acceptodds
Under review as a conference paper at ICLR 2027

PerturbBridge: Reference-Process Design for Heterogeneity-Aware Single-Cell Perturbation Prediction

Abstract

Predicting single-cell perturbation responses requires transferring a known perturbation to new cellular contexts, where control and perturbed populations are measured separately and remain unpaired. Coupling-based prediction faces two connected challenges. First, standard transport costs can let heterogeneous response amplitudes dominate state correspondence. Second, endpoint optimal transport provides a static coupling but no process over intermediate states, so the coupling alone cannot ensure consistency with the paths required for training and generation. Changing the coupling cost while retaining a fixed bridge further separates the pairing geometry from the reference process used for training and generation. We therefore recast the problem as designing a reference process that jointly defines the coupling and supplies the intermediate states. We propose PerturbBridge, a heterogeneity-aware count reference whose two-parameter negative-binomial-difference family separates displacement scale from tail shape at fixed variance. Each member of the family induces a Schrödinger endpoint coupling and supplies its matching tractable integer bridge, so that coupling and generation remain reference-consistent as the geometry changes. Theoretically, we analyze two baseline states with different response amplitudes and establish conditions under which the reference preserves state correspondence regardless of response magnitude, and we connect this pairing preference to prediction error under composition shift. Empirically, PerturbBridge achieves strong overall rankings across 80 metrics spanning four expression views, ranking first among six learned models in all three PBMC settings. These results establish reference-process design as a principled approach to unpaired perturbation prediction under heterogeneous response amplitudes.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell perturbation prediction, Schrödinger bridges, optimal transport, count data, conditional generative modeling, out-of-distribution generalization, cell correspondence, single-cell RNA sequencing

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_iLZUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PerturbBridge predicts single-cell perturbation responses from unpaired populations using a negative-binomial-difference reference process. The reference jointly defines endpoint coupling and integer bridge generation, with separate controls for displacement scale and shape. Experiments cover genetic and cytokine perturbations, including transfer to unseen cellular contexts.

Critical strengths

The method is Well motivated. It addresses state-dependent response heterogeneity and provides targeted analysis, component ablations, and evaluation across multiple perturbation settings.

Critical weaknesses

Limited conceptual novelty. NB-difference processes, Schrödinger bridge decompositions, and integer count-bridge generation have established precedents. The contribution is primarily a reference-design extension, with a useful but restricted two-state analysis, rather than a substantially new theoretical framework.

Incomplete attribution of improvements. The closest Poisson-reference comparison is limited to a PBMC subset, and key component ablations use single training runs. A direct conditional generator using the same NB coupling is also missing, leaving the additional value of bridge generation insufficiently isolated.

Questions for authors

Appendix K fixes (\delta=1), giving (H_1(r)=|r|-1/2) for every nonzero integer residual. Why was this threshold chosen? Does the NB advantage persist when the Huber threshold and entropy temperature are selected using the same validation protocol, allowing curvature to extend across moderate count differences?

The authors select reference parameters using validation Sinkhorn-(W_2), whereas Table 7 shows different rankings for count transport error and response recovery. Could you report the validation scores and selected reference parameters for each main task, and clarify how the highlighted NB configuration and its Poisson comparator were chosen under this criterion?

6 comments
AuthorsOpenReview ↗

Response to Reviewer iLZU 1/4

Thank you for the specific questions about attribution, robust-cost tuning, and validation selection. They helped make the methodological contribution and the actual model-selection procedure more precise.

Q1. Methodological contribution and established components (Weakness 1)

Limited conceptual novelty. NB-difference processes, Schrödinger bridge decompositions, and integer count-bridge generation have established precedents. The contribution is primarily a reference-design extension, with a useful but restricted two-state analysis, rather than a substantially new theoretical framework.

A. We agree that the established mathematical components should be distinguished from the design capability introduced here. We build on established integer processes and the Schrödinger decomposition, with explicit attribution. The new design capability is to adjust the pairing geometry's shape at fixed displacement variance while retaining a tractable matching integer bridge throughout the family. A robust static cost alone specifies a coupling but does not supply the matching conditional path law. A fixed Poisson reference supplies a path law but ties its shape to its scale. Our construction connects these two choices through one family, and the correspondence analysis identifies why its bounded marginal penalties can protect state information from large legitimate responses. The controlled tests below assess this design capability directly.

AuthorsOpenReview ↗

Response to Reviewer iLZU 2/4

Q2. Attribution to the reference, bridge training, and generation (Weakness 2)

Incomplete attribution of improvements. The closest Poisson-reference comparison is limited to a PBMC subset, and key component ablations use single training runs. A direct conditional generator using the same NB coupling is also missing, leaving the additional value of bridge generation insufficiently isolated.

A. We agree that the reference, training path, and inference steps should be assessed separately. We independently train a direct conditional predictor with the same NB coupling, condition inputs, backbone capacity, training budget, ZINB endpoint likelihood, and validation-based checkpoint selection rule. It predicts a count distribution from a control cell and is a probabilistic model, not deterministic regression. It is trained at the control endpoint t=1, whereas NB is trained on bridge states across times. Direct prediction and one-step NB each use one network call, testing the training construction; NB 1/5/20/100-step inference reuses one checkpoint to test the inference procedure.

The table reports all path and step comparisons on the same 1,080 unseen-cell-type groups. Direct prediction and interpolation use one inference step; both Gaussian bridges use 100 steps with their respective Gaussian updates. All four NB step settings reuse the same checkpoint.

