acceptodds
Under review as a conference paper at ICLR 2027

NicheTime: Spatiotemporal Generative Modeling of Cellular Microenvironment Evolution via Dual-Scale Optimal Transport

Abstract

Learning continuous-time evolution of structured systems from unpaired observations remains a fundamental challenge in generative modeling. Existing trajectory inference and generative approaches often assume conservative dynamics or overlook the hierarchical organization and directional constraints of spatiotemporal biological systems. We introduce NicheTime, a continuous-time generative framework for modeling non-conservative spatiotemporal dynamics through transport-guided flow matching. NicheTime addresses three key challenges: unknown correspondence across temporal observations, multi-scale structural evolution, and irreversible dynamics with variable mass flow. NicheTime develops a dual-scale optimal transport coupling that first aligns coarse-grained structures and then refines local correspondences, while allowing flexible mass variation to capture non-conservative transitions. Based on these transport plans, we propose a direction-aware Graph Transformer flow matching architecture that incorporates temporal ordering into message propagation and jointly generates system states and spatial configurations through neural ODE integration. We evaluate NicheTime on three time-resolved spatial transcriptomics datasets covering embryonic development, neural development, and aging. Extensive experiments show that NicheTime improves trajectory reconstruction, state fidelity, and spatial consistency over existing approaches, highlighting the potential of transport-guided flow matching for learning complex structured dynamics.

Primary area:
generative models
Keywords:
flow matching, optimal transport, spatiotemporal dynamics, graph transformers, spatial transcriptomics

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_McGNOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

NicheTime proposes a framework for modeling cellular microenvironment evolution using dual-scale optimal transport and pseudotime-guided graph flow matching. It extends niche-level OT with cell-level correspondence estimation and introduces a graph-based architecture that models cellular neighborhoods and developmental directionality. Experiments on three spatial transcriptomics datasets show improvements over NicheFlow on several metrics.

Critical strengths
  1. Biologically motivated optimal transport. The proposed dual-scale OT formulation is well motivated. Matching microenvironments first and then refining correspondences at the cellular level reflects the hierarchical organization of tissues. Semi-relaxed OT is also a sensible choice, as cellular populations change over time and strict source marginal constraints may be inappropriate. Table 2 provides a useful comparison of transport formulations with different marginal constraints at both scales.
  2. Explicit modeling of cellular neighborhoods. The PTG architecture incorporates spatial relationships between neighboring cells and uses pseudotime to guide directional message passing. These inductive biases are relevant for modeling tissue development, where local cellular interactions and developmental ordering matter. Tables 3 and 4 provide convincing ablations showing improvements in expression reconstruction over standard graph architectures and variants without pseudotime guidance.
Critical weaknesses
  1. The contribution of cell-level OT is not sufficiently isolated. The architectural ablations provide strong evidence for PTG, but Tables 1–4 do not evaluate PTG with niche-level OT alone. Consequently, it remains unclear how much the additional cell-level transport stage contributes to the reported improvements. This is an important missing comparison given that dual-scale OT is a central contribution of the paper.
  2. Unclear evaluation protocol and incomplete spatial evaluation. Appendix C defines the evaluation metrics but provides limited details about the held-out observations and whether evaluation is performed on individual microenvironments or complete tissue reconstructions. Furthermore, although PSD and SPD are reported for internal ablations, a direct comparison against NicheFlow is missing. This is particularly relevant because Figure 5 shows that PTG improves spatial precision but substantially reduces tissue coverage compared with standard graph architectures on ABD and MBA.
  3. Insufficient methodological details. The Post-GNN refinement used in the main experiments is not described in the main methodology, and its contribution is not isolated through an ablation. This makes it difficult to assess how much the refinement contributes to the reported improvements.
Questions for authors
  1. Contribution of cell-level OT. Could the authors provide an ablation retaining niche-level OT and the PTG architecture, but omitting the additional cell-level OT stage? For example, source and target cells could be sampled independently within matched niches rather than according to a cell-level transport plan, while retaining the same architecture and conditioning mechanism. This would help isolate the contribution of cell-level OT.
  2. Evaluation protocol and spatial reconstruction. Appendix C states that metrics are computed on held-out target snapshots using a shared OT matching protocol. Could the authors clarify which observations are used during training and which are held out for evaluation? How are generated microenvironments assembled into complete tissue predictions? Are MSE, MAE, and 1NN-F1 computed on complete reconstructed tissue slices or individual generated microenvironments? Could the authors also report PSD and SPD against NicheFlow and clarify whether the baseline was retrained or evaluated using existing checkpoints?
  3. Post-GNN refinement. Appendix F mentions Post-GNN refinement using a dynamic k-NN graph, and Table 9 indicates that it is enabled in the main experiments. Could the authors specify its architecture, neighborhood size, and graph-rebuilding frequency? Since this component is part of the final model, an ablation without Post-GNN refinement would help establish how much it contributes to the reported improvements.
2 comments
AuthorsOpenReview ↗

Response to Reviewer McGN

We appreciate the reviewer's specific suggestions. The revised manuscript includes both requested component ablations and a direct spatial comparison with NicheFlow.

