Pith. sign in

REVIEW 3 major objections 3 minor 25 references

Zero-Cost Virtual RNA: Approximating Immunotherapy Signatures via Cross-Modal WSI Retrieval

T0 review · 3 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read By aligning tissue slides and RNA in a shared latent space during training, VITA retrieves similar historical cases from an H&E slide alone and imputes the continuous 10-gene RNA signature, reaching 0.72 classification accuracy and 0.66 Spe

desk verdict VITA is a plausible proof-of-concept for retrieval-based virtual RNA from H&E, but the reported numbers don't yet show that the cross-modal training beats a frozen CONCH baseline. read the letter →

arxiv 2608.00544 v1 pith:I7QN76YW submitted 2026-08-01 cs.CV

Sigrid Vila-Bagaria , Mar Teixidó , Miquel Piñol , Felip Vilardell , Robert Montal , Veronica Vilaplana This is my paper
classification cs.CV
keywords computationalpathologycross-modalretrievalgastricadenocarcinomaimmunotherapyresponseRNAsignatureimputationwhole-slideimagescontrastivelearningvirtualtranscriptomics
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that the continuous 10-gene RNA signature used to predict immunotherapy response in gastric cancer can be approximated without running genomic sequencing. The proposed method, VITA, trains a dual-stream model on paired whole-slide images and RNA values so that both modalities land in a shared latent space. At inference the RNA stream is removed; a new tissue slide is embedded, its nearest neighbors are found among historical cases with known RNA, and averaging their signatures gives an imputed continuous profile. The paper reports 0.72 classification accuracy and 0.66 Spearman correlation between imputed and real RNA, matching a strong slide-level classifier while adding the continuous output. If this holds, pathologists get an interpretable, zero-cost pre-screening tool that preserves the biological spectrum instead of forcing a binary inflamed/non-inflamed label.

What carries the argument

The shared latent space is the load-bearing object: a dual-stream encoder—a frozen visual backbone with Fourier positional encodings and histology-conditioned attention for whole-slide patches, and a self-normalizing network over the 10-gene RNA values—projected into one common space by a supervised contrastive loss and an auxiliary binary cross-entropy loss. At inference, the machinery is k-nearest-neighbor retrieval with k=5: a query slide is embedded into that space, the five nearest historical cases are retrieved, and the mean of their known RNA signatures becomes the imputed continuous profile, with a linear probe optionally providing the binary inflamed/non-inflamed call.

What would settle it

Run a permutation test on the retrieval database: randomly shuffle the RNA signatures across the reference cases and repeat the same k-nearest-neighbor imputation on the test slides. If the shuffled database still yields a Spearman correlation near 0.66, the imputation is not actually exploiting morphology–RNA alignment; if the correlation collapses, the alignment is doing the work. A second check is external validation on slides from a different hospital or scanner: a sharp drop in imputed-signature correlation would show the retrieval generalizes only within the original cohort.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that cross-modal retrieval is a viable substitute for direct fusion or binary classification when predicting the 'Inflamed' immunophenotype in gastric adenocarcinoma. Instead of learning a single decision boundary over continuous RNA data, VITA aligns H&E and RNA embeddings during training with a supervised contrastive loss plus an auxiliary classification loss, then uses the aligned geometry at inference to retrieve five morphologically similar cases from the training database. The averaged RNA profiles of those neighbors constitute the imputed continuous signature. The paper reports that this retrieval-based imputation achieves 0.72 accuracy and 0.72 F1 o

Load-bearing premise

The load-bearing premise is that, once aligned in the joint space, morphologically similar H&E slides carry similar 10-gene RNA signatures; the paper does not independently validate this and it is not guaranteed by the training objective.

Editorial extensions

If this is right

  • If VITA's claim holds, gastric cancer immunotherapy pre-screening becomes possible from standard H&E slides alone, eliminating the sequencing cost of the 10-gene RNA test.
  • Because the output is a continuous imputed signature rather than a hard label, clinicians could track a patient's position on the inflamed-to-cold spectrum and feed it into other models that expect continuous molecular inputs.
  • Retrieval makes the prediction interpretable: each imputed value is traceable to specific historical slides whose RNA profiles were averaged, so a pathologist can audit the result.
  • The same training-time alignment / inference-time retrieval recipe could be reapplied to any cancer type with paired histology and molecular measurements, not just gastric cancer.
  • The binary classification can be derived either from the imputed signature or from the concurrently trained linear probe, so the approach does not sacrifice classification performance to gain the continuous output.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The fixed k=5 neighborhood is a hand-set hyperparameter; distance-weighted or adaptively sized neighborhoods could improve imputation fidelity, a direction the paper only gestures at in future work.
  • If morphological similarity genuinely carries transcriptomic signal, a large and diverse retrieval database is essential: on morphologically rare gastric cancer subtypes, the five nearest neighbors in a small cohort are unlikely to be truly similar, so imputation quality will degrade.
  • The 'zero-cost' framing holds at deployment, not during development: training still requires paired RNA measurements, so the economic benefit accrues to institutions that adopt the model rather than to the cohort that generated the paired data.
  • The continuous imputed signature could serve as a synthetic biomarker for exploratory analyses such as correlating morphology-derived RNA with survival or drug response, but any such use inherits the errors of the retrieval approximation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. The paper proposes VITA, a cross-modal retrieval framework that aligns H&E whole-slide images and a 10-gene RNA signature in a shared latent space during training, then at inference uses only H&E to retrieve k=5 nearest neighbors from a labeled database and imputes the continuous RNA signature by averaging their ground-truth profiles. The method is evaluated on a TCGA gastric adenocarcinoma cohort (N=265) with 5-fold cross-validation, reporting 0.72 classification accuracy and 0.66 Spearman correlation for imputation, with comparisons to CLAM, MCAT, and a frozen CONCH retrieval baseline.

