acceptodds
Under review as a conference paper at ICLR 2027

Self-Supervised Editable Transport for Counterfactual Spatiotemporal Gene Perturbations

12 views

Abstract

Predicting spatiotemporal tissue responses to localized gene edits from destructive, unpaired spatial snapshots is crucial for virtual perturbation screening. However, existing methods isolate temporal dynamics from interventional capacity. Bridging this dichotomy requires targeted edits that remain physically coherent and on-manifold throughout temporal propagation. We introduce **STEP** (**S**elf-supervised **T**ransport of **E**ditable **P**erturbations), a generative framework trained exclusively on unperturbed snapshots to forecast counterfactual spatial dynamics. STEP formalizes gene knockouts as *decoder-consistent proximal projections* within an editable, mask-supervised latent manifold. These interventional states propagate continuously across unpaired snapshots via a shared wild-type path law, governed by mass-calibrated flow matching and unbalanced optimal transport. Extensive benchmarks across static and dynamic spatial CRISPR assays confirm that STEP outperforms existing baselines in counterfactual forecasting and response-gene prioritization. Mechanistic analyses further validate that decoder-consistent proximal projections are crucial for maintaining manifold integrity during latent transport. Applying STEP to a zebrafish embryogenesis atlas uncovers stage-dependent *in silico* knockout heterogeneity, dissecting dynamic developmental mechanisms.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Spatial transcriptomics, Optimal Transport, Self-supervised representation learning, in silico perturbation

Reviews

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

The paper introduces STEP as a framework for spatiotemporal perturbation prediction using wildtype snapshots of spatial transcriptiomics. It combines a mask-aware autoencoder with unbalanced optimal transport and flow matching in order to learn wildtype temporal dynamics. The method performs interventions through a decoder-consistent proximal edit operator, ensuring counterfactual states remain biologically coherent. The manuscript validates STEP using both synthetic data, wildtype inter-/extrapolation, static spatial CRISPR, dynamic SPAC-seq, and a developmental case study on zebrafish.

Critical strengths

Setting. The paper is very well-motivated for a difficult and relevant problem.

Approach. The mechanism is conceptually appealing.

Breadth of benchmarking. Empirical benchmarking is diverse with wildtype dynamics, static perturbations, dynamic perturbations, and so on.

Critical weaknesses

The extrapolation results are difficult to interpret without simple sanity baselines. The paper argues that STEP maintains "positive rank alignment," but 0.010 Spearman is practically zero. I would not view this result as meaningful evidence of a successful extrapolation. The paper lacks simple sanity baselines such as: no-change, a linear gene-wise extrapolation, constant latent velocity, or an OT-only coupling (to disentangle the benefit of STEP's learned vector field). It is hard to see whether the better RMSE is explained by meaningful or more conservative predictions. Since the authors make strong claims about extrapolation, even when correlation is essentially zero, it is important to explore the predictions more.

The provided code raises substantial reproducibility concerns. The GitHub code has large inconsistencies with the paper, making reproducibility a major concern. It doesn't seem to contain any of the mask-aware neural encoders/decoders, unbalanced optimal transport, nor their flow-matching objective. It instead shows a very basic demo using PCA and nearest neighbors. Noticeably, it doesn't even include PyTorch or POT as dependencies -- thus not a single neural network -- despite the paper explicitly stating these packages.

Questions for authors

Clarification on basic assumptions. Disease-related populations may have GRNs and dynamics that differ substantially from wild-type populations. Does it make sense for knockouts to stay on the wild-type manifold?

Clarification about reproducibility. Is there any code for the experiments and results in the paper?

Baselines for the machine learning-based methods. Could you add some simple sanity baselines such that readers could better understand the strength of the methods? These could for example be relevant when interpreting results of Table 1.

Minor:

Line 161 seems to have an error with a sudden "(...), with full replicate integration and preprocessing protocols detailed in Appendix H" while discussing the destructive nature of spatial profiling.

Appendix D should be "Extended" rather than "Entended"

6 comments
AuthorsOpenReview ↗

Rebuttal Part 1/2

We thank the reviewer for the valuable suggestions on baseline evaluation, extrapolation assessment, and code reproducibility. Our point-by-point responses are provided below.

1. What do simple sanity baselines reveal about STEP's extrapolation claim (W1, Q3)?

We added four sanity baselines—no-change persistence, gene-wise linear extrapolation, latent OT-only constant-velocity continuation, and expression-space OT-only continuation, to compare with stVCR and DeepRUOT.

On ARTISTA and MOSTA, expression-space OT-only leads all four gene-level interpolation metrics (Pearson, Spearman, RMSE, and DE AUPRC), likely benefiting from observed snapshots bracketing held-out stages. This advantage does not consistently extend to extrapolation, which uses only preceding snapshots. Temporal sampling also matters: no-change achieves the lowest RMSE and SWD on densely sampled MOSTA, illustrating that conservative persistence can favor endpoint and distributional metrics without providing a directional forecast. ARTISTA's adult forecasting task comprises only one juvenile-to-adult transition across five non-uniform stages. Although expression-space OT performs strongly on this transition, STEP's Spearman is effectively zero and does not support meaningful rank alignment. These results underscore the need to distinguish endpoint accuracy, response-gene prioritization, and directional recovery.

To assess generalizability beyond developmental atlases, we evaluated axolotl limb regeneration profiled by Xenium (García-García et al., 2025, bioRxiv) and zebrafish heart regeneration profiled by Stereo-seq (Li et al., 2025, Nat. Commun.), spanning distinct dynamic processes and platforms. Across four datasets, two tasks, and five metrics (40 comparisons), STEP ranks first in 12 and among the top two in 26, the broadest top coverage among seven methods. On both regeneration datasets, STEP leads forecasting Pearson, Spearman, and RMSE, while achieving the highest DE AUPRC on zebrafish heart. Together, these results highlight STEP's directional forecasting strengths beyond developmental atlases, while simpler methods remain competitive on selected metrics. Table R3.1 reports forecasting; revised Table 1 covers both tasks (dataset and baseline details: Appendices H and J).

