acceptodds
Under review as a conference paper at ICLR 2027

HyperBridge: From Sparse Spatial Omics Observations to Dynamic 3D Tissue States

Abstract

Spatial omics profiles molecular states within their native tissue context, yet dense measurements across three-dimensional space and biological time remain costly and experimentally challenging. Available datasets therefore contain sparse tissue sections or observations from only a few biological stages, leaving spatial regions and intermediate tissue states unmeasured. Although these forms of missingness arise in different experimental settings, both require inference of unobserved spatial–molecular distributions from sparse tissue snapshots. We introduce HyperBridge, a generative method that reconstructs unobserved tissue states as an unbalanced stochastic bridge between endpoint distributions. A dynamic three-dimensional hypergraph captures higher-order relationships among observed cells or spots and provides endpoint-conditioned structural signals for the bridge control, while activity probabilities allow represented mass to vary without requiring particle-wise correspondence. The same formulation supports reconstruction between tissue sections and between three-dimensional states at different biological stages. Experiments across diverse spatial-omics platforms show that HyperBridge improves held-out section reconstruction and preserves tissue organization. Large-scale reconstruction further produces coherent dense tissue atlases, while experiments on planarian regeneration demonstrate reconstruction and forecasting across unobserved biological stages.

Primary area:
generative models
Keywords:
Spatial Omics, Schrödinger bridges, hypergraph learning, generative modeling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_YDNvOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The authors present HyperBridge, a generative framework for reconstructing unobserved tissue states from sparse spatial-omics snapshots, either missing sections between measured sections or unobserved biological stages between measured 3D tissue states. Reconstruction is formulated as an unbalanced Schrödinger bridge between endpoint point clouds of cells or spots, simulated with particles (“slots”) whose activity probabilities follow a learned on/off process, allowing the represented mass to vary without particle-wise correspondence. A dynamic 3D hypergraph, built from spatial and molecular affinities, supplies local tissue context to an endpoint-conditioned drift. The model is trained with control-energy, activity, mass-relaxed endpoint, and unbalanced-OT reconstruction losses. Empirical results suggest improved held-out section reconstruction across four spatial-omics platforms, coherent dense atlas generation from Allen MERFISH, and temporal interpolation and forecasting on planarian regeneration data.

Critical strengths

Overall, I commend the authors for formulating such an ambitious question. Reconstructing unmeasured spatial transcriptomics sections is a compelling direction for the field, as it enables saving costs for dense measurements with potential in-silico counterparts. In an ideal future, this could enable generalization across patients and enhance clinical research.

I also appreciate such a wide choice of datasets and spatial settings explored. In a widely expanding field, both technologically and from a methods development perspective, I think it is really important to explore method robustness across morphologies and biological systems.

Finally, dynamical graph evolution combined with diffusion bridges is a novel and creative direction for solving the problem; I believe it has potential for future iterations.

Critical weaknesses

Despite finding the research direction valid, I unfortunately think the current version of the paper is not ready for publication. I will explain my reasons below.

Overall

In my opinion, the way the paper is organized requires revisions.

  1. The math is very rushed and significantly lacks rigor. I hypothesized that it could be a space issue and that more room would be dedicated to results. But then the results are extremely summarized and underexplained. A whole page from the budget is left blank, so I believe the authors should have used it to expand both the methods description and experiments.
  2. The journal-style figures hinder the narrative quite a bit. Such a multi-panel choice requires careful elaboration, which a conference format does not allow for. I recommend that the authors restructure the paper to highlight key findings or consider moving to a journal submission. This would allow much more flexibility.

Methods

  1. L143-144: I would be more specific about what contribute means here. It's a switch-on/off mechanism for a certain position. More specificity is required to understand the binary evolution of the probability path later on.
  2. L167-168: There is no explanation of how one or more hyperedges are created for an anchor node. Also, it would be interesting to add intuition of when promoting either state distance or spatial distance in the graph construction is advised.
  3. L173-174: Minor, but a hypergraph operator is not directly intuitive to me. An appendix reference or a citation to a more rigorous description would have helped.
  4. Eq. (7). The notation is a bit counterintuitive. First, H\mathcal{H} is indexed by time only; then it becomes a function of the whole state, including positions, latents, and switch probabilities.
  5. Eq. (8). Agg is undefined. What aggregation is used? That is not immediately intuitive from standard GNN knowledge, as no information is provided on how the probabilities contribute to such aggregation.
  6. Eq. (10). The distance is not defined, nor is the parameter η\eta. There is no mention of how state distances are computed: whether coordinates and latents are concatenated.
  7. Eq. (11). Similar to Agg, also Msg remains undefined. It is not clear how you get from the second equality step to the third. I would add a detailed appendix description of the math.
  8. Eq (18). If someone is not familiar with the controlled-bridge objective, this equation is arbitrary. Even an Appendix section or a citation connecting this to the reference process would help.

