identity floor5/7
Unseen gene–context pairsLatentAdditive, CPA, SAMS-VAE, biolord, State, PT-RAG, scDFM, CellFlow, scGen, CellOT, GenePert, Decoder-only, identity floor5/7

HEFR ranks first in Cos-FC, MMD-G and MMD-P in every setting and leads most metrics.

  • Aligned comparisons: We retain the official splits and evaluate HEFR on each comparator’s eligible queries, determined by target availability and representation validity. For genetic context transfer, we retrain HEFR under GEARS’s stricter holdout protocol and evaluate both methods on the same eligible test queries. HEFR retains its advantage after alignment, as shown below:

Entries report comparator / HEFR (ours).

SettingComparatorShared queriesCos-FC ↑R-Cos ↓RMSE ↓R-RMSE ↓MMD-G ↓MMD-P ↓DEG-R ↑
Unseen genesGEARS9710.1488 / 0.31960.4056 / 0.19020.0633 / 0.05210.3950 / 0.30210.4407 / 0.33420.4788 / 0.39760.0929 / 0.1401
Unseen genesGenePert9710.3084 / 0.31930.2405 / 0.19080.0547 / 0.05220.2457 / 0.3030———
Unseen genetic combinationsGEARS430.3974 / 0.81360.1877 / 0.00220.0725 / 0.03870.1229 / 0.01110.6386 / 0.36490.8748 / 0.24320.0381 / 0.2906
Unseen gene–context pairsscGPT830.0731 / 0.19510.2086 / 0.20200.0338 / 0.02560.2440 / 0.19715.4142 / 0.26782.5811 / 0.14830.0000 / 0.0831
Unseen gene–context pairs, matched training accessGEARS640.0282 / 0.18760.4673 / 0.21960.0369 / 0.02840.4891 / 0.20380.3842 / 0.27840.2728 / 0.13360.0059 / 0.0638
  • Representation-limited configurations: Partial representation gaps are handled through shared eligible queries. Where no fully represented subset exists, we mark the limitation, for example: catalogue-based LatentAdditive/SAMS-VAE on unseen drug combinations, biolord’s missing-entry fallbacks on unseen drug combinations and cellular contexts, and chemCPA’s untrained context parameters for a held-out line. However, HEFR also outperforms these methods where their representations are supported. On genetic combinations, with both singletons in training, its Cos-FC is 0.8153 versus 0.6461/0.4846 for LatentAdditive/SAMS-VAE. On drug–context pairs, where both identities occur in training, HEFR exceeds biolord (0.5149 vs. 0.3443). On unseen drugs, it exceeds chemCPA (0.6986 vs. 0.6565).
    We have marked methods requiring adaptations in Tables 1–3 and 19–20, and added the aligned comparisons to Appendix E (Table 23) of the revised paper.
AuthorsOpenReview ↗

Part 2/4 of the responses

W1.3&Q1: Ambiguity between the stated ablation split and the reported pooled score.

Response: Thank you for pointing out this ambiguity. At the beginning of Section 5.3 in the initial manuscript (lines 407–408), we stated: “Using the unseen-perturbation setting on sciPlex3, we assess the contributions of PROPOSE, REASON, and REALIZE”. That means the main ablation follows this benchmark setting, including the pooled reporting convention specified in Section 5.1. Specifically, Full HEFR and all ablation variants use the same sciPlex3 unseen-perturbation splits (seeds 1–3) and the same pooling protocol. Accordingly, the Full HEFR score of 0.6986 in Table 4 is the pooled result also reported in Table 15 of the initial manuscript.
We apologize for the singular wording “the main sciPlex3 ablation split” in Appendix D.1. It was imprecise and could suggest a single-split evaluation. We have revised it to “the main sciPlex3 ablation splits (three seeds)” to avoid ambiguity, and provided per-seed results for all ablation variants below.

The per-seed results:

Cos-FC ↑

VariantSeed 1Seed 2Seed 3Pooled
Full HEFR0.61410.69980.71840.6986
Mean0.57040.66020.67780.6581
T = 10.52870.62540.64140.6218
Fixed-state0.56790.65980.67930.6583
Interpol.0.57120.66130.68100.6601
γ = 00.58630.67410.69050.6708
Control0.61410.69980.71840.6986

MMD-G ↓

VariantSeed 1Seed 2Seed 3Pooled
Full HEFR0.12830.10580.09780.1046
Mean0.19240.15860.15170.1591
T = 10.24760.20230.19580.2042
Fixed-state0.26980.22540.21970.2276
Interpol.0.21940.18060.17560.1825
γ = 00.20470.16580.16060.1676
Control0.33760.28910.28030.2903

W2.1: Mechanism analyses limited to chemical perturbations.