MethodW2 ↓MMD² ↓Pearson ↑Spearman ↑PDS ↑ES ↑
Direct conditional predictor96.83590.01710.25300.40210.67620.4921
Interpolation96.26100.01920.19900.30400.61820.3061
Conditional-moment Gaussian96.44190.01210.30750.46720.71940.5738
Fixed-scale Gaussian95.21650.01290.27620.42220.68230.5343
NB 1 step97.17030.01220.32050.47460.72360.5335
NB 5 steps94.21520.01040.33930.49680.71750.5552
NB 20 steps89.22460.00820.39900.54710.79250.5795
NB 100 steps89.16490.00940.39490.55250.78340.6231

One-step NB improves five of the six metrics, including MMD², Pearson, Spearman, PDS, and ES. With the checkpoint fixed, 20 steps further improve all six. Learning the same endpoint relation from multiple intermediate states couples overall response prediction with local response reconstruction; iterative sampling then applies this learned recovery along the matching path. Interpolation and Gaussian controls further distinguish the effects of the intermediate-state distribution. Matching the conditional variance improves all four response-oriented metrics over a fixed Gaussian scale, indicating that endpoint-dependent uncertainty contributes to learning useful response structure. The exact NB bridge combines this state-dependent uncertainty with a matching count-valued path and achieves the strongest overall performance among the 100-step variants.

Poisson, CellOT, and PerturbDiff now enter the main benchmark and its task-level ranks. PerturbBridge ranks first on all three PBMC tasks; on unseen types its mean rank is 1.6125, compared with 3.7625 for Poisson. The appendix table reports W2, MMD², Pearson, Spearman, PDS, and ES alongside the main figure’s ranks. We also independently trained the selected NB (κ,h)=(8,0.5) configuration with three additional seeds on the same subset. Their complete six-metric results are below; the mean and sample SD are calculated over these three additional runs.

Training seedW2 ↓MMD² ↓Pearson ↑Spearman ↑PDS ↑ES ↑
8888.17530.00940.39660.55090.78690.6732
202488.21680.00940.40210.56130.79710.6504
340788.48910.00980.38800.54280.78150.6097
Mean88.29370.00950.39560.55170.78850.6444
Sample SD0.17050.00030.00710.00920.00790.0322

Pearson remains within 0.3880–0.4021 and PDS within 0.7815–0.7971 across these runs, supporting the repeatability of response recovery for the selected configuration. Figure 5 also shows the range over the original run and these three additional runs. Section 4.3 and Appendix N.3, Tables 38–39 report the added path and seed experiments; Appendix N.6 gives all four-space results.

AuthorsOpenReview ↗

Response to Reviewer iLZU 3/4

Q3. Huber threshold and entropy-temperature selection (Question 1)

Appendix K fixes (\delta=1), giving (H_1(r)=|r|-1/2) for every nonzero integer residual. Why was this threshold chosen? Does the NB advantage persist when the Huber threshold and entropy temperature are selected using the same validation protocol, allowing curvature to extend across moderate count differences?

A. We agree with your observation that delta=1 places all nonzero integer residuals in the linear regime, so the original setting serves as a linear-tail control. The expanded comparison tests curvature across moderate count differences. We added thresholds {1,2,5,10,20} at tau=1, temperatures {10,30,100,300} at delta=1, and the crossed candidate (5,100). The same validation criterion selects (5,100), with score 95.0279. The ten candidates cover threshold changes, temperature changes, and a crossed setting.

The table compares the validation-selected Huber configuration with NB on the same subset. All Huber candidates retain the NB (8,0.5) bridge. Appendix N.4, Table 40 reports the validation scores and six key test metrics for all ten candidates; Appendix N.6, Tables 44, 47, 50 and 53 gives their complete metrics in the four expression spaces.

MethodValidation score ↓W2 ↓MMD² ↓Pearson ↑Spearman ↑PDS ↑ES ↑
Huber delta=5 tau=10095.027989.41850.01400.29990.43410.69300.5362
NB (8,0.5)89.172589.16490.00940.39490.55250.78340.6231

The threshold sigma in the experiment files is the Huber delta, which controls where the residual penalty changes from quadratic to linear. Tau controls the softness of the transport plan through exp(-cost/tau). They affect different aspects of supervision: delta changes which residual magnitudes are distinguished by curvature, while tau changes how strongly the resulting preferences concentrate transport mass. At tau=1, increasing delta from 1 to 5 raises Pearson/PDS from 0.2199/0.5865 to 0.3028/0.7006. At delta=1, tau=100 gives 0.2165/0.5912. This pattern identifies preserving moderate-residual geometry as a useful ingredient. The joint candidate (5,100) is selected by its lower validation score, 95.0279 versus 95.1946 for (5,1), and gives 0.2999/0.6930. NB achieves MMD²/Pearson/PDS of 0.0094/0.3949/0.7834. Its compact scale–shape family controls displacement variance and marginal penalties together; the known-state studies further verify the correspondence benefit of that geometry. The expanded Huber comparison is discussed in Section 4.3 and reported in Appendix N.4, Table 40.

Reviewer_aV4UOpenReview ↗
Rating
3
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies unpaired single-cell perturbation prediction, with an emphasis on transferring training-observed perturbations to new cellular contexts for which control cells are available. PerturbBridge introduces a negative-binomial-difference reference family that separates displacement variance from tail shape. The same reference defines both the entropic optimal-transport coupling used to construct pseudo-pairs and the integer-valued conditional bridges used for training and generation. A conditional ZINB endpoint denoiser is trained on these bridge states.

