Abstract

High-content imaging assays quantify cellular responses to chemical and genetic perturbations, yet continuous trajectories of individual cells are unobservable because cells are chemically fixed at acquisition. Perturbation modeling therefore reduces to inferring stochastic transport between control and treated populations observed only as separate marginals. While recent generative models achieve strong endpoint alignment, boundary consistency does not determine intermediate evolution: multiple stochastic processes may connect identical marginals while traversing regions unsupported by observed single-cell morphologies. We introduce FreeBridge, a Schrödinger Bridge formulation for single-cell transition modeling under endpoint-only supervision. FreeBridge defines atomic states as instance-segmented single-cell representations, establishing a fixed cellular manifold, and learns stochastic transport constrained within this geometry via empirical latent support regularization. Across BBBC021, RxRx1, and JUMP, FreeBridge maintains competitive or improved endpoint fidelity under a unified evaluation protocol; on BBBC021, it further reduces intermediate support violations and improves mechanism-of-action retention. These findings highlight the importance of geometric grounding for biologically interpretable perturbation dynamics.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2873_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/curioWang/FreeBridge

Link to the Dataset(s)

BBBC021: https://bbbc.broadinstitute.org/BBBC021 RxRx1: https://www.rxrx.ai/rxrx1 JUMP Cell Painting datasets: https://github.com/jump-cellpainting/datasets

BibTex

@InProceedings{WanXur_FreeBridge_MICCAI2026,
        author = { Wang, Xurui AND Ren, Qin AND Ma, Jun AND Ling, Haibin AND You, Chenyu},
        title = { { FreeBridge: Variational Schrödinger Bridges for Cellular Transition Dynamics } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16891},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The authors propose a new framework, termed FreeBridge, for modeling single-cell state transitions. To this end, they define atomic states as instance-segmented single-cell representations, which are used to construct a fixed cellular manifold and to learn stochastic transport between states. The proposed method is evaluated on several datasets and achieves competitive results.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.The authors tackle an interesting problem, namely modeling single-cell state transitions using generative models, which has several applications in biology.

    2.The authors introduce a novel loss function to constrain the cellular manifold, in contrast to previous works.

    3.Overall, the paper is well written and easy to follow.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.The Schrödinger bridge problem is not introduced with sufficient clarity, and there are several inconsistencies in the notation and equations.

    2.Equation (3) corresponds to the forward diffusion process used in standard diffusion models. It is unclear how the proposed approach differs from or extends existing diffusion-based methods.

    3.Equation (2), the construction of the support bank is not clearly explained.

    4.It is unclear how the authors assess trajectory consistency, given that intermediate states are not available in the data.

    5.The numerical approximation of the loss function is not clearly described.

    6.The paper does not specify the architecture of the neural networks used, nor does it provide sufficient training details.

    7.The qualitative evaluation is limited, relying on only two images. It would be beneficial to include visualizations of intermediate states, as modeling such states is a central objective of the work.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    While the motivation of this work is interesting, the proposed method lacks sufficient theoretical grounding, and the experimental evaluation is not adequate to convincingly demonstrate its effectiveness.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    FreeBridge introduces a Schrödinger Bridge framework to model single-cell transition dynamics from endpoint-only imaging data, where individual trajectories are unobservable. Its primary contribution is the Cell Engine architecture, which decouples cellular state specification from stochastic transport to ensure intermediate transitions remain biologically plausible. By regularizing latent trajectories with an empirical support cost, FreeBridge prevents the model from traversing non-biological voids in the latent space. Across BBBC021, RxRx1, and JUMP benchmarks, the method achieves superior endpoint fidelity and trajectory consistency compared to existing distribution-alignment approaches.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1) The introduction provides a well-structured motivation, correctly identifying that because infinitely many stochastic processes can connect identical marginals, endpoint alignment does not determine intermediate evolution.

