acceptodds
Under review as a conference paper at ICLR 2027

Constructive CytoBridge: Simulation-Free Stochastic Control of Interacting Particles

Abstract

Learning continuous cell-state dynamics from unpaired single-cell snapshots is essential for understanding cellular transitions and interactions. Achieving this without trajectory simulation becomes particularly challenging in the presence of pairwise cell–cell interactions, because they couple endpoint matching with the conditional paths connecting matched cells. Here we introduce Constructive CytoBridge (CCB), a simulation-free framework for stochastic optimal control with prescribed external and pairwise potentials. By conditioning on endpoint pairs and decomposing the control cost, CCB reveals an entropy-regularized quadratic OT structure in stochastic optimal control problems with pairwise interactions. We solve the resulting coupled problem by alternating KL-proximal transport updates with optimization of neural-network-parameterized conditional paths, and then regress a velocity field and a score function to support both deterministic and stochastic inference without requiring trajectory simulation during training. Experiments on synthetic and real datasets show that CCB accurately recovers the optimal control law and reconstructs marginal distributions, while incorporating external and pairwise potentials improves interpolation at unobserved time points.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
stochastic optimal control, optimal transport, single-cell dynamics

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_A4pJOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

CCB reformulates stochastic control with prescribed external and pairwise potentials as joint optimization of endpoint coupling and conditional paths. It alternates KL-proximal transport updates with learning Gaussian conditional paths, then regresses velocity and score for PF-ODE or SDE inference. Experiments cover analytic control, observed-marginal reconstruction, and held-out-time interpolation.

Critical strengths

The paper has certain novelties.

Critical weaknesses
  1. Proposition 2 concerns exact transport updates with fixed paths and full support; it does not establish convergence of the sparse joint implementation. Monte Carlo costs, regression, and numerical integration introduce additional errors. The analytic control benchmarks have Gaussian optima (Appendix B.3.1), limiting evidence on approximation in general control problems. KS/Coulomb test non-Gaussian interpolation but do not supply an optimal-control reference.

  2. Evidence for potential-dependent benefits is limited. Table 2 evaluates observed marginals and lacks CCB without potentials, so it does not isolate interaction benefits. In Coulomb, external-only accounts for most of the W1 gain: 0.217 without potentials, 0.083 external-only, and 0.079 full (Table 3). The spatial prior worsens spatial W1 at t = 3 from 0.2472 to 0.3283 (Table 19), although expression and neighborhood metrics improve. Prior sensitivity and training-seed uncertainty are not reported for the spatial result.

Questions for authors
  1. Could you quantify action and marginal errors against a small dense reference, including a non-Gaussian optimal-control problem?

  2. Could you add CCB without potentials to Table 2 and evaluate the potential choices at a held-out time?

  3. Clarification: How was the spatial penalty strength of 7 selected, and what explains the spatial-W1 deterioration at t = 3?

9 comments

We thank the reviewer for the constructive comments. We address all of your questions below, with clarifications and additional experiments in the revised manuscript.

[W1, Q1]

Could you quantify action and marginal errors against a small dense reference, including a non-Gaussian optimal-control problem?

We thank the reviewer for suggesting a comparison of the Dense and Sparse implementations and quantification of Action and marginal errors on a non-Gaussian optimal-control problem. We have added Dense results on the GM10D benchmarks to Section 4.1 and Table 1, with the ground-truth construction and coupling and conditional-path error analyses in Appendices B.3.1 and D.1. Below, we directly compare the Dense and Sparse implementations on the same GM10D benchmarks.

1. Both CCB Dense and Sparse achieve high accuracy on non-Gaussian optimal-control problems. GM10D has three Gaussian clusters in the initial distribution and four in the terminal distribution, covering linear and quadratic external potentials and quadratic attractive and repulsive interactions. The data construction and computation of the optimal control law are described in Appendix B.3.1. The table below compares CCB Dense, CCB Sparse, and other methods on the same problems, reporting Action MAE, control-law MSE, and terminal W1W_1.

PotentialMethodAction MAE ↓Control-law MSE (×10−1\times10^{-1}) ↓Terminal W1W_1 ↓
Linear external——
Modified WLF0.08892.15130.9794
Modified CytoBridge0.91612.83130.5269
GSBM0.48251.46210.7110
CCB Dense0.06340.21030.5758
CCB Sparse0.05230.23780.5668
Quadratic external——
Modified WLF4.629018.38381.1744
Modified CytoBridge3.53866.61250.5285
GSBM2.50027.66820.7522
CCB Dense0.16710.37270.5474
CCB Sparse0.12830.41370.5442
Quadratic attraction——
Modified WLF2.29698.31731.2453
Modified CytoBridge3.28166.21230.5380
GSBM2.56247.52900.7180
CCB Dense0.16300.41810.5474
CCB Sparse0.14520.43140.5479
Quadratic repulsion——
Modified WLF3.99104.99551.3853
Modified CytoBridge0.77392.82390.5258
GSBM0.73291.52210.7085
CCB Dense0.10230.19120.5994
CCB Sparse0.10200.20410.5964

CCB Dense achieves lower Action MAE and control-law MSE than all comparison methods under all four potentials. Sparse achieves slightly lower Action MAE than Dense in every case. Its control-law MSE increases slightly but remains substantially below the other methods, preserving Dense's accuracy advantage on multimodal non-Gaussian problems and maintaining accurate recovery of the optimal control law while improving efficiency.