The theoretical analysis identifies conditions under which the proposed cost preserves state-consistent pairing despite heterogeneous response amplitudes in a two-state model, and connects pairing errors to prediction errors under composition shift. Experiments cover NADIG in-distribution prediction and three PBMC settings, including unseen cell types and donors. The method achieves the lowest raw-count Wasserstein-2 distance across the four tasks and the best aggregate rank among the six learned models in the three PBMC settings, with the strongest response-recovery gains on unseen cell types.

Critical strengths
  1. A coherent and technically grounded reference-process design. The main contribution goes beyond modifying an OT cost: the paper constructs a count-valued reference family with separate scale and shape controls and derives both its endpoint coupling and tractable conditional bridges. The Poisson-reference limit provides a clear connection to prior work, while the correspondence-stability analysis explains how bounded marginal penalties can protect state correspondence under heterogeneous responses. The discussion of reference consistency further clarifies why the pairing relation and the paths used to learn it should be considered jointly. This combination provides a substantive methodological contribution and useful insight into conditional supervision from unpaired populations.

  2. Convincing improvements on the most relevant context-transfer setting. The unseen-cell-type results show improvements in response recovery across multiple expression views. For example, raw-count response Pearson reaches 0.4405, compared with 0.2218 for the strongest alternative in Table 27, and the advantages in response Pearson, cosine PDS, and ES persist after normalization. The fixed-variance shape sweep and matched-versus-mismatched bridge comparisons provide additional evidence that the reference design matters. These results support the practical relevance of the proposed approach for transferring known perturbations across cellular contexts.

Critical weaknesses
  1. The evidence isolating the contribution of the new reference is limited in scope. The closest same-backbone alternative, the h = 0 Poisson reference, is evaluated in the 1,080-group sensitivity study rather than in the full main benchmark.
  2. Limited comparison with recent population-level perturbation models. The manuscript does not discuss or evaluate several relevant recent approaches, including PerturbDiff (Yuan et al., 2026), scDFM (Yu et al., 2026), and SCALE (Chen et al., 2026). These methods offer alternative approaches to modeling perturbation-response distributions through functional diffusion, distributional flow matching, or set-based conditional transport. Given the emphasis on recovering heterogeneous responses and transferring perturbation effects to unseen cellular contexts, comparisons with task-compatible representatives would strengthen the empirical assessment, particularly on the PBMC unseen-cell-type task.
Questions for authors
  1. Contribution of the proposed reference. Could the authors extend the same-backbone comparison with the h=0h=0 Poisson reference to the full main benchmark, particularly the complete PBMC unseen-cell-type task?
  2. Comparison with recent population-level models. Could the authors compare PerturbBridge with task-compatible representatives among PerturbDiff, scDFM, and SCALE on the same PBMC unseen-cell-type split?
5 comments
AuthorsOpenReview ↗

Response to Reviewer aV4U 1/2

Thank you for the detailed assessment and the suggestions to extend the closest-reference comparison and engage with recent population models. Both additions sharpen the empirical scope of the contribution.

Q1. Poisson reference on the complete benchmark (Question 1)

Contribution of the proposed reference. Could the authors extend the same-backbone comparison with the (h=0) Poisson reference to the full main benchmark, particularly the complete PBMC unseen-cell-type task?

A. Following your suggestion, we have incorporated the selected Poisson reference into all four main tasks and official CellOT and PerturbDiff into both complete PBMC OOD tasks. Below we reproduce the main figure's normalized relative scores for all nine learned models on the two OOD tasks. Each expression-space cell lists distribution / response / DE / perturbation-level family scores, in that order, with higher scores better. Each score averages direction-aligned, per-metric min–max values within its family, using all evaluated methods, including Control and Linear, for normalization. The last column is the mean rank over the 80 original metrics among learned models, as in the main figure; lower is better.