    2) The most compelling contribution is the Empirical Latent Support Regularization, which addresses a key limitation of standard generative models that often produce shortest-path trajectories. By penalizing transitions that enter non-biological voids, FreeBridge represents a significant advancement toward achieving realistic in silico virtual cell simulations.

    3) The Cell Engine formulation, which decouples the State Unit from the Transition Rule, offers a highly effective and modular architectural design. This separation ensures that the underlying definition of a cellular state remains stable while learning complex transport dynamics. Furthermore, Figure 1 is excellently designed, providing a clear and accessible overview of the entire framework.

    4) A critical technical contribution is the implementation of a nearest-neighbor-based support cost to regularize intermediate states. This effectively penalizes trajectories traversing regions of the latent space unsupported by observed morphologies, solving the problem of stochastic processes connecting endpoints through biologically implausible states. Figure 2 is instrumental in building intuition for this mechanism, while the results in Table 2 and Figure 4 provide empirical evidence regarding the necessity of the support bank.

    5) Through extensive benchmarking in Table 1, the authors demonstrate that FreeBridge consistently outperforms recent distribution-alignment baselines, including PhenDiff, IMPA, and CellFlux, across three major datasets: BBBC021, RxRx1, and JUMP. The method achieves superior endpoint fidelity, as measured by FID and KID, while simultaneously yielding a significant reduction in intermediate support violations.

    6) The qualitative samples presented in Figure 3 are highly persuasive, demonstrating the model’s ability to preserve compound-specific morphological structures with high fidelity.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1) Implementation details are significantly underspecified, hindering the reproducibility of the work. Given the space constraints of the submission template, the release of open-source code is essential to make the framework fully replicable. Specifically, the manuscript does not disclose the architecture or pre-training regime of the frozen encoder used within the State Unit, nor does it specify the dimensionality of the resulting feature embeddings. Furthermore, the architecture of the FreeBridge model itself remains unknown; while Figure 3 displays generated images, the necessary decoder components are not described in the text.

    2) While the fixed geometry strategy is conceptually elegant, the transition rule is entirely dependent on the fidelity of the frozen encoder. If this encoder fails to capture subtle but biologically critical morphological features during its initial training, the subsequent Schrödinger Bridge transport will be fundamentally constrained to a sub-optimal manifold that cannot be corrected during the transition learning phase.

    3) As the paper acknowledges that true single-cell trajectories are unobservable due to chemical fixation, the evaluation of trajectory consistency relies heavily on the Support Violation Rate. While this metric confirms that intermediate states align with known cellular morphologies, it does not strictly prove that the learned sequence of states is biologically accurate, only that it is realistic according to the empirical bank.

    4) Although the authors utilize a minibatch approximation to mitigate the computational cost of calculating the empirical support distance, the approach still requires B x B distance comparisons at every discretization step. For high-dimensional embeddings or larger batch sizes, this requirement could significantly impact training efficiency compared to unconstrained generative models.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I believe the innovation of FreeBridge greatly outweighs its current limitations. The proposed geometric grounding is simple and original and offers a principled path for future expansion in virtual cell modeling. While the reliance on a frozen encoder and the computational cost of nearest-neighbor checks are noted weaknesses, these are addressable in future work. However, the lack of implementation details is a critical barrier to reproducibility; the authors must release the source code and full architectural specifications upon acceptance to ensure this work is replicable

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    Authors develop a constrained stochastic transport framework for single-cell transition modeling under endpoint-only supervision. Results demonstrate reduced intermediate support violations alongside competitive or improved generative performance relative to prior distribution-alignment approaches.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    Well written. The path constraint based on empirical data is an interesting approach. The manuscript is well written and the key ideas are clearly explained.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    Code not provided

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The results are clearly presented, and the key ideas, especially the aspect of including trajectory-levelconsistency within empirically observed morphology space is interesting.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Author Feedback

We thank the reviewers and the Area Chair. We use this response to clarify the theoretical framing, which addresses several of R1’s concerns, and to outline camera-ready revisions for the reproducibility issues raised across reviews.