Table R3.1. One-step forecasting; bold: best, underline: second.

DatasetMethodPearson↑Spearman↑RMSE↓DE AUPRC↑SWD↓
ARTISTASTEP−0.1058‾\underline{-0.1058}0.01030.11620.18550.4176
stVCR-0.2811-0.28890.23170.08590.4353
DeepRUOT-0.1816-0.25750.25890.09660.3566
No-changeN/AN/A0.10850.10000.1763‾\underline{0.1763}
Linear-0.3584-0.23530.11610.36590.2051
Latent OT-0.12110.0512‾\underline{0.0512}0.1064‾\underline{0.1064}0.18060.3757
Expr. OT0.45720.23330.10590.3658‾\underline{0.3658}0.1596
MOSTASTEP0.1326‾\underline{0.1326}0.25520.0987‾\underline{0.0987}0.32220.2985
stVCR-0.02610.05280.24390.23360.3897
DeepRUOT0.04500.00230.22180.12160.2659
No-changeN/AN/A0.08800.10000.1141
Linear-0.0965-0.27020.11690.3554‾\underline{0.3554}0.1408‾\underline{0.1408}
Latent OT0.13270.1888‾\underline{0.1888}0.11020.17970.2395
Expr. OT-0.0201-0.14090.10480.45510.1530
Limb XeniumSTEP0.29130.23520.12850.29250.2137
stVCR0.1855‾\underline{0.1855}0.20940.15150.35550.5040
DeepRUOT0.15940.2102‾\underline{0.2102}0.1358‾\underline{0.1358}0.18290.2131
No-changeN/AN/A0.13990.10200.1515
Linear0.10890.08530.14260.3011‾\underline{0.3011}0.1537‾\underline{0.1537}
Latent OT0.03670.07060.35150.20910.3728
Expr. OT-0.1408-0.01980.16810.25860.2132
Zebrafish heartSTEP0.56930.57070.04800.47910.2550
stVCR0.4703‾\underline{0.4703}0.4795‾\underline{0.4795}0.0530‾\underline{0.0530}0.4677‾\underline{0.4677}0.2971
DeepRUOT0.14580.15910.07700.22220.2232
No-changeN/AN/A0.06380.10000.0889
Linear-0.4304-0.49560.09950.39800.1084‾\underline{0.1084}
Latent OT0.08140.06240.21720.17510.2456
Expr. OT-0.2023-0.19550.08740.38980.1553

Importantly, STEP is not merely a WT forecaster; it unifies multi-snapshot spatial dynamics learning with target-specific perturbation prediction. It learns a continuous WT path law from unperturbed snapshots via mask-aware representation learning, relation-aware unbalanced OT, and neural flow matching. Decoder-consistent latent projection introduces localized gene interventions that are maintained during propagation through the learned WT field, generating spatially resolved, time-dependent counterfactual responses from WT-only training data.

AuthorsOpenReview ↗

Rebuttal Part 2/2

2. Is the actual neural implementation available (W2, Q2)?

Yes. We apologize for the incomplete code release: the previous repository contained only a demo. We have now updated the code repository linked in the manuscript with the full PyTorch implementation of STEP, including the mask-aware encoder/decoders, unbalanced optimal transport, flow-matching objective, and repeated projection operator. The repository provides the required dependencies, model configurations, and instructions for running STEP.

3. Why is a WT reference reasonable when disease dynamics differ (Q1)?

A WT reference provides a context-specific starting point, not a constraint that KO states must remain within the WT distribution. STEP assumes that WT-derived spatial context and local temporal organization provide useful information for short-horizon perturbation-response prediction, even when perturbations induce distributional shifts. This principle is also adopted by existing perturbation modeling methods, including CellOracle (Kamimoto et al., 2023, Nature) and Celcomen (Megas et al., 2025, ICLR), which use unperturbed reference data to model in silico perturbations.

In STEP, target-specific latent editing is mapped to expression space by the decoder. The decoder image (Eq. 29) defines representable expression states, whereas empirical WT support reflects the observed training-state distribution; these are not equivalent. We therefore assessed whether measured KO states and STEP-generated edits could extend beyond empirical WT support. The threshold was defined by the 95th percentile (q95) of WT leave-one-out latent kNN distances, using matched local tissue contexts.

Table R3.2. Percentage of states exceeding the empirical WT-reference threshold; means over three model initializations per assay.

AssayMatched WT above q95 (%)Measured KO above q95 (%)Generated edit above q95 (%)
HT291.195.950.00
Mouse lung8.3373.7512.08
SPAC-seq2.772.090.00

In mouse lung, 73.75% of measured KO states but only 8.33% of matched WT states exceed q95. Importantly, 12.08% of STEP-generated edits also exceed this threshold, demonstrating that STEP does not enforce hard membership in the empirical WT support. However, the smaller fraction relative to measured KO states indicates that STEP does not fully reproduce the observed distributional shift.

We further tested whether the predicted responses depend on the identity of the targeted gene. With the initial tissue context fixed, we compared correct-target edits against wrong-gene edits and 20 random directions matched to the final latent displacement norm. Candidate edited genes were excluded from response scoring.

Table R3.3. Gene-identity and final-displacement controls; AUPRC ↑, means over three model initializations.

AssayCorrect edit AUPRC ↑Wrong-gene edit AUPRC ↑Final-norm random AUPRC ↑
HT290.14100.13200.1258
Mouse lung0.26930.12470.1536
SPAC-seq0.16940.16260.1652

Correct-target editing achieves the highest AUPRC in all three assays, with the clearest separation in mouse lung (0.2693 vs. 0.1247 for wrong-gene edits and 0.1536 for random directions). The smaller differences in HT29 and SPAC-seq indicate that target-specific prioritization is assay-dependent.

