- 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.
- 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.
- 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
- 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.
- 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.
- 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.
- Eq. (7). The notation is a bit counterintuitive. First, is indexed by time only; then it becomes a function of the whole state, including positions, latents, and switch probabilities.
- 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.
- Eq. (10). The distance is not defined, nor is the parameter . There is no mention of how state distances are computed: whether coordinates and latents are concatenated.
- 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.
- 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
- As mentioned above, the amount of written detail describing the experiments is insufficient. Panels are barely described and left to interpretation.
- 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.
- 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.
- 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?
- 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
-
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.
-
Was any heuristic required to avoid that optimization pays too much attention to matching latent states instead of coordinates?
-
L208-214. The reference process is not defined, and it is not self-explanatory. Is it standard Brownian motion?
-
It is not clear to me how the rates and 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.
-
What is the difference between KL$_+$ and a standard KL?
11 comments
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 . 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 , as stated in Sec. 2.3. The Wiener driver is standard Brownian motion, with noise scaled by . 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 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 and 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 and , 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 , 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.
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.
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), where c=max(N0,N1)/K. Thus, K can be smaller than the observed cell/spot count. The implementation integrates continuous pi(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 kH highest-affinity neighbors satisfying ϵH, including candidates from other registered sections. If none qualifies, the anchor forms a singleton hyperedge. Smaller κr penalizes spatial separation more strongly, while smaller κz 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 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.