acceptodds
Withdrawn from ICLR 2027

COSMIC: Joint Representation Learning and Conditional Multi-Slice Alignment for Spatial Multi-Omics

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multi-slice spatial multi-omics integration seeks coherent tissue representations that preserve complementary molecular signals and biological heterogeneity. However, cross-modal fusion and cross-slice alignment are tightly coupled: reliable correspondence requires robust multimodal semantics, while learning such semantics is itself confounded by slice-specific shifts. Existing methods largely optimize one side first, propagating fusion noise or overcorrecting noncorresponding tissue regions. To address this limitation, we propose COSMIC, an end-to-end framework that couples target-conditioned consensus learning with cluster-guided conditional alignment in a shared latent space. Multimodal semantics guide cross-slice correspondence and are jointly refined by alignment, while target-conditioned cross-modal interaction forms a reliability-weighted consensus that preserves modality-private information. Based on the resulting representation, cluster-guided conditional alignment uses soft semantic assignments and confidence filtering to guide the alignment of well-supported slice-specific semantic centers with their global semantic counterparts. This conditional strategy avoids indiscriminate matching of entire slice distributions and reduces overcorrection of noncorresponding tissue structures. Through comprehensive benchmarks across multiple species, COSMIC achieves highly competitive performance while maintaining a favorable balance between biological structure preservation and slice-effect removal, thereby facilitating comparative analyses across heterogeneous spatial multi-omics tissue sections.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Spatial Multi-Omics、Multi-Slice Integration、Representation Learning、Multimodal Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_azQBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

COSMIC presents a framework for learning a unified latent representation from multi-slice spatial multi-omics data, while preserving biological structure and reducing batch effects.

It uses target-conditioned cross-modal interaction (TCCI) to integrate complementary information across modalities and cluster-guided conditional alignment (CGCA) to selectively align reliable biological structures. COSMIC is evaluated using four datasets, and with ablations showing generally modest contributions from its proposed TCCI and CGCA components.

Critical strengths
  • The paper is well motivated to jointly address cross-modal fusion and cross-slice alignment instead of treating them as two independent problems.
  • The evaluation is performed against seven existing methods, in Table 1 and 2, although I have some concerns about the reported results, please see weakness section.
Critical weaknesses
  • From table 1 and 2, I can see that COSMIC is not always the best for many individual metrics yet achieves the best Total score through a custom Total metric defined 0.6 times conservation + 0.4 times batch correction. I cannot see any justification or sensitivity analysis for this weighting and raises concerns about whether the reported overall ranking is robust to alternative weighting choices.
  • The paper content is difficult to follow with several technical abbreviations are introduced without being explicitly defined, such as MNN, and FFN. Also, some terminologies are inconsistent for example, E(m) is referred to as both the “topology-aware representation” and “target representation,” making the method unnecessarily confusing.
  • There are also inconsistencies while describing the results where the numbers in the text do not match with the tables, for example In table-3, text states that removing TCCI reduces “Bio Conservation and Total from 0.618 to 0.600”; however, Table 3 shows that Bio Conservation decreases from 0.485 to 0.464, while only the Total score decreases from 0.618 to 0.600.
Questions for authors
  • What is the rational of using Total metric as 0.6 times conservation + 0.4 times batch correction?
  • Would the relative ranking of methods will remain same under different alternative weighting choices?
  • Is the unified joint multi-slice topology graph is constructed only once and fixed during training or it is dynamically updated during optimization?
Reviewer_8g5yOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

COSMIC is an end-to-end framework for multi-slice spatial multi-omics integration that couples target-conditioned consensus learning with cluster-guided conditional alignment in a shared latent space. Target-conditioned cross-modal interaction forms a reliability-weighted consensus while preserving modality-private information. Based on the resulting representation, cluster-guided conditional alignment uses soft semantic assignments and confidence filtering to align well-supported slice-specific semantic centers with their global counterparts. Benchmarks across human and mouse tissues evaluate the balance between biological structure preservation and slice-effect removal.

Critical strengths
  1. The coupling of consensus learning and conditional alignment is well motivated. Using multimodal semantics to guide cross-slice correspondence addresses an important limitation of indiscriminate distribution matching. Retaining modality-private information also makes the formulation relevant to heterogeneous spatial multi-omics data.

  2. The experimental comparison covers several relevant settings. The benchmarks include different tissues and modality combinations, with competitive aggregate performance. Component ablations support the usefulness of TCCI and CGCA, while RNA and ADT marker analyses provide biological context for the inferred tonsil domains.

Critical weaknesses
  1. The optimization of the semantic assignments and joint representation is unclear. Q is detached in the alignment loss, and no listed objective appears to train the clustering projection head. Reconstruction also uses the shared and private representations before fusion, rather than the final representation Z. As written, a constant Z could satisfy the alignment objective without incurring a reconstruction penalty. How semantic assignments are learned and representation collapse is prevented needs to be explained.

  2. Cumulative assignment mass does not establish alignment confidence. A large slice can exceed the threshold even when assignments are diffuse, so the filtering rule does not clearly exclude uncertain or noncorresponding tissue structures. The ablation removes CGCA entirely but does not compare conditional alignment with global alignment or disable confidence filtering. The claimed reduction in overcorrection therefore lacks direct validation.

  3. The effect of Harmony preprocessing is not adequately controlled. Harmony already removes slice-specific batch effects from the RNA features before representation learning. It is unclear whether the baselines receive equivalent preprocessing, and no Harmony-only comparison or COSMIC variant without Harmony is reported. This makes the additional contribution to batch correction difficult to assess.

  4. The paper does not include a Reproducibility Statement.

Questions for authors
  1. What objective or update rule trains the clustering projection head when Q is detached? Does any loss directly constrain Z to prevent collapse, given that reconstruction uses representations before fusion?
  2. How does cumulative assignment mass distinguish confident correspondence from diffuse assignments? Do comparisons with global alignment and a variant without confidence filtering support the claimed reduction in overcorrection?
  3. Were the baselines evaluated with equivalent Harmony preprocessing? How does COSMIC perform without Harmony and relative to Harmony-corrected features alone?
Reviewer_11G2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes an improved approach for addressing multi-modal challenges in spatial omics technology and validates it using real-world data.

Critical strengths

This manuscript provides a clear description of the research motivation and methodological details, and both the qualitative and quantitative experiments have been well executed.

Critical weaknesses

The methods proposed in the manuscript—such as multimodal alignment and cluster-guided approaches—appear to be well-established techniques. This makes the work resemble an application of existing methods to the biological domain rather than a contribution to machine learning research, and it also lacks designs specifically tailored to this particular field.

Questions for authors

This manuscript is written in a clear and fluid manner; if a convincing clarification regarding its novelty can be provided, I might be willing to raise the score.

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

Related papers

Discussion (0)

Sign in to comment.