Experiments

  1. As mentioned above, the amount of written detail describing the experiments is insufficient. Panels are barely described and left to interpretation.
  2. Fig. 2A. The held-out slide is not mentioned or indicated. I find it hard to assess the quality of the results. For example, in Fig. 2A, if the held-out section is S2, then I would argue the results are not great, and cluster 1 remains much more similar to S1. If the goal is to showcase an interpolation between the two states, nothing points to whether this result is good or not.
  3. The same is valid for the rest of the figure:
  • Results do not seem very good from the stacked barplot in panel b.
  • Metrics of panel b are not described, not are the baselines.
  • Nothing explains what makes results in panel c significant.
  • Panel c is a little bit better, but visual improvement over Spatial Z is very hard to appreciate. Also, a visualization of whether S1 and S3 are significantly different from S2 would be helpful to assess whether Hyperbridge is not simply copying the two endpoints over.
  1. Similar lack of detail concerns Fig. 3. What slide is shown in d? What do the individual metrics represent? And the baselines? How does Fig. 3f show results are good?
  2. In Fig. 4, it is unclear if the correlations, the stacked bar plots, and the line plots are good without a reference baseline beyond linear (e.g., you could add the baselines from the previous figures). Interpolation in Fig. 4e seems a stretched version of 6hpa and 12hpa rather than recapitulating 7dpa. This corroborates my above hypothesis that there is nothing in the method's supervision approach that stimulates real biological constraints between slides. Logically, the best the model can achieve is a standard interpolation of two subsequent time points. Please correct me if I am wrong.

I am sorry I cannot be more positive at this stage; I hope the authors find my points useful, and I am happy to engage with them during the rebuttal phase. However, as things stand at the moment, I find it unlikely that I will increase my score beyond a weak reject, as I think acceptance requires a full rewrite of multiple sections.

Questions for authors
  1. Eq. (9). How can a model that is so heavily dependent on the endpoints figure out intermediate biology? Are there constraints on the system that this model can recapitulate? For example, I expect brain ageing to be a good candidate, as changes are gradual. But I doubt any more significant time lag in, e.g., development can be modeled solidly.

  2. Was any heuristic required to avoid that optimization pays too much attention to matching latent states instead of coordinates?

  3. L208-214. The reference process is not defined, and it is not self-explanatory. Is it standard Brownian motion?

  4. It is not clear to me how the rates λ+\lambda^+ and λ−\lambda^- are learnt. What is the supervision signal for them? An algorithm, even in the Appendix, would have helped here. The reference transition process in Eq. (19) is undefined.

  5. What is the difference between KL$_+$ and a standard KL?

11 comments
AuthorsOpenReview ↗

Response to Comments on Model Design

We sincerely thank the reviewer for the time and care devoted to this detailed assessment and for recognizing the value of our research direction. We apologize for the shortcomings in the original presentation, which made parts of the method and experiments difficult to follow. The reviewer’s comments have helped us greatly improve the quality of the manuscript, particularly the mathematical explanations and the presentation of the experimental results. We have expanded the model definitions and training details, and revised the experimental descriptions to clarify each task’s purpose, inputs, and evaluation setting. We have also revised the figure captions to explain the individual panels and added further details of training, baseline implementations, and evaluation protocols to the appendix. These revisions clarify the main findings and how the individual panels support them. We have submitted the revised manuscript alongside our responses and welcome further discussion.

W1 & W2. The mathematical presentation lacks rigor, the results are underexplained, and the multi-panel figures hinder the narrative.

A1 & A2. We have expanded Sections 2.1–2.4 to define the slot contributions, hyperedge construction, propagation and aggregation operators, terminal-context messages, and training objective. Appendices A and B provide the notation and computational workflow. Sections 3.1–3.3 and the captions of Figs. 2–4 now explain each experiment’s purpose, inputs, target, and main findings. Appendix C supplies the task definitions, preprocessing, evaluation procedures, metrics, and baseline implementations.

W3. What does a slot’s contribution mean, and how does its on/off mechanism determine the binary activity path?

A3. A slot is a computational representative, not an individually identified biological cell. Sec. 2.1, Eqs. (1)–(2), defines its contribution as cpi(s)c p_i(s), where c=max⁡(N0,N1)/Kc=\max(N_0,N_1)/K. Thus, KK can be smaller than the observed cell/spot count. The implementation integrates continuous pi(s)p_i(s) without sampling a binary on/off path. Low-activity slots continue evolving as shadow trajectories while contributing little mass, as clarified in Sec. 2.3, Eqs. (13)–(16), and Algorithm 1.

W4. How are hyperedges constructed around an anchor, and when should spatial or molecular distance receive greater weight?

A4. Sec. 2.2, Eqs. (4)–(5), defines each hyperedge as its anchor plus up to kHk_{\mathcal H} highest-affinity neighbors satisfying ϵH\epsilon_{\mathcal H}, including candidates from other registered sections. If none qualifies, the anchor forms a singleton hyperedge. Smaller κr\kappa_r penalizes spatial separation more strongly, while smaller κz\kappa_z emphasizes molecular dissimilarity. Spatial proximity is useful under reliable registration. Molecular compatibility helps distinguish nearby populations and associate related observations across sections. Appendix C.2 describes the scaling used before these distances are computed.

W5. What is the precise hypergraph operator, and where is its rigorous definition or supporting reference?

A5. Sec. 2.2, Eqs. (6)–(8), now specifies incidence-based propagation following Feng et al., Hypergraph Neural Networks (2019). The revision defines the incidence matrix, hyperedge weights, and node and hyperedge degree matrices, followed by the residual update and layer normalization. Features are aggregated from nodes through hyperedges and back to nodes.

