{"id":"1a4ee447-cc99-4f93-b33d-b927b37ddbe2","arxiv_id":"2608.08366","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"VOICE is a single-cell foundation model that fuses morphology regression and reference-cell retrieval to predict gene expression from H&E images, trained on 23 million Xenium cells.","lead":"VOICE predicts a single cell's gene expression from its H&E image by combining direct image regression with retrieval of similar reference cells that have measured expression. It is trained on 23 million cells from 75 Xenium slides, and in the paper's benchmarks it reports higher Pearson correlations than prior single-cell prediction methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Patient-level holdout is not demonstrated; cross-slide gains may be inflated by same-patient slides in the retrieval bank, undermining the held-out-patient claim.","rationale":"VOICE is a substantial contribution: 23M cells, 75 slides, 15 tissues, a principled two-branch design, and thorough ablations. The in-slide protocol with five-fold bands is well specified, and the ablation study (Table 6) demonstrates that both training stages and the fusion contribute. The paper also discloses limitations honestly, including the dependence on a reference bank and the lack of count calibration. However, the central claim of generalization to held-out patients is not supported by the described evaluation. Section 4.1 defines the cross-slide protocol as withholding slides, not patients. If a test slide is from a patient whose other slides are in the training set, the retrieval bank and the trained encoder have already seen that patient's tissue. This is a classic leakage path that can inflate correlations. The paper's abstract explicitly says 'held-out patients,' so the authors intended this claim; the absence of any patient-level split in the methods is a gap that must be closed. The reader's weakest assumption concerned the requirement of a same-tissue reference bank. That is a transparent limitation of the retrieval branch, and the paper acknowledges it in Section 3.3. In contrast, the patient-level leakage risk is not acknowledged and, if present, would invalidate the headline result. I therefore disagree with the reader's choice of weakest assumption. The concrete test is straightforward: report patient identifiers and re-run under a patient split. This is a standard check for medical imaging foundation models. If the authors can show no overlap, the conditional acceptance stands. If not, the central claim needs substantial revision.","tokens_in":29898,"tokens_out":9910,"duration_ms":91970,"concrete_test":"Obtain patient metadata for the 75 training slides and 5 cross-slide test slides (HEST-1k v1.3.0 and 10x Genomics portal). Check for patient overlap. If any overlap exists, re-run the cross-slide evaluation (Tables 2, 11-13) under a strict patient-level split, placing all slides from each patient in either training or test. Confirm whether VOICE-23M still achieves the highest scores on all seven metrics (modulo the acknowledged kidney H20 tie). If the wins disappear or shrink, the held-out-patients claim must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract claims VOICE generalizes to held-out patients, but Section 4.1 only defines slide-level separation: five vertical bands for in-slide evaluation and 'unseen slides' for cross-slide evaluation. No patient-level split is described. The retrieval branch (Eq. 3) constructs a reference bank from 'training slides from the same tissue' (Section 4.1). If any cross-slide test slide shares a patient with a training slide, the bank contains same-patient cells with measured expression, and the direct branch has been trained on the same patient's morphology. This could inflate all cross-slide metrics, including the claimed consistent outperformance. The paper does not report patient identifiers or state that training and test patients are disjoint. This is a concrete, checkable omission that directly affects the central generalization claim.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces VOICE, a foundation model that predicts single-cell gene expression from H&E images using paired Xenium spatial transcriptomics data. VOICE first contrastively aligns cell-mask-pooled H&E morphology features from UNI2-h with scFoundation expression embeddings over 23 million cells, then continues fine-tuning with an SE(2)-equivariant spatial decoder and a shared count head over a 6,029-gene vocabulary. At inference, a direct branch regresses expression from morphology, a retrieval branch averages measured expression from the K nearest reference cells under cosine similarity, and a per-gene gate fuses the two predictions with weights fit per tissue. The paper claims that VOICE generalizes to held-out patients, slides, and partially overlapping gene panels, consistently outperforming GHIST, sCellST, DeepSpot2Cell, and DeepSpot-M on seven PCC-based metrics in in-slide and cross-slide benchmarks, and that its direct branch can transfer genes measured only in other tissues. A zero-shot Xenium-to-CosMx transfer is also reported.","tokens_in":30071,"tokens_out":9841,"duration_ms":79613,"significance":"VOICE is a technically strong contribution: it is trained on a large corpus (23M cells, 75 slides, 15 tissues), includes careful ablations of the contrastive objective (Tables 8–9), cell-feature pooling (Table 10), training stages (Table 7), and fusion (Table 6), and it reports an honest cross-platform evaluation with a novel registration procedure. The per-gene fusion of direct and retrieval prediction is a sensible