{"id":"7cf5955e-dfd6-4096-aea3-165982387a7f","arxiv_id":"2411.18391","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A QA-style model that takes gene text descriptions as queries and histology images as context predicts spatial gene expression for both known and unseen genes.","lead":"GeneQuery reframes the task of predicting gene expression from tissue images as a question-answering problem: the histology image is the context and a text description of a gene is the query. The authors report that this design predicts both known and unseen genes better than three prior methods on liver and breast spatial transcriptomics datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The unseen-gene claim is untested against a gene-agnostic baseline; if the model ignores the gene query, the QA reformulation collapses.","rationale":"The reader's weakest_assumption identifies the same load-bearing premise: a short text description must be sufficient, when fused with image features, to determine a gene's spatial expression pattern for unseen genes. I agree this is the decisive assumption. The paper provides no ablation that isolates the contribution of the gene query to unseen-gene performance; Table 3 only shows absolute PCC values, which could be achieved by a gene-agnostic model predicting common spatial trends. The observation that unseen genes sometimes outperform seen genes (GSE, 20%: HEG 0.137 vs 0.117) is a red flag that the model may not be conditioning on gene identity at all. The GPT-4 experiment in Section 4.6 is suggestive but lacks statistical rigor, so it does not settle the question. The experimental protocol also has a secondary issue: HER2+ random WSI-level splits likely leak patient identity across train/test, which would inflate both known- and unseen-gene results; this reinforces the need for the proposed ablation under patient-stratified folds. Because the central conceptual novelty is at stake and the current evidence does not rule out the gene-agnostic alternative, the appropriate response is to keep the reader's CONDITIONAL verdict and add the query-ablation control as an explicit condition. I therefore recommend UNCHANGED rather than strengthening or weakening the verdict.","tokens_in":10759,"tokens_out":5443,"duration_ms":53465,"concrete_test":"Run the Table 3 unseen-gene protocol on GSE and HER2+ with three query conditions: (1) real gene metadata, (2) a constant gene embedding shared by all genes, and (3) gene-metadata pairs randomly permuted during training. If conditions (2) or (3) yield statistically indistinguishable unseen-gene PCC from condition (1) on the same folds (with paired significance tests and patient-stratified splits for HER2+), the model is not using gene-specific metadata and the central zero-shot generalization claim is unsupported. Report per-gene PCC mean and variance for each condition.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that reformulating gene expression prediction as QA with gene metadata as queries enables zero-shot prediction for unseen genes. The load-bearing premise, introduced in Section 3.2 (Eqs. 3-7), is that the gene text embedding, when fused with image features, determines a gene-specific spatial expression pattern for genes never seen during training. Table 3 reports positive PCC for unseen genes, but there is no comparison to a gene-agnostic control: a model with a constant/empty gene embedding, or one trained with permuted gene-metadata pairs. If such a control achieves similar unseen-gene PCC, then predictions are driven by a generic spatial expression trend common to HEG/HVG genes, not by the gene query. Supporting this worry, Table 3 shows unseen-gene HEG PCC at 20% seen genes on GSE (0.137) is actually higher than seen-gene HEG PCC (0.117), which is difficult to explain if the query is informative and easier to explain if the model is predicting a dataset-average spatial pattern. The GPT-4 metadata improvements in Table 5 are evaluated on one fold without error bars or significance tests, so they do not establish that the query is necessary. Additionally, the HER2+ dataset has 8 patients and 36 slides; the 5-fold random split of histology images almost surely places slides from the same patient in both training and test, which can inflate known- and unseen-gene PCC through patient-specific morphology. The primary conceptual concern is the untested sufficiency of the gene query, because if the model is gene-agnostic, the claimed QA-based zero-shot generalization is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GeneQuery, a framework that reformulates prediction of spatial gene expression from H&E whole-slide images as a question-answering task: a spot image is the context, gene metadata is the query, and the model regresses the queried gene's expression. Two fusion architectures are presented, spot-aware GeneQuery and gene-aware GeneQuery, and the model is trained with an MSE loss over all spots and genes. The central empirical claims are (i) that this shared-weight formulation outperforms STNet, HistoGene, and BLEEP on known genes, and (ii) that it can predict genes never seen during training. Experiments are reported on GSE240429 (liver), HER2+ breast, and HBD breast datasets, together with transfer-learning results and an ablation using GPT-4-generated gene descriptions.","tokens_in":10933,"tokens_out":6169,"duration_ms":58716,"significance":"If the claims hold, the paper makes a useful contribution: replacing per-gene output heads with a model that conditions on gene identity would be a practical advance for spatial transcriptomics and would open up zero-shot prediction for new genes from histology alone. The paper's strengths are that it evaluates on