Together, these results support WT-derived tissue context as a useful reference rather than a hard constraint on perturbation states. By repeatedly enforcing target-specific edits during temporal propagation, STEP allows perturbation responses to evolve beyond empirical WT support while retaining the intended intervention. Its current formulation provides a framework for target-specific response prioritization, although large disease-specific distributional shifts are not yet fully recovered. These analyses are in Appendix F (Tables 8–10); Section 6 discusses the WT reference and limitations.

4. Minor corrections. We have removed the misplaced Appendix-H phrase in Section 3.1 and corrected the heading to “Extended Data Figures” in Appendix D.

Thanks for providing fixed code. I did not check it, but I trust that it has the method now!

It was great to have some baselines to put the other numbers in perspective. However I believe that the authors will need to rethink the narrative after all of these new things provided to the individual reviewers.

When reading the pdf right now, the story gets very overwhelming, and it does not leave the sense of a clean ICLR-quality paper with highly trustworthy results.

I am thankful for the large and hasty efforts during the review process, and the idea is good, so it's likely that the paper will eventually reach publication -- but at the current stage I choose to retain weak reject

Reviewer_apSSOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper propose STEP, a generative framework for prediction of spatiotemporal responses to gene perturbation trained zero-shot on unpaired and unperturbed snapshots. It combines masked representation learning to learn an editable latent manifold, unbalanced optimal transport and mass-calibrated flow matching to account for unpaired temporal dynamics, and proximal projection operator to execute gene edits while maintaining latent consistency. STEP is evaluated on wild-type snapshot dynamics, static and dynamic gene perturbation, and studied on zebrafish atlas to assess for biological insights.

Critical strengths

S1. Spatiotemporal gene perturbation prediction from unpaired and unperturbed snapshots is an ambitious and useful problem. Although largely built upon existing components, STEP elegantly integrates masked representation learning, optimal transport, flow matching, and latent editing to predict gene perturbations and model temporal flow.

S2. The paper evaluates and compares the model on wild-type dynamics, static spatial counterfactuals, and dynamic perturbation forecasting. STEP achieves strong results on several of these benchmarks and improves over the baselines. In addition, the zebrafish experiment demonstrates how STEP can generate interpretable hypotheses. The results are biologically plausible and valuable beyond numerical benchmarks.

Critical weaknesses

W1. From Figure 7b, the direct clamp baseline lacks nearly any edit signal in latent space, while STEP explicitly optimises the latent state. The improvements may at least partly result from a larger edit rather than from the benefits of manifold constrained projection. This concern is important when isolating the benefit of the proximal editing operator.

W2. The proximal objective encourages the edited state to remain close to the original latent state. However, being close to an observed state does not guarantee that Z' (eq. 15) is in the region of the training distribution. Similarly, (D(Z') - y*) term (eq. 15) does not guarantee that the decoded gene stays close to the requested gene value. Proposition 3 provides a conditional bound on target fidelity, and Figures 6c and 7c report target residuals, but these do not establish that the edited states remain in regions where the learned wild-type dynamics are reliable.

W3. The method applies projection after every numerical integration step but does not isolate the benefit of doing so. The paper would benefit from evaluating the performance of initial-only projection and sensitivity to solver step size (h). A comparison without temporal flow would also help establish whether temporal propagation contributes to the dynamic perturbation results beyond the editing operator alone.

W4. Although STEP improves over the baselines, the absolute accuracy of counterfactual predictions is limited on some real-data benchmarks. For example, on HT29, signed Spearman correlation is 0.006 and response AUPRC is 0.128 against a prevalence baseline of approximately 0.10. These results appear to support response-gene prioritisation more clearly than accurate prediction of counterfactual expression profiles.

Questions for authors

Q1. Figures 6c and 7c report target-gene residuals for selected examples. How consistently is the requested value y* achieved across all evaluated perturbations and maintained throughout rollout? Can you analyse whether the edited states remain close to the training distribution?

Q2. Can you compare initial-only projection with repeated projection and quantify the sensitivity to integration step size? It would help to report the residuals before and after each projection and include a baseline with repeated projection but without temporal flow. An adaptive or scheduled projection strategy might also improve the performance.

3 comments
AuthorsOpenReview ↗

Rebuttal Part 1/3

Thank you for the constructive comments on the rigor and reliability of STEP's perturbation modeling. Our responses address edit specificity, projection fidelity, temporal propagation, and predictive accuracy.

1. Are STEP's gains explained only by a larger latent displacement (W1)?

To determine whether STEP's gains arise from displacement magnitude alone or from target-specific latent editing, we conducted two complementary analyses.

First, we compared direct expression clamping, mask-aware encoding without projection, and two projection strategies under identical models, queries, spatial contexts, and rollout settings (Table R2.1). Direct clamping modifies the expression input before encoding, whereas the no-projection variant retains the explicit target mask but disables decoder-guided latent optimization. The two projection strategies additionally optimize the latent state against the decoder target, either at initialization only or repeatedly during rollout. Notably, the target mask provides explicit intervention identity even when the target gene is already unexpressed at onset (96.4% of HT29 and 93.8% of SPAC-seq queries).

Table R2.1. Matched expression-clamp and latent-projection controls; response AUPRC ↑, means over three model initializations per assay.

Editing strategyHT29 AUPRC ↑Mouse lung AUPRC ↑SPAC-seq AUPRC ↑
Direct expression clamp0.10650.20770.1206
No projection0.13470.12240.1561
Initialization-only projection0.13870.24010.1565
Initialization + per-step projection0.14610.26920.1673

To directly isolate the effect of edit magnitude, we additionally randomized the direction of STEP's final latent displacement while preserving its per-cell norm, using 20 randomizations per query. We also performed wrong-gene controls by applying the same editing procedure to alternative target genes while keeping the original query and spatial context fixed. Both controls use the same WT endpoints and decoder.

