Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Spatial transcriptomics (ST) is pivotal for deciphering molecular organization, yet cross-modal variability challenges accurate H&E based profiling. Existing models struggle to generalize to unseen genes and lack clinical interpretability. We propose GeneRAG, a model-agnostic Retrieval-Augmented Generation framework with a Dual-Constrained Retrieval module. Unlike conventional black-box networks that rely solely on fixed parameters, GeneRAG explicitly decouples knowledge storage from model training. By optimizing an Elastic Net-based sparse sampling matrix, GeneRAG integrates morphological and biological constraints to fetch relevant samples from a pre-constructed bank. Leveraging conserved gene correlations, this enables accurate reconstruction of comprehensive profiles, including entirely unseen genes. On the HEST-1k dataset, GeneRAG seamlessly enhances state-of-the-art models in a plug-and-play manner, improving Stem’s PCC-10 from 0.8322 to 0.8711(Breast dataset). For zero-shot generalization (5,000 genes), Stem+GeneRAG achieves a PCC-5000 of 0.5188, vastly outperforming DeepSpot (0.0748). GeneRAG provides robust, transparent predictions, highlighting its potential for clinical deployment.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2142_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/HyeongSubKim/GeneRAG

Link to the Dataset(s)

HEST-1k dataset: https://github.com/mahmoodlab/hest

BibTex

@InProceedings{KimHye_GeneRAG_MICCAI2026,
        author = { Kim, Hyeongsub AND Kim, Sihyun AND Cho, Minyoung AND Jo, Sanghyun AND Lee, Minhyeong AND Kim, Kyungsu},
        title = { { GeneRAG: A Retrieval-Augmented Framework for Spatially Resolved Gene Expression Prediction } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16895},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The authors propose GeneRAG, a retrieval-augmented approach for spatial transcriptomics prediction from histology images. The method combines morphological similarity and partial gene expression predictions to retrieve relevant samples from a reference bank and reconstruct full gene expression profiles via a sparse optimization framework. The approach is model-agnostic and demonstrates improvements across multiple datasets, including imputing the expression of genes.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • Interpretability: The retrieval-based formulation provides an intuitive and transparent mechanism for prediction. By explicitly selecting and weighting similar reference patches, the model allows one to trace predictions back to specific morphological patterns and corresponding gene expression profiles. This is a meaningful advantage over black-box deep learning approaches.
    • The method is conceptually straightforward and avoids heavy architectural complexity. The use of a sparse optimization framework when imputing the expression.
    • Generalization to unseen genes is a addressing the closed-set limitation
    • Model-agnostic and can be integrated with different backbone models
    • Despite its simplicity the method shows improved predictive performance across datasets
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • Limited methodological novelty: The core idea of retrieval-based reconstruction and bimodal dictionary-like modelling has been explored in prior multimodal and biomedical settings. The methodological contribution is minimal.
    • The framework relies on sample-level similarity to approximate gene–gene dependencies. It would be interesting to compare against more direct approaches, such as conditional modelling or latent factor-based methods, which explicitly capture gene relationships.
    • The multi-stage pipeline (initial prediction -> retrieval -> expression aggregation) may accumulate errors, particularly when the initial gene prediction is inaccurate.
    • Dependence on reference bank: Performance is likely sensitive to the diversity and coverage of the training dataset, raising questions about generalization to out-of-distribution samples.
    • The method assigns per-sample weights over a reference bank, which may limit its ability to generalize beyond the training distribution. This design makes the method inherently dependent on dataset coverage and raises concerns about robustness under domain shift as no cross-dataset validation results are reported.
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The method is simple, interpretable, and clearly presented, but offers limited methodological novelty. Its reliance on per-sample retrieval raises concerns about generalization under domain shift, especially without cross-dataset evaluation.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    fThe rebuttal reframes reference-bank dependency as a strength for domain adaptation, but this does not fully address the OOD concern. Since predictions are reconstructed from retrieved samples, the method still fundamentally depends on the coverage and diversity of the reference bank, making robustness under distribution shift insufficiently validated.