Task / methodRaw counts ↑Log counts ↑Normalized counts ↑Normalized log counts ↑Mean rank ↓
(OOD cell type) / PerturbBridge0.98 / 0.99 / 0.94 / 0.950.88 / 0.99 / 0.94 / 0.960.75 / 0.94 / 0.95 / 1.000.57 / 0.99 / 0.96 / 1.001.6125
(OOD cell type) / CellFlow0.95 / 0.73 / 0.32 / 0.080.85 / 0.73 / 0.29 / 0.100.63 / 0.69 / 0.26 / 0.150.36 / 0.73 / 0.28 / 0.165.7125
(OOD cell type) / STATE0.82 / 0.77 / 0.59 / 0.430.76 / 0.78 / 0.56 / 0.460.28 / 0.50 / 0.51 / 0.300.38 / 0.73 / 0.53 / 0.355.4750
(OOD cell type) / DOLORIS0.85 / 0.79 / 0.54 / 0.070.78 / 0.74 / 0.56 / 0.140.70 / 0.70 / 0.33 / 0.100.54 / 0.70 / 0.39 / 0.145.7125
(OOD cell type) / DEPARTURES0.88 / 0.78 / 0.67 / 0.370.64 / 0.69 / 0.69 / 0.280.94 / 0.72 / 0.52 / 0.210.66 / 0.70 / 0.52 / 0.304.9250
(OOD cell type) / Lingshu-Cell0.07 / 0.35 / 0.68 / 0.490.18 / 0.26 / 0.67 / 0.410.29 / 0.17 / 0.67 / 0.600.30 / 0.44 / 0.65 / 0.625.5500
(OOD cell type) / Poisson0.96 / 0.79 / 0.84 / 0.520.85 / 0.84 / 0.83 / 0.710.62 / 0.67 / 0.72 / 0.600.33 / 0.80 / 0.79 / 0.653.7625
(OOD cell type) / CellOT0.68 / 0.51 / 0.09 / 0.250.67 / 0.07 / 0.08 / 0.220.58 / 0.15 / 0.12 / 0.250.67 / 0.08 / 0.09 / 0.366.4500
(OOD cell type) / PerturbDiff0.86 / 0.29 / 0.53 / 0.380.83 / 0.72 / 0.51 / 0.420.51 / 0.41 / 0.37 / 0.270.35 / 0.70 / 0.46 / 0.265.8000
(OOD donor) / PerturbBridge0.97 / 0.86 / 0.73 / 0.840.77 / 0.92 / 0.70 / 1.000.71 / 0.79 / 0.78 / 0.830.43 / 0.91 / 0.82 / 0.903.0125
(OOD donor) / CellFlow0.97 / 0.76 / 0.25 / 0.110.74 / 0.75 / 0.22 / 0.100.77 / 0.67 / 0.25 / 0.160.46 / 0.75 / 0.26 / 0.145.2375
(OOD donor) / STATE0.75 / 0.80 / 0.47 / 0.610.71 / 0.85 / 0.46 / 0.640.29 / 0.69 / 0.48 / 0.310.46 / 0.83 / 0.54 / 0.345.7125
(OOD donor) / DOLORIS0.77 / 0.78 / 0.40 / 0.190.55 / 0.83 / 0.40 / 0.360.58 / 0.72 / 0.44 / 0.240.45 / 0.82 / 0.52 / 0.275.7875
(OOD donor) / DEPARTURES0.88 / 0.85 / 0.75 / 0.590.46 / 0.74 / 0.78 / 0.590.71 / 0.71 / 0.50 / 0.340.30 / 0.67 / 0.70 / 0.394.9750
(OOD donor) / Lingshu-Cell0.55 / 0.86 / 0.74 / 0.920.29 / 0.85 / 0.74 / 0.880.60 / 0.71 / 0.91 / 0.970.50 / 0.91 / 0.92 / 0.974.1875
(OOD donor) / Poisson0.37 / 0.37 / 0.69 / 0.780.72 / 0.92 / 0.68 / 0.910.70 / 0.79 / 0.77 / 0.740.42 / 0.90 / 0.83 / 0.804.3500
(OOD donor) / CellOT0.67 / 0.53 / 0.12 / 0.390.67 / 0.10 / 0.10 / 0.360.67 / 0.09 / 0.15 / 0.340.67 / 0.09 / 0.10 / 0.496.4375
(OOD donor) / PerturbDiff0.89 / 0.64 / 0.45 / 0.490.73 / 0.74 / 0.46 / 0.550.61 / 0.46 / 0.46 / 0.380.45 / 0.74 / 0.54 / 0.365.3000

PerturbBridge ranks first in all three PBMC tasks, with mean ranks 2.0625, 1.6125, and 3.0125 for ID, unseen types, and unseen donors. On unseen types, its response-family scores are 0.99/0.99/0.94/0.99 across the four spaces, compared with 0.79/0.84/0.67/0.80 for Poisson; its perturbation-level scores are 0.95/0.96/1.00/1.00, compared with 0.52/0.71/0.60/0.65. The advantage persists after expression normalization and spans both gene-level responses and perturbation-level structure. This is the role of useful state–response supervision: it supports a coherent response pattern across representations and evaluation targets. The overall ranks combine that evidence across the full metric panel. These expanded comparisons appear in Sections 4.1–4.2, Figure 3 and Appendix M, Table 20.

AuthorsOpenReview ↗

Response to Reviewer aV4U 2/2

Q2. Recent population models and task-compatible comparisons (Question 2)

Comparison with recent population-level models. Could the authors compare PerturbBridge with task-compatible representatives among PerturbDiff, scDFM, and SCALE on the same PBMC unseen-cell-type split?

A. We agree that recent population models are important related advances and should be represented in task-compatible empirical comparisons. We have added PerturbDiff, scDFM, and SCALE to Related Work (Section 2). They offer valuable ways to model heterogeneous populations and reduce dependence on assumed cell correspondences. PerturbDiff's reported unseen-perturbation results are especially relevant to the broader goal of virtual-cell generalization. scDFM combines control-conditioned flow matching and population MMD supervision; SCALE develops set-aware endpoint transport. These approaches incorporate cellular context through different representations and learning objectives.

We evaluated the official PerturbDiff implementation on both complete PBMC OOD splits under the same data and conditioning protocol as our baseline panel; official CellOT is included as well. The comparison uses no additional external pretraining data; the protocol and results are included in Appendix N.2 and Figure 3.

The distinction between generalizing to a new perturbation and to a new cellular context is central here. The latter requires extrapolating how a known perturbation depends on cellular state. Our split contains ten training types and eight held-out types: millions of cells do not make these context identities densely covered. Matching training marginals can leave multiple state–response relations compatible with the observations; the reference supplies a low-dimensional restriction on that ambiguity. This motivates using the measured expression of each control cell to constrain the response relation when context coverage is sparse.