and well-motivated mechanism. If the generalization claims are confirmed, VOICE could enable cost-effective molecular analysis of large H&E archives. However, the lack of a patient-level evaluation and the absence of uncertainty estimates are significant gaps that must be addressed before the central claims can be accepted.","major_comments":[{"comment":"The abstract claims that VOICE generalizes to held-out patients, but the evaluation protocol in Section 4.1 only defines slide-level separation: five vertical bands for the in-slide evaluation and 'unseen slides' for the cross-slide evaluation. No patient identifiers are reported, and the paper does not state whether training and test slides come from disjoint patients. Since the retrieval branch constructs its reference bank from training slides of the same tissue (Section 3.3, Eq. 3; Section 4.1), any overlap in patients between the training set and the cross-slide test slides could inflate the reported cross-slide gains for both the retrieval and direct branches. This is a concrete, checkable omission that directly bears on the central generalization claim. Please provide patient-level metadata and verify that the cross-slide test slides are patient-disjoint from all training slides, or re-run the evaluation under an explicit patient-level split.","section":"Section 4.1, Table 4"},{"comment":"All reported results are point estimates of per-gene Pearson correlation averaged over gene sets, with no confidence intervals, standard errors, or significance tests. For the small gene sets (H20, S20), the mean PCC may have substantial uncertainty, and many of the observed differences between VOICE-23M and the strongest baseline are small (e.g., Table 2, kidney H20 in the Xenium-shared panel: 0.525 vs 0.527). Without measures of uncertainty, the claim of 'consistently outperforms' on all seven metrics is not fully supported. Please report per-gene PCC distributions, bootstrap confidence intervals, or paired tests across genes for the key comparisons.","section":"Tables 1 and 2"},{"comment":"The per-gene fusion weights beta_g are selected by grid search on training slides to maximize the correlation between the fused prediction and ground-truth expression, which is the same metric used for the main evaluation. With a small number of training slides per tissue (e.g., pancreas uses four slides in Appendix A.3), this selection procedure may overfit and lead to optimistic fusion gains. The paper should report the variability of beta_g across training slides or select weights using a metric-neutral criterion (e.g., log-likelihood) to demonstrate that the fusion benefit in Table 6 is not an artifact of fitting the evaluation metric.","section":"Section 3.5, Eq. (7)"},{"comment":"The abstract states that VOICE 'consistently outperforms prior single-cell expression prediction methods on seven metrics,' but Table 2 contains a tie (kidney H20 in the full target panel, 0.554 for both VOICE-23M and fine-tuned DeepSpot-M) and a loss (kidney H20 in the Xenium-shared panel, 0.525 vs 0.527). While Section 4.3 acknowledges these exceptions, the abstract overstates the result. Please soften the claim or incorporate the exceptions into the abstract.","section":"Abstract and Section 4.3"}],"minor_comments":[{"comment":"The SE(2)-equivariant transformer is attributed to reference [31], Uni-Mol, which is a 3D molecular representation model, not an SE(2)-equivariant spatial transformer for cell neighborhoods; please provide the correct citation or describe the architecture in the text.","section":"Section 3.4"},{"comment":"Section 3.3 states that K=200 neighbors and tau=0.03 are used in all experiments, but the contrastive-objective experiments in Appendix B.3 use K=20; please clarify which retrieval settings apply to the objective comparison.","section":"Section 3.3"},{"comment":"Table 2, full target panel, kidney column for DeepSpot-M (fine-tuned) contains run-together values ('0.4790.5540.473'); please fix the table formatting.","section":"Table 2"},{"comment":"Appendix D.2 contains a typo: 'fix the registration. between the H&E and DAPI images' should read 'fix the registration between the H&E and DAPI images.'","section":"Appendix D.2"},{"comment":"The selection criteria for the 24-slide subset used to train VOICE-7M are not described; please state whether the subset is stratified by tissue and how it was chosen.","section":"Appendix A.1"},{"comment":"Section 1: 'which cannot learn morphology–molecular relationships that transfer across datasets' would be clearer as 'which cannot learn transferable morphology–molecular relationships across datasets.'","section":"Section 1"}],"recommendation":"major_revision","confidential_remarks":"The patient-level holdout issue is the main risk to the paper's central claim; I recommend requesting a clear statement of patient IDs and, if necessary, a rerun with patient-disjoint splits. The paper does not provide a code repository or data access link; please request these at revision. The abstract overstatement about 'consistently outperforms' should be corrected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"VOICE is a well-built engineering contribution: CLIP-style alignment of cell-mask-pooled H&E features to scFoundation embeddings, a direct regression branch, a retrieval branch, and a per-gene fusion. The ablations are transparent, and the cell-mask pooling ablation is convincing (0.086 All-gene PCC drop). The cross-tissue gene transfer analysis (Table 3) is a genuinely useful evaluation, and the CosMx zero-shot attempt with its registration pipeline is a good-faith extension even though performance collapses to 0.07 macro All PCC. The limitations section is also honest: Xenium-only, correlation metrics, no count calibration. Fine.