Review #2

  • Please describe the contribution of the paper

    The paper proposes a retrieval-and-reconstruction approach for H&E-to-spatial-transcriptomics prediction. Given an initial prediction from a frozen backbone, the method solves a dual-constrained ElasticNet optimization to find a sparse set of training samples that best reconstruct the query in both morphological and gene expression spaces. The resulting sparse weights are then applied to the full gene bank, enabling prediction of genes unseen during backbone training. This replaces the kNN retrieval used in prior work (e.g., BLEEP) with a more principled optimization-based formulation, and in the authors’ experiments improves prediction performance, particularly for zero-shot gene extrapolation.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.The retrieval module introduces no additional trainable parameters. The sparse weight vector α is obtained by solving a convex ElasticNet optimization at test time, which guarantees a global optimum and eliminates the risk of overfitting in the retrieval step. This is a meaningful practical advantage over methods that require training additional neural network components for retrieval or alignment. 2.The decoupling between the optimization target (anchor genes and morphological features) and the reconstruction target (full gene panel) is a well-motivated design choice. By leveraging conserved gene-gene co-expression structure, the method enables prediction of genes that were entirely unseen during backbone training. The zero-shot extrapolation results on 5,000 genes, while requiring more rigorous baselines for proper evaluation, demonstrate a capability that most existing methods cannot offer by design. 3.The framework is genuinely model-agnostic. The authors demonstrate consistent improvements across three architecturally distinct backbones — a pure vision model (UNI), a multimodal foundation model (EXAONE Path 2.5), and a diffusion generative model (Stem). This plug-and-play flexibility lowers the barrier to adoption and allows practitioners to benefit from the retrieval module regardless of their choice of backbone.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.The paper claims to be “the first universal retrieval-augmented framework for spatial transcriptomics,” but this is factually incorrect. BLEEP (Xie et al., NeurIPS 2023) already performs retrieval-augmented gene expression prediction from a reference bank of paired image-expression entries. 2.The zero-shot comparison against DeepSpot in Table 2 is fundamentally asymmetric. DeepSpot must predict 5,000 genes entirely from learned parameters, whereas GeneRAG retrieves ground-truth expression profiles from the training set. This is not a fair evaluation of the retrieval formulation. The appropriate baselines would be alternative retrieval strategies with access to the same reference bank — such as kNN in the image embedding space, kNN in the gene space, or kNN with a weighted dual-modality similarity — which would isolate the contribution of the ElasticNet sparse reconstruction over simpler retrieval mechanisms. 3.Computational cost is never discussed. The method solves an N-dimensional ElasticNet optimization for every test spot individually. A typical Visium slide contains ~5,000 spots, and the reference bank may contain tens of thousands of entries. The scalability of this approach to large cohorts or whole-slide-level prediction is unclear and could be a significant practical limitation that undermines the claimed ease of clinical deployment. 4.The foundation models used as backbones (UNI, CONCH, EXAONE Path 2.5) were pretrained on massive histopathology datasets that may overlap with the HEST-1k evaluation data. The paper does not discuss or rule out this potential contamination. Since the retrieval module relies on the frozen encoder’s feature space, pretraining data overlap would make the morphological retrieval artificially easy, inflating results in a way that is difficult to disentangle from the method’s actual contribution. 5.The “interpretability” claim is overstated. Showing top retrieved patches and their weights is a property shared by any retrieval-based method, including simple kNN. The sparse weights do not provide meaningfully richer explanations than similarity scores for the purpose of clinical reasoning.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    1.The paper positions GeneRAG as “the first universal retrieval-augmented framework” for spatial transcriptomics. However, prior methods such as BLEEP and other retrieval-based approaches already explore related ideas in this space. I encourage the authors to soften and better localize this claim. The sparse reconstruction formulation and zero-shot extrapolation results appear sufficiently novel on their own and do not require a broad priority claim.

    2.For Table 1, it would be helpful to clarify how predictions are generated for the foundation-model baselines (UNI, CONCH, EXAONE Path 2.5) in the without-GeneRAG setting. Since these models are primarily encoders, more detail on the decoder/prediction head, training strategy, and whether the encoder is frozen or fine-tuned would help readers assess baseline strength and reproducibility.