Table R2.2. Gene-identity and final-displacement controls; AUPRC ↑, means over the same three initializations.

AssayCorrect edit AUPRC ↑Wrong-gene edit AUPRC ↑Final-norm random AUPRC ↑
HT290.14100.13200.1258
Mouse lung0.26930.12470.1536
SPAC-seq0.16940.16260.1652

Table R2.1 excludes the queried target gene, whereas Table R2.2 excludes all candidate edited genes identically across conditions. The correct-edit predictions are identical across both tables; differences in AUPRC arise solely from the scoring masks.

The norm-matched controls demonstrate that STEP's gains cannot be explained by displacement magnitude alone. On mouse lung, the correct edit achieves an AUPRC of 0.2693, substantially exceeding both random directions (0.1536) and wrong-gene edits (0.1247), with corresponding Precision@20 values of 0.2222, 0.1193, and 0.0833. HT29 and SPAC-seq also achieve their highest mean AUPRC with the correct edit, although the separations are smaller.

Table R2.1 further shows that decoder-guided latent projection provides additional benefits beyond mask-aware encoding alone. While mask-aware encoding without projection improves over direct clamping on HT29 and SPAC-seq, this advantage is not observed on mouse lung. Nevertheless, latent projection achieves the highest AUPRC across all three assays, with repeated projection further improving over initialization-only editing. Together, these results support the contribution of target-specific latent editing beyond displacement magnitude, with particularly clear evidence from mouse lung. These controls are included in Appendix F (Tables 6 and 8).

AuthorsOpenReview ↗

Rebuttal Part 2/3

2. Is the requested target maintained, and do edited states remain near empirical WT support (W2, Q1)?

We assessed target fidelity throughout rollout and proximity to empirical WT support separately. Proposition 3 provides conditional bounds for an ideal global proximal solution, but does not guarantee these properties under finite-step optimization. Likewise, decoder-manifold membership (Eq. 29) does not imply empirical WT-support membership. We therefore evaluated both properties directly.

For target fidelity, we compared initialization-only and repeated projection under matched initial states, queries, spatial graphs, frozen WT dynamics, and solvers. Across 552 query–model trajectories, we tracked raw decoder-scale target residuals at edited anchors before output clipping or target-coordinate overwriting, with the KO target set to zero.

Table R2.3. Complete rollout target residuals; each entry is mean / median / q95 in assay-specific decoder units.

AssayTimeInitialization-only projectionRepeated projection: beforeRepeated projection: after
HT2900.0389 / 0.0215 / 0.14290.0535 / 0.0230 / 0.25690.0389 / 0.0215 / 0.1429
HT290.250.0388 / 0.0210 / 0.14220.0388 / 0.0210 / 0.14220.0302 / 0.0200 / 0.0856
HT290.50.0388 / 0.0209 / 0.14050.0305 / 0.0197 / 0.08860.0250 / 0.0187 / 0.0672
HT290.750.0388 / 0.0209 / 0.14000.0251 / 0.0190 / 0.06830.0214 / 0.0172 / 0.0547
HT2910.0389 / 0.0210 / 0.13360.0213 / 0.0173 / 0.05480.0190 / 0.0158 / 0.0491
Mouse lung00.6609 / 0.6172 / 0.97070.7097 / 0.6697 / 1.02380.6609 / 0.6172 / 0.9707
Mouse lung0.250.6661 / 0.6268 / 0.95110.6661 / 0.6268 / 0.95110.6207 / 0.5799 / 0.9002
Mouse lung0.50.6696 / 0.6331 / 0.94060.6266 / 0.5893 / 0.89200.5841 / 0.5452 / 0.8448
Mouse lung0.750.6688 / 0.6367 / 0.93880.5846 / 0.5470 / 0.84610.5445 / 0.5095 / 0.8033
Mouse lung10.6640 / 0.6303 / 0.94550.5372 / 0.5056 / 0.80420.4992 / 0.4755 / 0.7611
SPAC-seq40.0688 / 0.0496 / 0.18310.0712 / 0.0510 / 0.19090.0688 / 0.0496 / 0.1831
SPAC-seq4.50.0632 / 0.0487 / 0.16300.0632 / 0.0487 / 0.16300.0611 / 0.0474 / 0.1571
SPAC-seq50.0588 / 0.0456 / 0.15110.0568 / 0.0445 / 0.14640.0550 / 0.0434 / 0.1417
SPAC-seq5.50.0553 / 0.0404 / 0.14500.0518 / 0.0378 / 0.13620.0502 / 0.0368 / 0.1319
SPAC-seq60.0530 / 0.0365 / 0.14000.0482 / 0.0333 / 0.12760.0467 / 0.0324 / 0.1237
SPAC-seq6.50.0513 / 0.0364 / 0.13800.0454 / 0.0336 / 0.12330.0440 / 0.0328 / 0.1197
SPAC-seq70.0503 / 0.0360 / 0.13850.0434 / 0.0319 / 0.12130.0421 / 0.0311 / 0.1178

All 3,048 projection events reduce or preserve target residuals. Repeated projection also yields lower mean, median, and q95 residuals than initialization-only projection at every post-onset time. The final mean residual decreases by 51.2%, 24.8%, and 16.3% across HT29, mouse lung, and SPAC-seq, respectively, confirming improved target fidelity throughout rollout.

We next assessed empirical WT support using latent five-nearest-neighbor distances calibrated against WT leave-one-out q95 thresholds. We evaluated matched WT and KO anchors, WT rollout inputs, edited states supplied to the WT temporal field, and final edited endpoints.

Table R2.4. Latent states exceeding the empirical WT-reference threshold (%); means over three initializations.

AssayMatched WT anchorsMeasured KO anchorsWT rollout inputsEdited rollout inputsEdited endpoints
HT291.195.950.150.000.00
Mouse lung8.3373.752.081.3512.08
SPAC-seq2.772.091.270.350.00

Most edited states supplied to the WT temporal field remain within empirical WT support, with only 0.00–1.35% of rollout inputs exceeding the WT threshold. Importantly, projection does not enforce strict WT-support membership: 12.08% of mouse-lung edited endpoints extend beyond the threshold, compared with 73.75% of measured KO anchors.