2. The appendix further quantifies reconstruction errors in the endpoint coupling and conditional paths separately. Appendix D.1, Figures 6–7 compare CCB's learned endpoint coupling and conditional bridges with the GM10D ground truth under all four potentials.

Endpoint coupling error. Figure 6 shows mass allocation from three source peaks to four target peaks. Each heatmap cell corresponds to a source–target peak pair, and errors are reported in percentage points of total mass. Under all four potentials, the absolute mass-allocation error for every pair is less than one percentage point.

Conditional-path error. For each potential, Figure 7 uses the same 100 endpoint pairs drawn from the final CCB coupling to compare the means and standard deviations of CCB and ground-truth conditional bridges. Except for the conditional mean under quadratic attraction, the absolute values of all displayed conditional mean and standard deviation errors are below 1%.

[W2, Q2]

Could you add CCB without potentials to Table 2, evaluate the potential choices at a held-out time, and report the missing spatial training-seed uncertainty?

We thank the reviewer for pointing out the lack of clarity between the experimental purpose and the contribution of potentials. We apologize that the original manuscript did not clearly distinguish generation quality with potentials from the improvement that potentials bring to held-out interpolation. We have rewritten the motivation of Experiment 2, added no-potential CCB, and introduced potential ablations on the HMF Model and held-out-time interpolation experiments using estimated potentials.

1. Experiment 2 tests whether generation quality can be maintained when conditional paths become curved, rather than whether potentials improve generation. Reference [1] shows that straightening flow trajectories enables accurate generation with fewer discretization steps; nearly straight flows can even achieve high generation quality with a single Euler step. Potentials curve the conditional mean paths, so we ask whether CCB can maintain high generation quality when potentials introduce path curvature. Experiment 2 uses attractive and repulsive interactions as controlled settings to evaluate reconstruction of observed marginal distributions; these potentials are not biological priors prescribed for the expression datasets. The table below lists the results from Table 2 without potentials and with attractive or repulsive potentials, measured by W1W_1 averaged over observed times:

CCB setting / Inference modeEMT 10DEB 5DEB 50DEB 100DCITE 50DWeinreb 50D
Without potentials / PF-ODE0.1801.4837.63810.97919.9045.141
Without potentials / SDE0.193 ± 0.00091.420 ± 0.00737.831 ± 0.012711.127 ± 0.015420.284 ± 0.01335.376 ± 0.0247
Attractive potential / PF-ODE0.1781.5507.85211.04921.9255.171
Attractive potential / SDE0.193 ± 0.00081.452 ± 0.01507.723 ± 0.017011.283 ± 0.011022.342 ± 0.03205.409 ± 0.0075
Repulsive potential / PF-ODE0.1821.5477.79910.95621.9645.166
Repulsive potential / SDE0.197 ± 0.00181.446 ± 0.01407.758 ± 0.018010.842 ± 0.027022.031 ± 0.02805.339 ± 0.0140

The results show that CCB maintains marginal reconstruction quality close to its no-potential version even when potentials curve the conditional paths.

2. The newly added HMF Model experiment shows that interaction potentials alone can substantially improve held-out interpolation.

First, the small improvement in empirical W1W_1 on Coulomb should be interpreted alongside cluster-scale recovery. Figure 2 shows that the external potential primarily constrains cluster locations, while the interaction potential controls their spread; removing the interaction potential produces overly contracted clusters. With finite samples in high dimensions, even two point clouds drawn from the same distribution can have a large matching distance, and a contracted point cloud may instead obtain a lower empirical W1W_1. To test this effect, we independently draw 512 points each from N(0,I20)\mathcal N(0,I_{20}) and N(0,αI20)\mathcal N(0,\alpha I_{20}), compute empirical W1W_1, and repeat the experiment 20 times:

Covariance ratio α\alphaEmpirical W1W_1, mean ± standard deviation
1.04.016 ± 0.022
0.53.614 ± 0.024
0.33.575 ± 0.027

The population W1W_1 is zero at α=1\alpha=1, but empirical W1W_1 at α=0.3\alpha=0.3 is lower in all 20 repetitions. This shows that empirical W1W_1 with finite samples in high dimensions can favor overly contracted distributions, thereby underestimating the benefit of interaction potentials in recovering the correct cluster scale.

To further directly assess the contribution of interaction potentials, we have added a 20D HMF Model dataset. Its particles evolve under external and pairwise interaction potentials according to second-order dynamics. All methods are trained using only the position distributions at t=0t=0 and t=1t=1 and evaluated at t=0.25,0.5,0.75t=0.25,0.5,0.75, which are not used in training. The data construction is described in Appendix B.3.2, and the following results have been added to Table 3.

