acceptodds
Withdrawn from ICLR 2027

BSPP predicts protein expression from spatial transcriptomics across tissue sections

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Spatial transcriptomics (ST) has emerged as a powerful method for linking gene expression to tissue architecture. While recent efforts have focused on ST platforms that measure both RNA and protein levels simultaneously, these approaches suffer from several limitations: they are often restricted by the availability of protein markers, high cost, and the inability to capture large-scale tissue information due to memory and computational constraints. Furthermore, existing models either fail to incorporate spatial dependencies or oversimplify the relationship between RNA and protein expression by treating them as independent variables. In this work, we propose BSPP, a framework combining a frozen RNA-derived proteomic prior with benefit-aware spatial residual routing. BSPP leverages Universal Cell Embedding (UCE) to incorporate cross-species biological knowledge while requiring neither manual annotations nor anatomical registration. We evaluate BSPP on multiple benchmark datasets, where it outperforms baselines by achieving better improvement in protein expression prediction accuracy and recovering spatially organized protein patterns. Our results highlight the potential of BSPP in bridging the gap between RNA expression and protein prediction in spatial transcriptomics.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Spatial omics, Cross-modal prediction, Adaptive routing, Mixture of experts, Spatial representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_wy2ROpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies protein-expression prediction from spatial RNA measurements, using a source section with paired RNA and protein data to predict the same protein panel in a target section. BSPP combines an RNA encoder with pretrained UCE features through a learned gate, then predicts a protein latent representation that is decoded by a frozen VAE. The authors evaluate the method on four datasets covering replicate, cross-section, and within-section regional prediction, and report improvements over several RNA-to-protein prediction methods.

Critical strengths
  1. The problem is relevant and useful. Joint RNA and protein measurements are not always available, so transferring protein information from a paired reference to an RNA-only target has practical value.

  2. Combining RNA features with a pretrained cell representation is a reasonable approach. Predicting through a protein latent space could also help capture relationships among the measured proteins.

  3. The evaluation covers several tissues and prediction settings, with both non-spatial and spatial baselines. I also appreciate the stated separation between source-side checkpoint selection and target-protein evaluation.

Critical weaknesses
  1. I could not work out whether spatial routing is actually part of the evaluated model. The abstract describes benefit-aware spatial residual routing, but Sections 3.2 and 3.6 explicitly state that only the intrinsic pathway is used and the spatial gate is fixed to zero. Table 2 nevertheless reports changes after removing spatial context. These descriptions need to be reconciled, since they concern a central part of the claimed contribution.

  2. Some results are inconsistent across tables. For Human Lymphnode 1→2, BSPP has a protein-wise PCC of 0.354 in Table 1, while “BSPP (full)” has 0.5542 in Table 2. The reverse direction also differs substantially, at 0.350 versus 0.5457. If these are different model configurations or evaluation protocols, this needs to be clearly explained. In the current presentation, it is difficult to interpret the benchmark and ablation results together.

  3. The pretraining procedure for the frozen components is not clear enough. The protein VAE and UCE projection are fixed during predictor training, but this does not explain how they were obtained or which samples were used. Since the paper emphasizes a strictly target-blind setting, the data access for these earlier stages needs to be explained as well.

Questions for authors
  1. Which exact model produced Table 1 and Table 2? If spatial routing is disabled, why does changing the spatial graph affect predictions?

  2. Could you explain the differences between the main benchmark and full-model ablation results, especially on Human Lymphnode? Are the data splits, preprocessing, checkpoints, and scoring procedures identical?

  3. How were the protein VAE and frozen UCE projection trained or initialized? Did any stage use target-section protein measurements, including pretraining and preprocessing?

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

The paper tackles protein prediction from spatial transcriptomics across tissue sections. A model is trained on one section where both RNA and protein are measured, and is then used to predict the same protein panel on another section where only RNA is available. The proposed model, BSPP, combines an RNA encoder with frozen embeddings from a pretrained single-cell foundation model (UCE), mixed through a learned gate. Instead of predicting proteins directly, the model predicts the latent representation of a pretrained protein autoencoder and decodes it into protein expression. The paper also describes a spatial component that adds neighbourhood context, but the evaluated version of the model does not use it.

The model is validated on four paired spatial RNA-protein datasets against single-cell and spatial baselines, using correlation and structural similarity metrics. Ablations show that the UCE embeddings drive most of the protein-wise correlation performance, while the spatial context has little effect.

Critical strengths

None

Critical weaknesses

The method is inconsistent with its own description.

  • Figure 1 shows a Delaunay graph, RelativeSpatialAttention, a context head and a spatial residual gate, but none is defined in Sec. 3. Appendix A uses a kNN graph, and Sec. 2.2/Appendices refer to "STPFlow", which looks strange. The evaluated model is intrinsic-only (Sec. 3.2, 3.6), but the abstract, Fig. 1, Sec. 4.1 and the Table 2 caption ("permuted node labels") describe spatial routing. It is therefore unclear which model variant produced which results.
  • Design choices have limited described rationale:
  1. Why is the scalar cic_i, a summary of the UCE embedding magnitude (Eq. 8), concatenated in Eq. 9?
  2. Why are losses applied at both latent and protein level?
  3. What does LPearsonL_{Pearson} add? In Table 3, setting λPearson\lambda_{Pearson} = 0 gives similar results.