Together with the improved response AUPRC over initialization-only projection across all three assays (Table R2.1), these results demonstrate that STEP maintains target fidelity throughout rollout while keeping most WT-field evaluation inputs within empirical WT support, without constraining final perturbation states to the observed WT distribution. Although this proximity assessment cannot guarantee vector-field accuracy far outside WT support, such states constitute only a small fraction of the inputs encountered during propagation. Complete results are included in revised Appendix C and F (Tables 7 and 9).

AuthorsOpenReview ↗

Rebuttal Part 3/3

3. Do the conclusions depend on solver step size, and does temporal flow add value (W3, Q2)?

We conducted matched SPAC-seq Day-4→Day-7 controls for projection placement, solver resolution, and temporal flow. The initialization-only, repeated-projection, and no-flow conditions share identical initial projected states. We additionally included a budget-matched initialization control (14 gradient updates at onset) and a no-flow control retaining the same repeated-projection schedule and optimization budget but disabling temporal velocity.

Table R2.5. SPAC-seq projection placement, solver resolution, and temporal-flow controls; means over three initializations.

ControlSolver step (days)Projection interval (days)Events / gradient updatesTemporal flowAUPRC ↑Final raw residual ↓
Initialization-only projection0.5Onset only1 / 2Yes0.15650.0503
Initialization-only, matched budget0.5Onset only1 / 14Yes0.16400.0424
Repeated projection (STEP)0.50.57 / 14Yes0.16730.0421
Repeated projection, no flowN/A0.57 / 14No0.15890.0566
Repeated projection0.250.57 / 14Yes0.16700.0426
Repeated projection0.1250.57 / 14Yes0.16750.0429
Scheduled projection0.2514 / 8Yes0.16380.0465
Scheduled projection0.250.2513 / 26Yes0.16790.0362

The gains are not explained solely by a larger optimization budget. With equal budgets (14 updates), distributing optimization across the trajectory improves AUPRC over onset-only updates (0.1673 vs. 0.1640), with comparable residuals (0.0421 vs. 0.0424). STEP also outperforms standard initialization-only projection (AUPRC: 0.1565 vs. 0.1673).

STEP is also robust to solver resolution. With projection intervals and optimization budgets fixed, varying the integration step from 0.5 to 0.25 and 0.125 days produces nearly identical AUPRC (0.1673, 0.1670, and 0.1675, respectively). Projection scheduling, however, can influence performance: at the same solver resolution (0.25 days), increasing projection frequency from every 1 day to every 0.25 days improves AUPRC from 0.1638 to 0.1679 and reduces the final residual from 0.0465 to 0.0362, albeit with additional optimization cost. We thank the reviewer for suggesting adaptive projection, which we consider a promising extension. These results provide preliminary evidence that projection timing influences target fidelity, suggesting potential benefits from adapting projection frequency to local residuals or dynamical changes. The current experiments use prespecified intervals; Section 6 discusses adaptive scheduling as future work.

To isolate the contribution of temporal propagation further, we compared the residual trajectories with and without temporal flow under the same repeated-projection schedule.

Table R2.6. Raw target residual before and after every SPAC-seq projection, with and without temporal flow; means over the same three initializations.

DayWith flow: beforeWith flow: afterWithout flow: beforeWithout flow: after
40.07120.06880.07120.0688
4.50.06320.06110.06880.0665
50.05680.05500.06650.0644
5.50.05180.05020.06440.0623
60.04820.04670.06230.0603
6.50.04540.04400.06030.0584
70.04340.04210.05840.0566

Both conditions reduce or preserve target residuals at all 1,008 projection events per condition. Importantly, restoring temporal flow improves response AUPRC from 0.1589 to 0.1673 and reduces final residual from 0.0566 to 0.0421 relative to repeated projection without flow, with AUPRC gains observed across all three initializations. Together, these matched controls establish that STEP's performance reflects not only repeated target optimization but also its temporal propagation, while remaining robust to numerical integration resolution. These controls are included in Appendix F (Table 11).

4. Do the real-data results support accurate counterfactual expression profiles (W4)?

On real data, STEP primarily supports response-gene prioritization from WT-only spatial snapshots, rather than accurate reconstruction of full counterfactual expression profiles. HT29 illustrates this limitation, with near-zero signed Spearman and AUPRC only modestly above prevalence (~0.10). Nevertheless, under matched comparisons (Table R2.1), STEP improves response AUPRC over direct clamping by 37%, 30%, and 39% on HT29, mouse lung, and SPAC-seq, respectively. These gains highlight the value of combining target-specific latent editing with temporal propagation and WT-derived spatial context. We revised the manuscript to clarify this scope and discuss limitations in Section 6.

Reviewer_S6RNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

STEP predicts spatiotemporal responses to localized gene perturbations using models trained only on unperturbed observations. It combines masked reconstruction, unbalanced optimal transport, and conditional flow matching, with decoder-consistent latent projection applied at initialization and after each integration step.

Critical strengths

1.The paper addresses a relevant problem: predicting spatially localized, sustained gene perturbation responses without perturbation labels during training.

2.The evaluation covers synthetic systems and real static and temporal perturbation datasets.

Critical weaknesses

1.The incremental benefit of per-step projection is unclear. Existing comparisons do not separate the effects of initialization projection and repeated projection, making it difficult to assess their respective contributions to the reported gains.

2.The benefits of several auxiliary terms are unclear. In Figure 9e–h, setting some weights to zero generally yields performance close to the default, with some curves slightly higher. Their added value on the reported metrics is therefore not clearly established.

Questions for authors

1.Some perturbations may produce biologically valid states outside the learned manifold. Could the authors empirically assess this possibility and whether projection suppresses real perturbation responses?