Method / Inference modeW1↓W_1\downarrowMMD2;(×10−3)↓\mathrm{MMD}^2;(\times10^{-3})\downarrow
OT-CFM0.506 ± 0.00123.972 ± 0.017
WFR-FM0.510 ± 0.00113.979 ± 0.010
SF²M0.520 ± 0.00243.961 ± 0.037
VGFM0.958 ± 0.006210.310 ± 0.090
DeepRUOT1.433 ± 0.029314.793 ± 0.185
WLF (OT)0.578 ± 0.00853.606 ± 0.117
Modified WLF (with potentials)0.825 ± 0.10102.374 ± 1.026
Modified CytoBridge (with potentials)1.114 ± 0.005212.743 ± 0.257
GSBM (without potentials)1.304 ± 0.02096.292 ± 0.394
GSBM (with potentials)1.251 ± 0.01573.731 ± 0.092
CCB (without potentials, PF-ODE)0.506 ± 0.00243.966 ± 0.035
CCB (without potentials, SDE)0.523 ± 0.000143.996 ± 0.004
CCB (external potential only, PF-ODE)0.471 ± 0.00273.433 ± 0.038
CCB (interaction potential only, PF-ODE)0.063 ± 0.00330.040 ± 0.006
CCB (full model, PF-ODE)0.049 ± 0.00280.011 ± 0.004
CCB (full model, SDE)0.127 ± 0.00090.012 ± 0.005

Full CCB outperforms all comparison methods on both metrics. More importantly, the ablation isolates the contribution of the interaction potential: adding only the external potential reduces W1W_1 from 0.5060.506 to 0.4710.471, whereas adding only the interaction potential reduces W1W_1 to 0.0630.063, a reduction of approximately 87.5%, and MMD2\mathrm{MMD}^2 from 3.966×10−33.966\times10^{-3} to 0.040×10−30.040\times10^{-3}, a reduction of approximately 99.0%. Adding the external potential as well further reduces the errors to 0.0490.049 and 0.011×10−30.011\times10^{-3}. Thus, the interaction potential is the main source of interpolation improvement on the HMF Model. This complements the Coulomb results, where the external potential contributes more, and demonstrates the roles of the two potentials in different dynamics.

Reviewer_h2vNOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper authors consider stochastic transport between fixed endpoint distributions under prescribed external and pairwise interaction potentials. They decompose the control objective into an entropy-regularised quadratic transport problem coupled to learnable Gaussian conditional paths. The method alternates transport-plan and path updates without trajectory simulation during training, then fits velocity and score networks for generation. Experiments evaluate synthetic optimal-control recovery, reconstruction of observed single-cell distributions, and held-out-time interpolation on synthetic and spatial transcriptomics data.

Critical strengths
  • Jointly adapting the transport plan and conditional probability paths is well motivated: pairwise interactions make each route's cost depend on the other routes and their transport weights. The alternating optimisation provides a principled way to address this dependence without trajectory rollouts during training.
  • The problem formulation is interesting and the mathematical treatment provides useful support for the method.
  • The spatial transcriptomics interpolation experiment provides useful evidence for the prior through evaluation at held-out times and comparison with CCB without the spatial prior.
Critical weaknesses
  • The biological value of the inferred single-cell dynamics remains insufficiently established. Table 2 evaluates distributions at observed training times: improved reconstruction does not, by itself, establish that the intermediate dynamics or interaction assumptions are more accurate.

  • The reviewer finds it difficult to assess how much of the real-data improvement comes from the interaction prior rather than model capacity, coupling optimisation or tuning.

  • The author might consider including Curly Flow Matching into related works, which also learns conditional paths, calculates transport costs from those paths, and then performs flow/score matching.

  • Table 2 serves its stated purpose of checking reconstruction at observed training times. However, this evaluation does not validate the intermediate dynamics on those expression datasets, leaving the biological usefulness of the chosen interaction priors insufficiently tested.

  • The within-CCB spatial-prior ablation and synthetic coupling ablation provide useful evidence. The reviewer nevertheless finds the model-selection procedure and the comparability of capacity, tuning budgets and prior information across methods insufficiently clear, limiting interpretation of the real-data gains.

References

Questions for authors
  1. How much of the improvement can be attributed to the interaction prior, rather than model capacity, coupling optimisation or hyperparameter-search budget? Could the paper authors clarify which comparisons isolate these contributions in the real-data experiments?

  2. Table 2 evaluates distributions at observed training times. What do these improvements establish about the inferred dynamics beyond better endpoint reconstruction, and what additional evidence supports the chosen interaction assumptions?

  3. Are competing methods that can incorporate the spatial prior given the same prior and initial-position information? Could the paper authors clarify whether the comparisons isolate the optimisation method from the benefit of supplying this prior?

  4. How were the potential forms, signs and strengths, diffusion levels, and training checkpoints selected? Was a separate validation set used, and did any model-selection decisions use the held-out evaluation snapshots? How sensitive are the real-data results to these choices?

  5. What is the biological justification for attractive versus repulsive quadratic interactions in expression space? Both separately trained variants reconstruct the observed marginals well; what independent evidence would distinguish their biological plausibility?