multiple datasets, holds out entire genes for the unseen-gene experiments, releases code, and includes transfer experiments that are concrete and falsifiable. However, the current evidence is not yet sufficient: the evaluation lacks patient-level split control, an ablation showing that the gene query rather than a generic image trend drives unseen-gene predictions, and error-bounded comparisons for several headline claims.","major_comments":[{"comment":"The cross-validation splits are by whole-slide image, not by patient. The HER2+ set is described as collected from 8 patients with 36 WSIs, so a random 5-fold WSI split will almost certainly place slides from the same patient in both training and test. Same-patient slides share patient-specific morphology and expression, which can inflate the reported PCCs through memorization rather than generalizable prediction. Please use patient-stratified folds (all slides from one patient in the same fold) and report patient-level summary statistics. The manuscript should also state the patient/slide mapping for HBD, since the same concern applies if it contains repeated patients.","section":"§4.1–4.2, Table 2"},{"comment":"The unseen-gene result is the paper's most distinctive claim, but Table 3 contains no gene-agnostic control. The load-bearing premise, introduced in Eqs. (3)–(7), is that the gene text embedding, when fused with image features, determines a gene-specific spatial expression pattern for genes never seen during training. To test this, the authors should evaluate the same architecture with a constant or empty gene embedding, or with randomly permuted gene-metadata pairs. If such a control reaches similar unseen-gene PCC, the predictions are driven by a generic spatial-expression trend shared by HEG/HVG genes rather than by the gene query. Table 3 itself contains a warning sign: on GSE at 20% seen genes, unseen-gene HEG PCC (0.137) is larger than seen-gene HEG PCC (0.117), which is difficult to explain if the query is informative.","section":"§3.2, Eqs. (3)–(7); §4.4, Table 3"},{"comment":"Several headline claims of consistent superiority are not supported by the reported uncertainty. For example, on HER2+ HVG, GeneQuery_gene (0.318±0.025) and BLEEP (0.322±0.100) have strongly overlapping standard deviations, and on HBD ALL, GeneQuery_gene (0.061±0.007) is numerically below STNet (0.073±0.011). Please report paired significance tests across folds or across genes for every comparison, and revise the blanket statement in §4.3 that GeneQuery achieves the best performance across all datasets and settings.","section":"§4.3, Table 2"},{"comment":"The GPT-4 enhancement results are reported on a single validation fold with no standard deviations or significance tests, so the conclusion that GPT-4 metadata improves GeneQuery by 3.6% and 1.8% is not reliable. In the HER2+ rows of Table 5, spot-aware GeneQuery with GPT-4 is actually lower on all three metrics (0.342→0.336 for HEG, 0.347→0.341 for HVG, and 0.183→0.164 for ALL). Please run the ablation across the same multiple folds as the main results and report paired differences with uncertainty.","section":"§4.6, Table 5"}],"minor_comments":[{"comment":"The table header contains the typo 'Unseeen'; it should read 'Unseen'.","section":"§4.4, Table 3"},{"comment":"The section heading 'Conlusion' should be 'Conclusion'.","section":"§5"},{"comment":"Several entries in Table 4 appear fused in the text, for example '0.1170.071' and '0.0900.041'; the column separation should be fixed so each mean and standard deviation is readable.","section":"§4.5, Table 4"},{"comment":"Equation (1) writes P(y_j|x_i) for a real-valued quantity, and Eq. (2) presents an MLE-style objective that is then used as an MSE loss without an explicit likelihood or noise model; the formulation should be cleaned up or the claims about 'estimating the gene distribution' should be softened.","section":"§3.1, Eqs. (1)–(2)"},{"comment":"The clinical BERT encoder is cited to a reinforcement-learning diabetes trial; the reference appears to be incorrect and should be replaced with the actual clinical BERT publication.","section":"References, [14]"}],"recommendation":"major_revision","confidential_remarks":"The QA idea is timely and potentially useful, but the verification is currently too weak for publication. The most important fixes are patient-stratified cross-validation, a query-ablated baseline for the unseen-gene experiments, and error-bounded comparisons for the main and GPT-4 tables. I would also ask the authors to include exact gene split indices and patient/slide metadata, since the baselines and split protocols differ across prior works."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper deserves a serious look because it actually attempts something the STNet/HistoGene/BLEEP line does not: predicting expression for genes held out of training, by conditioning on text metadata. The unseen-gene experiments are genuinely external (gene names to GPT-4 descriptions, no expression data used to build queries), so the circularity concern the stress-test raised is not real. But the central claim is under-supported: there is no gene-agnostic baseline, and the reported gains over baselines are often inside the noise.\n\nWhat is new is the reformulation: instead of per-gene heads, GeneQuery shares weights across genes and fuses gene text embeddings with image features. That is a reasonable step, and the transfer results (same tissue better than cross-tissue) are believable. The two implementations, spot-aware and gene-aware, are competently described. The UMAP segmentation figures are suggestive but not quantitative.