W6. How should the time-indexed notation in Eq. (7) be reconciled with its dependence on positions, latent states, and activity probabilities?

A6. The subscript in Hs\mathcal H_s identifies the hypergraph associated with the current spatial–molecular bridge state. Sec. 2.3 clarifies that memberships are rebuilt from current coordinates and latent states every four solver steps. Activities enter aggregation through Eq. (9). Between membership updates, contexts are recomputed from the evolving states and activities.

W7. How is Agg defined, and how do activity probabilities contribute to aggregation?

A7. Sec. 2.3, Eq. (9), defines activity-weighted averaging within each hyperedge, followed by weighted averaging over incident hyperedges. A slot’s activity therefore determines its contribution to the hyperedge context.

AuthorsOpenReview ↗

Response to Comments on Formulation and Experimental Setup

W8 & Q2. How is the joint-state distance defined, what does its weighting parameter mean, and are heuristics needed to balance coordinates and latent states during optimization?

A8&A2. The distance is now defined in Sec. 2.3, Eq. (10), as dη(q,q′)=ηr∥r−r′∥22+ηz∥z−z′∥22d_{\boldsymbol{\eta}}(\mathbf q,\mathbf q')=\eta_r\lVert\mathbf r-\mathbf r'\rVert_2^2+\eta_z\lVert\mathbf z-\mathbf z'\rVert_2^2. The weights control the spatial–molecular balance, and the temperature in Eq. (11) controls attention concentration. Appendix C.2 explains that both feature blocks are standardized using statistics computed from the observed training inputs, with these transformations fixed during bridge fitting. Standardization does not guarantee equal contributions, particularly when the blocks differ in dimension.

W9. How is Msg defined, and how is the original Eq. (11) obtained?

A9. Equations (11)–(12) in Sec. 2.3 align observed terminal contexts to each current slot and concatenate its current context with the aligned context difference. The original Eq. (11), now Eq. (12), is an architectural definition supplying structural conditioning, rather than an identity derived from the bridge objective.

W10 & Q3. What reference process underlies the controlled-bridge objective, is it standard Brownian motion, and how is the objective connected to this process?

A10 & A3. Although Eq. (13) allows a general reference drift, the experiments use bref=0\mathbf b_{\mathrm{ref}}=\mathbf0, as stated in Sec. 2.3. The Wiener driver is standard Brownian motion, with noise scaled by σs\sigma_s. The quadratic penalty in Sec. 2.4, Eq. (19), is motivated by the relative-entropy control energy for diffusions sharing the same noise, but additionally weights this energy by represented mass. Since shadow slots also undergo controlled diffusion, it is a mass-weighted regularizer, not the exact path-space KL of the complete slot process.

Q4. How are activation and deactivation rates learned and supervised, and what reference transition process is used in the original Eq. (19)?

A4. The activity network receives (qi,mi,s)(\mathbf q_i,\mathbf m_i,s) and predicts positive activation and deactivation rates, as defined in Sec. 2.3, Eqs. (14)–(15). These rates are learned through endpoint matching and path regularization, without activity labels or biological birth/death annotations.

In Sec. 2.4, Eq. (20), the positive reference rates λref+\lambda_{\mathrm{ref}}^+ and λref−\lambda_{\mathrm{ref}}^- specify the inactive-to-active and active-to-inactive transitions of a two-state reference process. The activity regularizer compares the learned rates with these reference rates, weighting the activation and deactivation terms by 1−pi1-p_i and pip_i, respectively. It is evaluated using continuously integrated activities, without sampling binary jump paths.

Q5. How does the generalized KL term differ from standard KL divergence?

A5. Generalized KL compares finite nonnegative measures and includes the correction −M(α)+M(β)-M(\alpha)+M(\beta), as defined in Sec. 2.4, Eq. (21). It reduces to ordinary probability KL when both measures have unit mass and remains sensitive to total-mass differences otherwise. Appendix C.5 specifies that both endpoint measures are smoothed with the same Gaussian kernel before evaluating the generalized-KL approximation. The numerical comparison therefore uses smoothed densities, allowing initially nonoverlapping Dirac supports to be compared.

W11. Can the experimental protocols and multi-panel results be explained more fully, including what each panel shows and how it supports the conclusions?

A11. Sections 3.1–3.3 and the captions of Figs. 2–4 now explain each panel’s purpose, inputs, target, and main findings. Appendix C.1 defines the tasks and training scopes, C.3–C.4 cover evaluation procedures and metrics, and C.5 describes training and baselines.

W12. Which sections are inputs and which is the generated target in Fig. 2a? If S2 is held out, why does cluster 1 resemble S1 more closely; if the panel instead illustrates interpolation, what supports the quality of the result?

A12. We apologize for the labeling error in Fig. 2a and have corrected S0.5 to S1.5. Figure 2a examines spatial gap filling across a large inter-section gap in spot-level DLPFC Visium data. S1 and S2 are measured bridge endpoints, and S1.5 is generated at an unmeasured intermediate depth. The shared encoder–decoder is pretrained on four measured sections and fixed before fitting the S1–S2 bridge.

HyperBridge generates fewer irregular internal voids than isoST and SpatialZ, providing qualitative evidence of cortical tissue continuity within the gap. Section 3.1 and the revised Fig. 2 caption explain this setting and comparison. Appendices C.1 and C.3 describe the training inputs and visualization.