6 comments

We thank the reviewer for the careful review and for recognizing the mathematical formulation and the motivation for jointly optimizing the transport plan and conditional paths. We first emphasize that the core of this paper is stochastic optimal control under prescribed external and interaction potentials. Our main methodological contributions are: an entropy-regularized quadratic OT control objective for fixed conditional paths, derived through an endpoint-conditioned decomposition of path-space KL; alternating updates of the coupling and conditional probability paths, addressing the dependence of each path's cost on other paths and their transport weights; and mean and covariance parameterizations satisfying the endpoint boundary conditions that ensure finite conditional-path KL, preventing endpoint singularities from making transport costs diverge. Together, these constructions enable a solver that does not simulate controlled trajectories during training. Below, we address all of your questions and describe the experiments and analyses added in the revision.

[W1]

Table 2 evaluates distributions at observed training times: improved reconstruction does not, by itself, establish that the intermediate dynamics or interaction assumptions are more accurate.

We apologize that the original manuscript did not clearly explain the motivation and comparison setup of Experiment 2. We have rewritten its motivation in Section 4.2. Reference [2] shows that straighter flow trajectories help maintain generation quality with fewer integration steps. Potentials curve the conditional mean paths and may make regression and numerical integration more difficult. Experiment 2 therefore asks whether CCB can maintain generation fidelity comparable to existing methods when using potential-induced curved conditional paths. Attractive and repulsive potentials serve as controlled experimental settings here, rather than biological priors prescribed for these expression datasets. The reviewer is correct that generation fidelity also depends on training procedures, hyperparameters, and training recipes, such as the simulation-based training of DeepRUOT and VGFM, the growth penalty in WFR-FM, and how supervision samples are drawn. We use the same MLP architecture as the comparison methods to parameterize the velocity field and score, and adapt each method's official implementation. Experiment 2 shows that our CCB implementation achieves comparable generation quality. These results involved almost no hyperparameter search, including for noise and potential strengths.

[W2, W4, Q1, Q2]

How much of the improvement can be attributed to the interaction prior rather than model capacity, coupling optimisation or tuning, and what evidence supports the inferred dynamics beyond better reconstruction at observed training times?

We thank the reviewer for the suggestions concerning the contribution of potentials on real data. We have added experiments and analysis in Section 4.3 and Table 5 that estimate potentials from visible single-cell snapshots and then use them for held-out-time interpolation, with methodological details and per-time-point results in Appendices B.6 and D.5. Single-cell expression snapshots lack the positional information available in spatial transcriptomics and thus do not offer such a prior. We therefore developed CCB Bootstrap to learn a shared potential landscape that explains distributional changes across visible times. Specifically, we construct triples (a,c,b)(a,c,b) with a<c<ba<c<b, fix the OT coupling between the initial and final snapshots, obtain conditional bridges by minimizing the Action under the current potentials, and supervise potential learning using the snapshot at the intermediate time cc. Denoting the bridge marginal at cc by ρca,b;ω\rho_c^{a,b;\omega}, the marginal-fitting term is W22(ρca,b;ω,ρ^c)W_2^2(\rho_c^{a,b;\omega},\widehat{\rho}_c), where ω\omega denotes the parameters of the external and interaction potentials. All visible triples share the same potentials, so the learned landscape must simultaneously explain distributional changes over different time intervals. We then freeze the potentials for CCB interpolation at held-out times. Potential estimation uses only visible snapshots; details are provided in Appendix B.6.1.

With and without potentials. For a fair comparison, we reimplemented OT-CFM and SF²M within exactly the same experimental pipeline as CCB. We also compare CCB with and without potentials under matched inference modes, so that their difference directly reflects the effect of adding estimated potentials. The experiments cover seven dataset settings. The table below compares CCB without potentials and CCB with jointly learned V+WV+W, alongside OT-CFM and SF²M, using the mean W1W_1 results from Table 5. Held-out times are weighted equally within each dataset. To save space in this response, we omit standard deviations here; Table 5 and Appendix D.5 retain the corresponding standard deviations for stochastic inference.

Method / Inference modeEMT 10DCITE 50DEB 5DEB 50DEB 100DWOT 30DAxolotl 30D
OT-CFM0.29937.0903.4189.96312.0933.9774.377
SF²M0.32537.2473.43710.51812.9373.9924.380
CCB without potentials / PF-ODE0.29837.0333.4199.90212.0433.9464.334
CCB without potentials / SDE0.32337.1953.43310.49312.9123.9634.340
CCB with potentials / PF-ODE0.29739.0543.2919.46311.3133.7964.269
CCB with potentials / SDE0.30339.1693.3069.81311.9803.8124.274

Under matched inference modes, CCB with potentials improves interpolation over no-potential CCB in 6/7 settings; its PF-ODE and SDE results also outperform OT-CFM and SF²M, respectively, in those six settings. The per-time-point results in Appendix D.5 further show that PF-ODE outperforms both no-potential CCB and OT-CFM at 20/22 held-out times, while SDE outperforms both no-potential CCB and SF²M at 21/22 held-out times. These results show that a shared potential landscape learned from visible snapshots captures single-cell dynamical structure that transfers across time and consistently improves interpolation across most datasets and held-out times.

External-only and interaction-only ablations. To separately assess the effect of interactions, we evaluate CCB interaction only, which fixes V=0V=0 and estimates only WW from visible snapshots, alongside CCB external only, which fixes W=0W=0 and estimates only VV. The table below compares both variants with no-potential CCB.

