{"id":"afce84f4-3fff-42bd-bbc3-6122b905dd5f","arxiv_id":"2608.00544","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"VITA aligns H&E slides and RNA signatures in a shared latent space, then imputes a patient's RNA signature by retrieving morphologically similar historical cases, reaching 0.72 classification accuracy and 0.66 Spearman correlation.","lead":"This paper introduces VITA, a method that uses deep learning on H&E tissue slides to estimate a 10-gene RNA immune signature without doing genetic sequencing. It retrieves similar historical cases from a database and averages their RNA profiles, potentially offering a cheap pre-screening approach for immunotherapy response in gastric cancer.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"VITA's claimed advantage over CONCH retrieval is within noise; cross-modal alignment's contribution to virtual RNA is unestablished.","rationale":"The reader's weakest assumption was that the retrieval hypothesis—morphologically similar H&E slides carry similar RNA signatures—is unvalidated. Table 2 is in fact a direct validation attempt: it correlates imputed RNA from retrieved neighbors with ground truth. The more precise and load-bearing weakness is that this validation does not show VITA's cross-modal alignment outperforms a trivial baseline. The Spearman difference (0.66 vs 0.64) is tiny, and the R@5 difference has overlapping standard deviations. Without a significance test, the observed numbers are consistent with chance, so the paper's central novelty—the joint embedding trained with supervised contrastive loss—is not demonstrated to add value. This is not an external-consensus disagreement; it is an internal statistical-evidence weakness. The paper has strengths: a clear architecture, a sensible problem, and an explicit limitations section. But the empirical support for the headline claim is currently insufficient. I agree with the reader's CONDITIONAL verdict; the condition should specifically include a significance test against the CONCH retrieval baseline. Hence no change to the reader's verdict.","tokens_in":3587,"tokens_out":5720,"duration_ms":75698,"concrete_test":"Perform a paired permutation test on the Test→Train Spearman correlation difference between VITA and CONCH. Concretely: for each test slide, compute the Spearman correlation between the imputed 10-gene RNA profile (k-NN average of retrieved train profiles) and the ground-truth RNA profile, once using VITA's latent space and once using frozen CONCH features. Then randomly permute the method labels within each fold 10,000 times to build a null distribution of the mean paired difference. Report the two-sided p-value and a bootstrap 95% CI. If p>0.05 or the CI includes 0, the claimed superiority over CONCH is unsupported and the central claim should be weakened or rejected.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that VITA's joint H&E-RNA alignment enables accurate retrieval-based RNA imputation. Table 2 is the direct test: Test→Train Spearman is 0.66±0.03 for VITA versus 0.64±0.04 for a frozen CONCH feature baseline; Test→Test Spearman is identical (0.62±0.03 vs 0.62±0.08), and R@5 differs 0.54±0.12 vs 0.42±0.05. No significance test, confidence interval, or per-fold comparison is provided. With only 5 folds from N=265, the observed differences are within one standard deviation, so the claimed superiority of VITA over CONCH is not established. If the true difference is zero or negative, the joint alignment provides no measurable benefit over an off-the-shelf morphology encoder, and the paper's core contribution—that cross-modal training is what makes zero-cost virtual RNA feasible—collapses. The paper's own Limitations section concedes 'gains are modest,' but that does not replace a formal test of whether the gain is real.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":3890,"tokens_out":2982,"duration_ms":36306,"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":[{"comment":"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.","section":"Section 3, Table 2"},{"comment":"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.","section":"Table 2, Section 3"},{"comment":"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.","section":"Section 2.1, Fig. 1"}],"minor_comments":[{"comment":"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.","section":"Throughout"},{"comment":"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.","section":"Section 2.2"},{"comment":"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.","section":"Section 2.