Cell-level OT ablation. Table 1 provides the requested controlled comparison between dual-scale OT and niche-only OT. Both variants retain niche-scale SemiOT, PTG, Post-GNN, and the same training settings; only the full model additionally enables cell-level SemiOT. On ABD, adding cell-level OT reduces MSE from 1.5989 to 1.4189 and increases 1NN-F1 from 0.4736 to 0.6659. All nine reported metric estimates favor the full model, although the magnitude of improvement varies across datasets. This supports a contribution from cell-level OT beyond niche-level alignment under the same graph architecture.

Evaluation and spatial reconstruction. Appendix C now clarifies the evaluation protocol. Training uses online spatially balanced niche sampling, whereas evaluation uses fixed grid centers. Both access the same observed snapshots, so the experiments assess observed-stage reconstruction rather than generalization to independently held-out tissues. At inference, target niche cardinality determines the generated cell count, but target expression and coordinates are not used as generator conditioning. MSE/MAE are computed index-wise without additional OT matching at evaluation time.

NicheFlow was trained from scratch on the same preprocessed datasets. Both methods share identical evaluation centers, generated cell counts, and metric implementations, while retaining model-specific training configurations. Spatial predictions are aggregated without deduplication; PSD uses the complete observed target snapshot as reference, whereas SPD uses observed evaluation-niche coordinates. Table 5 shows lower PSD for NicheTime on ABD and MBA, but higher SPD across all three datasets, indicating weaker coverage of the sampled evaluation niches.

Post-GNN refinement. Section 3.4 and Appendices F–G detail the one-layer Post-GNN with k=10 spatial neighbors and hidden dimension 128. The target-side backbone graph is refreshed every five Euler steps, whereas Post-GNN rebuilds its spatial graph from the current decoder predictions at each forward pass. Table 3 directly isolates the refinement effect: removing Post-GNN increases ABD MSE from 1.4189 to 1.5996 and reduces 1NN-F1 from 0.6659 to 0.4122. All nine reported metric estimates favor the inclusion of Post-GNN.

Thank you for the additional experiments and clarifications!

The new Tables 1 and 3 provide convincing evidence that both cell-level OT and the Post-GNN refinement contribute meaningfully to the overall improvements. I also appreciate the clarification of the evaluation protocol and the additional spatial comparison with NicheFlow.

Overall, I believe my main concerns have been addressed, and I have raised my score to 3.

Reviewer_FJRhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a continuous-time generative modeling method for the evolution of cellular microenvironments from unpaired spatial transcriptomics snapshots. It based on: (0) dual-scale optimal transport (OT), (1) a graph-based representation of cellular niches, and (2) a pseudotime-guided (PTG) message-passing architecture. The framework is evaluated on three datasets with improvements over NicheFlow on reconstruction metrics.

Critical strengths

The paper addresses an interesting problem: learning how cellular microenvironments evolve when observations from different time points cannot be directly matched. The combination of hierarchical optimal transport and graph-based modeling looks like a right solution for it, from the intuition point. In general the solution provides a reasonable way to incorporate both spatial organization and temporal information.

Critical weaknesses

0). While relaxed OT constraints accommodate unequal transport weights, it is not clear whether the generative model can actually predict changes in cell population sizes.

1). Reported improvements appear to depend heavily on the introduction of pseudotime information.

2). Reported results demonstrate generation of target states associated with observed time points rather than generalization to unseen stages.

3). Related recent methods are missing from the baselines: MUST-FM (multiscale supervised unbalanced OT + flow matching), scFM (OT-informed bi-directional flow matching), ContextFlow (context-aware OT couplings for spatial priors), and MIOFlow 2.0 (spatially conditioned, non-conservative NDE/SDE dynamics)

Questions for authors
  1. Can the model predict changes in cell population sizes or cell-type abundances, and how is the number of generated cells determined during inference? Also, is there any evidence of generalization to previously unseen biological time points?

  2. How are pseudotime values ptipt_i obtained for each dataset during training? Report sensitivity to noisy/misspecified pseudotime (e.g., shuffled, perturbed) and whether PTG still helps when pseudotime is weak, I think is important.

  3. MUST-FM, scFM, or ContextFlow, in detail: why such comparisons were not feasible?This seems crucial.

  4. Could the authors compare single-scale and dual-scale OT while keeping the full PTG architecture unchanged? Additionally, could they compare PTG against standard graph architectures that receive the same pseudotime information?