The headline results are not reliable as currently reported.

  • Most of the protein-wise PCC gain comes from UCE, but no baseline in Table 1 receives UCE features. Without UCE, BSPP's protein PCC is well below several baselines (e.g. thymus about 0.35 vs totalVI about 0.75). The comparison therefore shows the value of UCE features for protein-wise PCC, not of the BSPP framework.
  • Spatial component is not shown to help: differences with and without it are insignificant on most tasks, with no statistical tests.
  • No statistics: every number is a single run, with no seeds, confidence intervals or significance tests. The 8 tasks are 4 correlated pairs, so "first in 21 of 32 columns" is weak evidence.
  • SSIM metric is unclear: as written, Eq. 21 has no window, so it is a global SSIM, which penalizes any shift in the mean or spread of predictions, even when the spatial pattern is correct. How exactly is SSIM computed? This matters because the main claimed advantage rests on it.
  • BSPP's results differ between tables. Table 1 and Table 2 report different numbers for the same tasks (e.g., Lymph node 1→2 PCC (protein): 0.354 vs. 0.554).
Questions for authors

Clarifications

  • On what data was the protein VAE trained? If target-section proteins were included at any stage, the setup is no longer target-blind, and this would directly affect the main claim of the paper?
  • Is SSIM computed over local windows or globally as in Eq. 21? What data range and constants (C1, C2) are used?
  • How is UCE computed for spot-level data? UCE was trained on single cells, but the evaluated datasets are spot-level, where each spot mixes several cells.
  • What is the motivation for the Pearson loss, given its very small weight and the almost unchanged results across its values in Table 3?

Suggested experiments

  • Give the baselines the same UCE features as BSPP: at minimum an MLP or a linear model on RNA+UCE, and ideally totalVI with UCE as additional input. This is needed to show that BSPP adds value beyond the features themselves.
  • Replace UCE with other input features, such as another single-cell foundation model (e.g., scGPT or Geneformer) or plain PCA of the RNA. This would show whether the gain is specific to UCE or comes from any pretrained embedding.
  • Add a trivial baseline that predicts the mean source protein profile for every spot. This baseline would show how much of the PCC (cell) values is explained by abundant proteins.
  • Report mean and standard deviation over several random seeds, add statistical tests for the ablations.
Reviewer_D18wOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces BSPP, a framework for predicting protein expression from spatial transcriptomics data across tissue sections. The method combines RNA-derived features with pretrained UCE representations through a learned fusion gate, followed by a protein latent predictor and a pretrained VAE decoder. Although the framework includes a spatial routing component, the evaluated model does not use spatial information. The authors evaluate BSPP on four datasets across eight prediction tasks, comparing it with five baselines, including RNA-to-protein prediction and multi-omics integration methods.

Critical strengths
  1. Predicting protein expression using source datasets is important, particularly when the measurements are unavailable in target tissue sections.
  2. The authors have performed ablation analyses to examine the contribution of pretrained UCE features and have benchmarked with several baseline methods.
Critical weaknesses

1. A major concern is the inconsistency between the proposed framework and the model actually evaluated. The abstract and Figure 1 emphasize benefit-aware spatial routing, but Section 3.2 explicitly states that the evaluated configuration uses only the intrinsic pathway, without spatial residual corrections. Thus, the reported improvements appear to come from RNA and UCE features rather than the proposed spatial modeling components. Moreover, Table 2 reports differences between the full model and the w/o spatial variant, which is very small.

2. Most of the methods the paper benchmarked are not specifically designed for the protein prediction task. Without stronger protein-prediction baseline methods, it is difficult to establish the claimed advantage of BSPP.

Questions for authors
  1. Which version of BSPP was used to generate the results in Tables 1 and 2? If the spatial component was disabled during evaluation, why does removing spatial information lead to different results in Table 2?

  2. Can the authors compare with more methods specifically designed for protein prediction, for example:

  • SpaMIE: Spatial multi-omics imputation and embedding with SpaMIE (Cell Reports Methods, 2026).
  • scTranslator: A pre-trained large generative model for translating single-cell transcriptomes to proteomes (Nature Biomedical Engineering, 2025).
  • CAPTAIN: a multimodal foundation model pretrained on co-assayed single-cell RNA and protein (Nature Communications, 2026).
  1. SSIM is the metric supporting BSPP’s significant performance gain, however, it is unclear how it was calculated. Some results are also difficult to interpret, for example high SSIM despite negative PCC for MLP baseline.

  2. There are some duplicated entries in the references section.

  3. The current ablation is limited, they mainly removed UCE and spatial information. There is no ablation studies isolating the contributions of the fusion gate, protein VAE, and latent prediction framework. A simpler model combining RNA and UCE features with direct protein regression, as well as an AE-based alternative would help determine whether there are improvements beyond pretrained features alone.

  4. Figure 2 only shows spatial maps for two predicted proteins in one dataset. The authors should provide more examples across different proteins and tissues to demonstrate that BSPP consistently recovers meaningful biologically informed spatial protein patterns.

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

Related papers

Discussion (0)

Sign in to comment.