acceptodds
Under review as a conference paper at ICLR 2027

LoCoFlow: Local-Context Flow Matching for Single-Cell Dynamics and Fate Inference

37 views

Abstract

Single-cell sequencing profiles cellular states at discrete time points, providing population-level snapshots from which developmental and disease-associated trajectories can be reconstructed. Existing optimal-transport and flow-matching approaches typically infer temporal couplings at the level of individual cells, with limited representation of the local manifold geometry of cell-state space and the uncertainty inherent in lineage diversification. Here, we introduce LoCoFlow, a population-level framework that represents each cell together with its local neighborhood as a point-cloud patch, thereby modeling cellular dynamics on local manifolds rather than between isolated states. This representation supports probabilistic temporal coupling while preserving local cell-state geometry, separates collective population displacement from within-neighborhood variation to capture both coordinated state transitions and cellular heterogeneity, and provides a natural distributional representation of fate uncertainty through changes in the shape and dispersion of local cell populations. We evaluate LoCoFlow across multiple longitudinal single-cell datasets, including the LARRY dataset with experimentally measured lineage-tracing information, and show that it captures complex cellular dynamics and fate uncertainty in both trajectory reconstruction and cell-fate inference tasks.

Primary area:
generative models
Keywords:
Flow Matching, Single-Cell Dynamics, Optimal Transport, Trajectory Inference, Cell Fate Inference

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_mejVOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The authors present LoCoFlow, a flow matching framework for single-cell trajectory and fate inference that models the local neighborhood of each cell as an evolving point cloud. Training has two stages. First, a hierarchical OT, between overlapping patches and then between cells within matched patches, yields probabilistic cell correspondences, and an endpoint-preserving neural bridge regularized by kinetic energy, a manifold-normal penalty, and a radial deformation term generates teacher velocities. Second, these velocities are distilled into a conditional velocity field driven by codes from a hierarchical latent model, with patch-shared and cell-specific Gaussian latents, so that repeated latent draws yield a distribution of outcomes for each starting cell without future observations. Empirical results suggest improved distributional distance on held-out time-point reconstruction (EB, iPSC, C. elegans) and spatial interpolation (Axolotl), and a lower raw Brier score for fate prediction on LARRY, complemented by in silico focal-cell and neighborhood perturbation analyses.

Critical strengths

I think the hierarchical approach is quite interesting, especially in the spatial setting. I also appreciate the fact that multiple datasets were used and the attention to provide some carefully designed biological use cases, rather than only reporting raw scores. I think this aspect is really important for the credibility of the method. Other choices are interesting, such as the generative model on the endpoint context encoding.

Critical weaknesses

Despite the clear strengths, I, unfortunately, believe the paper is not ready for publication. Overall, the clearest miss is the absence of a background section, which makes some choices and methodological descriptions not very clearly motivated. I think another iteration would make the contribution more signficant.

Introduction

  1. L53-54. I personally do not consider constraining trajectory inference point cloud evolution so interesting beyond the spatial transcriptomics application, and would frame that as an ultimate goal. The concept of neighborhood in dissociated transcriptomics is unclear, as measured cells are not physically related. I do see why one would want to move point clouds in time in spatial transcriptomics; I’d find that more fitting as a motivation ground for this paragraph and the whole paper in general.

Methods

  1. Connected to my point above, it is not very clear to me what the advantage of a patch-level OT term is. What is the advantage of matching coarser views of a cell state in dissociated transcriptomics where proximity does not mean co-evolution? I assume that this hierarchical modeling may provide some practical advantage, but it is not clear from the text.
  2. In my opinion, the interpolant in Eq. 4 is not sufficiently described. Both fθf_{\theta} and yy are not described. This gap impairs the understanding of the paper.
  3. Similarly, Eq. 6 is not very clear. What is λ\lambda here? What is ww set to, or how is it derived? In general, I think the paper misses a crucial general intro on manifolds and geometry-informed flow matching. Without it, most of the choices appear arbitrary. I recommend that the authors significantly shorten the Related Works section and dedicate the gained space to a more rigorous methods description.
  4. Similarly, there is no introduction to bridge-based generative models or interpolations. When VωV_{\omega} is first mentioned, it is not clear what it refers to.
  5. The losses should be rewritten as functions of the parameters. This is important to understand what each loss is optimizing. For example, it is not immediately clear that Eq. 6 enables optimization over θ\theta.
  6. I think the paper really requires an algorithmic box or a summary in the main text. In my opinion, the methods sections are disjoints and they should refer to each other. For example, Stage 1 has to explain how subsequent patches are drawn from the above coupling to build the bridge.

Experiments

  1. LoCoFlow is inconsistently spelled. Sometimes it is introduced as Locoflow.
  2. The sources of the different datasets are not cited, except for iPSC reprogramming. W1 and W2 should be addressed in full in their first instance. I assumed they stand for Wasserstein. Baselines are not described.
  3. Baselines differ across experiments. Fate prediction has more of them. This stands out as inconsistent. Is there anything preventing their widespread application?
  4. scDiffEq seems consistently better than LoCoFlow on most of the metrics. Is there any reason why Raw Brier is more relevant than the others? As of now, this result makes the method a little bit less convincing.
  5. L319-323. In my opinion, this requires some introduction to what we expect from the edits. What fate do we expect from different ranges of these genes? It’s hard to assess the result if someone is not an expert in developmental biology. I believe the paper should be self-contained. This also extends to Fig. 3’s bars.
  6. I apologise if this may sound like a hard assessment, but I am sceptical about the results in Table 4. To believe that predictions are reliable, I would need some qualitative layout of the predicted brain slides with compositional labels (eg inferred cell types) on top. I tried OT-CFM on spatiotemporal modeling before, and I found it significantly underperforming. In my opinion, it is a bit worrying that LoCoFlow cannot overcome it. I appreciate the authors’ awareness and honesty in suggesting that this application needs future work, but currently I find the task potentially ill-posed, as interpolating spatial biology is hard without intermediate supervision.
  7. Also, in the axolotl example, the setting and modeling strategy are not introduced. The paper is not cited.
Questions for authors
  1. L 64-65: Why are point clouds interesting for modeling bifurcation? I feel that could be done effectively at a particle level. Am I missing something crucial?

