acceptodds
Under review as a conference paper at ICLR 2027

FlowMOS: Continuous Latent Transport for Spatial Multi-Omics Integration

Abstract

Spatial multi-omics integration aims to combine molecular measurements from different modalities with spatial information to learn a unified representation that can be used for downstream biological analysis. However, significant differences among modalities in terms of dimensionality, sparsity, noise, and biological semantics make it difficult to obtain stable and discriminative unified representations, while the transformation from modality specific representations to the shared space is typically modeled implicitly. To address these challenges, we proposes FlowMOS, a flow matching representation learning framework for spatial multi-omics integration. FlowMOS introduces shared space flow matching to explicitly model the continuous transfer process from each modality specific representation to a shared latent space. To further improve the robustness of representation learning against incomplete and noisy measurements, mask consistency regularization is employed to simulate local feature missingness and sequencing noise in the input data. Extensive experiments on several real world spatial multi-omics datasets demonstrate that FlowMOS effectively captures modality to shared transformations, while the unified representation achieve robust performance in downstream task.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Spatial multi-omics, FlowMOS, Flow Matching, Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ncVvOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposed a flow matching based method for spatial multi-omic clustering.

Critical strengths

This paper proposed a flow matching based method for spatial data integration while the idea is interesting.

Critical weaknesses

This paper proposed a flow matching based method for spatial data integration while the idea is interesting, I’m doubting usability of method and the definition of problems.

  1. Special multi-omic data is known with high noise level, but the authors do not discuss them. Please design experiments to justify the data quality and human annotation match your expectation and standard for computational method, if not, all learned signals can be noise.
  2. Clustering is a very trivial problem, and PRAGA shows very similar performance FlowMOS, and thus the method contribution seems not informative and helpful. I’m also doubting the performances of other baselines. Do you tune their method to achieve the best performances? If not, this is an unfair comparison.
  3. Ablation study is also very weird. If the authors remove the reconstruction loss, then the performance is quite close, compared with FlowMOS. Can the authors explain the reasons are how to justify the loss function is optimal setting?
  4. These UMAPs also cannot distinguish their methods versus other baselines, every distribution seems very similar to me and biologists
  5. There is no biological interpretation about cluster labels and how do the authors justify the method and preserve the biological information such as gene pathway?
Questions for authors

See my questions above, mainly related the definition of problem and soundness of method.

Reviewer_eV8LOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

FlowMOS addresses multi-omics data for spatial pairing. It combines spatial neighbors and feature neighbors to encode each modality separately, then fuses them into a shared representation. During training, flow matching is introduced to learn the velocity from modality representations toward a smooth shared target; simultaneously, masked consistency enhances the robustness of representations against minor feature perturbations. Final clustering uses the directly fused shared representation. The authors report improvements on mouse brain data, but further validation is needed regarding statistical stability, partial implementation details, and cross-modal recovery generalization.

Critical strengths
  1. Integrating heterogeneous molecular measurements while preserving spatial organization is an important problem.
  2. Learning modality-specific velocity fields toward a shared EMA target is a potentially useful way to regularize multimodal embeddings.
Critical weaknesses
  1. This paper only conducted experiments and reported results on two datasets, and the proposed method was completely inferior to the existing methods on one of the datasets. Therefore, the experimental results are very weak and not convincing.
  2. From the ablation studies, the contribution of each component is tiny, as the gaps between most variants and the best implementation are even narrower than the stds reported in Table 1. So, the necessity of combining these components is not clear, which raises doubts about the innovation and rationality of this paper. Besides, the results of the best implementation in Table 2 are inconsistent with those in Table 1.
  3. There are three key hyperparameters in Equation 10, while only two are explored in Figure 4. In Table 3, what is the relationship between λcons\lambda_{cons} and λ1∼λ3\lambda_1 \sim \lambda_3?
  4. From the perspective of reproducibility, the author should present all the parameter settings used in the experiments, especially the hyperparameters for different components on different datasets.
Questions for authors

See above weaknesses.

Reviewer_5pXfOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes FlowMOS, a flow-matching-based framework for spatial multi-omics integration. The method uses graph encoders to learn modality-specific latent representations, which are concatenated and passed through an MLP to obtain a shared representation. Flow matching is then employed to model continuous latent transport from each modality-specific representation toward the shared space. Additionally, mask consistency regularization is introduced to improve robustness against missing or noisy features. The framework is evaluated on spatial multi-omics datasets using clustering metrics, ablation studies, and qualitative latent transport analyses. The results demonstrate competitive performance against existing methods, with improvements on certain datasets.

Critical strengths

The incorporation of mask consistency regularization is a useful aspect of the proposed framework. By randomly masking input features during training and encouraging consistency between the original and masked latent representations, the method aims to improve robustness against missing features and incomplete measurements in spatial multi-omics data.

Critical weaknesses

(1) The introduction claims that explicitly modeling continuous transitions provides a new perspective beyond conventional latent integration methods. However, FlowMOS already generates its shared representation by concatenating modality-specific embeddings followed by an MLP, similar to the endpoint-based integration approaches it criticizes. Flow matching is subsequently introduced as an auxiliary objective rather than the mechanism that produces the final representation.

(2) The benefit of modeling continuous latent transport using flow matching is not sufficiently justified. Given the endpoints, a continuous trajectory can also be constructed by interpolating between the input and output of a conventional neural network. The manuscript does not clearly establish what additional interpretable information FlowMOS learns through its velocity field.

(3) STAGATE is primarily a single-modality spatial transcriptomics method based on graph attention and denoising. Its use as a baseline for multi-omics integration requires clarification, particularly regarding how the different modalities are incorporated and integrated.

(4) Although FlowMOS outperforms several baselines, it does not consistently achieve the best performance, particularly on the human lymph node dataset. More importantly, the claimed advantage of continuous latent transport is not supported by sufficient biological interpretability experiments. The presented latent-space visualizations do not establish whether the learned trajectories represent biologically meaningful transformations between modalities.

Questions for authors

(1) Since the shared latent representation is already obtained by concatenating modality-specific embeddings followed by an MLP (Equation 2), what additional benefit does flow matching provide beyond conventional endpoint-based integration?

(2) A continuous transformation can also be visualized by interpolating between the input and output embeddings of a conventional neural network. What makes the trajectories learned by FlowMOS more meaningful or interpretable than such interpolation?

(3) STAGATE is primarily designed for denoising single-modality spatial transcriptomics data. How was STAGATE adapted to integrate multiple spatial omics modalities, and how was a fair comparison ensured?

(4) Since continuous latent transport is a central contribution of the paper, can the authors provide biological evidence demonstrating that the intermediate latent states represent meaningful cross-modal transformations rather than simply mathematical interpolation?

(5) FlowMOS does not consistently outperform the strongest baseline across datasets. Can the authors provide additional ablation experiments to clarify when flow matching improves integration performance and whether these improvements justify its additional computational complexity?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.