Population modeling is an important direction for virtual cells: it learns from the form in which experimental supervision is naturally available and can represent heterogeneous responses at the population level. Strong generalization across perturbations makes these approaches particularly valuable. Our cross-cell-type setting highlights a complementary data regime: many cells are observed, but only ten training cell-type identities cover the contexts to which response laws must generalize. An explicit expression-conditioned reference makes use of this fine-grained supervision in that regime. As context coverage grows, population objectives may learn these relations with progressively less need for prescribed pairing. This scaling perspective connects the two approaches and motivates their joint study.

Area_Chair_BWgnOpenReview ↗

AC Request

Dear Reviewer,

Thank you very much for your time and feedback. I would greatly appreciate if you could confirm if the rebuttal addresses your concerns and let them know about any remaining comments. I would appreciate if you could avoid using generative AI (not even for editing) in your response to the rebuttal.

Best regards,

AC

Reviewer_zSVkOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

PerturbBridge defines a negative-binomial-difference reference process whose scale and tail shape can vary separately. Its endpoint kernel supplies a Schrödinger coupling, while its conditional paths provide tractable integer bridges for training and generation. A two-state analysis connects robust pairing to composition-shift error. Experiments on NADIG and PBMC, including unseen cell types and donors, evaluate prediction and isolate reference shape and coupling–bridge matching.

Critical strengths
  1. The method has a coherent and economical organizing principle. One reference law determines both pairing and bridge dynamics. Shape control at fixed variance offers a meaningful extension of the Poisson reference, rather than an unrelated addition to an existing generator. The tractable sampler and correspondence analysis explain the design.
  2. The ablations test the proposed principle directly. Shape variation improves response recovery beyond the tested Poisson scale sweep, and matched versus mismatched coupling–bridge comparisons support joint reference design (Figure 5; Appendix J.1). Strong unseen-cell-type response results across normalized and unnormalized views provide evidence beyond raw-count transport scores.
Critical weaknesses
  1. The biological correspondence mechanism is less established than predictive performance. Theorem 3.2 concerns two fixed states under composition shift, whereas the real transfer tasks can change state-specific responses. The selected shape-changing populations do not supply known cell correspondences. Consequently, better biological state matching remains a plausible explanation rather than an experimentally demonstrated mechanism.
  2. The need for multistep bridge generation is not isolated from the selected coupling. The paper motivates direct conditional likelihood on pseudo-pairs (§3.1; Appendix F.1), but does not compare that simpler predictor with the full bridge under matched capacity and endpoint likelihood. Existing bridge-shape ablations establish a preference within the bridge framework, leaving its added value over directly learning the chosen conditional relation unresolved.
Questions for authors
  1. Assessment-critical evidence: In a controlled population with known state correspondences and varying response amplitudes/composition, does the proposed reference improve pairing accuracy and subsequent transfer error together?
  2. Assessment-critical comparison: With the selected pseudo-pairs and comparable capacity fixed, how does a direct conditional count predictor perform relative to multistep PerturbBridge?
3 comments
AuthorsOpenReview ↗

Response to Reviewer zSVk 1/2

Thank you for identifying the two assessment-critical questions. They distinguish predictive success from evidence about the proposed correspondence mechanism, and motivated separate controlled tests of correspondence and of bridge generation.

Q1. Experimental evaluation of state correspondence (Question 1)

Assessment-critical evidence: In a controlled population with known state correspondences and varying response amplitudes/composition, does the proposed reference improve pairing accuracy and subsequent transfer error together?

A. We agree that directly evaluating correspondence is necessary to connect the pairing mechanism to predictive performance. We use two approaches, synthetic and semi-synthetic populations, with known states to measure correct-state transport mass. These experiments vary competing state counts and response heterogeneity at balanced state composition; Theorem 3.2 analyzes the predictive consequence of composition changes, and the OOD benchmark evaluates real context transfer.

In the synthetic study, each population contains 4–32 baseline states with 128 expression features. Cells share a state-specific expression background and vary around it. A perturbation has a common direction for each feature but different response magnitudes across states; control and treated observations are generated and shuffled separately. In the semi-synthetic study, real PBS cells from CD4 Memory, CD8 Memory, and NK provide expression prototypes. We aggregate 2,048 genes into 64 expression-based groups, use each prototype to define a state, apply a state-dependent multiplicative response, and independently sample members at the two endpoints. Thus the background expression comes from measured cells while state identity and response heterogeneity remain controlled. Increasing the number of states creates more competing matches; increasing factor f increases the between-state log-response standard deviation.

The table shows correct-state transport mass (%) across state counts and heterogeneity levels. Entries are means ± sample SD over five seeds, after averaging two test panels within each seed.