Method / Inference modeEMT 10DCITE 50DEB 5DEB 50DEB 100DWOT 30DAxolotl 30D
CCB without potentials / PF-ODE0.29837.0333.4199.90212.0433.9464.334
CCB without potentials / SDE0.32337.1953.43310.49312.9123.9634.340
CCB external only / PF-ODE0.30540.0873.1639.72311.9263.8684.268
CCB external only / SDE0.31340.2273.18410.28612.8003.8844.274
CCB interaction only / PF-ODE0.30838.6473.3799.54011.6843.8014.217
CCB interaction only / SDE0.31238.7533.4129.93212.4823.8184.222

Relative to no-potential CCB under the same inference mode, interaction only improves 5/7 and 6/7 settings with PF-ODE and SDE, respectively, covering 18/22 and 20/22 held-out times (Appendix D.5, Tables 26–32). Learning only WW achieves lower mean W1W_1 than learning only VV in 5/7 dataset settings with PF-ODE and 6/7 with SDE. Together, these results show that interaction potentials yield consistent benefits on their own and further support the need to explicitly model interactions to obtain stronger interpolation gains.

In response to Reviewer NShB's request, we also constructed a new 20D HMF Model dataset with external and pairwise interaction potentials, where adding only the interaction potential substantially improves held-out interpolation: under PF-ODE inference, W1W_1 decreases from 0.506 to 0.063 and MMD2\mathrm{MMD}^2 from 3.966×10−33.966\times10^{-3} to 0.040×10−30.040\times10^{-3} (Table 3).

[Q3]

Are competing methods that can incorporate the spatial prior given the same prior and initial-position information?

We thank the reviewer for highlighting the importance of fair comparisons. We have supplied WLF and GSBM with the same spatial deformation potential, potential strength, and initial-position information as CCB, and added the results to Table 4. In our understanding, the starting position does not constitute additional information here. Because the spatial deformation potential depends on the starting position, conditioning the velocity field and score on it is necessary to ensure that regression after alternating optimization can recover the corresponding optimal control law, as discussed in Appendix B.5. All metrics below are averaged equally over held-out times:

Method / Inference modeSpatial W1W_1 ↓Expression W1W_1 ↓Cell-type neighborhood TV ↓
WLF (Spatial Potential)0.3725.0000.537
GSBM (Spatial Potential)0.494 ± 0.000625.076 ± 0.000480.576 ± 0.00110
CCB (Spatial Potential, PF-ODE)0.2564.8260.443
CCB (Spatial Potential, SDE)0.255 ± 0.000354.833 ± 0.000430.481 ± 0.00115

With the same spatial potential and initial-position information, both PF-ODE and SDE inference with CCB outperform WLF and GSBM on all three metrics.

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

This paper tackles the problem of learning cell dynamics between unpaired snapshots, with a model where cells are subject to external and pairwise interaction potentials specified by the user. Without potentials, the optimal dynamics with respect to the least-action principle are straight paths, hence existing simulation-free methods first match cells across snapshots by solving an optimal transport (OT) problem, then train a velocity network along the resulting paths. However, with potentials, paths that minimize action are no longer straight lines, and in general do not have an analytical form. Even worse, including a pairwise interaction potential makes optimal paths depend on all other paths and on the matching, so cell matching and path finding have to be optimized jointly. The main idea behind the authors' method is to alternate between optimizing matching cells between snapshots (minimizing an entropy-regularized OT problem) and path finding (minimizing the action). They model the learned paths by a neural network, and once optimized, the resulting paths are then used to train a velocity network to learn the dynamics, enabling simulation-free training with potentials. The authors evaluate the whole pipeline on synthetic benchmarks, with known ground truth potentials, and on real single-cell and spatial transcriptomics data, with generic / hand-crafted potentials, and report better results than existing methods.

Critical strengths

The paper tackles a difficult problem: learning dynamics with pairwise interactions without simulating trajectories. I find that the authors clearly explain why existing simulation-free methods do not apply. I also find their main idea to be elegant: by conditioning on the endpoints, the action splits into an OT problem on the matching and a cost on each path, which naturally leads to an alternating scheme. The authors also show that both steps decrease the same objective, which I find to provide a welcome theoretical framing. Additionally, the authors show that their method recovers the optimal control law more accurately than other baselines on synthetic benchmarks with known ground truth (Table 1).

Critical weaknesses

Although I acknowledge the authors’ methodological contribution to simulation-free optimal control, I find the evidence supporting the biological application insufficient.

Choice of potentials in practical applications. The main idea of the paper is to extend simulation-free methods to pairwise potentials. However, from an application point of view, the obvious question is how to choose V and W in practice. To the best of my knowledge, models of cell-cell interactions are usually not available as potentials in gene space, and it is unclear how one should construct them (Q1). The authors use generic quadratic potentials or hand-crafted priors in their experiment, but do not discuss how such potentials should be selected in practice. Consequently, it is unclear whether the gains reported in Table 3, where the true potentials are known, would transfer to real data (Q2).

