acceptodds
Under review as a conference paper at ICLR 2027

ARCADE: Cross-Section Multimodal Learning for Mosaic Spatial Deconvolution

Abstract

Spatial multi-omics profiles tissues across multiple molecular modalities, but real-world experiments often exhibit heterogeneous modality coverage across sections, giving rise to mosaic spatial multi-omics, where different sections contain different subsets of molecular modalities. Deconvolving such data creates an incomplete multi-view latent mixture inference problem, where cell-type compositions must be recovered under heterogeneous modality coverage. Existing spatial deconvolution methods typically require RNA measurements, whereas multimodal integration methods mainly focus on representation alignment without explicitly modeling cell-type compositions. We propose ARCADE, a unified framework for mosaic spatial deconvolution that integrates cross-section multimodal information, spatial structure, and single-cell RNA references through heterogeneous graph learning and reference-guided mixture inference. A directional–magnitude factorization further separates directional information for cross-modal alignment from abundance-related information for composition inference. Experiments on real spatial multi-omics datasets demonstrate consistent improvements over widely used spatial deconvolution methods, with ARI gains over the best competing method averaging 18% across RNA-observed datasets and reaching 98% in the section where RNA is entirely unavailable. Reproducible code is available at https://anonymous.4open.science/r/submission-D956.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Mosaic spatial multi-omics, Spatial deconvolution, Incomplete multi-view learning, Multimodal learning, Graph representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_Vk4ZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposed ARCADE for mosaic spatial deconvolution, where different tissue sections profile disparate subsets of molecular modalities. By integrating multi-relational graph learning, bridge section-guided cross-modality alignment, and reference-guided NB likelihood, ARCADE demonstrates superior performance on two real-world datasets.

Critical strengths

Inferring the cell-type composition of spots in spatial multi-omics is an important and timely task. In the context of mosaic spatial multi-omics, the formulation of deconvolution via joint representation learning is both novel and practically significant.

The article is well-motivated, the writing is well-organized, and the model design is very clear and easy to follow.

The paper also presents a highly detailed ablation and sensitivity analysis of the model, clearly demonstrating the design rationale and function of each component. The authors also provide well-organized codes to reproduce the results.

Critical weaknesses

My primary concern lies in the author's evaluation: the use of non-standard metrics, insufficient datasets, and the absence of a synthetic gold standard. (That's why I gave a low soundness score)

  1. All eight quantitative evaluation metrics in the paper (LISI, SC, PAS, CHAOS, SMS, BEAR, ARI, NMI) seem to evaluate hard, discrete partition accuracy after reducing the continuous proportion vector to a single dominant cell type, which may induce severe bias. If an algorithm predicts pure hard 0-1 assignments, it may also achieve superior ARI or boundary agreement scores while failing as a reliable proportion estimator. In fact, they're more like evaluating spatial clustering instead of deconvolution.
  2. As real-world spatial spots lack ground-truth cellular mixture proportions, why not benchmarking on synthetic pseudo-spots generated by mixing known single-cell profiles, like what did in CARD [1], cell2location [2], RCTD [3], and benchmark article [4]? The authors only evaluate on two real datasets with eight spatial-clustering style metrics, and it is insufficient to show that ARCADE can accurately estimate mixing proportions.

In addition, the method proposed in this paper also has several theoretical flaws.

  1. In section 2.2, the authors claim that constraining the cell-type size factor vector (b1,…,bC)(b_1, \dots, b_C) to have a unit geometric mean ensures model identifiability, but provide no theorectical justification. In fact, I believe it theoretically flawed: pbsp_{bs} remains unidentifiable because the uncentered nuisance parameters (αbs,γg)(\alpha_{bs}, \gamma_g) and the convex cone cone(V)\text{cone}(V) induce a scale-shift gauge symmetry that permits non-trivial affine deformations of pbsp_{bs} within the simplex while preserving identical NB likelihood.
  2. While the intuition that "direction u reflects relative proportions and magnitude r reflects molecular abundance" seems reasonable, It's not clear for me why GNN and MLP processed graph representation zz can automatically achive this. Did the authors consider principled disentangled representation learning to achieve this?