  2. L137-138: Why doesn't matching the marginals enable forecasting giving an initial particle? In my opinion, this paragraph needs a better motivation for using point clouds beyond standard flow point-level flow matching.

  3. Eq. (2). It looks like some Gaussian assumption or approximation is required to justify this cost. Is it correct? How reliable are covariance estimates in high dimensions.

  4. How are cells paired up between patches? If I understand correctly, OT only runs at a global patch level, but the bridge is pointwise.

  5. Why is the velocity first trained in stage 2a, and then further trained with eq 8? Shouldn't stage a suffice if the decoded codes approximate the real ones correctly?

  6. Why not comparing with something like metric flow matching, as it also addresses manifold constraints? Or meta flow matching, which act at a population level? What drives the choice of these baselines compared to the standard benchmarks in the field presented by other papers (e.g., in https://arxiv.org/abs/2505.11823).

Reviewer_34iSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2(−1)
Soundness
1
Read review
Paper summary

The paper proposes LoCoFlow, which combines geometry-aware correspondence, regularized bridge paths, and latent conditioned dynamics to model the cell state transition. It evaluates the method on the differentiation and reprogramming datasets and on LARRY cell fate prediction, and reports lower W2 on reconstruction in most settings  and lower Brier on cell fate prediction.

Critical strengths

As a minor positive, the authors clearly explain the endpoint preservation mechanism of the bridge, where the geometric term is a ‘prior‘ rather than an intermediate observation, and the latent is not an identified fate label. The evaluation conventions in Appendix D are also relatively complete. This supports the presentation, but doesn’t offset the main issue of soundness.

As a second minor positive factor, the paper proposes a coherent framework that integrates local patch transport, geometry-regularized bridges, and latent-conditioned dynamics, and reports trade-offs between metrics rather than just favorable numbers.

Critical weaknesses

The core concern is that the incremental contribution of combining local geometry, bridge regularization, and latent-conditioned dynamics remains unclear, and the current evidence does not support consistent gains.

First, the evidence that local geometry provides a consistent benefit is not clean. In Table 6, A2 deteriorates overall, indicating that neighbor information helps. A6 shows that random mini-batch achieves better Brier (0.6212 vs. Full 0.6833), which weakens the claim that “local geometry can help establish better probability vectors”. At the same time, the BAcc of A6 was worse (18.2403 vs 26.1934), indicating that the metrics move in opposite directions. The authors should provide a matched ablation and separate ablations for the Bures and radial terms.

Second, the reported gains are metric-dependent. On LARRY, Brier improves but BAcc deteriorates; in Table 1, W2 improves while MMD worsens; On Axolotl, W2 improves but MMD²/Energy worsens. This makes the incremental contribution unclear.

Questions for authors
  1. Matched ablation: Could the authors provide a strict comparison between geometry patch and random mini-batch under the same coupling, seed and tuning parameters?

  2. Could the authors provide separate ablations for the Bures term in Eq. (2) and the radial term in Eq. (6)?

  3. I notice that in the A3 experiment, U decreases (which is better), but Acc, BAcc, and Brier are worse. What is the benefit of hierarchical latent? Could the authors provide the calibration diagnosis?

  4. W2 in Table 1 improved but MMD deteriorated; LARRY’s Brier improved but BAcc was lower than scDiffEq; Axolotl's W2 has improved, but MMD²/Energy are not optimal and visual spatial organization has not improved. Please explain the inconsistent direction of these indicators and indicate the source of W2 improvement.

Reviewer_aRwqOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper explores an interesting combination of local-population transport and hierarchical latent dynamics. However, the empirical gains are mixed, and some important implementation details are missing. The current version does not yet provide convincing evidence that the added geometric and hierarchical components improve over simpler alternatives.

Critical strengths
  1. The paper matches local populations before computing cell-level couplings within each patch pair. This allows neighborhood location and shape to influence the transport, rather than treating every cell as an isolated point.

  2. The evaluation is fairly broad, covering continuous-dynamics reconstruction, lineage-traced fate prediction on LARRY, component ablations, perturbations, and spatial data.

Critical weaknesses
  1. The method does not consistently outperform existing approaches (As claimed in the Conclusion). Improvements in W2W_2 often come with worse MMD. I also found linear interpolation or random-minibatch OT achieves better raw Brier scores in the test-set ablation.

  2. The paper combines many components without convincingly isolating their individual contributions, and most components have been seen before.

Questions for authors
  1. Please provide the exact definitions of wk(τ)w_k(\tau) and PN⊥(zi,t)P_N^\perp(z_i,t), together with the sampling procedure and the values of λN\lambda_N and λR\lambda_R. Equation (6) explains the purpose of each term, but the main text and appendices do not fully specify how to implement it.
  2. Could the authors directly evaluate the quality of the inferred couplings using lineage information, independently of downstream model training? I am interested in understanding the effect of the proposed matching method alone.
  3. As the authors mention in the last paragraph of the Related Work section, what are the main differences between the proposed method and existing approaches, such as Wasserstein Flow Matching, Metric Flow Matching, and PACE? And what is the precise conceptual advantage over these baseline methods.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.