Given that no established potentials for practical application are available, learning them (as in CytoBridge) is a natural idea. Unfortunately, as acknowledged by the authors in their limitations (L485), the design of CCB prevents this approach. Hence, another idea would be to reuse previously learned potentials, but doing so does not seem straightforward: for example, CytoBridge learns forces rather than action potentials. The authors mention Guan et al. (2026) as an alternative when potentials must be estimated (L485), but this is not explored in the paper (Q3).

Benefit of pairwise interactions.
The alternating optimization scheme is the direct consequence of the additional pairwise interactions (L065), so I think evidence of their benefit is central to the paper. However, such evidence is limited to the KS dataset in Table 3, which is a synthetic dataset with oracle potentials, and to the TV results in Table 4. By contrast, the ablation on the Coulomb dataset in Table 3 shows little gain on both metrics: CCB without potentials and CCB with the pairwise potential only perform equally (W1=0.217 in both cases), and the full model (W1=0.079) is close to the external-only model (W1=0.083). Likewise, performance gains in the first two columns of Table 4 remain unclear in the asbence of standard deviations (Q4): CCB seems almost on-par with WFR-FM on Spatial W1 (0.256 vs 0.258), and loses to DeepRUOT on Exp. W1. I am also surprised CytoBridge is absent from Table 4: I find it to be an important baseline, given that this method's paper reports strong results on spatial transcriptomics. (Q5) All evidence considered, I believe that pairwise interactions are not sufficiently supported, despite being the cornerstone of the authors' narrative. (Q6)

Missing interpolation on scRNA-seq. The abstract says that “incorporating external and pairwise potentials improves interpolation at unobserved time points” (L025). However, this task is only evaluated on synthetic data in Table 3 and spatial transcriptomics in Table 4. By contrast, scRNA-seq datasets are only evaluated at observed time points (Table 2). Held-out time point benchmarks on scRNA-seq are standard in the literature, hence such results would directly support the relevance of the method for biological applications. (Q7)

Baseline results. I cross-checked the baseline results in Table 2, and I found that the CytoBridge paper reports 8.217 (t=1) and 11.086 (t=2) for SF2M over Weinreb (50D), while the authors report 11.579. (Q8)

Conclusion. Overall, the authors bring an elegant, well-crafted methodological improvement to simulation-free optimal control, but their experiments do not yet support the biological application the paper is centered on. In particular, I believe the gap between the narrative and the evidence should be addressed: either the paper should provide stronger evidence that the method brings an important contribution to modeling cell-state dynamics (which is the direction my questions try to explore), or it should be substantially rewritten so that its storyline reads as a method paper on simulation-free optimal control, with the biological experiments as illustration. Currently, I lean toward a weak reject based on the submitted version of the paper, but I am open to adapting my score based on the authors' responses to my questions and additional evidence they can provide to address the issues I raise.

Questions for authors

Q1. Could you discuss how V and W should be selected in practice? Could you recommend potentials informed by biological knowledge, or derived from an external estimation method?

Q2. Could you assess how robust CCB is when the potentials differ from the ground truth on the Table 3 datasets?

Q3. Could you conduct experiments on ways to leverage potentials learned by other methods in CCB?

Q4. Could you report standard deviations for experiments in Table 4?

Q5. Could you add CytoBridge to Table 4?

Q6. Could you provide an experiment, ideally on real data, where the pairwise potential brings a clear gain over the no-potential and external-only CCB variants?

Q7. Could you report held-out time point interpolation on the scRNA-seq datasets, with and without potentials?

Q8. Could you explain the gap between the SF2M results in Table 2 and those reported in the CytoBridge paper?

13 comments

We thank the reviewer for carefully reading our paper and providing constructive comments. First, we clarify that Constructive CytoBridge is not a sequel to CytoBridge. CytoBridge focuses on inferring cell dynamics and interaction potentials from data, whereas our central problem is stochastic optimal control under prescribed external and pairwise interaction potentials. If the reviewer finds the name confusing, we are willing to change the paper title and method name.

We address all of your questions below, with clarifications and additional experiments in the revised manuscript.

[W1, Q1]

How should V and W be selected in practice, using biological knowledge or an external estimation method?

We thank the reviewer for highlighting the importance of potential selection in practical applications. For single-cell expression data without an explicit potential prior, we recommend external estimation methods that infer potentials from visible snapshots and then freeze them for CCB interpolation. We have therefore added the potential estimator CCB Bootstrap, together with its interpolation results, to Section 4.3, and provided implementation details in Appendix B.6. It learns an external potential VωV_{\omega} and a pairwise interaction potential WωW_{\omega}, with the total potential energy given by:

Uω[ρ]=EX∼ρ[Vω(X)]+12E(X,X′)∼ρ⊗ρ[Wω(X−X′)].U_{\omega}[\rho]=E_{X\sim\rho}[V_{\omega}(X)]+\frac{1}{2}E_{(X,X')\sim\rho\otimes\rho}[W_{\omega}(X-X')].

We parameterize both potentials with neural networks and enforce Wω(δ)=Wω(−δ)W_{\omega}(\delta)=W_{\omega}(-\delta) by symmetrization. CCB Bootstrap estimates these potentials using only the visible snapshots in each held-out fold. Our aim is to learn a shared landscape that explains changes across visible times and then test its transfer to the held-out time. To learn this landscape, we construct all triples of visible times:

Q={(a,c,b)∣a<c<b,  a,c,b∈Tvis}.\mathcal{Q} = \{(a,c,b)\mid a<c<b,\; a,c,b\in\mathcal{T}_{\mathrm{vis}}\}.