Significance. If the central claim were established, VITA would be a useful and practical contribution: it addresses a real clinical need (avoiding expensive RNA sequencing) and does so with an interpretable retrieval-based framework that preserves the continuous phenotypic spectrum. The use of a frozen foundation-model backbone, a simple MLP projection, and a supervised contrastive objective is clean and reproducible in principle. The paper is also honest in its limitations section, conceding that gains are modest and validation is single-cohort. However, the paper's core claim—that the joint cross-modal alignment provides a measurable benefit over off-the-shelf morphological features for RNA imputation—is not statistically supported by the presented results, and the main retrieval evaluation uses a clinically unrealistic Test→Test setting. These issues are load-bearing for the claimed contribution.

major comments (3)
  1. [Section 3, Table 2] The clinically relevant setting for a pre-screening tool is retrieval from a historical database of known cases, i.e., Test→Train. Table 2 reports R@5 and Spearman for Test→Test, but only Spearman for Test→Train. In Test→Test, the query and the database come from the same partition, which can include the query's own slide or near-duplicates and is not the scenario described in Section 2.1 ('new queries against a database of known cases'). The sentence in Section 3 claiming to 'emulate a real-world clinical scenario' is therefore contradicted by the actual evaluation. The authors should report R@5 and Spearman for Test→Train, and ensure the database excludes the query.
  2. [Table 2, Section 3] The claimed superiority of VITA over the frozen CONCH retrieval baseline is not supported by any significance test, confidence interval, or per-fold comparison. With only 5 folds (N=265), the differences are within one standard deviation (Test→Train Spearman 0.66±0.03 vs 0.64±0.04; Test→Test Spearman 0.62±0.03 vs 0.62±0.08; R@5 0.54±0.12 vs 0.42±0.05). The phrase 'outperforms... with superior recall' is not justified by these numbers. A paired test across folds (e.g., Wilcoxon or permutation test) or reporting per-fold paired differences is needed. If no significant improvement is found, the conclusion that cross-modal alignment is what enables zero-cost virtual RNA should be softened or removed.
  3. [Section 2.1, Fig. 1] The retrieval neighborhood size k=5 is described as 'determined empirically,' but no sensitivity analysis is shown. Since the entire imputation mechanism depends on averaging the RNA profiles of retrieved neighbors, the result may be sensitive to k. The authors should report Spearman and R@5 for a range of k values (e.g., k=1,3,5,10,20) or provide a justification for k=5. Without this, the reader cannot assess the robustness of the central zero-cost RNA approximation.
minor comments (3)
  1. [Throughout] There are several formatting artifacts: 'T raining' in Section 2.1, unusual diacritics in author names ('V er´ onica'), and inconsistent use of 'ρs' in Table 2. A final proofread is recommended.
  2. [Section 2.2] The description of the ground-truth label derivation is terse. It would be helpful to state explicitly how the continuous 10-gene signature is thresholded into 'Inflamed'/'Non-Inflamed' (e.g., median split or a predefined cutoff from Rodriguez et al.). Currently the cohort split (142/123) appears without justification.
  3. [Section 2.1] The histological stream includes 'Fourier Positional Encodings' and 'tumour histology-conditioned attention,' but these components are not further specified. Since the paper is short, a brief reference or architectural detail would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: retrieval-based RNA imputation is evaluated on held-out ground truth.