\n\nSoft spots, in order of severity. First, the evaluation splits by whole-slide image, not by patient. HER2+ has 8 patients and 36 slides; a 5-fold random slide split almost guarantees same-patient slides on both sides of the train/test boundary. That can inflate PCC. Patient-level cross-validation or at least per-fold patient reporting is needed. Second, Table 2 has overlapping error bars on several headline comparisons. HER2+ HVG: BLEEP is 0.322 +/- 0.100, GeneQuery_gene is 0.318 +/- 0.025; the text claims superiority that the numbers do not support. HBD ALL: STNet is actually ahead of gene-aware GeneQuery. No significance tests are reported. Third, the unseen-gene table has no variance or significance information, and the GSE 20% result shows unseen HEG PCC (0.137) above seen HEG PCC (0.117). The paper calls this surprising but does not explain it. If the gene query were informative, you would expect seen to beat unseen; this pattern is easier to explain if the model is predicting a generic spatial trend. A constant-query or permuted-metadata control would settle it, and it is missing. Fourth, the GPT-4 enhancement is evaluated on one fold with no error bars, so it is not convincing. Finally, the \"gene random variable\" language does no work in Eq. 2; that is ordinary conditional regression. This is a rhetorical issue, not a technical flaw.\n\nFor whom: this is a methods paper for the spatial-transcriptomics-from-histology community. If the controls are added and the evaluation is redone at patient level, the unseen-gene capability would be a real step forward. As it stands, the paper should go to review, not be desk-rejected, but the reviewers should ask for the missing baseline and patient-level splits. I would cite it as prior work on text-conditioned prediction next year.","headline":"Useful attempt at text-conditioned zero-shot gene prediction; evaluation needs patient-level splits and a no-query control before the headline claims hold.","tokens_in":11641,"tokens_out":2219,"would_cite":true,"duration_ms":21788,"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":"GeneQuery recasts gene-expression prediction as a question-answering task, letting one model predict known and unseen genes from H&E slides.","keywords":["spatial transcriptomics","gene expression prediction","histology images","question answering","gene metadata","unseen gene generalization","whole-slide images","gene-image fusion"],"falsifier":"Hold out entire patients rather than random slides when splitting HER2+ (8 patients, 36 slides) and check whether unseen-gene Pearson correlation survives; or shuffle the gene metadata among genes at test time. If predicted expression maps barely change under shuffled text, the model is not using the text query and the claimed generalization mechanism is absent.","tokens_in":10433,"feed_emoji":"🧬","tokens_out":5402,"duration_ms":46037,"temperature":0.7,"pith_summary":"The paper argues that spatial gene-expression prediction from histology images should not be treated as a separate regression problem per gene. It reformulates the task as question answering: the histology image is the context, a text description of a gene is the query, and the model outputs that gene's expression at each spot. Because gene identity enters through the query rather than through per-gene weights, one trained model can answer for any gene in the library, including genes never seen in training. The authors report that this framing outperforms per-gene baselines on known genes and gives usable Pearson correlations on unseen genes, with richer LLM-generated gene descriptions helping further. If the framing works as claimed, histology-based expression prediction becomes a general queryable capability instead of a fixed gene panel.","feed_headline":"Ask an H&E slide about any gene and get its expression","feed_subtitle":"Reformulating the task as question answering with gene metadata as the query lets one model generalize to unseen genes.","key_machinery":"The load-bearing object is the gene random variable $g$, which turns each prediction into a conditional distribution $P(y \\mid x, g)$ rather than a per-gene regressor $P(y_j \\mid x)$. This is operationalized by a gene encoder (clinical BERT) that converts gene metadata into a query vector and a fusion module that cross-attends image and gene features; because the same image encoder and regressor handle all genes, the only thing that changes between genes is the query. The two architecture variants choose which modality becomes the sequence dimension: spot-aware GeneQuery sequences spots and adds the queried gene's embedding to every spot, while gene-aware GeneQuery sequences genes and adds the spot embedding to every gene.","core_discovery":"The central claim is that introducing a gene random variable and feeding gene metadata as a query allows a single model to estimate a conditional distribution $P(y \\mid x, g)$ and thereby generalize across both spots and genes. Concretely, GeneQuery encodes each 224-by-224 spot with ResNet50, encodes gene metadata with clinical BERT, projects both to a common dimension, fuses them with a simple addition followed by two transformer blocks, and regresses expression values with mean-squared error. Two variants are proposed: spot-aware GeneQuery treats all spot images of a slide as a sequence and adds the queried gene's feature to each spot, while gene-aware GeneQuery treats the gene list as a sequence and adds the spot feature to each gene. The paper reports that this method beats STNet, HistoGene, and BLEEP on known-gene Pearson correlation and gives competitive predictions for genes held out from training, along with stronger transfer between datasets of the same tissue type.","pith_inferences":["A direct test of the mechanism would