For each triple, we fix the OT coupling between snapshots aa and bb, obtain conditional bridges by minimizing the Action under the current potentials, and use the observed distribution at the intermediate time cc to supervise potential learning. Specifically, we write the conditional mean as mt=(1−αt)xa+αtxb+rtm_t=(1-\alpha_t)x_a+\alpha_t x_b+r_t, where αt=(t−a)/(b−a)\alpha_t=(t-a)/(b-a) and ra=rb=0r_a=r_b=0. During estimation, we fix the Brownian bridge covariance and optimize the correction rtr_t to obtain conditional paths under the current potentials. Denoting the intermediate marginal obtained by mixing these bridges by ρca,b;ω\rho_c^{a,b;\omega}, the marginal-matching term for potential learning is:

Lmatch(ω)=1∣Q∣∑(a,c,b)∈QW22(ρca,b;ω,ρ^c).L_{\mathrm{match}}(\omega)=\frac{1}{|Q|}\sum_{(a,c,b)\in Q}W_2^2(\rho_c^{a,b;\omega},\widehat{\rho}_c).

We update the potentials using implicit gradients through the inner Action-minimization problem. All visible triples share the same parameters ω\omega. After estimation, we freeze VωV_{\omega} and WωW_{\omega} and supply them to CCB for interpolation at held-out times.

[W1, Q2]

Could you assess how robust CCB is when the potentials differ from the ground truth on the Table 3 datasets?

We thank the reviewer for suggesting a robustness evaluation when potentials deviate from the ground truth. We have added a sensitivity analysis of known potentials in Appendix B.8.3. The results show that even when both external and interaction potential strengths deviate by 10%, CCB still clearly outperforms its no-potential counterpart under the same inference mode. Specifically, we decrease or increase potential strengths by 10% on KS, Coulomb, and the newly added HMF Model. For datasets with both external and interaction potentials, the tables below show the results of weakening or strengthening both together: V~=(1+ϵ)V\widetilde V=(1+\epsilon)V and W~=(1+ϵ)W\widetilde W=(1+\epsilon)W, where ϵ∈{−0.1,0.1}\epsilon\in\{-0.1,0.1\}. KS has no external potential, so only WW is perturbed. Results are averaged over three held-out times, and all MMD2\mathrm{MMD}^2 values are reported in units of 10−310^{-3}. Appendix B.8.3 provides the full results for individual perturbations and other combinations.

PF-ODE inference

CCB settingKS W1W_1 ↓KS MMD2\mathrm{MMD}^2 ↓Coulomb W1W_1 ↓Coulomb MMD2\mathrm{MMD}^2 ↓HMF W1W_1 ↓HMF MMD2\mathrm{MMD}^2 ↓
Without potentials0.157 ± 0.000310.163 ± 0.0210.217 ± 0.005147.574 ± 2.0490.506 ± 0.00243.966 ± 0.035
Correct potentials0.117 ± 0.00041.791 ± 0.0310.079 ± 0.0115.169 ± 0.5220.049 ± 0.00280.011 ± 0.004
Potential strength −10%0.1171.8820.0897.3850.0690.056
Potential strength +10%0.1161.8070.0653.4330.0760.064

SDE inference

CCB settingKS W1W_1 ↓KS MMD2\mathrm{MMD}^2 ↓Coulomb W1W_1 ↓Coulomb MMD2\mathrm{MMD}^2 ↓HMF W1W_1 ↓HMF MMD2\mathrm{MMD}^2 ↓
Without potentials0.195 ± 0.000469.389 ± 0.0580.251 ± 0.0009244.635 ± 0.3900.523 ± 0.000143.996 ± 0.004
Correct potentials0.168 ± 0.00161.229 ± 0.0380.146 ± 0.00033.934 ± 0.0350.127 ± 0.00090.012 ± 0.005
Potential strength −10%0.1671.3320.1526.0400.1350.052
Potential strength +10%0.1681.2100.1412.5780.1400.072

Even with ±10% deviations in potential strength, CCB substantially reduces both W1W_1 and MMD2\mathrm{MMD}^2 relative to the corresponding no-potential baselines across all three datasets and both inference modes. This shows that CCB retains a clear interpolation advantage under imperfect potentials and is robust to the tested changes in potential strength.

[W1, W4, Q3, Q7]

Could you evaluate potentials learned by other methods in CCB and report held-out time point interpolation on the scRNA-seq datasets, with and without potentials?

We thank the reviewer for suggesting an evaluation of externally estimated potentials and held-out-time interpolation on single-cell data. We have added a complete HOO evaluation of CCB Bootstrap, comparing it with CCB without potentials, OT-CFM, and SF²M; results appear in Table 5 of the main text. The estimator uses only visible snapshots; we then freeze the potentials for held-out interpolation. Held-out snapshots are used only for evaluation; methodological details are provided in Appendix B.6.