AuthorsOpenReview ↗

Response to Comments on Spatial Reconstruction

W13. The remaining panels of Fig. 2 need clearer explanations of the composition bars, metrics, baselines, IMC results, and fine-structure reconstruction. Can the S1/S3-to-S2 comparison distinguish target recovery from resemblance to the endpoints?

A13. Figure 2b evaluates single-cell reconstruction in human cSCC Deep-STARmap data. Four observed layers train a shared encoder–decoder, which is then fixed. Three bridges between adjacent observed layers reconstruct three intervening targets, all excluded from training. isoST and SpatialZ use their official implementations and native generation procedures.

The stacked bars compare cell-population proportions pooled across the three reconstructed targets with their measured counterparts. HyperBridge shows the closest compositional agreement. Cell PCC and Gene PCC evaluate cellular expression profiles and gene-wise patterns, respectively. MSE measures expression error, and Chamfer distance measures spatial discrepancy. HyperBridge achieves the highest mean Cell PCC and lowest mean MSE and Chamfer distance. Each dot represents one held-out target layer, and the bars report mean ± SD across the three target layers from the same volume.

Figure 2c examines spatial proteomics using human breast-cancer IMC measurements. We generated intermediate layers between 15 measured sections to form a dense protein volume. The central views show denser CK5-high and SMA-high structures. XZ views of the same region show more continuous signals across depth and less banding than isoST and SpatialZ, using a common spatial window and protein-color mapping. These views qualitatively demonstrate volume densification and cross-section protein-signal continuity.

Figure 2d evaluates fine-structure reconstruction in mouse-brain Xenium data. Only S1 and S3 are used for representation learning and bridge fitting, with measured S2 serving as the evaluation reference. The hippocampal enlargements highlight the curved arrangement of cell populations. HyperBridge more closely reproduces this organization and achieves the highest Cell PCC, 0.522, among the compared methods.

For visualization, expression centroids derived from measured S2 assign common population colors to the predicted profiles after generation. The generated coordinates and expression values remain fixed during coloring. Figure 2d compares the reconstructed hippocampal cell arrangement with measured S2, while Cell PCC quantifies agreement between the reconstructed and measured expression profiles.

Section 3.1 and the Fig. 2 caption explain these experiments. Appendix C.1 specifies training inputs and targets. Appendix C.3 describes composition pooling, coordinate matching, and visualization. Appendix C.4 defines the metrics, and Appendix C.5 describes the baseline implementations.

W14. Which section is shown in Fig. 3d, what reference is used for comparison, what do the metrics and baselines represent, and how does Fig. 3f demonstrate reconstruction quality?

A14. The reconstructed atlas contains 25,722,672 points—2,616,328 measured cells and 23,106,344 generated points—across 129 measured and 1,182 virtual sections.

Figure 3d shows a digital sagittal section through this atlas. The upper row presents the Allen CCFv3 anatomical reference and measured marker-expression maps, and the lower row shows the reconstructions. The reconstructed Slc17a7 map reproduces cortical and hippocampal organization, while Penk recovers localized striatal enrichment. Appendix D, Fig. 5, provides coronal and horizontal views.

Figure 3e is a separate held-out S100 benchmark. HyperBridge, Linear interpolation, isoST, and SpatialZ all use S99 and S101 as inputs. In contrast, final atlas construction uses all 129 measured sections for representation learning. Cell PCC and Gene PCC evaluate cellular expression profiles and gene-wise patterns, while MSE measures expression error. ACC, NMI, and ARI evaluate annotation agreement. Relative mass error measures total represented-mass discrepancy.

Figure 3f displays predicted cell distributions in the striatum, hippocampus, and cerebellum. Their regional shapes and continuous extension across depth illustrate organization within the dense atlas. This qualitative view complements the anatomical comparison in Fig. 3d and the held-out evaluation in Fig. 3e.

Section 3.2 and the Fig. 3 caption distinguish these panels. Appendix C.1 separates the atlas-construction and held-out settings, and Appendices C.3–C.5 describe the evaluation procedures, metrics, and baselines.

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

In this paper, the authors propose HyperBridge, a generative framework for reconstructing unobserved 3D tissue states from sparse spatial omics observations. HyperBridge combines a dynamic 3D hypergraph with an unbalanced stochastic bridge to model higher-order cellular relationships and changes in represented tissue mass. It supports both spatial reconstruction between tissue sections and temporal reconstruction across biological stages. Experiments on multiple spatial omics datasets demonstrate improvements in tissue reconstruction, molecular profile preservation, and temporal state prediction.

Critical strengths
  • HyperBridge provides a unified framework for spatial and temporal tissue reconstruction. The combination of dynamic hypergraphs and unbalanced stochastic bridges enables the modeling of higher-order tissue relationships and variations in represented tissue mass.

  • The proposed method is evaluated on multiple spatial omics datasets, demonstrating improved reconstruction accuracy and preservation of tissue structure. Ablation studies further support its effectiveness.

Critical weaknesses
  • The temporal forecasting mechanism requires further clarification. Equations (9) - (12) rely on endpoint-conditioned information, while the forecasting experiment predicts the 7 dpa state using observations only up to 3 dpa. It is unclear how the proposed bridge operates when the future endpoint is unavailable during forecasting.