    3.The manuscript would be strengthened by clarifying whether any overlap exists, or can be ruled out, between the pretraining corpora of the foundation-model backbones and the HEST-1k evaluation data. At present, this is difficult to assess from the paper. If overlap cannot be ruled out, that limitation should be discussed explicitly.

    4.The comparison against DeepSpot is informative, but the two approaches have different information access at inference time: DeepSpot performs purely parametric prediction, whereas GeneRAG retrieves and recombines training-set reference profiles. Acknowledging this asymmetry would help readers interpret the reported margins more appropriately.

    5.Since the method solves an ElasticNet-style retrieval optimization for each test spot, runtime and scaling analysis would be valuable. Reporting average inference time per spot or slide, and how performance/runtime vary with reference bank size, would improve the paper’s practical relevance.

    The core idea is interesting, and my concerns are mostly about positioning and evaluation clarity rather than a fatal flaw in the method itself. If the authors address these issues convincingly in rebuttal, I would be open to revising my score upward.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I assigned a score of 3 because, while the paper presents an interesting idea and reports promising empirical results, I do not yet find the current presentation strong enough for a higher score. In particular, the novelty claim appears overstated: positioning GeneRAG as “the first universal retrieval-augmented framework” for spatial transcriptomics is difficult to reconcile with prior retrieval-based methods already cited or closely related in the literature. In addition, several important experimental details remain unclear, especially how the foundation-model baselines without GeneRAG (e.g., UNI, CONCH, EXAONE Path 2.5) are converted into gene-expression predictors, which makes it hard to fully assess baseline strength and reproducibility. The comparison to DeepSpot is also not entirely like-for-like, since GeneRAG has access to a training-set reference bank with ground-truth profiles at inference time, whereas DeepSpot performs purely parametric prediction. Finally, the paper would benefit from a clearer discussion of possible pretraining/evaluation overlap for the pathology foundation models and from runtime/scalability analysis, given that the proposed method solves an ElasticNet optimization for each test spot. Overall, I view these as meaningful weaknesses in novelty positioning, evaluation clarity, and practical reporting rather than fatal flaws in the core idea, which is why I settled on a 3 rather than a lower score.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The response and promised changes address most of my concerns, I would like to raise the score to 5.actually I think the largest concern is the overclaim of the first RAG system, but since they promise to change the wording, I am good with the paper, I think the novelty of the method is good despite some minor weaknesses.



Review #3

  • Please describe the contribution of the paper
    • This work introduces GeneRAG, a retrieval-augmented framework for spatial transcriptomics that aims to mitigate cross-modal variability.
    • The proposed approach improves interpretability by addressing the black-box nature of deep learning models, providing visual and transcriptomic rationales based on retrieved reference samples.
    • The method demonstrates zero-shot generalization to unseen genes and exhibits a model-agnostic design, allowing integration with a variety of existing state-of-the-art architectures in a plug-and-play manner.
  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The concept of performing retrieval using both image features and gene expression features appears to be a well-founded and reasonable design choice.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • The paper does not provide details on how the dataset is split; could the authors clarify what experimental setting is assumed? In recent work, it has become common to evaluate under both intra-patient and inter-patient settings. However, when generalizing across patients, gene expression profiles can differ substantially due to batch effects. Does the proposed method still operate effectively under such conditions? It seems likely that certain assumptions are made, and therefore clearly specifying the scope of applicability and underlying assumptions of the method is very important.
    • The definitions of each component in Equation (1) appear to be unclear and somewhat inconsistent. It seems that $D_{\text{img}}$ and $D_{\text{anchor}}$ are defined as a set, yet the output appears to have the same dimensionality as $f_{\text{img}}$. My understanding is that D is matrix $D_{\text{img}} = ( f_{\text{img}}^(1) … )$, and that a weighted aggregation over samples in the dataset is performed using $\alpha$, resulting in a single feature representation.
  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    As noted in the weaknesses section, I am concerned that retrieval-based approaches are often evaluated with insufficiently validated or inaccurately specified data pool requirements. However, this issue is common across other retrieval-based methods as well. I agree with the overall concept of the approach, and since it has the potential to advance the exploration of retrieval-based methods, I believe it is acceptable to accept it.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The reviewers were convinced by the rebuttal. They thought that using both genetic features and image features makes sense and is a good design choice.



