- 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
- 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
- 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.
- In my opinion, the interpolant in Eq. 4 is not sufficiently described. Both and are not described. This gap impairs the understanding of the paper.
- Similarly, Eq. 6 is not very clear. What is here? What is 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.
- Similarly, there is no introduction to bridge-based generative models or interpolations. When is first mentioned, it is not clear what it refers to.
- 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 .
- 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
- LoCoFlow is inconsistently spelled. Sometimes it is introduced as Locoflow.
- 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.
- Baselines differ across experiments. Fate prediction has more of them. This stands out as inconsistent. Is there anything preventing their widespread application?
- 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.
- 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.
- 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.
- Also, in the axolotl example, the setting and modeling strategy are not introduced. The paper is not cited.
Questions for authors
-
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?
-
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.
-
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.
-
How are cells paired up between patches? If I understand correctly, OT only runs at a global patch level, but the bridge is pointwise.
-
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?
-
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).