  • The experimental comparisons are insufficient to fully demonstrate the advantages of the proposed framework. Temporal reconstruction is mainly compared with linear interpolation, while the balanced-transport ablation assumes equal endpoint masses, which does not reflect the unequal-mass setting targeted by the method.

Questions for authors
  1. The proposed bridge relies on endpoint-conditioned information, but the forecasting experiment predicts the 7 dpa state without observing this endpoint. Could the authors explain how endpoint conditioning works in this setting and how the model predicts an unseen future state?

  2. The temporal reconstruction experiments mainly compare HyperBridge with linear interpolation. Could the authors include stronger baselines, such as optimal transport or other generative methods, to provide a more convincing evaluation?

  3. The balanced-transport ablation in Equation (27) assumes equal endpoint masses, while the proposed framework allows unequal endpoint masses. How is this difference addressed in the experiments? A comparison with a mass-normalized balanced baseline would help clarify the benefit of the unbalanced mechanism.

5 comments
AuthorsOpenReview ↗

Response to Comments on Endpoint Conditioning and Experimental Comparisons

We sincerely thank the reviewer for the time and care devoted to our work. We appreciate the recognition of our unified framework and the helpful suggestions. We apologize for the unclear forecasting explanation and limited baseline comparisons. The reviewer’s comments have helped us greatly improve the manuscript’s clarity and quality. We have clarified the endpoint choice and post-generation calibration, expanded temporal comparisons, and added a mass-normalized balanced control. We have submitted the revised manuscript alongside these responses and welcome further discussion.

Q1 & W1. How do the endpoint-conditioned equations operate when the 7 dpa endpoint is unobserved, and what mechanism predicts the future state from observations only up to 3 dpa?

A1. Figure 4e examines reference-guided recovery in planarian regeneration. The regenerative setting motivates using the observed 0 hpa state as a biological reference for restoration of the original tissue architecture. We therefore take 3 dpa, the last observed stage before 7 dpa, as the starting state and 0 hpa as the recovery pseudo-endpoint. The bridge evolves from 3 dpa, with the endpoint-conditioned terms supplied by the observed 0 hpa reference. Only these two stages are used for representation learning and bridge fitting.

The fitted bridge generates spatial–molecular states along this recovery path at normalized progress ss. After generation, comparison with measured 7 dpa selects the displayed state at s=0.55s=0.55 and provides approximate time calibration along the path. The reported 7 dpa result therefore measures agreement after calibration. The approximately 8.5 and 9.5 dpa outputs are generated farther along the same path, with their times assigned through this calibration. They are conditional predictions without measured counterparts for validation.

Section 3.3, the Fig. 4 caption, and Appendix C.1 clarify the endpoint choice and post-generation calibration. They also distinguish this reference-guided recovery experiment from the independent held-out reconstruction tasks in Fig. 4b–d.

Q2 & W2. Can stronger temporal reconstruction baselines, such as optimal transport or other generative methods, be included beyond linear interpolation?

A2. We thank the reviewer for this suggestion. We have added OT, Partial OT, and WGAN-GP alongside Linear interpolation in Fig. 4d. The comparison covers four held-out targets at 6 hpa, 12 hpa, 24 hpa, and 3 dpa. For each target, all methods use the same pair of adjacent observed stages. The target is excluded from model fitting, including HyperBridge’s representation learning and bridge fitting.

HyperBridge achieves the highest program PCC and lowest normalized RMSE at all four targets, supporting improved intermediate-stage reconstruction against the expanded baseline set. Section 3.3 presents these results, and Appendices C.1, C.4, and C.5 describe the tasks, metrics, and baseline implementations. This held-out comparison is separate from the reference-guided recovery experiment in Fig. 4e.

Q3 & W2. How is the equal-mass balanced ablation reconciled with unequal endpoint masses, and what does a mass-normalized balanced comparison show?

A3. We thank the reviewer for this helpful suggestion. We agree that a mass-normalized balanced baseline provides a more appropriate comparison in the unequal-mass setting because it retains all endpoint observations. We have added this control to the ablation study in the main text, alongside the original Fixed Mass control, and report the full results in Appendix C.6.

The two controls address endpoint-mass differences in distinct ways. Fixed Mass resamples the endpoints to a common mass before balanced bridge fitting. Mass-normalized retains all endpoint observations and normalizes each endpoint measure to unit mass. We compare both controls with full HyperBridge on Allen S100 under the same training budget and evaluation protocol. For mass evaluation, the Mass-normalized output is rescaled using a cell-count estimate derived from the visible S99 and S101 sections.

Table 1. Unbalanced-dynamics ablations on Allen S100

VariantCell PCC ↑Gene PCC ↑MSE ↓ACC ↑NMI ↑ARI ↑Mass err. ↓
Full HyperBridge0.49370.18740.14840.35710.18720.10170.0522
Fixed Mass0.41910.13790.17830.31670.10110.07440.6975
Mass-normalized0.44970.16800.15910.34030.09720.09720.2362

The Mass-normalized control retains all endpoint observations and improves expression reconstruction and mass error relative to Fixed Mass, providing a more informative balanced comparison. Overall, the expanded comparison supports the effectiveness of HyperBridge’s unbalanced dynamics, with the full model outperforming both balanced controls across all reported metrics.

