acceptodds
Withdrawn from ICLR 2027

GASP: Gene-Aware Spatial Predictor for Histology-based Transcriptomics

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Spatial transcriptomics (ST) provides unprecedented insights into tissue architecture but remains constrained by high costs and complex protocols. Predicting ST directly from ubiquitous Hematoxylin and Eosin (H&E) stained images offers a scalable alternative. However, existing computational methods predominantly adopt a ”vision-centric” paradigm, projecting passively extracted visual features into the gene space via a shared decoder. This approach overlooks the inherent heterogeneity of genes, failing to capture their unique morphological dependencies. In this paper, we propose the Gene-Aware Spatial Predictor (GASP), a novel framework that introduces a paradigm shift towards “gene-centric” modeling. GASP employs learnable Gene Queries that actively probe a multi-scale visual context pool—comprising local, neighborhood, and global features—via a Vision-Gene Cross-Attention mechanism. This allows each gene to adaptively retrieve its preferred morphological cues. Furthermore, a Gene-Specific Residual Decoding strategy is introduced to balance global expression trends with gene-specific variations. Extensive experiments on benchmark datasets demonstrate that GASP outperforms state-of-the-art methods, offering superior predictive accuracy for spatial gene expression modeling.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Spatial Transcriptomics, Histology Image Analysis, Cross-Attention, Gene-Centric Modeling, Computational Pathology, Multi-Scale Feature Fusion

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_mj9BOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GASP, a framework for predicting spatial gene expression from histology images. The authors argue that existing approaches mainly follow a vision-centric paradigm, where visual representations are extracted and mapped to gene expression through shared decoders, which overlooks the heterogeneous morphological dependencies of different genes. To address this issue, GASP introduces learnable gene queries that interact with multi-scale visual contexts through a vision-gene cross-attention mechanism. Furthermore, a gene-specific residual decoding strategy is proposed to model gene-level variations. The paper evaluates GASP on several spatial transcriptomics benchmarks and reports improvements over existing approaches. While the motivation of modeling gene-specific morphology relationships is interesting, I have concerns regarding the fundamental novelty, biological interpretation of the proposed query mechanism, and the completeness of experimental validation.

Critical strengths

1.The paper identifies an important limitation of existing spatial transcriptomics prediction methods. Different genes may indeed correspond to different morphological patterns, and modeling such heterogeneity is a meaningful direction. The idea of moving beyond a shared decoder toward gene-conditioned prediction is intuitively appealing. 2.The proposed use of multi-scale visual contexts, including local morphology, neighborhood information, and global tissue structures, is reasonable for spatial transcriptomics prediction. The cross-attention mechanism provides a flexible way to integrate gene representations with visual features.

Critical weaknesses

1.The paper describes gene queries as actively retrieving preferred morphological cues. However, these queries are randomly initialized learnable embeddings optimized only through downstream prediction loss. Without explicit biological priors, pathway information, or structural constraints, it remains unclear whether these queries represent meaningful biological entities or simply task-specific latent variables. The authors should provide stronger evidence demonstrating that the learned queries capture genuine gene-specific morphological dependencies. 2.The design assigns an independent learnable query to each target gene. However, it remains unclear whether this leads to meaningful specialization or simply introduces redundant parameters. Many genes share related biological pathways and morphological patterns, suggesting that independent queries may not be the optimal representation strategy. A more principled design might require modeling relationships among genes rather than learning isolated gene-specific embeddings.

Questions for authors

The central claim of GASP is that gene queries can actively retrieve gene-specific morphological cues. However, these queries are learned solely through the downstream gene expression prediction objective without explicit biological priors or semantic constraints. How do the authors demonstrate that the learned gene queries represent genuine biological morphology–gene relationships?

Reviewer_56xTOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes GASP for spatial gene-expression prediction from H&E images. It combines target-patch, neighborhood, and global features, applies cross-view fusion, and uses learnable gene queries with residual decoding. Experiments cover two breast cancer datasets and one squamous cell carcinoma dataset.

Critical strengths
  1. Spatial gene-expression prediction from routine histology is practically important.
  2. The model uses multiple tissue scales and is evaluated on three public datasets.
  3. The paper includes component ablations and gene-level visualizations.
Critical weaknesses
  1. The overall novelty is limited. Multi-scale local and global modeling closely resembles TRIPLEX, the pathology representation comes from UNI, and gene-specific querying overlaps with GeneQuery [1]. Combining these components may be useful, but the paper does not identify a new mechanism or explain the technical difference from the closest methods. The current contribution therefore appears to be an integration of established designs.
  2. The term gene-aware is not well supported by the model. Each gene query is a free learnable vector tied to one of 250 fixed output genes; no gene names, functions, pathways, or other biological information are encoded, and unseen genes are not evaluated. This is closer to using gene-specific output tokens than modeling gene knowledge. Comparisons with a standard per-gene prediction head and a gene-ID embedding are needed.
  3. The source of the reported improvement remains difficult to isolate. GASP combines CIGAR and UNI features, multi-scale context encoders, cross-view fusion, gene-query attention, and residual decoding, giving it more components than many baselines. Removing one module at a time does not show whether the gene-centric design is responsible for the gain. Parameter-matched controls using the same backbone, a shared query, and a conventional per-gene MLP would be more informative.
  4. Table 2 reports only averaged cross-validation results, without fold-level variation, confidence intervals, or significance tests. This is important because several gains, especially on BC2, are modest, while the paper describes them as consistent or significant. Report each fold, the mean and standard deviation, and an appropriate paired statistical test.
Questions for authors
  1. What differs technically from combining TRIPLEX-style features, UNI, and GeneQuery [1]?
  2. How do the gene queries differ from a conventional gene-specific head, and can they handle unseen genes?
  3. Do the gains remain under matched backbones, parameter counts, and decoder capacity?
Reviewer_kmEUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper presents GASP, which uses learnable gene queries to extract gene-relevant information from local, neighborhood, and global histological features, and applies residual corrections to base gene expression predictions. The model achieves better average predictive performance than the listed baselines on three spatial transcriptomics datasets.