\n\nWhat is actually new: single-cell resolution that combines both branches with per-gene gating, training one shared count head across heterogeneous panels, and evaluating genes never measured in same-tissue training. That combination warrants attention.\n\nSoft spots, in proportion. The biggest is the gap between the abstract and the protocol. The abstract says 'held-out patients,' but the paper only describes slide-level separation. Cross-slide test slides are never stated to come from patients disjoint from the 75 training slides. The retrieval bank is built from 'training slides from the same tissue,' so if a test slide shares a patient with a training slide, the bank contains same-patient cells with measured expression, and the direct branch has also been trained on that patient's morphology. That could inflate cross-slide metrics, and the reported 0.01–0.05 PCC margins over fine-tuned DeepSpot-M are the size same-patient leakage could produce. This is a concrete, checkable omission. The authors need to state patient IDs or otherwise demonstrate no patient overlap before the headline claim stands.\n\nSecond, the per-gene fusion weights are selected on training slides by maximizing the same PCC metric used for evaluation. That is selection on the validation set; it does not invalidate the results, but a nested protocol or an estimate of the selection effect is needed. Third, no error bars anywhere, and the kidney H20 exception (VOICE-23M 0.525 vs DeepSpot-M 0.527) is a small crack in the 'consistently outperforms' claim; the paper does acknowledge it. No code release yet, which matters for a pipeline with this many moving parts.\n\nThe in-slide comparisons are the hardest to leak and they hold up. The central argument mostly holds. The 'held-out patients' wording is over-strong as written, but the fix is straightforward: report a patient-level split, add error bars, and release code.\n\nWho it is for: computational pathology and spatial transcriptomics researchers who want a practical H&E-to-expression predictor for archives. It deserves a serious referee, with a specific request for the patient-level split. I would bring it to a reading group.","headline":"Solid single-cell expression prediction pipeline with honest ablations; the 'held-out patients' claim needs a patient-level split before the cross-slide numbers can be trusted.","tokens_in":30610,"tokens_out":2265,"would_cite":true,"duration_ms":20776,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"VOICE predicts single-cell gene expression from H&E images alone, transferring to genes never measured in the target tissue.","keywords":["spatial transcriptomics","H&E histology","gene expression prediction","foundation model","contrastive alignment","retrieval-augmented prediction","per-gene fusion","single-cell"],"falsifier":"Hold out an entire tissue that appears in no training slide, build the retrieval bank exclusively from other tissues, and measure the direct branch's per-gene Pearson correlation on that tissue's H&E test slide; if the all-gene PCC falls to near zero rather than remaining in the 0.2 range reported for same-tissue genes, the claimed cross-tissue transfer of morphology-to-expression knowledge fails beyond the trained tissue set.","tokens_in":29719,"feed_emoji":"🧬","tokens_out":6894,"duration_ms":55484,"temperature":0.7,"pith_summary":"VOICE is a foundation model that predicts a single cell's gene expression purely from its appearance in an H&E-stained image, trained on paired Xenium measurements. The paper's central claim is that this prediction generalizes to held-out patients, slides, and partially overlapping gene panels, and that fusing a direct morphology-to-expression regression with a retrieval branch that copies measured expression from similar reference cells beats either strategy alone. If the claim holds, archived H&E slides—costing tens of dollars and collected routinely—could be mined for single-cell molecular information that currently requires spatial transcriptomics assays costing thousands of dollars per slide. The model also reports predicting genes that were never measured in the target tissue, by transferring morphological knowledge from other tissues.","feed_headline":"H&E images alone predict single-cell gene expression","feed_subtitle":"VOICE matches spatial transcriptomics accuracy on unseen tissues at a fraction of the cost.","key_machinery":"The central object is the shared cell-mask-pooled morphology embedding that feeds both prediction branches, together with the per-gene gate that weights their outputs. This embedding comes from a LoRA-adapted pathology encoder aligned to a transcriptome foundation model's space in Stage 1 and refined by an SE(2)-equivariant spatial decoder in Stage 2, so both branches read the same representation while making partly independent errors; the gate picks the more reliable branch gene by gene.","core_discovery":"VOICE aligns cell-mask-pooled H&E morphology features from a pathology foundation model with single-cell transcriptomic embeddings from a transcriptome foundation model, using LoRA adapters and a symmetric InfoNCE loss over 23 million cells from 15 tissues. From the resulting shared space, a direct branch predicts expression through an SE(2)-equivariant spatial decoder with a shared count head over a 6,029-gene vocabulary, while a retrieval branch averages measured expression from the nearest reference cells in morphology