replace each gene's real description with another gene's description at evaluation; if predicted expression maps barely change, the model is not using text and unseen-gene gains must come from shared image features.","The same query-conditioned design could accept other text modalities, including pathways, protein domains, disease annotations, or free-form questions about a region, because anything that correlates with expression can enter through the gene encoder.","If the text-query mechanism holds, the architecture points toward foundation-model-style scaling: keep one image encoder and one gene encoder and treat the gene library as a prompt, replacing task-specific regression heads in other spatial-omics prediction problems."],"forward_implications":["One trained GeneQuery model can produce expression predictions for any gene in a library, so adding a new gene to the panel no longer requires training a separate per-gene head.","Unseen-gene performance improves as the number of seen genes grows, suggesting the model learns transferable gene-morphology associations from the training gene set.","Richer metadata, such as GPT-4-generated descriptions, improves predictions, implying that gene text quality is a controllable input to accuracy.","The learned spot representations separate tissue structures such as invasive cancer, connective tissue, and adipose tissue, so the same model may double as a tissue-segmentation tool.","Within-tissue transfer (HER2+ to HBD and back) exceeds cross-tissue transfer, indicating the query mechanism generalizes most reliably when tissue context is similar."],"supporting_citations":[{"why":"Defines the per-gene regression baseline STNet that GeneQuery must beat and supplies the ALL-gene evaluation setting.","marker":"[7]"},{"why":"HistoGene baseline uses a vision transformer and MLP; its spot-sequence idea is adapted into spot-aware GeneQuery, and its gene-selection protocol is used for HER2+ and HBD.","marker":"[8]"},{"why":"BLEEP baseline is the contrastive-learning retrieval method; its HEG/HVG evaluation setting and 1,000 highly variable gene selection are adopted for GSE.","marker":"[9]"},{"why":"ResNet50 is the image encoder used inside GeneQuery and in the latent-space visualization.","marker":"[10]"},{"why":"Provides the biological rationale that gene expressions are interdependent, motivating the gene random variable.","marker":"[13]"},{"why":"Cited as the clinical BERT gene encoder that converts gene descriptions into query vectors.","marker":"[14]"}],"fun_headline_variants":["GeneQuery: QA-based gene expression prediction from H&E slides","Ask a slide about a gene: QA framework predicts expression","GeneQuery generalizes to unseen genes via question answering","Reformulating gene prediction as QA outperforms regression methods","Spatial gene expression: one model answers for any gene"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument stands on the premise that a short text description of a gene carries enough information about that gene's spatial expression pattern that, when fused with image features, it can determine where and how strongly the gene is expressed; if the text is just a name or a description with no morphology-expression link, unseen-gene generalization would have no mechanism and the model would only memorize gene identities seen in training.","fun_headline_variants_meta":{"raw":{"variants":["GeneQuery: QA-based gene expression prediction from H&E slides","Ask a slide about a gene: QA framework predicts expression","GeneQuery generalizes to unseen genes via question answering","Reformulating gene prediction as QA outperforms regression methods","Spatial gene expression: one model answers for any gene"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000199,"raw_usage":{"total_tokens":1408,"prompt_tokens":1019,"completion_tokens":389,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":635,"completion_tokens_details":{"reasoning_tokens":307}},"tokens_in":635,"tokens_out":389,"duration_ms":4210,"temperature":1.0,"reasoning_tokens":307,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:15:07.495203+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out entire patients rather than random slides when splitting HER2+ (8 patients, 36 slides) and check whether unseen-gene Pearson correlation survives; or shuffle the gene metadata among genes at test time. If predicted expression maps barely change under shuffled text, the model is not using the text query and the claimed generalization mechanism is absent.","supporting_citations":[{"cited_title":"Integrating spatial gene expression and breast tumour morphology via deep learning","cited_arxiv_id":null,"evidence_quote":"Defines the per-gene regression baseline STNet that GeneQuery must beat and supplies the ALL-gene evaluation setting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"BLEEP baseline is the contrastive-learning retrieval method; its HEG/HVG evaluation setting and 1,000 highly variable gene selection are adopted for GSE."},{"cited_title":"Deep residual learning for image recognition","cited_arxiv_id":null,"evidence_quote":"ResNet50 is the image encoder used inside GeneQuery and in the latent-space visualization."},{"cited_title":"Detecting gene–gene interactions that underlie human diseases","cited_arxiv_id":null,"evidence_quote":"Provides the biological rationale that gene expressions are interdependent, motivating the gene random variable."},{"cited_title":"Optimized glycemic control of type 2 diabetes with reinforcement learning: a proof-of-concept trial","cited_arxiv_id":null,"evidence_quote":"Cited as the clinical BERT gene encoder that converts gene descriptions into query vectors."}],"review_version":1}