Reference:

[1] Ma, Y., Zhou, X. Spatially informed cell-type deconvolution for spatial transcriptomics. Nat Biotechnol 40, 1349–1359 (2022). https://doi.org/10.1038/s41587-022-01273-7

[2] Kleshchevnikov, V., Shmatko, A., Dann, E. et al. Cell2location maps fine-grained cell types in spatial transcriptomics. Nat Biotechnol 40, 661–671 (2022). https://doi.org/10.1038/s41587-021-01139-4

[3] Cable, D.M., Murray, E., Zou, L.S. et al. Robust decomposition of cell type mixtures in spatial transcriptomics. Nat Biotechnol 40, 517–526 (2022). https://doi.org/10.1038/s41587-021-00830-w

[4] Li, H., Zhou, J., Li, Z. et al. A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics. Nat Commun 14, 1548 (2023). https://doi.org/10.1038/s41467-023-37168-7

Questions for authors

When evaluating the deconvolution performance on sections without RNA data, why not compare the results obtained with and without the inclusion of RNA measurements on paired data to better demonstrate the model's performance on RNA-missing section? For example, compare the deconvolution results on slice 3 obtained by training on RNA+protein data from slice 1 and protein data from slice 3 of the human tonsil dataset, against the results obtained by using both modalities from both slices 1 and 3.

ARCADE aggregates modality-specific representations using unweighted average. Why not using a weighted one to accounting for the different different signal-to-noise ratios, capture efficiencies, biological informativeness or uncertainties of different modalities?

Also, the authors need to provide more justifications on the identifiability and direction-magnitude representation disentanglement (either theorectical or empirical).

Reviewer_cW5bOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The authors developed ARCADE, a framework for mosaic spatial deconvolution using multimodality data across sections. Their architecture jointly learns multimodal representations and performs reference-guided cell-type composition inference. The authors carried out the experiments in two real spatial mult-omics datasets from mouse brain and human fossil, benchmarking with existing methods which were mainly designed for RNA measurements and achieved good performance across multiple evaluation metrics. The authors also did the ablation study to show the contributions of additional modality and the components that they designed in the model, as well as the experiments of scalability.

Critical strengths

The mixture architecture to consider the multimodal data and the bridge sections to improve the cell type composition performance. The modeling itself seems novel.

Critical weaknesses
  1. The two datasets are too limited to convince readers that the method can be generally and robustly used.
  2. The cell type composition is the main task for the devolution, but there is no direct evaluation.
  3. There is no synthetic or semisynthetic data for benchmarking
  4. There is no discussion about how and why the bridge section contributes to the performance.
  5. The application situation might not be very common, because it’s rare to measure multiple modalities across bridging sections in the real biological context. This limits the impact of the work.
  6. There is no robustness and uncertainty related experiments.
Questions for authors
  1. The authors mentioned that the cell type composition doesn’t exist as ground truth, however, there are many high-resolution spatial datasets from which the author could generate the semisynthetic data for benchmarking and ablation experiments.
  2. In addition to the semisynthetic data, the authors could also generate synthetic datasets for benchmarking and ablation experiments.
  3. In the experiments, the authors showed whether additional modality is useful for the performance, so I was wondering if the bridge section itself is useful or not. Moreover, whether the physical distance of sections will impact the performance. For example, if the two sections are quite far away, whether the information provided by distant sections will negatively influence the performance?
  4. As point 3, the author should decouple the effect from modality and bridge section.
  5. In one experiment to show the non-RNA modality improves the deconvolution accuracy by replacing it with duplicated RNA inputs, I wonder why the original RNA input can't be used? Does the “duplicated” mean the same RNA as the targeted section?
  6. Whether the methods can be used for the RNA-only sections individually without the bridge section?
  7. For Figure 3, there is a lot of high-resolution spatial data which should be able to cover the sections presented in the paper, so the cell type composition for each manually annotated region region can be obtained from those data to evaluate the results. Additionally, the cell type annotation is very rough, so the conclusion from this analysis is not convincing.
Reviewer_utDJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ARCADE addresses spatial deconvolution when tissue sections contain different subsets of molecular modalities, including sections without RNA measurements. It combines spatial, molecular-similarity, and cross-section graphs with contrastive alignment and an RNA-reference-guided negative-binomial model for cell-type composition inference. Modality-specific representations are decomposed into direction and magnitude, with direction used for alignment and both components used for composition prediction. Paired bridge sections connect RNA-observed and RNA-missing sections. The method is evaluated on mouse brain and human tonsil datasets, including one protein-only tonsil section, using spatial-domain agreement metrics, RNA-based and representation-transfer baselines, ablations, and runtime measurements.