Theoretical framing and relation to diffusion (R1.W1, W2, W5) We agree the Schrödinger Bridge positioning was not stated clearly enough. Eq. (3) specifies the controlled SDE for the candidate path measure, using a Brownian noising process as reference. While this resembles the forward process in diffusion models, FreeBridge does not learn a singlemarginal diffusion score; instead, it optimizes a bridge path measure with fixed initial and terminal marginal constraints and an empirical state cost. Eq. (9) can be written as a Generalized Schrödinger Bridge objective with state cost Vt(z) = λdemp(z)2, where demp is the nearest-neighbor distance to the observed cellular morphology support. The camera-ready will make this GSB positioning explicit, add the GSBM reference, and correct notation inconsistencies in Sec. 2.We will also clarify the previously imprecise statement on underdetermination: (i) endpoint-only marginal matching is underdetermined before specifying a reference process; (ii) classical SB is well-defined once a Brownian reference is fixed; (iii) FreeBridge augments this with a data-driven state cost anchoring intermediate paths to the empirical manifold.

Support bank and numerical approximation (R1.W3, W5) We will distinguish the population-level support in Eq. (2) from the minibatch estimator used in training, in which demp is computed against the union of control and perturbed embeddings in the current batch — a stochastic minibatch approximation to the population state-cost term under i.i.d. sampling. The camera-ready will provide the full discrete-time estimator, including time discretization, nearest-neighbor computation, normalization, and which quantities receive gradients.

Trajectory evaluation and qualitative results (R1.W4, W7, R2.W3) Because fixed-cell assays do not provide observed single-cell trajectories, true temporal ordering cannot be directly verified. The Support Violation Rate measures whether intermediate states remain within the empirical morphology manifold — a support-feasibility metric, not a direct proof of biological temporal correctness. On this metric, endpoint-only baselines perform worse than FreeBridge (Table 1: Rviol = 0.11 for FreeBridge vs 0.31 for the strongest baseline; lower is better). Wewill state this limitation explicitly and retain the Mechanism-of-Action ablation (Table 3: 72.08% full model vs 65.08% without geometric regularization) as an orthogonal evaluation independent of the support-violation metric. The camera-ready will also include intermediate-state visualizations beyond endpoint samples.

Implementation and reproducibility (R1.W6, R2.W1, R3.W1) We agree key implementation details were underspecified. The camera-ready supplementary will include a full appendix covering the frozen encoder architecture and pretraining, embedding dimensionality, drift network, decoder used in Fig. 3, discretization scheme, hyperparameters, batch sizes, random seeds, and evaluation protocol. Source code and checkpoints will be released upon publication

Frozen-encoder limitation and computational cost (R2.W2, W4) We acknowledge the Cell Engine design fixes the latent geometry through a frozen encoder, so features absent from its pretraining cannot be recovered by transport learning alone. The B × B support-distance computation uses batched torch.cdist; this does not remove the O(B2) scaling but reduces constant factors. The camera-ready will report measured wall-clock and memory at the batch sizes used. We thank R2 and R3 for the supportive assessments and R1 for the careful critique. These clarifications will be incorporated into the camera-ready to improve clarity, reproducibility, and theoretical positioning.




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    There are strong discrepancies among the three reviewers, consisting of one reject and two accept. The paper tackles an important task of modelling single cell state transitions, for which FreeBridge, a constrained diffusion and transport framework is proposed, with favorable empirical results. Meanwhile, there are a number of aspects of the work that raise concerns. It would be beneficial if the authors provide further input and clarifications in the areas specifically brought out by the reviewers, including but not limited to, Schrödinger bridge problem; theoretical justifications and analyses; implementation details such as details of the engaged NNs, training details, and reproducibility; empirical evaluations esp. qualitative results; the computational cost of nearest-neighbor check.



back to top