Author Feedback

We sincerely appreciate the Area Chairs for managing the review process, and the reviewers for their constructive feedback recognizing our framework’s zero-shot scalability and ability to address closed-set limits. Following MICCAI policy prohibiting new experimental results, we address the major concerns grouped by topic below.

1.Phrasing, Notation & Claims [R1,R2,M] “First universal retrieval”: To prevent misinterpretation, as in our intro, we don’t claim priority over the retrieval paradigm. By “first universal,” we specifically meant the first model-agnostic framework—a plug-and-play module integrating with any backbone. We will refine positioning to a “model-agnostic retrieval-augmented framework.” [R2] Interpretability: We do not assert superior clinical reasoning, but simply emphasize our approach’s transparency compared to SOTA black-box models (e.g., Stem, TRIPLEX, Foundation-only). We will precisely articulate this as the “inherent transparency of reference-based models.” [R3] Eq.(1) Notation: For mathematical rigor, we will explicitly redefine D_img∈R^{dxN} and D_anchor∈R^{kxN} as matrices.

2.Scalability & Bank Dependency [R2,R3,M] Cost & Scalability: Test-spot optimization concerns are resolved by our joint structure. Eq.(1) is algebraically equivalent to a multi-output ElasticNet problem. Using FISTA solvers, it processes via batched matrix operations for fast convergence, completing inference within minutes per slide. Since large-scale ST profiling is typically run as offline batches clinically, we expect this to be highly practical. [R1,M] Bank Dependency & OOD: While OOD vulnerability is a universal challenge, we expect our bank dependency to turn it into a core strength. Just as traditional RAG dynamically adapts by updating reference documents, GeneRAG enables instant domain adaptation simply by replacing the bank with target data, without weight updates. This parallel inspired our naming.

3.Baselines & Experimental Setup [R2,M] Asymmetric baseline (DeepSpot) & ElasticNet Rationale: Comparing against DeepSpot was inevitable. No prior retrieval work attempted prediction on such a massive scale (5,000 genes). To our knowledge, DeepSpot is the most recent baseline attempting 5,000-gene prediction, making it the necessary benchmark to show how our zero-shot framework breaks their parametric “closed-set limits.” Furthermore, a direct comparison against the existing kNN-based SOTA model (BLEEP) is already provided in Table 1.Therefore, replacing ElasticNet with kNN here is an internal ablation, not a SOTA comparison. Simple kNN is inherently limited by heuristic constraints (e.g., fixed k) and redundancy. Adopting ElasticNet was a fundamental evolution; its L1 penalty actively penalizes redundancy, dynamically determining a sparse, complementary reference set that kNN cannot achieve. [R2,R3,M] Foundation models set up & Contamination: We strictly applied linear probing (frozen encoder) to assess pure representational power. While image-level overlap with HEST-1k is possible during pretraining, models were blind to ST labels (0% contamination). Low standalone backbone performance vs. the massive leap upon integrating GeneRAG confirms our pure contribution. [R3,M] Dataset Split & Batch Effects: We did not employ arbitrary splits, strictly adhering to the SOTA Stem (ICLR 2025) Leave-One-Slide-Out (LOSO) protocol (p.6, ‘~strictly following the experimental design [21]’). The datasets’ diverse compositions naturally subjected our framework to both scenarios: (Intra-patient): In Prostate and HER2ST, LOSO inherently tests intra-patient generalization. (Inter-patient & Batch Effects): In the Kidney dataset (23 slides, 22 patients), LOSO naturally forms a strict inter-patient test; test slides (e.g., ‘20-0038’) belong to completely unseen patients. To ensure the evaluation context is entirely transparent, we will explicitly clarify these setups in the text and release all codes via our GitHub upon publication.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    Paper Summary The paper proposes GeneRAG, a retrieval-augmented framework for predicting spatial transcriptomics from histology images. It aims to address limitations of existing approaches, particularly lack of interpretability and inability to generalize to unseen genes. The method first generates initial gene expression predictions and morphological features using a pretrained backbone, then formulates a dual-constrained retrieval process that leverages both image similarity and partial gene predictions to retrieve relevant samples from a reference bank. A sparse ElasticNet optimization is used to weight retrieved samples, and the final full gene expression profile is reconstructed as a weighted combination of these references, enabling prediction of unseen genes and improving performance across multiple backbones .