Critical strengths

1. Predictive gains are observed across all three datasets under patient-level evaluation.

GASP achieves mean gene-wise PCCs of 0.2946, 0.1612, and 0.2489 on BC1, BC2, and SCC, respectively, exceeding the listed baselines. The evaluation also uses patient-level cross-validation, preventing sections from the same patient from appearing in both the training and test sets. These results support the framework's practical predictive value within the evaluated cohorts and are the strongest basis for my positive assessment of the work. Although the statistical stability of the gains requires further reporting, the consistent average improvements across datasets are encouraging.

2. Gene-conditioned feature retrieval is reasonably aligned with the heterogeneity of the task.

The model allows different genes to retrieve different features from local, neighborhood, and global visual information, providing a clear and testable modeling hypothesis. The performance decrease after removing gene queries in Table 4 also provides preliminary support for the practical utility of this design. This matters because the architecture offers a concrete way to investigate relationships between individual genes and tissue morphology. If the independent value of target-specific retrieval can be further established, the design could have methodological relevance beyond the current setting.

Critical weaknesses

1. The core novelty is insufficiently positioned relative to prior work and therefore does not adequately support the claim of a "paradigm shift."

Query2Label [1] already uses class queries and cross-attention to retrieve target-relevant visual features, while GeneQuery [2], in the same application area, has proposed gene-conditioned expression prediction. These methods differ from GASP in their specific implementations, so their existence does not establish that this paper duplicates prior work. Nevertheless, they limit the conceptual originality of the claimed transition from "vision-centric" to "gene-centric" modeling.

The more clearly identifiable contribution of this paper is the combination of learnable gene queries with a pool of multi-scale visual features updated through cross-view interactions. However, the manuscript does not sufficiently explain the distinctive capabilities or design challenges associated with this difference. This directly affects my assessment of its methodological contribution: the authors should clearly distinguish established mechanisms, adaptations to the task, and genuinely new elements, and revise their novelty claims accordingly.

2. The fairness of the experimental comparisons is not sufficiently established, and the independent benefit of per-gene retrieval remains unclear.

GASP uses visual features extracted by both ResNet and UNI, but the manuscript does not sufficiently report the feature extraction models, pretrained weights, freezing strategies, or input configurations used by each baseline. It is therefore unclear whether the methods use comparable visual representations. The UNI ablation shows that these additional features contribute to performance, so the full model's improvement over the baselines cannot be directly attributed to its per-gene query architecture.

Moreover, a shared fused representation can still distinguish genes through different output weights and therefore should not be equated with ignoring gene heterogeneity. The static representation used in the "w/o Gene Queries" ablation in Table 4 is also insufficiently defined. Consequently, this ablation does not clearly separate the value of per-gene retrieval from the benefits of additional capacity or the retrieval branch itself.

These issues directly affect my assessment of the central contribution. The results support average performance gains for the complete framework under the reported configurations, but do not yet establish the independent value of gene-specific retrieval. I recommend first clarifying the visual feature configurations of all methods and conducting the key architectural comparisons using the same visual features. Furthermore, the authors should compare "per-gene queries" against "a shared query followed by multi-output prediction," while keeping the information pool, base prediction branch, residual formulation, loss, and training budget the same, and matching parameter counts as closely as possible. Such a comparison would more directly address whether the improvement actually comes from different genes retrieving visual information differently.

[1] Shilong Liu et al. Query2Label: A Simple Transformer Way to Multi-Label Classification. arXiv preprint arXiv:2107.10834, 2021.

[2] Ying Xiong et al. GeneQuery: A General QA-based Framework for Spatial Gene Expression Predictions from Histology Images. arXiv preprint arXiv:2411.18391, 2024.

Questions for authors
  1. How are the scale preferences in Figure 2 defined and estimated? Please describe the computation, data splits, and validation procedure. Could the local, neighborhood, and global inputs be perturbed separately before fusion, with per-gene performance changes reported to directly examine scale dependence? The interpretation should account for overlapping information across branches and the possibility that masking produces atypical inputs.

  2. Are the average gains stable across patients or splits? Please provide paired per-fold results and appropriate uncertainty estimates, particularly for the smaller average improvement on BC2. The current reporting is insufficient to determine whether the claim of "significantly outperforming" the baselines has statistical support.

  3. What visual features does each method use? Please report the backbone, pretrained weights, freezing strategy, and input configuration. The gain from adding UNI supports its incremental value within the current model. If the hybrid backbone itself is also considered a contribution, its value relative to a comparable all-UNI configuration should be further explained.

  4. What exactly is the static representation used in "w/o Gene Queries"? Does this variant retain the same retrieval procedure, output branch, and approximately the same parameter count? This detail directly affects the interpretation of the ablation.

  5. Are the readout weights in Section 3.4 independent for each gene? The text describes independent readout layers, but the formula does not include a gene index on the readout weights. Please clarify how the weights are shared and where gene specificity arises.

  6. How is "patient specificity" in Section 4.2 computed? Please provide its definition, the unit of computation, and the procedure used to estimate its correlation with gene expression variance.

  7. How were the examples in Figure 3 selected? Please explain the selection criteria for the displayed section and genes so that their representativeness can be assessed, and distinguish prediction visualizations from mechanistic evidence.

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

Related papers

Discussion (0)

Sign in to comment.