2.Could the authors compare no projection, initialization-only projection, and initialization plus per-step projection, while holding the trained model and rollout settings fixed?

3.What measurable benefits do the additional terms in Figure 9e–h provide?

3 comments
AuthorsOpenReview ↗

Rebuttal Part 1/3

Thank you for the valuable suggestions on WT-guided perturbation modeling and evaluation. Our point-by-point responses are provided below.

1. What is the incremental contribution of per-step projection beyond initialization (W1, Q2)?

We conducted the requested three-way ablation under matched conditions, keeping the frozen model, masked inputs, perturbation queries, and rollout settings identical. The no-projection variant retains mask-aware encoding but disables decoder-target latent optimization.

Table R1.1. Comparison of projection strategies (response AUPRC ↑; means over three matched model initializations).

Projection strategyHT29Mouse lungSPAC-seq
No projection0.13470.12240.1561
Initialization-only0.13870.24010.1565
Initialization + per-step0.14610.26920.1673

Per-step projection consistently improves AUPRC over initialization-only projection across all three assays, with gains of 0.0074, 0.0291, and 0.0108 on HT29, mouse lung, and SPAC-seq, respectively. Initialization provides the main improvement on mouse lung, while repeated projection yields additional gains across all assays, demonstrating distinct contributions from initial editing and subsequent constraint maintenance.

To assess this mechanism directly, we measured the raw target-gene residual immediately before and after each projection across all evaluated queries (Table R1.2). All projection events reduce or preserve the residual for every query–model pair (100% non-increasing). Compared with initialization-only projection, repeated projection reduces the final mean residual by 51%, 25%, and 16% across HT29, mouse lung, and SPAC-seq, respectively. These results demonstrate that per-step projection not only improves response prioritization but also maintains target fidelity throughout temporal propagation. These matched comparisons and complete rollout diagnostics are included in Appendix F (Tables 6–7).

Table R1.2. Target residuals before and after each projection event (means over the same three matched initializations).

AssayPairs/eventInit.-only final residual ↓Repeated projection: time (before → after) ↓Non-increasing
HT291680.03890: 0.0535→0.0389; 0.25: 0.0388→0.0302; 0.5: 0.0305→0.0250; 0.75: 0.0251→0.0214; 1: 0.0213→0.0190100%
Mouse lung2400.66400: 0.7097→0.6609; 0.25: 0.6661→0.6207; 0.5: 0.6266→0.5841; 0.75: 0.5846→0.5445; 1: 0.5372→0.4992100%
SPAC-seq1440.05034: 0.0712→0.0688; 4.5: 0.0632→0.0611; 5: 0.0568→0.0550; 5.5: 0.0518→0.0502; 6: 0.0482→0.0467; 6.5: 0.0454→0.0440; 7: 0.0434→0.0421100%
AuthorsOpenReview ↗

Rebuttal Part 2/3

2. What measurable benefits do the auxiliary objectives provide (W2, Q3)?

We agree that the original ablations in Figure 9e–h, conducted on MOSTA, do not fully establish the contributions of individual auxiliary objectives, as some terms yield only marginal changes on the evaluated metrics. However, their contributions may be dataset- and task-dependent, given that they regularize distinct aspects of spatial transport and temporal prediction. To better assess their roles. we extended the ablation analysis to four additional settings (ARTISTA, Limb Xenium, and SPAC-seq, alongside MOSTA), covering both interpolation and extrapolation as well as perturbation-response evaluation. Each ablation removes only the corresponding loss term, while keeping the architecture, WT data, remaining loss weights, and training budget fixed. Table R1.3 summarizes their task-specific contributions.

Table R1.3. Cross-task auxiliary-objective ablations.

ObjectiveMatched assay and metricFull STEPWithout term
Relation-aware transportARTISTA forecast AUPRC ↑0.17310.1627
Growth regularizationARTISTA forecast relative design-mass error ↓0.13463.8507
Endpoint consistencyARTISTA interpolation AUPRC ↑0.40970.3378
Endpoint consistencySPAC-seq final target residual ↓0.04210.0514
Endpoint consistencyLimb Xenium interpolation AUPRC ↑0.35580.3282
Moment consistencyMOSTA forecast AUPRC ↑0.33100.3272
Moment consistencyARTISTA forecast AUPRC ↑0.17310.1661
Moment consistencyARTISTA interpolation SWD ↓0.31790.3312
Moment consistencyLimb Xenium forecast Spearman ↑0.29720.0819
Moment consistencyLimb Xenium forecast SWD ↓0.23630.3374

The results reveal complementary, task-dependent contributions of the auxiliary objectives. Relation-aware transport improves temporal coupling, while growth regularization is important for calibrating transported design mass (relative error: 0.1346 vs. 3.8507). Endpoint consistency improves interpolation AUPRC on ARTISTA (0.3378→0.4097) and Limb Xenium (0.3282→0.3558), while reducing SPAC-seq target residual (0.0514→0.0421). Moment consistency is particularly beneficial for Limb Xenium forecasting, where its removal decreases Spearman from 0.2972 to 0.0819 and increases SWD from 0.2363 to 0.3374. Thus, the auxiliary objectives provide complementary benefits across datasets and prediction tasks, even when their individual effects are modest on particular benchmarks. The representative comparisons are included in Appendix F (Table 5).

AuthorsOpenReview ↗

Rebuttal Part 3/3

3. Does projection suppress valid responses outside the WT manifold (Q1)?

To examine whether projection restricts biologically valid perturbation responses, we quantified the distributional shift of measured KO states relative to matched WT anchors and STEP-generated edits, using latent kNN distances calibrated against the WT leave-one-out 95th percentile.

Table R1.4. Percentage of states exceeding the empirical WT-reference threshold; means over the same three initializations.