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for MIDL as a short paper and addresses an important problem. However, the central quantitative claim rests on a retrieval evaluation that is not clinically realistic and on differences that are not statistically tested. The authors should be asked to provide a proper Test→Train evaluation and a paired significance test. If the comparison against frozen CONCH features remains within noise, the paper should be reframed as an exploration rather than a demonstrated improvement, or additional evidence (e.g., sensitivity analysis, larger cohort) is needed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a sensible, well-written proof of concept, and the retrieval idea is genuinely new for this problem. The catch is that the central quantitative claim—that joint H&E-RNA alignment is what makes the imputation work—is not actually supported by the numbers in Table 2. Test→Train Spearman is 0.66±0.03 vs 0.64±0.04 for frozen CONCH features; Test→Test Spearman is identical at 0.62; and R@5 is 0.54±0.12 vs 0.42±0.05. With 5 folds on N=265, those differences are within one standard deviation. No significance test or per-fold comparison is given. So the core contribution is plausible but unestablished. The paper's own Limitations section concedes 'gains are modest,' but modest is not the same as shown.\n\nWhat is actually new: framing immunophenotype classification as cross-modal retrieval, so you can impute a continuous 10-gene signature from H&E at inference time without a genomic stream. That is a real application-level contribution, and the evaluation is mostly honest: they test against a CONCH-feature retrieval baseline, which is exactly the right control, and they report both classification and correlation. The use of a single TCGA cohort and the self-cited ground-truth label (Rodriguez et al., overlapping authors) are limitations, but not fatal; the label is what it is.\n\nSoft spots beyond the noise issue: k=5 is said to be 'determined empirically' with no sensitivity analysis. The Test→Test retrieval setup is not clinically realistic; the real scenario is Test→Train, which is reported but without R@5. And reproducibility: no code or data release is mentioned, which is disappointing for a short paper. Minor: 'superior recall' is an overstatement given the confidence intervals.\n\nWho this is for: anyone working on virtual transcriptomics or H&E-based biomarker prediction. It's a useful proof of concept, but it needs a significance test, sensitivity analysis on k, and ideally external validation before the claims are taken as established. I'd send it to peer review—the idea is worth engaging with—but I'd expect revision.","headline":"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.","tokens_in":4315,"tokens_out":1951,"would_cite":false,"duration_ms":20247,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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","keywords":["computational pathology","cross-modal retrieval","gastric adenocarcinoma","immunotherapy response","RNA signature imputation","whole-slide images","contrastive learning","virtual transcriptomics"],"falsifier":"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.","tokens_in":3574,"feed_emoji":"🧬","tokens_out":10416,"duration_ms":90259,"temperature":0.7,"pith_summary":"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.","feed_headline":"Tissue slides alone approximate the immunotherapy RNA signature","feed_subtitle":"A new model imputes the immunotherapy signature from ordinary slides, matching classifier accuracy without sequencing.","key_machinery":"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.","core_discovery":"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","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the unimodal whole-slide classifier baseline that VITA must match and the weakly-supervised formulation the task builds on.","marker":"Lu et al., 2021"},{"why":"Supplies the multimodal-fusion baseline that VITA compares against to show direct fusion struggles in data-scarce regimes.","marker":"Chen et al., 2021"},{"why":"Provides the frozen visual backbone used in the histological stream to extract patch features.","marker":"Lu et al., 2024"},{"why":"Supplies the self-normalizing network architecture used to encode the 10-gene RNA signature.","marker":"Klambauer et al., 2017"},{"why":"Defines the 'Inflamed' immunophenotype and the 10-gene RNA signature that VITA imputes and classifies.","marker":"Rodriguez et al., 2025"},{"why":"Supplies the histology classification used to stratify cases and condition the attention in the histological stream.","marker":"Lauren, 1965"}],"fun_headline_variants":["H&E slides alone impute RNA immunotherapy signature","No sequencing: retrieval predicts inflamed immunophenotype","Virtual transcriptomics from standard slides via retrieval","Cross-modal WSI retrieval approximates RNA signature","Slides-only model imputes immunotherapy signature"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["H&E slides alone impute RNA immunotherapy signature","No sequencing: retrieval predicts inflamed immunophenotype","Virtual transcriptomics from standard slides via retrieval","Cross-modal WSI retrieval approximates RNA signature","Slides-only model imputes immunotherapy signature"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000194,"raw_usage":{"total_tokens":1139,"prompt_tokens":643,"completion_tokens":496,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":387,"completion_tokens_details":{"reasoning_tokens":426}},"tokens_in":387,"tokens_out":496,"duration_ms":5770,"temperature":1.0,"reasoning_tokens":426,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T00:39:47.509050+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[],"review_version":1}