PopulationStatesfNBTuned HuberSquared L2Euclidean
Synthetic4187.70 ± 0.8477.18 ± 1.4660.45 ± 0.9960.55 ± 1.09
Synthetic8176.55 ± 0.6060.79 ± 0.4941.95 ± 0.5242.12 ± 0.92
Synthetic16164.60 ± 0.6646.43 ± 0.5028.23 ± 0.5728.23 ± 0.60
Synthetic320.598.80 ± 0.0774.72 ± 0.6349.80 ± 0.2749.23 ± 0.46
Synthetic32152.99 ± 0.4334.51 ± 0.2719.06 ± 0.3518.86 ± 0.26
CD4 Memory81.583.33 ± 1.5167.34 ± 2.5746.48 ± 2.2829.74 ± 2.01
CD4 Memory161.571.72 ± 1.6053.53 ± 1.4130.20 ± 1.1419.08 ± 1.07
CD4 Memory320.588.30 ± 0.8067.85 ± 0.8139.46 ± 0.9428.01 ± 1.06
CD4 Memory321.559.02 ± 0.9740.24 ± 0.9419.95 ± 0.9312.15 ± 0.55
CD8 Memory81.581.02 ± 1.7767.07 ± 2.8242.69 ± 1.2626.95 ± 0.76
CD8 Memory161.569.44 ± 1.3348.88 ± 1.6628.75 ± 0.8917.06 ± 0.98
CD8 Memory320.584.34 ± 1.5262.98 ± 1.2534.99 ± 0.7423.94 ± 0.82
CD8 Memory321.555.53 ± 2.1037.60 ± 1.3818.06 ± 0.9110.54 ± 0.69
NK81.576.01 ± 1.7265.35 ± 2.3143.90 ± 1.7027.76 ± 1.01
NK161.563.28 ± 1.1247.98 ± 0.7729.55 ± 1.1717.57 ± 0.60
NK320.581.15 ± 0.7361.82 ± 1.0138.35 ± 0.5826.47 ± 0.29
NK321.550.08 ± 0.8334.54 ± 1.1519.22 ± 0.7210.95 ± 0.34

Each cost family is selected on the same validation populations and then fixed for testing; state labels enter the scoring criterion. NB has the highest mean against L1, Huber, squared L2, and Euclidean costs in all 12 synthetic and 36 semi-synthetic conditions. At the largest tested state count and heterogeneity, NB versus tuned Huber is 52.99/34.51 in the synthetic study and 59.02/40.24, 55.53/37.60, and 50.08/34.54 for the three semi-synthetic types. The advantage therefore persists as both the number of competing states and response variation increase. Section 4.3, Table 2 summarizes the semi-synthetic results; Appendix N.5, Tables 41–42 reports the complete grids.

These results explain why correspondence matters for transfer. In a balanced training population, exchanging the responses assigned to two states can leave the treated marginal unchanged, even though the conditional relation is wrong. When their frequencies change, those assignment errors receive different weights and become visible in the predicted marginal. Theorem 3.2 quantifies this mechanism under fixed state-specific responses. The controlled studies show that the proposed geometry preserves state identity across many-state, heterogeneous populations; the real OOD tasks show that the resulting conditional predictor transfers to new cellular contexts. Together, they support the central claim that preserving useful state–response supervision is consequential for transfer from unpaired populations.

AuthorsOpenReview ↗

Response to Reviewer zSVk 2/2

Q2. Direct prediction and bridge generation with fixed coupling (Question 2)

Assessment-critical comparison: With the selected pseudo-pairs and comparable capacity fixed, how does a direct conditional count predictor perform relative to multistep PerturbBridge?

A. Following your suggestion, we added a fixed-coupling comparison to identify what the bridge contributes beyond direct endpoint learning. The direct predictor is trained exclusively from the control endpoint (t=1); the NB model learns endpoint recovery from bridge states across times. Both direct prediction and one-step NB start from a control cell and use one network call to predict a ZINB distribution, isolating the training construction. Different NB step counts then reuse the same checkpoint to examine iterative generation.

We independently train the direct conditional predictor with the same NB coupling, condition inputs, backbone capacity, training budget, ZINB endpoint likelihood, and validation-based checkpoint selection rule. It predicts a count distribution from a control cell and is a probabilistic model, not deterministic regression. On the 1,080-group subset:

MethodW2 ↓MMD² ↓Pearson ↑Spearman ↑PDS ↑ES ↑
Direct conditional predictor96.83590.01710.25300.40210.67620.4921
NB 1 step97.17030.01220.32050.47460.72360.5335
NB 20 steps89.22460.00820.39900.54710.79250.5795
NB 100 steps89.16490.00940.39490.55250.78340.6231

The one-step NB model improves five of the six metrics over the directly trained predictor, identifying the benefit of intermediate-state supervision. Each selected pseudo-pair supplies a family of endpoint-recovery tasks with different prediction distances and uncertainties. The shared denoiser learns the same response relation across those states, strengthening the one-step control-to-response prediction. The fixed-checkpoint step sweep then shows how repeatedly applying that learned denoiser improves generation further.

Twenty steps already give strong distribution and response recovery; 100 steps provide additional improvement in ES. Section 4.3 and Appendix N.3 analyze how training paths and iterative generation contribute, alongside Gaussian controls and additional NB training seeds.

AuthorsOpenReview ↗

Thank you for your time and feedback

We appreciate how clearly you identified the two questions that mattered most to your assessment. Do the known-state correspondence experiments and the comparison with a direct predictor using the same coupling address those questions? If the additional evidence gives you greater confidence in the paper, we would be grateful if you would consider updating your rating. Thank you again for helping us distinguish what the pairing achieves from what the bridge contributes.

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

This paper proposes PerturbBridge, a framework for predicting single-cell perturbation responses from unpaired control and treated populations. The central contribution is a negative-binomial-difference reference process that jointly recovers couplings and generates stochastic intermediate states that are used for conditional generation. The authors argue that this construction improves single-cell response prediction under heterogeneous perturbation, including transfer to unseen cell types and donors.

Critical strengths

The paper provides a principled math that connects endpoint coupling and intermediate-state generation through a reference process. The proposed reference independently controls displacement variance and tail shape. The theoretical results show how squared Euclidean transport may lead to incorrect correspondences. The paper shows that PerturbBridge achieves strong performance when transferring perturbation responses to unseen cell types.