1 comment
AuthorsOpenReview ↗

Response to Reviewer FJRh

We thank the reviewer for raising important questions about the scope of the model and the role of pseudotime.

Population size and unseen stages. Semi-relaxed OT allows unequal transport weights but does not explicitly predict future cell counts. In our evaluation, the generated cell count is determined by the observed target niche size. Likewise, the continuous flow parameter describes generation between observed endpoints, rather than demonstrated interpolation or extrapolation to unseen biological stages. We clarify these distinctions in Appendix C and the limitations.

Pseudotime sensitivity. Pseudotime is estimated using Monocle 3 and rank-normalized before guiding message passing. New experiments in Table 4 examine Gaussian perturbations, shuffling, and removal of pseudotime guidance, with each modification applied consistently during training and evaluation. Shuffling increases MSE on ABD (1.4189 to 1.6658) and MBA (1.2637 to 1.4261), supporting the value of informative temporal ordering. MED behaves differently, with improved expression MSE under shuffling. We interpret these results as dataset-dependent sensitivity rather than uniform robustness to pseudotime errors.

Related methods. MUST-FM, scFM, ContextFlow, and MIOFlow 2.0 are relevant approaches to transport-based dynamics modeling. Their native formulations and evaluation protocols are not identical to our joint niche-level expression-and-coordinate generation task. A harmonized end-to-end comparison would require task-specific adaptation, which we have not conducted. We therefore restrict our empirical claims to the evaluated baseline and controlled ablations.

Controlled comparisons. Revised Table 1 isolates cell-level SemiOT under the same PTG and Post-GNN architecture, supporting a contribution from correspondence modeling beyond pseudotime guidance. Table 4 separately examines the effect of pseudotime within PTG. Table 3 compares PTG with standard GCN, GAT, and GIN architectures, providing an architectural comparison rather than an equal-pseudotime-input control.

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

This paper proposes NicheTime for modeling cellular microenvironments from unpaired spatial transcriptomics snapshots. The method constructs correspondences through optimal transport at both the niche and cell levels, encodes source microenvironments with a pseudotime-guided graph network, and uses conditional flow matching to jointly generate expression features and spatial coordinates. Experiments cover mouse embryogenesis, axolotl brain development, and mouse brain aging. The strongest part of the paper is the evidence for the PTG module. It improves expression reconstruction across datasets, and removing pseudotime guidance substantially reduces performance. However, the relationship to NicheFlow and several aspects of the experimental protocol need clarification before the contribution and generalization results can be assessed confidently.

Critical strengths
  1. The problem is relevant. Modeling expression and spatial context together is a reasonable approach to tissue development. The proposed components address two important aspects of this problem: correspondence across snapshots and interactions within local microenvironments.

  2. The ablations support the PTG module. In Table 3, PTG improves over NONE in eight of the nine dataset–metric combinations, including MSE and MAE on all three datasets. Table 4 provides further evidence: removing pseudotime guidance increases MSE from 1.4189 to 1.8746 on ABD and from 1.2637 to 1.6896 on MBA. Comparisons with GCN, GAT, and GIN also suggest that simply adding a standard graph network does not provide the same reconstruction gains.

  3. The method achieves meaningful improvements in some settings. Compared with NicheFlow, the full model improves all three reported metrics on ABD and reduces expression reconstruction errors on MBA. These results motivate further investigation of developmental ordering in microenvironment generation.

Critical weaknesses
  1. The relationship to NicheFlow needs clarification. Section 4.2 uses NicheFlow versus NicheTime_NONE to assess dual-scale OT, but does not establish whether other components and training settings are identical. Moreover, NONE has worse MSE and MAE on all three datasets, contrary to the positive interpretation in the text.

  2. The external comparison is narrow, and is the major weakness. NicheFlow is an appropriate baseline, but the paper should explain why related spatiotemporal OT methods, such as moscot and DeST-OT, are excluded and whether they can help evaluate the correspondence component.

  3. The evaluation protocol is underspecified. The main text does not clearly explain the train/test split or how generated and real cells are paired for MSE and MAE. This makes the level of generalization difficult to assess.

  4. The improvements are mixed. NicheTime improves ABD results, but worsens expression errors on MED and 1NN-F1 on MBA. These differences need explanation. Given the emphasis on spatial organization, the main experiments also need a direct quantitative comparison of spatial reconstruction with NicheFlow.

Questions for authors
  1. Is NicheTime implemented directly on top of NicheFlow? What exactly changes between NicheFlow, NicheTime_NONE, and NicheTime_PTG, and are their data splits and training settings identical?

  2. Why is NicheFlow the only baseline? Could other spatial or temporal modeling methods be compared on this task?

  3. How are training and test observations separated, and how are generated and real cells paired when computing MSE and MAE?