Critical strengths
  1. Relevant problem formulation. Estimating cell-type compositions under section-level modality missingness addresses a practical limitation of RNA-based deconvolution and provides a meaningful application for multimodal learning.
  2. Coherent joint framework. The approach connects cross-modal representation learning with explicit reference-guided composition inference. The RNA-surrogate control and alternative transfer baselines are useful attempts to assess the benefits of complementary modalities and joint learning.
Critical weaknesses
  1. The RNA-missing comparison requires correction. The supplied transfer code removes section identifiers before matching barcodes, causing collisions across sections. A local check reproduced the archived predictions; correcting only this association increased CARD’s section-4 ARI from 0.081 to 0.323, compared with ARCADE’s 0.299. Although this check does not involve retraining, it indicates that Table 3 and the claimed 98% improvement require reassessment.
  2. Accurate cell proportions remain insufficiently established. The reported metrics primarily assess spatial-domain agreement rather than errors in cell-type fractions. Moreover, the observation model permits compensating changes in proportions, cell-type scale factors, and spot-specific efficiency that preserve its predicted RNA counts despite the stated geometric-mean constraint. Additional constraints or validation against known compositions are needed to substantiate the biological interpretation of the inferred proportions.
Questions for authors
  1. Transfer baseline correctness. Could you clarify how section identity is preserved when matching reference embeddings to composition estimates? The supplied code removes section identifiers before deduplication, which creates cross-section barcode collisions. Please report the RNA-missing comparisons using both section identity and barcode for matching, and clarify whether the main conclusions remain supported.
  2. Interpretation of inferred proportions. Beyond the unit-geometric-mean constraint, what prevents compensating changes in cell-type scale factors, spot-specific efficiency, and cell proportions from producing identical expected RNA counts? Is there quantitative validation against known compositions or independent cell-level measurements that supports interpreting the outputs as accurate cell fractions, beyond their agreement with spatial domains?
1 comment

Some of my concerns remain unresolved after the authors’ response, so I am maintaining my original rating.

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

In this paper, the authors tackle the problem of estimating which cell types are in each spot of a tissue slice when different slices were measured with different technologies, for example RNA in one slice and only proteins in another. Standard tools need RNA, so they cannot handle the protein-only slice. The authors built ARCADE, which connects the slices through those that share measurements, learns a common representation across slices and data types, and uses a single-cell RNA reference to estimate cell-type mixtures even where RNA is missing. It also splits each learned representation into a direction part, used to align the data types, and a size part, kept to preserve abundance information. On a mouse brain dataset and a human tonsil dataset it scored better than five RNA-only methods, with a large gain reported for the protein-only slice.

Critical strengths
  1. First of all, the authors tackles a real and neglected problem. Many labs measure different things in different tissue slices. Tools that need RNA cannot handle a slice without RNA. ARCADE is built for exactly this case, and the authors describe it clearly as learning from incomplete views of the same tissue.
  2. The different components of the model seem to work well together. The same output layer works for any mix of measurements, so one model covers slices with RNA, protein, or both.
  3. The authors performed informed ablation studies. They made a version called ARCADE-RS that copies RNA in place of the other data type. This shows that the gain comes from new information, not from simply having more input. They also removed each part of the model in turn, and they compared against a natural alternative: align the slices first, then copy over the results of RNA-based tools.
  4. The analysis is thorough and the appendices provide lot of useful information for the experiments.
  5. The results are promising and the gains are consistent. ARCADE won most metrics on the mouse brain data. It also did well on the tonsil slices that have RNA.
Critical weaknesses
  1. The paper is missing "ground truth" data. It scores how well results match region labels, and those labels were made with a related method (SpatialGlue) that also learns from the combined data. Tests with simulated mixtures, where the "ground truth" is known, are needed.
  2. The evidence is small. There are only two datasets, and the claim “98% gain” comes from one protein-only slice. The paper shows no repeated runs, error bars, or significance tests, so we cannot tell how stable that number is.
  3. Comparisons are somewhat weak. The authors compared ARCADE with two-step pipelines they built themselves, where RNA results are copied over with a simple nearest-neighbor model and no tuning. What about comparisons with related works/prior works or a more stronger baseline like predicting RNA from protein and then running a standard tool?
  4. The novelty is at the best "modest" and some wording is bold. Most building blocks already exist: graph propagation, nearest-neighbor links across slices, contrastive learning, and the RNA count model from earlier tools. The "new" part is the combination.
Questions for authors
  1. Can you do failure analysis?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.