full rationale

VITA's inference-time RNA imputation is a k-nearest-neighbor average of the ground-truth RNA signatures of retrieved training cases. Although the retrieval space is learned with a supervised contrastive loss on the same RNA signature, the test-fold RNA labels are held out and never enter either the embedding training or the retrieval/imputation step. The reported Spearman correlation and classification accuracy are therefore measured against external ground truth, not against the training objective. The 10-gene signature is defined in Rodriguez et al. (2025), which shares an author with the present paper, but this is a label-definition provenance, not a load-bearing circular argument: the model would train and evaluate identically against any externally defined label set, and the paper makes no attempt to derive the signature from the model. The comparison vs. CONCH is statistically underpowered (gains within one standard deviation) and the paper concedes modest gains, but statistical weakness is a correctness risk, not circularity. No equation or construction in the paper defines the target in terms of the prediction, no fitted parameter is renamed as a prediction, and no uniqueness theorem is imported from the authors' prior work. The central derivation—paired alignment during training, unimodal H&E retrieval at inference—is self-contained and externally evaluated on held-out data.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

All experimental outcomes depend on the retrieval hypothesis and the validity of the cited RNA signature; the only explicit fitted parameter is k=5. The ground-truth signature is taken from a prior paper by overlapping authors, and no independent validation is provided.

free parameters (1)
  • k (number of nearest neighbors for retrieval) = 5
    Determined empirically on the dataset (Section 2.1); no sensitivity analysis or independent validation of this choice is provided.
assumptions (4)
  • domain assumption Morphologically similar H&E slides (in the aligned space) share similar 10-gene RNA signatures.
    This is the retrieval hypothesis of VITA, used in Section 2.1 to justify KNN RNA imputation at inference. If false, retrieval-based imputation cannot work.
  • domain assumption The 10-gene 'Inflamed' RNA signature is a valid proxy for immunotherapy response and its binarization into Inflamed/Non-Inflamed is correct.
    Used as ground truth for training and evaluation. The signature comes from Rodriguez et al. 2025, which shares co-authors with this paper, and no independent validation is provided here.
  • domain assumption TCGA cases are representative and the WSI-RNA pairing is correct.
    All experiments rely on TCGA data; curation criteria for the 265 cases are not described, so selection effects are possible.
  • standard math Supervised contrastive loss with the given positive/negative sampling yields a well-behaved embedding space.
    Standard optimization assumption; no analysis of convergence or embedding quality is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Zero-Cost Virtual RNA: Approximating Immunotherapy Signatures via Cross-Modal WSI Retrieval." pith.science (2026). https://pith.science/paper/I7QN76YW

@misc{pith2026260800544,
  author       = {Pith},
  title        = {Pith review of: Zero-Cost Virtual RNA: Approximating Immunotherapy Signatures via Cross-Modal WSI Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I7QN76YW}},
  note         = {Machine review of arXiv:2608.00544}
}
read the original abstract

Identifying the ``Inflamed'' immunophenotype in Gastric Adenocarcinoma predicts immunotherapy response but requires an expensive 10-gene RNA signature. While deep learning on standard H\&E slides offers a scalable alternative, conventional binary classifiers oversimplify continuous RNA data and introduce label noise. To resolve this, we propose VITA (VIrtual Transcriptomic Approximation). By aligning H\&E and RNA into a joint latent space during training, VITA requires only standard H\&E at inference to retrieve morphologically similar historical cases and approximate the continuous RNA signature. Achieving 0.72 classification accuracy and a 0.66 Spearman correlation, VITA provides a cost-effective ``virtual transcriptomics'' pre-screening tool that preserves the continuous phenotypic spectrum without requiring genomic sequencing.

Figures

Figures reproduced from arXiv: 2608.00544 by the authors.

Figure 1
Figure 1. Cross-modal retrieval training and inference framework. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Cross-modal retrieval. Overcoming the Domain Gap: Direct mul￾timodal fusion (MCAT) struggles in data-scarce regimes (Tab. 1). While unimodal baselines (CLAM) perform well, they rigidly binarize con￾tinuous data. VITA addresses this by matching unimodal classification performance while aligning H&E and RNA modalities, setting the foundation for downstream transcriptomic imputation. Zero-Cost RNA via Retrieval: Operat… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 15 canonical work pages

  1. [1]

    International conference on machine learning , pages=

    Attention-based deep multiple instance learning , author=. International conference on machine learning , pages=. 2018 , organization=

    2018

  2. [2]

    Advances in neural information processing systems , volume=

    Transmil: Transformer based correlated multiple instance learning for whole slide image classification , author=. Advances in neural information processing systems , volume=

  3. [3]

    arXiv preprint arXiv:2502.06750 , year=

    Accelerating Data Processing and Benchmarking of AI Models for Pathology , author=. arXiv preprint arXiv:2502.06750 , year=

    arXiv

  4. [4]

    and Chen, Bowen and Zhang, Andrew and Williamson, Drew F

    Lu, Ming Y. and Chen, Bowen and Zhang, Andrew and Williamson, Drew F. K. and Chen, Richard J. and Ding, Tong and Le, Long Phi and Chuang, Yung-Sung and Mahmood, Faisal , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2023 , pages =

    work page 2023

  5. [5]