    Key Strengths The paper presents a conceptually simple and interpretable framework that provides transparent predictions through explicit retrieval and weighting of reference samples. The approach is flexible and model-agnostic, allowing integration with different backbone architectures without additional training. It demonstrates consistent empirical improvements and introduces a principled optimization-based retrieval mechanism that avoids additional trainable parameters. The ability to generalize to unseen genes through leveraging gene correlations is viewed as a meaningful capability, and the combination of morphological and gene-based retrieval is considered a reasonable and well-motivated design.

    Key Weaknesses The novelty is considered limited due to similarities with prior retrieval-based approaches, and some claims about being the first of its kind are seen as overstated. Concerns are raised about fairness and clarity of experimental comparisons, particularly in zero-shot settings where the method benefits from access to a reference bank. The evaluation lacks sufficient detail regarding dataset splits, baseline implementation, and potential data overlap from pretrained models. There are also concerns about scalability due to per-sample optimization, as well as robustness and generalization under distribution shifts given reliance on a reference bank. Some methodological descriptions and formulations are unclear, and reproducibility details are insufficient.

    Review Summary All reviewers agree that the paper addresses an important problem and proposes an interpretable and flexible retrieval-based framework with promising empirical performance and zero-shot capabilities. There is consensus that the design is intuitive and that combining morphological and gene information for retrieval is reasonable. However, reviewers also consistently raise concerns about limited methodological novelty and insufficient clarity in positioning relative to prior work. There is further agreement that evaluation details and experimental fairness need improvement, particularly regarding baseline comparisons, dataset assumptions, and potential data leakage. Given these mixed perspectives, the paper would benefit from clarification and stronger justification before a final decision.

    Decision Invite to rebuttal. The paper shows promise but has unresolved concerns regarding novelty, evaluation fairness, and clarity.

    Rebuttal Focus The rebuttal should focus on addressing concerns, particularly by clarifying the novelty and positioning relative to prior retrieval-based methods, including correcting or moderating claims. The authors should justify the fairness of experimental comparisons, especially in zero-shot settings, and provide clearer descriptions of baseline implementations and dataset splits, including assumptions about intra- and inter-patient generalization. It is also important to address potential data leakage from pretrained models, discuss computational cost and scalability of the optimization process, and clarify unclear methodological formulations.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The rebuttal successfully addressed most of the key concerns raised in the initial reviews, particularly regarding novelty positioning, experimental protocol, baseline implementation, and scalability. These clarifications convinced two reviewers to raise their assessments to acceptance. While concerns about reference-bank dependence and OOD robustness remain, they are not sufficient to outweigh the method’s demonstrated effectiveness and practical value. Therefore, I recommend acceptance.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The paper addresses the limitations and black-box nature of deep-learning regression models in histopathology-based spatial transcriptomics prediction. The approach implements a dual-constrained sparse optimization method that explicitly selects and weights paired reference tissue patches to predict high-dimensional expression panels. Its key strengths include an interpretable inference trail, and a plug-and-play design. The rebuttal deflated original overstatements regarding priority, clarified the FISTA matrix operations and the dataset splits. The technical soundness supports acceptance.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    This paper proposes GeneRAG, a retrieval-augmented framework for spatial transcriptomics prediction with the ability to generalize to unseen genes. Reviewers appreciated the model-agnostic design, the retrieval-and-reconstruction formulation, and the strong performance across multiple datasets and backbone models. I agree with Reviewers #2 and #3 that the proposed framework addresses an important limitation of existing methods and provides a practical and interpretable solution. The main concerns were related to novelty positioning, scalability, and dependence on the reference bank. The rebuttal satisfactorily clarified the novelty claims, experimental protocol, and computational considerations. While the dependence on the reference bank remains a limitation, this is a common challenge for retrieval-based approaches and does not undermine the main contribution. Overall, I find the work technically sound, well evaluated, and worthy of acceptance.



back to top