AssayMatched WT above q95 (%)Measured KO above q95 (%)Generated edit above q95 (%)
HT291.195.950.00
Mouse lung8.3373.7512.08
SPAC-seq2.772.090.00

Mouse lung exhibits the strongest WT–KO distribution shift, with 73.75% of measured KO anchors exceeding the WT-reference threshold, compared with 8.33% of matched WT anchors. We therefore further compared three projection strategies on this assay under matched model and rollout settings. Observed responses are target–section pseudobulk contrasts between lesion spots and spatially matched WT controls; all response metrics exclude the edited gene.

Table R1.5. Projection strategy and measured KO-response recovery in mouse lung; means over three initializations.

Projection strategyWT support exceedance (%)Norm ratioEffect Pearson ↑Signed Spearman ↑Effect RMSE ↓Response AUPRC ↑
No projection (mask-only)0.000.0366-0.1028-0.12110.49440.1224
Initialization-only projection0.000.10370.06560.03200.48920.2401
Repeated projection12.080.41510.10020.05960.50660.2692

Repeated projection improves recovery of measured KO responses beyond initialization-only editing, increasing the response norm ratio from 0.1037 to 0.4151 and AUPRC from 0.2401 to 0.2692, together with higher Pearson and signed Spearman correlations. Importantly, the response component aligned with the measured KO effect, quantified by the normalized inner product between predicted and observed response vectors, increases from 0.0166 to 0.0720, with a similar improvement among patches exceeding the WT-reference threshold (0.0161→0.0648). Moreover, 12.08% of generated edits extend beyond empirical WT support, indicating that projection does not impose strict membership in the observed WT distribution. Although this captures only part of the shift observed in measured KO states (12.08% vs. 73.75%), it demonstrates that STEP can generate perturbation-associated departures from empirical WT support.

The improvement in response prioritization is accompanied by a modest trade-off in expression-level calibration (RMSE: 0.4892→0.5066), with effects varying across targets and sections: Tgfbr2 in KP_1 improves in RMSE (0.4694→0.4120), whereas Jak2 shows the opposite trend (0.4819→0.5167). Together, these results suggest that repeated projection can partially alleviate the suppression of perturbation responses while maintaining target-specific editing and allowing predicted states to extend beyond empirical WT support, although the full extent of perturbation-induced distributional shifts is not recovered. These analyses are included in Appendix F (Tables 9–10), with the WT reference and response-calibration trade-off discussed in Section 6.

Comments on the paper

3 comments

Authors' Response Summary

We thank all reviewers for their careful reading and constructive feedback. The reviewers consistently recognized the importance and difficulty of predicting spatially localized, time-dependent perturbation responses from unpaired WT snapshots. Reviewer S6RN highlighted the relevance of the problem and the breadth of evaluation across synthetic, static, and temporal perturbation data. Reviewer apSS described STEP as an ambitious and useful framework that elegantly integrates masked representation learning, optimal transport, flow matching, and latent editing, and particularly recognized that the zebrafish experiment generates biologically plausible, interpretable hypotheses, demonstrating STEP's value beyond numerical benchmarking. Reviewer Xvcz likewise found the setting very well motivated, the mechanism conceptually appealing, and the empirical benchmarking diverse. Their suggestions prompted substantial new analyses that sharpen the interpretation of projection, temporal propagation, WT-guided prediction, and extrapolation.

New experiments. We conducted the following additional experiments in response to the reviewers:

  1. Projection contribution, target fidelity, and real-response recovery. We added matched comparisons of mask-only, initialization-only, and repeated projection, together with direct-clamp, wrong-gene, and norm-matched random-direction controls. Across 552 query–model trajectories, all 3,048 projection events reduced or preserved the raw target residual, and repeated projection improved response AUPRC over initialization-only editing in all three assays. Empirical WT-support and measured KO-response analyses on mouse lung further showed improved response magnitude and correlation while allowing final edits to extend beyond empirical WT support.
  2. Temporal-flow, solver, and projection-schedule controls. We compared repeated projection against budget-matched initialization and a repeated-projection/no-flow baseline, and varied both integration step size and projection frequency. These experiments isolate the contribution of temporal propagation, establish robustness to solver resolution, and motivate residual- or dynamics-guided adaptive scheduling.
  3. Cross-task auxiliary-objective ablations. We extended single-component removals beyond MOSTA to ARTISTA, Limb Xenium, and SPAC-seq. The results identify complementary, task-dependent roles for relation-aware transport, growth regularization, endpoint consistency, and moment consistency.
  4. Sanity baselines and broader WT-dynamics evaluation. We added no-change, gene-wise linear, latent OT-only, and expression-space OT-only baselines, and expanded evaluation to Limb Xenium and zebrafish heart Stereo-seq. Across four datasets, two tasks, and five metrics, STEP ranks first in 12 and among the top two in 26 of 40 comparisons, while the new controls also clarify when conservative or OT-only predictions are competitive.

Revisions to the manuscript and code. Guided by these analyses, we made the following revisions:

  1. We expanded Table 1 with the new baselines and regeneration datasets, and added the matched projection, identity, rollout-residual, WT-support, response-recovery, solver, flow, and auxiliary-objective analyses to Appendix F (Tables 5–11).
  2. We refined the theoretical presentation in Appendix C to distinguish decoder-manifold consistency from empirical WT-support membership and to connect the conditional bounds in Proposition 3 with the corresponding geometric, target-fidelity, and response-recovery experiments.
  3. We clarified throughout the manuscript that the real-data evidence primarily supports response-gene prioritization from WT-only spatial snapshots, while accurate reconstruction of full signed counterfactual expression profiles remains an open challenge.
  4. We expanded the discussion of the WT reference assumption, altered regulatory regimes, response-calibration trade-offs, disease-specific conditioning, and adaptive projection scheduling.
  5. We replaced the previous demo-only repository with the full PyTorch implementation of STEP, including the mask-aware encoder/decoders, unbalanced optimal transport, flow-matching objective, repeated projection operator, dependencies, configurations, and usage instructions. We also corrected the minor textual issues identified by Reviewer Xvcz.