    Advances in neural information processing systems , volume=

    Self-normalizing neural networks , author=. Advances in neural information processing systems , volume=

  6. [6]

    Journal for Immunotherapy of Cancer , volume=

    Comprehensive immunophenotyping of gastric adenocarcinoma identifies an inflamed class of tumors amenable to immunotherapies , author=. Journal for Immunotherapy of Cancer , volume=

    work page

  7. [7]

    CA: a cancer journal for clinicians , volume=

    Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries , author=. CA: a cancer journal for clinicians , volume=. 2021 , publisher=

    2020

  8. [8]

    Acta pathologica microbiologica Scandinavica , volume=

    The two histological main types of gastric carcinoma: diffuse and so-called intestinal-type carcinoma: an attempt at a histo-clinical classification , author=. Acta pathologica microbiologica Scandinavica , volume=. 1965 , publisher=

    work page 1965

Show all 25 references
  1. [9]

    Nature biomedical engineering , volume=

    Data-efficient and weakly supervised computational pathology on whole-slide images , author=. Nature biomedical engineering , volume=. 2021 , publisher=

    2021

  2. [10]

    2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06) , volume=

    Dimensionality reduction by learning an invariant mapping , author=. 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06) , volume=. 2006 , organization=

    2006

  3. [11]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  4. [12]

    Machine learning for healthcare conference , pages=

    Contrastive learning of medical visual representations from paired images and text , author=. Machine learning for healthcare conference , pages=. 2022 , organization=

    2022

  5. [13]

    Proceedings of the Conference on Empirical Methods in Natural Language Processing

    Medclip: Contrastive learning from unpaired medical images and text , author=. Proceedings of the Conference on Empirical Methods in Natural Language Processing. Conference on Empirical Methods in Natural Language Processing , volume=

  6. [14]

    Computers in Biology and Medicine , volume=

    Contrastive learning-based histopathological features infer molecular subtypes and clinical outcomes of breast cancer from unannotated whole slide images , author=. Computers in Biology and Medicine , volume=. 2024 , publisher=

    2024

  7. [15]

    Computer Methods and Programs in Biomedicine , volume=

    Cross-modal alignment and contrastive learning for enhanced cancer survival prediction , author=. Computer Methods and Programs in Biomedicine , volume=. 2025 , publisher=

    2025

  8. [16]

    Briefings in Bioinformatics , volume=

    Multimodal contrastive learning for spatial gene expression prediction using histology images , author=. Briefings in Bioinformatics , volume=. 2024 , publisher=

    2024

  9. [17]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Multimodal co-attention transformer for survival prediction in gigapixel whole slide images , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  10. [18]

    bioRxiv , pages=

    Contrastive Learning for Omics-guided Whole-slide Visual Embedding Representation , author=. bioRxiv , pages=. 2025 , publisher=

    2025

  11. [19]

    Nature Medicine , pages=

    A visual-language foundation model for computational pathology , author=. Nature Medicine , pages=. 2024 , publisher=

    2024

  12. [20]

    The Lancet , volume=

    First-line nivolumab plus chemotherapy versus chemotherapy alone for advanced gastric, gastro-oesophageal junction, and oesophageal adenocarcinoma (CheckMate 649): a randomised, open-label, phase 3 trial , author=. The Lancet , volume=. 2021 , publisher=

    2021

  13. [21]

    The Lancet , volume=

    Pembrolizumab plus trastuzumab and chemotherapy for HER2-positive gastric or gastro-oesophageal junction adenocarcinoma: interim analyses from the phase 3 KEYNOTE-811 randomised placebo-controlled trial , author=. The Lancet , volume=. 2023 , publisher=

    2023

  14. [22]

    The Lancet Oncology , volume=

    Pembrolizumab plus chemotherapy versus placebo plus chemotherapy for HER2-negative advanced gastric cancer (KEYNOTE-859): a multicentre, randomised, double-blind, phase 3 trial , author=. The Lancet Oncology , volume=. 2023 , publisher=

    2023

  15. [23]

    Annals of Oncology , volume=

    Novel patterns of response under immunotherapy , author=. Annals of Oncology , volume=. 2019 , publisher=

    2019

  16. [24]

    Nature reviews Disease primers , volume=

    Gastric adenocarcinoma , author=. Nature reviews Disease primers , volume=. 2017 , publisher=

    2017

  17. [25]

    Acta Pathologica Microbiologica Scandinavica , volume =

    LAURÉN, PEKKA , title =. Acta Pathologica Microbiologica Scandinavica , volume =. doi:https://doi.org/10.1111/apm.1965.64.1.31 , url =. https://onlinelibrary.wiley.com/doi/pdf/10.1111/apm.1965.64.1.31 , year =

    1965 doi

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.