space. A per-gene convex gate then fuses the two predictions, favoring direct regression for genes whose expression tracks visible morphology and favoring retrieval for genes that do not. On seven per-gene Pearson correlation metrics, the model reports the highest scores in all 21 in-slide comparisons and on nearly every cross-slide panel across breast, lung, kidney, ovary, and pancreas, and it predicts genes measured only in other tissues at levels comparable to same-tissue genes.","pith_inferences":["A tissue type entirely absent from the 15 training tissues would be the strictest test: if the direct branch's all-gene PCC collapses, the claimed cross-tissue gene transfer is partly a memorization of the trained tissue distribution.","The same-tissue requirement for retrieval banks implies that practical deployment depends on reference-slide availability; building batch-corrected cross-tissue banks and testing whether the per-gene gate transfers across tissues are natural next steps.","One could test the per-gene gate's semantics directly: genes with weights near 1 should be enriched for morphological markers like structural proteins, while genes near 0 should be enriched for non-morphological programs such as immune response, which would confirm the mechanism the gate is meant to implement."],"forward_implications":["Archived H&E slides can be re-analyzed as single-cell expression maps without running new spatial assays.","Genes absent from a slide's measured panel remain predictable, since the shared count vocabulary is trained across heterogeneous panels.","Per-gene gating lets the method be applied to any gene that appears in at least one reference tissue, not just genes with visible morphological correlates.","Scaling the training set from 7 million to 23 million cells improved every in-slide metric, suggesting further gains from larger cell corpora.","Zero-shot transfer to CosMx slides, though weaker than within-platform performance, indicates the morphology–expression link is not specific to Xenium."],"supporting_citations":[{"why":"Supplies the pathology encoder whose cell-mask-pooled features VOICE adapts and aligns.","marker":"[2]"},{"why":"Provides the frozen transcriptomic embeddings that define the alignment target and shared gene vocabulary.","marker":"[11]"},{"why":"Defines the symmetric InfoNCE (CLIP) objective used in Stage 1 contrastive alignment.","marker":"[22]"},{"why":"GHIST is the single-cell supervised baseline VOICE outperforms on in-slide and cross-slide panels.","marker":"[10]"},{"why":"DeepSpot-M is the strongest multimodal baseline VOICE is compared against on the full target panels.","marker":"[19]"},{"why":"sCellST is the spot-level supervised baseline VOICE beats on the Visium-shared panel.","marker":"[1]"},{"why":"BLEEP's soft-target contrastive objective is the ablation alternative that VOICE's harder CLIP target matches.","marker":"[28]"}],"fun_headline_variants":["H&E images alone predict single-cell gene expression","VOICE predicts single-cell gene expression from H&E alone","Routine H&E slides yield single-cell transcriptomes","Morphology-only model beats targeted panels","One image, thousands of gene predictions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Retrieval assumes that cells whose H&E appearance is similar also have similar expression for genes with no direct morphological signal, and that a same-tissue reference bank of measured cells exists at inference time; without such a bank, the retrieval branch and its cross-gene recovery stop working.","fun_headline_variants_meta":{"raw":{"variants":["H&E images alone predict single-cell gene expression","VOICE predicts single-cell gene expression from H&E alone","Routine H&E slides yield single-cell transcriptomes","Morphology-only model beats targeted panels","One image, thousands of gene predictions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1558,"prompt_tokens":952,"completion_tokens":606,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":535}},"tokens_in":568,"tokens_out":606,"duration_ms":5563,"temperature":1.0,"reasoning_tokens":535,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:06:07.636766+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out an entire tissue that appears in no training slide, build the retrieval bank exclusively from other tissues, and measure the direct branch's per-gene Pearson correlation on that tissue's H&E test slide; if the all-gene PCC falls to near zero rather than remaining in the 0.2 range reported for same-tissue genes, the claimed cross-tissue transfer of morphology-to-expression knowledge fails beyond the trained tissue set.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the frozen transcriptomic embeddings that define the alignment target and shared gene vocabulary."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the symmetric InfoNCE (CLIP) objective used in Stage 1 contrastive alignment."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"GHIST is the single-cell supervised baseline VOICE outperforms on in-slide and cross-slide panels."},{"cited_title":"Koelzer, and Gunnar Rätsch","cited_arxiv_id":null,"evidence_quote":"DeepSpot-M is the strongest multimodal baseline VOICE is compared against on the full target panels."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"sCellST is the spot-level supervised baseline VOICE beats on the Visium-shared panel."},{"cited_title":"Chung, Catia T","cited_arxiv_id":null,"evidence_quote":"BLEEP's soft-target contrastive objective is the ablation alternative that VOICE's harder CLIP target matches."}],"review_version":1}