Response: HEFR’s feedback mechanism is shared across chemical and genetic perturbations. We used sciPlex3 as a representative testbed for the original controlled analyses and now extend the component analysis to unseen gene perturbations, using the same ablation definitions as Table 4. All variants follow the genetic benchmark setting in Table 1 of the initial manuscript. The results below support the component contributions beyond chemical perturbations. Reasoning improves Cos-FC from 0.2741 to 0.3192, while fixing retrieval to the initial hypothesis reduces it to 0.2918. Learned correction outperforms interpolation, and bounded innovation provides a further gain. At identical terminal means, REALIZE reduces MMD-G from 0.4861 to 0.3347 relative to rescaled control cells, demonstrating a contribution to population fidelity beyond aggregate accuracy. We have included this genetic ablation in Appendix D.4 of the revised paper (Table 15).

Mechanism analyses under unseen gene perturbation setting:

MetricFull HEFRMeanT = 1Fixed-stateInterpol.γ = 0Control
Cos-FC ↑0.31920.28640.27410.29180.29530.30670.3192
MMD-G ↓0.33470.39420.43860.42150.40780.37890.4861
AuthorsOpenReview ↗

Part 3/4 of the responses

W2.2: Idealized theoretical assumptions not guaranteed by HEFR.

Response: The propositions identify when adapting evidence and reasoning depth offers an advantage. HEFR explicitly trains hypothesis-dependent retrieval to estimate evidence utility and its stopping critic to predict revision gains, aligning the learned policies with the underlying theoretical decision principles. The theoretical conditions define the scope of the guarantees, while Appendix B directly tests whether these learned mechanisms deliver the corresponding benefits.

  • Evidence adaptation: Proposition 1 establishes a strict advantage when different hypotheses benefit from complementary evidence, so no single evidence is optimal throughout. This is a condition on evidence usefulness, rather than an architectural constraint. HEFR exploits it by reassessing evidence after each hypothesis update. At identical states and with the revision network fixed, current-state selection yields a mean cosine gain of 0.0043 versus 0.0011 for retaining the initial record (Table 6 in the initial manuscript). Retrieval scores correlate with realized gains (within-state Spearman ρ=0.43\rho=0.43, Table 6). These results connect the theoretical value of evidence adaptation to actual prediction gains delivered by HEFR’s learned retriever.
  • Adaptive stopping: Although the sufficient persistence condition in Proposition 2 is not enforced architecturally, HEFR’s stopping policy is explicitly trained to follow the corresponding gain-based decision principle. Gain-supervised critic training and calibrated threshold selection encourage accepting beneficial revisions and stopping when the expected gain is insufficient. We evaluate the practical consequence of stopping by continuing trajectories offline: mean missed utility is 0.0012 among early-stopped queries (Table 7 in the initial manuscript). With upstream modules frozen, adaptive stopping also improves Cos-FC from 0.6867 at the validation-selected fixed depth to 0.6986 (Table 8 in the initial manuscript). Together, these results support the practical relevance of Proposition 2’s gain-based stopping principle, showing that HEFR’s learned critic captures the benefit of query-adaptive depth over a validation-selected fixed policy.

W2.3&Q4: Biological grounding linking retrieved evidence to predicted cell-state shifts.

Response: The original presentation emphasized prediction refinement in the CYC116–MCF7 example. Here, we make the biological interpretation explicit by connecting the retrieved AMG-900–MCF7 response to the shared Aurora-inhibition mechanism in this example. CYC116 inhibits Aurora A/B and induces mitotic failure and polyploidy [1], while AMG-900 inhibits Aurora kinases and causes failed cell division [2]. This shared pharmacology provides an independent biological rationale for the retrieved evidence’s relevance to the cell-cycle response.
The initial hypothesis overestimates cycling-associated transcription relative to the observed CYC116 response. The measured AMG-900 response guides revision toward the observed program activity. Consistent with this correction, the terminal population also corrects the overestimated cycling-state fraction, reducing composition total-variation distance from 0.102 to 0.055 (Fig. 3c). Thus, independent pharmacology explains the evidence’s relevance, while the expression data establish the direction of correction. We have clarified this connection between retrieved evidence, program correction, and improved cell-state composition in the revised paper (Line 479).

[1] Discovery of N-phenyl-4-(thiazol-5-yl)pyrimidin-2-amine aurora kinase inhibitors, Journal of Medicinal Chemistry, 2010.
[2] Preclinical evaluation of AMG 900, a novel potent and highly selective pan-aurora kinase inhibitor with activity in taxane-resistant tumor cell lines, Cancer Research, 2010.

Reviewer_eNvuOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes HEFR an approach for cellular perturbation prediction. The method iteratively retrieves experimental evidence, updates a latent response hypothesis and decides when to stop. A separate module generates the perturbed cell population.