All manuscript additions and revisions are marked in blue. We sincerely appreciate the reviewers' constructive feedback, which has led to a more rigorous evaluation, more precise claims, and a clearer presentation of STEP's contributions and limitations. We hope that these revisions adequately address the concerns raised and clarify the significance of our work.

Thank you for acknowledging the additional projection ablation. We would like to clarify two points regarding the remaining concerns.

1. Auxiliary objectives: cross-dataset consistency and trade-offs. To eliminate any metric-selection bias, we report the complete ablation matrix across all four datasets (MOSTA, ARTISTA, Limb Xenium, and zebrafish heart Stereo-seq), both tasks (interpolation and one-step forecasting), all folds, and three initializations. Each ablation zeroes out only the named loss coefficient while holding all else fixed.

Core finding: In the global cross-dataset aggregate, removing any single auxiliary objective degrades overall prediction quality (+0.75% to +2.23% composite loss increase), with average deterioration strictly exceeding improvement (D>ID > I for every term). While individual datasets exhibit task-dependent trade-offs (e.g., local improvements when removing specific terms in certain tissues), Full STEP achieves the lowest aggregate composite loss across tissues while preventing severe regime-specific failures.

Composite metric: We convert all five Table 1 metrics into unified loss ratios ℓ=(1−r, 1−ρ, RMSE, 1−AUPRC, SWD)\ell = (1-r,\, 1-\rho,\, \mathrm{RMSE},\, 1-\mathrm{AUPRC},\, \mathrm{SWD}) and compute d=15∑m=15log⁡(ℓmw/o/ℓmFull)d = \frac{1}{5}\sum_{m=1}^5 \log(\ell_m^{\mathrm{w/o}} / \ell_m^{\mathrm{Full}}) per matched fold and seed. Percentage change 100[exp⁡(d‾)−1]100[\exp(\overline{d})-1] is reported relative to Full STEP (0%0\%): positive values mean removal worsens prediction; negative values reflect local trade-offs.

Table R1.6a. Composite prediction-loss change (%) relative to Full STEP. Aggregated equally across metrics, folds (interp./forecast: MOSTA 6/4, ARTISTA 3/1, Limb 4/2, heart 6/4), and three initializations. Bold means indicate net degradation upon removal.

DatasetTaskw/o relation (%)w/o growth (%)w/o endpoint (%)w/o moment (%)
MOSTAInterpolation+0.141+0.022-0.021+0.323
MOSTAOne-step forecast+0.0430.000-0.319+0.183
MOSTAEqual-task mean+0.092+0.011-0.170+0.253
ARTISTAInterpolation+0.986+1.302+10.265-5.302
ARTISTAOne-step forecast+1.527+0.160+0.162+0.696
ARTISTAEqual-task mean+1.256+0.729+5.092-2.349
Limb XeniumInterpolation-0.537-1.695+3.512+6.455
Limb XeniumOne-step forecast+1.468+0.866-0.455+20.101
Limb XeniumEqual-task mean+0.461-0.423+1.509+13.073
Zebrafish heart Stereo-seqInterpolation+0.914+3.388+0.347-1.315
Zebrafish heart Stereo-seqOne-step forecast+1.485+2.969-0.349-1.352
Zebrafish heart Stereo-seqEqual-task mean+1.199+3.178-0.002-1.333
Equal-dataset meanEqual-task mean+0.751+0.864+1.586+2.229

Table R1.6b. Cross-dataset trade-off decomposition. DD and II denote 100×100\times the equally aggregated positive and negative log-loss ratios (100[exp⁡((D−I)/100)−1]100[\exp((D-I)/100)-1]). Physical design-mass error is evaluated separately.

Removed componentDeterioration DD (log pts)Improvement II (log pts)Net prediction loss change (%)Separate design-mass error change (%)
Relation1.6250.877+0.751-25.587
Growth2.1111.250+0.864+181.852
Endpoint2.3950.822+1.586+7.506
Moment4.8692.664+2.229+8.222

Takeaway on consistency and trade-offs:

  • Net gain across all terms: Every removal produces net deterioration (D>ID > I, Table R1.6b), which also holds under equal prediction-domain weights and an inverse-score loss transformation.
  • Complementary safeguards against collapse: Each objective prevents severe failure in specific biological regimes—removing endpoint degrades ARTISTA interpolation by +10.27%, moment degrades Limb forecasting by +20.10%, and growth degrades heart predictions by +3.18% while inflating design-mass error by +181.85%.
  • Principled regularization: Minor local trade-offs (e.g., moment smoothing ARTISTA interpolation by -5.30%, or relation trading -25.59% mass tightness for consistent expression gains across all four datasets) reflect expected multi-objective regularization without dataset-specific hyperparameter tuning.

2. Off-manifold coverage and response recovery. We acknowledge that the 12.08% versus 73.75% exceedance rates indicate incomplete recovery of the measured KO distributional shift. However, these rates quantify WT-support exceedance, not the proportion of biological response recovered. In matched mouse-lung comparisons, repeated projection increases the response norm ratio from 0.1037 to 0.4151, response alignment from 0.0166 to 0.0720, and AUPRC from 0.2401 to 0.2692. Correct-target edits also outperform wrong-gene and norm-matched random controls (AUPRC: 0.2693 vs. 0.1247 and 0.1536). Thus, STEP demonstrates target-specific response-prioritization gains beyond edit magnitude, although accurate recovery of large, signed perturbation responses remains challenging (as noted in our discussion section).

STEP's central contribution is a WT-trained framework integrating spatial representation learning, target-specific latent editing, and temporal propagation to prioritize perturbation-responsive genes from unpaired spatial snapshots. We respectfully ask that the final assessment consider these demonstrated methodological and empirical contributions alongside the acknowledged limitations.

Thank you for your time and participation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.