For a consistent comparison, we reimplemented OT-CFM and SF²M within the same experimental pipeline as CCB. The experiments cover the datasets from Experiment 2 except Weinreb, and additionally include expression-only Axolotl data and WOT reprogramming data, yielding seven dataset settings. Removing one time point from Weinreb leaves only two snapshots, which cannot provide visible triples for CCB Bootstrap potential estimation. The table below reproduces all results from Table 5, reporting W1W_1 averaged equally over held-out times within each dataset.

Method / Inference modeEMT 10DCITE 50DEB 5DEB 50DEB 100DWOT 30DAxolotl 30D
OT-CFM0.29937.0903.4189.96312.0933.9774.377
SF²M0.325 ± 0.002037.247 ± 0.00813.437 ± 0.005810.518 ± 0.006012.937 ± 0.00673.992 ± 0.00154.380 ± 0.0004
CCB without potentials / PF-ODE0.29837.0333.4199.90212.0433.9464.334
CCB without potentials / SDE0.323 ± 0.001037.195 ± 0.00603.433 ± 0.006810.493 ± 0.011712.912 ± 0.00433.963 ± 0.00104.340 ± 0.0003
CCB + CCB Bootstrap / PF-ODE0.29739.0543.2919.46311.3133.7964.269
CCB + CCB Bootstrap / SDE0.303 ± 0.000939.169 ± 0.00393.306 ± 0.00259.813 ± 0.013011.980 ± 0.00683.812 ± 0.00014.274 ± 0.0005

Under matched inference modes, CCB Bootstrap improves interpolation over no-potential CCB in 6/7 dataset settings; its PF-ODE and SDE results also outperform OT-CFM and SF²M, respectively, in those six settings. The per-time-point results in Appendix D.5 further show that PF-ODE outperforms both no-potential CCB and OT-CFM at 20/22 held-out times, while SDE outperforms both no-potential CCB and SF²M at 21/22 held-out times. These results show that a shared potential landscape learned from visible snapshots captures single-cell dynamical structure that transfers across time and consistently improves interpolation across most datasets and held-out times. CITE performs least well, and the PF-ODE improvement on EMT is also small. Both datasets contain only four snapshots, leaving three visible snapshots and a single training triple per fold. The limited temporal information available for estimating a shared potential landscape is consistent with the weaker transfer performance on these two datasets.

Comments on the paper

2 comments

We thank the reviewer for acknowledging the additional experiments and clarifications, and for raising the question of the computational cost of the complete pipeline. Mean-field optimal-control solvers and potential-estimation methods address two distinct tasks: the former solve for optimal controlled dynamics under prescribed potentials, whereas the latter infer potentials from observed snapshots. We therefore compare each stage with methods designed for the corresponding task.

First, we clarify that Table 15 was already included in the manuscript before this rebuttal round. It reports the time spent solving the optimal-control problem under prescribed potentials (including flow/score matching). It does not cover the CCB Bootstrap potential estimation introduced in this round. The table compares CCB with GSBM and Modified WLF, which can incorporate interaction potentials and solve mean-field optimal control problems, to assess the computational efficiency of solving the optimal-control problem itself.

Following the reviewer's suggestion, we further provide a comparison of peak GPU memory usage during the optimal-control stage, as shown below (in GiB):

DatasetCCBGSBMModified WLF
EB 5D0.4850.2990.802
EB 50D0.8841.0950.808
EB 100D1.3501.9810.817

The results show that CCB's peak GPU memory usage during the optimal-control stage is comparable to that of existing methods.

To further evaluate the computational efficiency of potential estimation, we compare CCB Bootstrap with the previously discussed alternative potential-estimation methods, CytoBridge and WLM [5]. The table below reports training time and peak GPU memory usage. The timing for CCB Bootstrap includes OT preprocessing and subsequent potential learning.

DatasetPotential-estimation methodTraining time (seconds)Peak GPU memory (GiB)
EB 5DCCB Bootstrap315.01.91
EB 5DCytoBridge925.81.51
EB 5DWLM919.81.86
EB 50DCCB Bootstrap858.02.01
EB 50DCytoBridge958.85.76
EB 50DWLM1003.81.88
EB 100DCCB Bootstrap911.42.12
EB 100DCytoBridge982.810.56
EB 100DWLM1042.81.90

The results show that CCB Bootstrap takes less time to estimate potentials than CytoBridge and WLM at all three dimensionalities, with comparable peak GPU memory usage. Its memory usage is close to that of WLM and lower than that of CytoBridge on EB 50D and EB 100D.

Compared with the other potential-estimation alternatives, CCB Bootstrap does not incur substantially greater time or memory costs because each potential update samples only B=128B=128 endpoint pairs and 16 interaction partners per sample, rather than using all initial particles. This limits the computational and memory requirements of the inner conditional-bridge optimization and implicit-gradient solve. Replacing full-data OT preprocessing with minibatch OT could further reduce preprocessing time and memory usage.

[5] A Call to Lagrangian Action: Learning Population Mechanics from Temporal Snapshots.

Dear Reviewer h2vN,

Thank you again for your constructive feedback. We have responded to your question about computational cost with timing clarifications and runtime and memory comparisons. As the discussion closes on November 18, we would appreciate your feedback on whether this addresses your remaining concern.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.