Dear Reviewer,

Thank you again for your feedback. As the discussion period is coming to a close, we would greatly appreciate it if you could let us know whether our responses and revisions have adequately addressed your concerns. Please let us know if there's anything further we can provide or discuss. We are thankful for your time and consideration and look forward to your response.

Best regards,

The Authors

AuthorsOpenReview ↗

Clarification on Figure Renumbering

Dear Reviewer,

Thank you again for your feedback.To improve readability, the previous Fig. 4 has been renumbered as Fig. 7 in the latest revised manuscript. Accordingly, Fig. 4d and Fig. 4e referenced in our earlier response now correspond to Fig. 7d and Fig. 7e. The experiments and conclusions remain unchanged.

We would greatly appreciate it if you could let us know whether our responses and revisions have adequately addressed your concerns. Please let us know if there is anything further we can clarify or discuss. We are grateful for your time and consideration.

Best regards,

The Authors

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

The paper proposes HyperBridge for reconstructing unobserved tissue states across anatomical depth and biological time. It combines an unbalanced stochastic bridge with activity probabilities over fixed computational slots and a dynamic spatial-molecular hypergraph. Experiments cover multiple spatial-omics platforms, dense Allen MERFISH atlas reconstruction, and planarian regeneration.

Critical strengths
  1. The paper presents a coherent generative formulation for reconstructing unpaired tissue distributions with unequal represented mass. Activity probabilities and endpoint-conditioned hypergraph messages offer a promising way to jointly model changing mass and spatial-molecular organization without requiring cell-wise correspondence.

  2. The experiments span several measurement technologies and both spatial and temporal reconstruction. The reported comparisons and component ablations provide encouraging evidence for the proposed formulation, while the paper appropriately distinguishes population-level reconstruction from individual cell-lineage inference.

Critical weaknesses
  1. The general machine-learning contribution could be articulated more clearly. The paper would benefit from distinguishing its methodological advances over existing unbalanced bridges and dynamic hypergraph models, and explaining which aspects of the formulation are applicable beyond spatial omics. This would make its relevance to the broader representation-learning audience easier to assess.

  2. The empirical scope is broader than the quantitative validation. Cross-platform results are not reported uniformly, the Allen quantitative evaluation centers on one held-out section, and performance as observation gaps increase is not systematically examined. The forecasting mechanism also needs clearer specification, together with a persistence baseline. These limitations qualify the broader claims but do not negate the promise of the approach.

Questions for authors
  1. What is the main methodological advance over existing unbalanced bridge and dynamic hypergraph approaches, and which assumptions or components are specific to spatial omics?

  2. How does reconstruction performance change with increasing gaps between observed sections? Are additional held-out-section results or comparisons with directly relevant generative reconstruction methods available?

  3. How is the 7 dpa forecast generated without a measured terminal endpoint, and how does it compare with retaining the last observed state?

4 comments
AuthorsOpenReview ↗

Response to Comments on Methodology and Reconstruction Robustness

We sincerely thank the reviewer for the thoughtful and encouraging assessment of our work. The comments have helped us greatly improve the clarity and quality of the manuscript. We have clarified the methodological contribution and expanded the reconstruction experiments across observation gaps and anatomical regions. We have submitted the revised manuscript alongside these responses and welcome further discussion.

Q1 & W1. What is the methodological advance over existing unbalanced bridges and dynamic hypergraph models, and which assumptions or components are specific to spatial omics or applicable to broader machine learning?

A1. The main methodological advance is a structure-conditioned unbalanced bridge for reconstructing dynamic 3D tissue states from sparse, unpaired spatial-omics observations. Each cell or spot belongs to a three-dimensional tissue context that cannot be captured by molecular similarity or within-section proximity alone. HyperBridge models these relationships through spatial–molecular hyperedges across registered sections. During bridge fitting, the hypergraph is rebuilt from evolving states, and its activity-weighted context is aligned with the observed terminal context to guide both state evolution and represented-mass changes.

Existing unbalanced bridge formulations accommodate distributional transport and unequal mass but do not explicitly encode evolving higher-order tissue organization. Dynamic hypergraph methods capture changing relationships but do not themselves define a generative bridge between unpaired tissue states. HyperBridge makes these relationships part of the reconstruction dynamics: structure guides generation, and generated states update structure. Its spatial-omics-specific assumptions are registered coordinates, comparable molecular features, and spatial-molecular compatibility as a tissue prior, without requiring cell-wise correspondence.

W2. Does the cross-platform quantitative validation adequately support the breadth of the empirical claims?

A2. The cross-platform results in Fig. 2 support our claims about HyperBridge’s reconstruction performance across molecular modalities, profiling technologies, and spatial resolutions. The four experiments address different reconstruction tasks, ranging from filling large inter-section gaps to recovering fine tissue structure.