  4. How should the NicheTime_NONE results be reconciled with the interpretation of dual-scale OT in Section 4.2?

  5. Can the authors provide a direct spatial reconstruction comparison with NicheFlow and explain the higher expression error on MED and lower 1NN-F1 on MBA?

2 comments
AuthorsOpenReview ↗

Response to Reviewer PTQN

We thank the reviewer for highlighting the need for more controlled comparisons and a clearer evaluation protocol.

NicheFlow and cell-level OT. NicheTime builds on the NicheFlow codebase, but comparing NicheFlow with NicheTime_NONE changes both correspondence modeling and network architecture. We agree that this comparison cannot isolate cell-level OT. Revised Table 1 therefore provides a controlled comparison between dual-scale OT (SemiOT at both scales) and niche-only OT (niche-scale SemiOT retained, cell-level OT disabled), while keeping PTG, Post-GNN, and the remaining training settings unchanged. Adding cell-level OT reduces ABD MSE from 1.5989 to 1.4189 and increases 1NN-F1 from 0.4736 to 0.6659. Improvements are also observed on MBA and MED, although the MED expression gain is modest.

Related baselines. We use NicheFlow as our main end-to-end baseline for joint expression-and-coordinate generation. Table 2 provides a complementary analysis by varying balanced, unbalanced, and semi-relaxed transport at both scales while retaining the same generative framework. This isolates the effect of transport formulations but is not an end-to-end comparison with moscot or DeST-OT.

Evaluation protocol. Appendix C now clarifies that we evaluate observed-stage reconstruction, not unseen-stage generalization. Training uses online spatially balanced niche sampling, while evaluation uses fixed grid centers; both access the same observed snapshots. Target niche size determines the number of generated cells, but target expression and coordinates are not used as generator conditioning. MSE/MAE are computed index-wise without evaluation-time OT matching. NicheFlow and NicheTime use the same preprocessed data, evaluation centers, and metric implementations.

Expression and spatial reconstruction. Table 5 provides the direct spatial comparison. NicheTime achieves lower PSD on ABD and MBA but higher SPD across all datasets. On MED, NicheFlow retains lower expression errors; on MBA, NicheTime improves expression reconstruction but has lower 1NN-F1. We therefore distinguish the benefits of cell-level correspondence from uniform superiority across datasets and metrics.

AuthorsOpenReview ↗

Clarification on OT Formulations and Related Methods

We would like to further clarify the relationship between our OT experiments and existing methods such as moscot and DeST-OT.

NicheTime independently implements established OT formulations within its dual-scale framework. Specifically, WOT uses balanced marginal constraints, corresponding to a transport formulation supported by moscot, while SemiOT relaxes the source marginal and preserves the target marginal, following the semi-relaxed OT principle employed in DeST-OT. These connections concern the underlying transport formulations, rather than reproducing the complete objectives or implementations of either method.

Regarding end-to-end comparisons, moscot and DeST-OT primarily address spatiotemporal alignment, whereas NicheTime jointly generates target-stage cellular expression and spatial coordinates. Evaluating their native outputs using our generative reconstruction metrics would require an additional generative model, introducing architectural choices beyond the original methods.

Table 2 is therefore an internal controlled comparison of OT formulations at both niche and cell scales, with the generative architecture and evaluation protocol held fixed. This allows us to examine how different transport constraints affect downstream generation without conflating their effects with changes in the generative model.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Additional Experiments

We thank all reviewers for their constructive feedback. We have revised the manuscript to address the main concerns regarding component contributions, evaluation protocol, and spatial reconstruction.

Controlled ablations. Revised Table 1 isolates cell-level OT by comparing dual-scale SemiOT with niche-only SemiOT under the same PTG and Post-GNN architecture. Enabling cell-level OT improves all nine reported metrics across three datasets. Table 3 further isolates Post-GNN refinement, also showing consistent improvements over its ablated counterpart. Together, these controlled comparisons strengthen the evidence for the contributions of correspondence modeling and graph refinement beyond pseudotime guidance alone.

Pseudotime and spatial evaluation. Table 4 examines sensitivity to perturbed and shuffled pseudotime, revealing dataset-dependent behavior. We also added a direct spatial comparison with NicheFlow (Table 5). While NicheTime improves PSD on ABD and MBA, its higher SPD reveals a spatial precision–coverage trade-off, which we now discuss explicitly.

Evaluation clarity. Appendix C now details the observed-stage reconstruction protocol, including training/evaluation sampling, target-informed niche sizing, and metric computation. We also clarify the scope of our comparisons and distinguish continuous generative flow from unseen-stage generalization or population-size prediction.

We hope these additional results and clarifications help address the reviewers' main concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.