The model is evaluated on genetic and chemical perturbations: unseen perturbations, combinations, and cellular contexts.

Critical strengths

I like the idea of using the current prediction to guide evidence retrieval. This idea is a natural way to approach perturbation prediction.

Strong experimental results across several benchmarks. The ablations are useful and show that feedback-based retrieval improves over fixed retrieval.

Critical weaknesses

Overall the system is quite complex. It has lots of components: kernel regression, PCA, retrieval, multiple neural networks, a critic, and population generation. It is not fully clear which parts drive the improvement.

The theoretical results are fairly general and rely on assumptions that are not enforced by the method.

Also, adaptive stopping saves only 3% in actual inference time over fixed depth 4. This is a rather small benefit for the additional complexity.

Questions for authors

How does performance change as the available evidence memory becomes smaller or less relevant?

Can the authors compare against a simpler iterative model without retrieval, under the same computational budget?

5 comments
AuthorsOpenReview ↗

Part 1/2 of the responses

We sincerely thank you for your time and thoughtful comments, which have helped us clarify and strengthen the manuscript. We respond point by point below.

W1: Method complexity and improvement of each component?

Response: Our method is built around hypothesis-evidence feedback reasoning, with a controlled ablation showing that enabling the REASON stage improves Cos-FC from 0.6218 to 0.6986 (Table 4 in the initial manuscript), demonstrating its core contribution beyond the supporting components. Below, we clarify the motivation behind each component and its contribution.

  • Initialization: Kernel ridge regression provides the query-specific initial response, while PCA provides a fixed low-dimensional response space shared by evidence retrieval and hypothesis revision. Table 4 isolates the contribution of the learned initialization: replacing it with the mean training response reduces Cos-FC from 0.6986 to 0.6581.
  • The two revision networks in REASON (GθG_\theta and HνH_\nu): The conditional revision MLP GθG_\theta adapts retrieved evidence to the current query and hypothesis. With innovation disabled, this learned correction improves Cos-FC from 0.6601 for direct interpolation to 0.6708. The residual MLP HνH_\nu learns a correction that is projected outside the local evidence–correction span and norm-bounded to form the innovation update, further improving Cos-FC to 0.6986 (Table 4). These comparisons separately establish the contributions of learned correction and residual innovation.
  • The critic: Evidence retrieval determines which evidence to use, and the critic evaluates whether the resulting revision is worth accepting. This distinction matters because even a well-selected evidence need not yield an improving update. With the upstream modules frozen, adaptive acceptance and stopping achieve 0.6986 Cos-FC, compared with 0.6807 at fixed depth (K=4) and 0.6867 at the validation-selected fixed depth (Table 8 in the initial manuscript).
  • Population realization: The final aggregate response specifies a mean, but does not determine cell-to-cell heterogeneity. REALIZE addresses this separate prediction requirement. At exactly the same terminal mean and Cos-FC, it reduces MMD-G from 0.2903 for rescaled control cells to 0.1046 (Table 4 in the initial manuscript), demonstrating a contribution to population fidelity beyond aggregate accuracy.

We have revised the presentation to make these functional distinctions and their corresponding ablation evidence more explicit. (Line 451 in the revised paper)

W2: Theoretical results relying on unenforced assumptions.

Response: The propositions identify when adapting evidence and reasoning depth offers an advantage. HEFR explicitly trains hypothesis-dependent retrieval to estimate evidence utility and its stopping critic to predict revision gains, aligning the learned policies with the underlying theoretical decision principles. The theoretical conditions define the scope of the guarantees, while Appendix B directly tests whether these learned mechanisms deliver the corresponding benefits.

  • Evidence adaptation: Proposition 1 establishes a strict advantage when different hypotheses benefit from complementary evidence, so no single evidence is optimal throughout. This is a condition on evidence usefulness, rather than an architectural constraint. HEFR exploits it by reassessing evidence after each hypothesis update. At identical states and with the revision network fixed, current-state selection yields a mean cosine gain of 0.0043 versus 0.0011 for retaining the initial record (Table 6 in the initial manuscript). Retrieval scores correlate with realized gains (within-state Spearman ρ=0.43\rho=0.43, Table 6). These results connect the theoretical value of evidence adaptation to actual prediction gains delivered by HEFR’s learned retriever.
  • Adaptive stopping: Although the sufficient persistence condition in Proposition 2 is not enforced architecturally, HEFR’s stopping policy is explicitly trained to follow the corresponding gain-based decision principle. Gain-supervised critic training and calibrated threshold selection encourage accepting beneficial revisions and stopping when the expected gain is insufficient. We evaluate the practical consequence of stopping by continuing trajectories offline: mean missed utility is 0.0012 among early-stopped queries (Table 7 in the initial manuscript). With upstream modules frozen, adaptive stopping also improves Cos-FC from 0.6867 at the validation-selected fixed depth to 0.6986 (Table 8 in the initial manuscript). Together, these results support the practical relevance of Proposition 2’s gain-based stopping principle, showing that HEFR’s learned critic captures the benefit of query-adaptive depth over a validation-selected fixed policy.
AuthorsOpenReview ↗

Part 2/2 of the responses

W3: Limited runtime savings of adaptive stopping.