In DLPFC Visium (Fig. 2a), virtual S1.5 bridges the wide S1–S2 gap with better cortical continuity and fewer irregular voids than isoST and SpatialZ. In cSCC Deep-STARmap (Fig. 2b), reconstruction of three held-out layers from four observed layers achieves the closest cell-population proportions, highest mean Cell PCC, and lowest mean MSE and Chamfer distance. In IMC (Fig. 2c), intermediate layers between 15 measured sections densify the protein volume, with more continuous CK5-high and SMA-high structures and less banding in XZ views than isoST and SpatialZ. In mouse-brain Xenium (Fig. 2d), reconstruction of held-out S2 from S1 and S3 better recovers the curved hippocampal cell-population arrangement and achieves the highest Cell PCC of 0.522. We have clarified these experiments in Section 3.1 and the Fig. 2 caption, with evaluation details in Appendices C.1 and C.3–C.5.

Q2 & W3. How does reconstruction performance change as the gaps between observed sections increase?

A3. To more fully assess how observation gap affects reconstruction performance, we added experiments at S60 and S47, keeping each target fixed while progressively widening the gap between its two observed flanking sections. Each target was excluded from training throughout. S60 is dominated by striatum (37%) and isocortex (32%), with relatively stable regional proportions across its flanking sections as the gap widens. S47 contains isocortex (39%), striatum (36%), and olfactory areas (11%), with greater changes in endpoint regional composition.

Table 1. S60 reconstruction across observation gaps (Cell PCC ↑)

MethodGap 1Gap 3Gap 5Gap 7
HyperBridge0.5480.5440.5300.515
isoST0.2730.2550.2190.184
SpatialZ0.3550.3480.3090.305
Linear0.3770.3740.3410.336

HyperBridge achieves the highest Cell PCC at every gap at S60 and shows the smallest relative decline.

Table 2. HyperBridge performance and relative change across observation gaps

GapS60 Cell PCC ↑Change from Gap 1S47 Cell PCC ↑Change from Gap 1
10.5480.0%0.5390.0%
30.544−0.7%0.484−10.2%
50.530−3.3%0.485−10.0%
70.515−6.0%0.462−14.3%

The two targets have similar accuracy at Gap 1, but S47 shows a larger decline as the gap widens. By Gap 7, its Cell PCC decreases by 14.3%, compared with 6.0% at S60. This difference is consistent with the greater variation in regional composition around S47, suggesting that gap sensitivity depends on both section spacing and tissue variation between the observed endpoints.

AuthorsOpenReview ↗

Response to Comments on Additional Evaluation and Temporal Prediction

Q3 & W4. Can additional held-out-section evaluations or comparisons with relevant generative reconstruction methods be provided?

A4. To strengthen the held-out evaluation and baseline comparison, we added S50, S85, and S125 to the existing S100 benchmark and introduced an optimal transport (OT) baseline. Each target section was excluded from model fitting.

SectionMain regionsMethodCell PCC ↑Gene PCC ↑MSE ↓
S50Striatum, isocortexHyperBridge0.5610.1890.128
isoST0.2620.0280.285
SpatialZ0.3620.0820.217
Linear0.3840.0850.194
OT0.3410.0740.221
S85Isocortex, thalamus, hypothalamusHyperBridge0.5020.2040.147
isoST0.2700.0520.246
SpatialZ0.3180.0950.235
Linear0.3230.0940.220
OT0.2820.0740.244
S100Hippocampal formation, midbrain, isocortexHyperBridge0.4940.1870.148
isoST0.2640.0460.247
SpatialZ0.3080.0800.239
Linear0.3200.0830.220
OT0.2730.0630.249
S125Cerebellum, medullaHyperBridge0.5130.1210.148
isoST0.3030.0280.268
SpatialZ0.3390.0560.231
Linear0.3470.0560.217
OT0.3000.0400.244

HyperBridge achieves the highest Cell PCC and Gene PCC and the lowest MSE at all four held-out sections, including against the added OT baseline. These results extend the evidence beyond S100 and support consistent improvements in molecular reconstruction across the evaluated targets.

Q4 & W5. How is the 7 dpa forecast generated without a measured terminal endpoint?

A5. The planarian dataset captures tissue regeneration, motivating a recovery path toward the original tissue architecture. In Fig. 4e, we therefore use the observed 3 dpa state as the starting state and the observed 0 hpa state as a recovery pseudo-endpoint. The bridge evolves from 3 dpa, with endpoint-conditioned information supplied by the 0 hpa reference. Only these two stages are used for representation learning and bridge fitting.

The fitted bridge generates spatial–molecular states along this recovery path at normalized progress ss. After generation, comparison with measured 7 dpa selects the displayed state at s=0.55s=0.55 and provides approximate time calibration along the path. The reported 7 dpa result therefore measures agreement after calibration. The approximately 8.5 and 9.5 dpa outputs are generated farther along the same path, with their times assigned through this calibration. They are conditional predictions without measured counterparts for validation.

Section 3.3, the Fig. 4 caption, and Appendix C.1 clarify the endpoint choice and post-generation calibration. They also distinguish this reference-guided recovery experiment from the independent held-out reconstruction tasks in Fig. 4b–d.

Q5 & W6. How does the 7 dpa forecast compare with a persistence baseline that retains the last observed state?

A6. We thank the reviewer for suggesting this comparison. We evaluated a persistence baseline that retains the observed 3 dpa state and compared it with the selected HyperBridge state against measured 7 dpa. Pearson correlation was calculated across the mean-expression values of 512 genes.