Critical weaknesses
  1. The proposed method samples intermediate states from a pre-defined reference bridge conditioned on control and treated endpoints, then trains a neural network to predict the treated endpoint from these states. However, the intermediate states are generated using the observed endpoints, and there is no additional biological information beyond the original training data. I would suggest adding an ablation study to compare the proposed model with an identical conditional neural network on the same couplings without the reference bridge. Also, including CellOT as a baseline would strengthen the empirical evidence.

  2. The reference displacement distribution K_\Delta^{\kappa,h}(d) depends on the displacement and the selected parameters (\kappa,h), and ignores the perturbation identity or cellular context. Therefore, two drugs that induce the same displacement receive the same reference bridge, even if they have fundamentally different biological mechanisms. Moreover, the proposed neural network is conditioned on perturbations and learns responses, so why is a reference distribution needed?

  3. The reference parameters are selected using validation data. In the OOD cell-type experiment, validation cells come from the training cell types, whereas evaluation involves entirely unseen cell types. It is not clear to me whether the selected reference geometry remains useful when the underlying response distribution changes.

Questions for authors

Please address the questions in the weaknesses section.

7 comments
AuthorsOpenReview ↗

Response to Reviewer yBCC 1/3

Thank you for the constructive suggestion to hold the coupling and predictor capacity fixed. This comparison clarifies that the benefit of a bridge has both a training component and an inference component.

Q1. Bridge contribution and CellOT (Weakness 1)

The proposed method samples intermediate states from a pre-defined reference bridge conditioned on control and treated endpoints, then trains a neural network to predict the treated endpoint from these states. However, the intermediate states are generated using the observed endpoints, and there is no additional biological information beyond the original training data. I would suggest adding an ablation study to compare the proposed model with an identical conditional neural network on the same couplings without the reference bridge. Also, including CellOT as a baseline would strengthen the empirical evidence.

A. We agree that fixing the coupling and capacity is essential for identifying the bridge contribution. Following your suggestion, we independently trained a direct conditional predictor and evaluated official CellOT. The predictor outputs a count distribution from a control cell: it is probabilistic, not deterministic regression. We added a fixed-checkpoint step sweep to separate training and inference benefits, plus interpolation and Gaussian controls to examine the intermediate-state distribution.

Direct prediction and one-step NB differ in training. The direct predictor trains only at the control endpoint (t=1, x_t=u); NB learns the same treated endpoint from bridge states sampled across times. At one-step inference, both start from a control cell and make one network call to predict a ZINB distribution. NB 1/5/20/100-step results reuse one checkpoint. All training variants share the NB coupling, condition inputs, backbone capacity, training budget, ZINB endpoint likelihood, and validation-based checkpoint selection rule, and are evaluated on the same 1,080 unseen-type groups.

MethodW2 ↓MMD² ↓Pearson ↑Spearman ↑PDS ↑ES ↑
Direct conditional predictor96.83590.01710.25300.40210.67620.4921
Interpolation96.26100.01920.19900.30400.61820.3061
Conditional-moment Gaussian96.44190.01210.30750.46720.71940.5738
Fixed-scale Gaussian95.21650.01290.27620.42220.68230.5343
NB 1 step97.17030.01220.32050.47460.72360.5335
NB 5 steps94.21520.01040.33930.49680.71750.5552
NB 20 steps89.22460.00820.39900.54710.79250.5795
NB 100 steps89.16490.00940.39490.55250.78340.6231

Even with one-step inference, NB improves five of six metrics: MMD² falls from 0.0171 to 0.0122, while Pearson, Spearman, PDS, and ES increase. This identifies a training benefit. With the same NB checkpoint, 20 steps further improve all six metrics over one step, identifying an additional inference benefit.

The key difference is how each pseudo-pair is used. Direct prediction learns the endpoint distribution from the control state. Bridge training presents the same target through states with different information and uncertainty, from the control end to states closer to the target. Shared parameters learn both the overall response and its recovery across these states. Local patterns learned near the treated endpoint can therefore help constrain the harder control-to-response task.

Iterative generation repeatedly applies this denoiser at the intermediate states covered in training. Each endpoint prediction guides a bridge update, whose output informs the next prediction. Training and inference thus organize response reconstruction using the same reference geometry; the fixed-checkpoint sweep supports this complementary benefit.

The Gaussian controls clarify the role of intermediate-state uncertainty. A fixed Gaussian scale gives every endpoint pair the same noise magnitude at a given time. Matching NB conditional variance adapts uncertainty to the endpoints and improves Pearson, Spearman, PDS, and ES. The coupling selects the endpoint association; the bridge determines how its uncertainty is presented during learning. The exact NB bridge performs best overall among the 100-step path variants. Conditional-moment approximations offer a direction for studying which path statistics suffice for efficient learning.

CellOT comparison. Under the common baseline protocol, official CellOT/PerturbBridge Pearson is 0.0693/0.4405 on complete unseen-type tasks and 0.0883/0.2192 on complete unseen-donor tasks. The revised main figure includes updated 80-metric ranks; the appendix reports six key metrics. The added evidence and analysis appear in Section 4.3 and Appendices N.2–N.3.

AuthorsOpenReview ↗

Response to Reviewer yBCC 2/3

Q2. Reference geometry and perturbation conditioning (Weakness 2)

The reference displacement distribution K_\Delta^{\kappa,h}(d) depends on the displacement and the selected parameters (\kappa,h), and ignores the perturbation identity or cellular context. Therefore, two drugs that induce the same displacement receive the same reference bridge, even if they have fundamentally different biological mechanisms. Moreover, the proposed neural network is conditioned on perturbations and learns responses, so why is a reference distribution needed?