Response: The benefit of adaptive stopping lies in improving prediction quality while controlling inference cost. Its purpose is to decide whether each candidate revision is worth accepting, since additional revisions can reduce accuracy, as shown by the fixed-depth comparison in Table 8 in the initial manuscript. With the shared modules frozen, adaptive stopping achieves 0.6986 Cos-FC, outperforming both the validation-selected fixed depth K=3K=3 (0.6867) and K=4K=4 (0.6807, Table 8). Compared with K=4K=4, it improves Cos-FC by 0.0179 while reducing end-to-end inference time by 3%, including the critic’s evaluation cost. The critic’s additional complexity therefore delivers higher predictive accuracy at lower inference cost, making runtime savings only one part of its benefit. We have emphasized this joint accuracy–cost benefit of adaptive stopping in the revised paper (Line 988 in the revised paper).

Q1: Performance sensitivity to evidence memory size and relevance.

Response: Thank you for your suggestion. To answer this question, we examine memory size and relevance separately, keeping the other settings fixed.

  • Memory size: On the sciPlex3 unseen-perturbation benchmark, we construct randomly sampled, nested subsets retaining 100%, 50%, 25%, and 10% of the eligible evidence memory. All conditions use the same test queries and evaluation protocol. From the table below we can see that the performance degrades as memory shrinks. Retaining half the memory reduces Cos-FC by 0.0268. Even with only 10%, HEFR improves over the initial prediction without revision (0.6218, Table 4 in the original manuscript), indicating that useful refinement remains possible with substantially fewer records.
  • Evidence relevance: We retain 50% of the memory and compare random removal with removal of the most or least relevant records for each query. Relevance is ranked using HEFR’s scorer at the initial hypothesis, and the retained memory is fixed throughout subsequent adaptive retrieval. Removing the most relevant records causes greater degradation than random removal, whereas removing the least relevant records preserves performance close to the full-memory result (shown below). Thus, performance depends on the information retained, beyond memory size alone. We have included these results in Appendix D.3 in the revised paper (Table 13 and 14).

Ablation on memory size:

Memory retainedCos-FC ↑MMD-G ↓
100%0.69860.1046
50%0.67180.1129
25%0.66150.1284
10%0.64080.1612

Ablation on memory relevance:

Construction of the retained 50% memoryCos-FC ↑MMD-G ↓
Remove least relevant records0.69610.1087
Random removal0.67180.1129
Remove most relevant records0.63890.1652

Q2: Comparison with a simpler retrieval-free iterative model under matched computational budgets.

Response: We compare HEFR with a single shared MLP that iteratively refines the current prediction without retrieval, allowing it to spend the saved computation on additional iterations.

  • Baseline: The model applies zk+1=zk+ηFθ(hp,c,b,zk,k)z_{k+1}=z_k+\eta F_\theta(h_p,c,b,z_k,k), using the perturbation descriptor, context, basal state, current hypothesis, and iteration index. It shares HEFR’s initialization, PCA basis, and decoder, but has no evidence inputs, separate innovation head, or critic. The MLP is trained independently on its own unrolled trajectories, with response supervision at each step and the same data partitions.
  • Matched computational budget: We select the best validation-performing MLP width and iteration count whose mean inference time fits within HEFR’s budget, measured under identical hardware, batching, and output cell counts.

At comparable computational budget, HEFR improves Cos-FC by 0.0674 and reduces MMD-G from 0.2316 to 0.1046 (shown below). These results suggest that hypothesis-dependent evidence provides useful correction signals beyond what the shared MLP can obtain through repeated refinement alone.

Comparison with simple iterative model under matched computational budgets:

MethodCos-FC ↑MMD-G ↓Relative inference latency
Shared iterative MLP, matched budget0.63120.23161.00
Full HEFR0.69860.10461.00

We hope our responses have addressed your concerns, and we welcome further discussion if any clarification is needed, thank you!

Dear Reviewer,

Thank you for your thoughtful review and contribution to the discussion.

As the discussion period approaches its conclusion, I would like to kindly remind you to review the authors' rebuttal and engage in the discussion if you have not already done so.

In particular, it would be helpful if you could indicate whether the authors have adequately addressed your concerns and, if appropriate, reconsider your assessment and score.

Thank you again for your time and effort!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.