StatePearson rr with measured 7 dpa
Persistence baseline (observed 3 dpa)0.897
HyperBridge (selected state at s=0.55s=0.55)0.915

The HyperBridge-generated state at s=0.55s=0.55 achieves a correlation of 0.915, compared with 0.897 for the unchanged 3 dpa state, showing closer agreement with the measured 7 dpa mean-expression profile. The bridge also generates later tissue states corresponding to approximately 8.5 and 9.5 dpa, illustrating HyperBridge’s capacity to model tissue recovery and predict subsequent states using a biological recovery reference.

Dear Reviewer,

Thank you again for your feedback. As the discussion period is coming to a close, we would greatly appreciate it if you could let us know whether our responses and revisions have adequately addressed your concerns. Please let us know if there's anything further we can provide or discuss. We are thankful for your time and consideration and look forward to your response.

Best regards,

The Authors

AuthorsOpenReview ↗

Clarification on Figure Renumbering

Dear Reviewer,

Thank you again for your helpful feedback. To improve readability, the four datasets previously presented in Fig. 2a–d are shown separately in the latest revised manuscript: DLPFC in Fig. 2, cSCC in Fig. 3, IMC in Fig. 4, and Xenium in Fig. 5. The previous Figs. 3 and 4 have consequently been renumbered as Figs. 6 and 7. Thus, the Fig. 4b–e references in our earlier response now correspond to Fig. 7b–e. This reorganization concerns presentation only, the experimental results and conclusions remain unchanged.

We would greatly appreciate it if you could let us know whether our responses and the latest revision have adequately addressed your concerns. Please let us know if there is anything further we can clarify or discuss. We are grateful for your time and consideration.

Best regards,

The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Area Chair (AC) and Reviewers,

We appreciate the AC's coordination of the review process and the reviewers' constructive feedback.

Our detailed rebuttal was submitted on November 15, together with additional experiments, clarifications, and a substantially revised manuscript.

Although the discussion period concluded on November 18, we would be grateful if the remaining reviewers could, if possible, briefly indicate whether our responses and revisions have addressed their concerns. We also respectfully ask the AC to consider the revised manuscript and the complete discussion record in the final assessment.

We remain available to provide any further clarification.

Best regards,

Authors of Submission #44394

AuthorsOpenReview ↗

Summary of Contributions and Responses to Reviewer Feedback

We thank the reviewers for their careful evaluation and constructive feedback. We are encouraged that they recognize the value of a unified generative framework for spatial and temporal tissue-state reconstruction and the novelty of using a dynamic 3D hypergraph to guide an unbalanced stochastic bridge between unpaired tissue distributions. These assessments highlight HyperBridge’s potential to move beyond pointwise interpolation and approaches designed for spatial reconstruction alone, using a shared formulation to recover entire tissue states at unmeasured anatomical depths or biological stages. We sincerely appreciate the reviewers’ recognition of the following contributions:

  • investigate an important problem in spatial omics by asking how to reconstruct unmeasured tissue states from sparse observations across anatomical depth or biological time. Recovering unmeasured tissue sections is “a compelling direction for the field” (Reviewer YDNv) given the cost and experimental difficulty of dense tissue profiling.

  • develop a mathematically principled and biologically motivated generative strategy that represents each tissue snapshot as a joint distribution of spatial positions and molecular profiles and learns an unbalanced stochastic bridge between unpaired observations. This “coherent generative formulation” (Reviewer nXNH) accommodates differences in represented mass and supports reconstruction of both missing sections and unmeasured biological stages (Reviewer obkU).

  • adopt a “novel and creative” approach (Reviewer YDNv) to embedding evolving higher-order tissue structure in the bridge dynamics. HyperBridge rebuilds the 3D hypergraph as the generated tissue state changes and uses the updated relationships among cells or spots to guide subsequent reconstruction.

  • provide strong experimental validation across diverse spatial-omics technologies and tissue morphologies (Reviewers nXNH and YDNv). For spatial reconstruction, held-out section tests showed “improved reconstruction accuracy and preservation of tissue structure” (Reviewer obkU). For temporal modeling, HyperBridge recovered molecular patterns at four held-out stages and generated reference-guided forecasts of later, unmeasured tissue states.

Regarding concerns about methodological clarity and experimental presentation, we revised the manuscript within the original HyperBridge formulation, whose core contribution is to use a dynamic 3D hypergraph to guide an unbalanced stochastic bridge. We expanded the description of the model and its computational workflow, clarified the evaluation protocols, and reorganized the figures to make the results easier to follow.

For issues related to model robustness and empirical validation, we evaluated how section spacing affects spatial reconstruction, broadened the baseline comparisons for both spatial and temporal reconstruction, added a balanced control to assess the role of unbalanced dynamics, and clarified the scope of HyperBridge’s contribution in the revised Conclusion.

In summary, HyperBridge advances an emerging direction in spatial omics by using evolving higher-order tissue relationships to reconstruct unobserved tissue states from sparse snapshots. The same formulation produced coherent 3D tissue maps across anatomical depth and supported held-out stage reconstruction and reference-guided forecasts of later tissue states across biological time.

We have incorporated the reviewers’ constructive feedback and are confident that the revisions and additional experiments directly address the questions and concerns raised during the discussion. We sincerely hope the final assessment will consider our contribution, its relevance to the ICLR community, and the evidence within the scope of the problem addressed.

Regards from the authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.