A. Your comment highlights the importance of perturbation-specific information for response prediction. It is present in the observed treated endpoints and in the network’s explicit perturbation condition. The reference family and selected global parameters (κ,h) are shared across contexts within each task, while each training bridge law R(· | Y=y, X=x) is conditioned on its observed endpoint pair. The shared geometry therefore produces endpoint-dependent intermediate states. At inference, an endpoint proposed by the conditional network and the current state determine the next bridge update.

The two components solve different parts of the unpaired problem. The reference selects which treated examples supervise each control state; the conditional network learns the resulting perturbation- and context-specific endpoint distribution. Conditioning the network on a drug does not by itself specify this missing state–response association. A shared reference can therefore assign the same displacement penalty to two drugs while their learned endpoint distributions differ. The bridge is a computational path, not an inferred biochemical trajectory. We make this distinction explicit in Section 3.1.

AuthorsOpenReview ↗

Response to Reviewer yBCC 3/3

Q3. Value of source-selected geometry under OOD shift (Weakness 3)

The reference parameters are selected using validation data. In the OOD cell-type experiment, validation cells come from the training cell types, whereas evaluation involves entirely unseen cell types. It is not clear to me whether the selected reference geometry remains useful when the underlying response distribution changes.

A. We agree that source-type validation gains should be explained together with the reference’s role across contexts. The selected geometry governs how state–response supervision balances large response displacements against baseline-state differences; the conditional network learns the specific response from control expression, perturbation, and context. A change in the response distribution need not invalidate the geometric principle used to construct supervision. The reference constrains how responses are associated with states; the network uses that supervision to learn how the response depends on the measured cellular context.

Theorem 3.2 gives a concrete condition for this mechanism. In its two-state construction, m and ℓ are baseline-state coordinate differences and M is the larger response amplitude. The swapped-minus-consistent cost gap for squared Euclidean pairing is 2m²−2Mℓ, which favors a swap as M grows. For the h=1/2 NB reference, the bounded marginal penalty of its cost ρ yields

Δρ≥δstate:=2ρ(m)−ρ(ℓ)−ηℓ,η=arcosh⁡(1+1/κ).\Delta_\rho\geq\delta_{\mathrm{state}}:=2\rho(m)-\rho(\ell)-\eta\ell,\qquad \eta=\operatorname{arcosh}(1+1/\kappa).

When state separation gives δ_state>0, this bound holds for every M≥ℓ. It identifies a pairing preference stable to changes in response amplitude: large legitimate responses remain in the treated population while their influence on correspondence selection is bounded. Synthetic and semi-synthetic studies examine this property across many states and heterogeneous responses; the complete OOD task tests its predictive value under real context shifts.

Useful geometry helps the network learn more informative state–response associations. If coupling assigns a large but valid response to the wrong baseline state, the endpoint likelihood rewards fitting that incorrect association. Preserving correspondence instead makes the training targets more consistent with their control-cell context. A conditional predictor can then learn response dependence on expression rather than absorb systematic state-assignment errors. This connects the cost-gap argument to learning: the geometry acts on the quality of conditional supervision, so its value can persist across contexts even when their response distributions differ. The many-state correspondence tests and OOD prediction results examine these two links empirically.

Selection protocol. The model is trained on cells from ten source cell types and evaluated for control-to-perturbed prediction on held-out cells from those same types. Training and validation share cell-type identities but have disjoint cells. Both the checkpoint and (κ,h) are selected with this source-domain validation score, then the model and reference configuration are fixed for unseen-type evaluation using controls from the target context. Treated cells from all eight target types are excluded from training and selection.

The Poisson stage selects (κ,h)=(8,0), and the shape stage selects (8,0.5), reducing validation score from 95.3923 to 89.1725. On the complete 8,545-group unseen-type task, the selected Poisson/NB response Pearson is 0.1745/0.4405 and PDS is 0.6170/0.8012. Thus the reference design selected on source types retains predictive value on target types. Appendix N.1 specifies selection, and Figure 3 and Appendix M report the complete-task comparison.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Revisions and Responses

We thank the reviewers for recognizing the coherence of the reference-process construction and for proposing concrete tests of its contribution. Their suggestions helped us separate three questions: which state–response relation the coupling selects, how intermediate-state supervision helps learn that relation, and what iterative generation adds at inference.

The revision adds a direct conditional predictor with the same NB coupling; deterministic and Gaussian path controls; an inference-step sweep; full-benchmark Poisson references; expanded Huber tuning; three additional NB seeds; CellOT and PerturbDiff on both complete PBMC OOD tasks; and synthetic/semi-synthetic tests with known state correspondence. We also specify the staged reference search and the stratified validation objective. New and revised manuscript text is blue. The main benchmark now includes Poisson on all four tasks and official CellOT and PerturbDiff on both OOD tasks, with recomputed task-level ranks.

AuthorsOpenReview ↗

Discussion update

We thank Reviewer yBCC for asking whether a reference geometry selected on one cell type preserves state correspondence on another. In a semi-synthetic experiment with known states, we selected each method’s parameters on one cell type, froze them, and tested all three cell types with newly sampled perturbations. We repeated this with each cell type as the selection type. Across all 36 settings, the NB reference assigned 11.9–28.9 percentage points more transport mass to the correct state than the best alternative. Appendix N.7, Table 55 and our follow-up to yBCC report the full results. We appreciate the reviewers’ suggestions and the Area Chair’s time in considering this additional evidence.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.