{"id":"72ff2c56-a657-4d10-8ee3-66a7ad803ed6","arxiv_id":"2607.20896","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":8,"one_line_summary":"HierarchicalDAEW predicts spatial gene expression with expression-derived domain edge typing and calibrated uncertainty, beating 13 baselines on breast Visium sections.","lead":"A new graph model predicts gene activity across tissue slides by combining tissue-domain structure, gene networks, and calibrated confidence estimates. It reports the best accuracy among fourteen methods on breast tissue, but the evaluation relies on gene-expression data that a real new slide would not provide.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-time edge typing uses expression (Eq. 4), so the reported H&E-only SOTA is unsupported; the paper's own Sec. 6.1.5 ablation shows morphology typing is worse than no typing.","rationale":"Reader's weakest assumption identified the same core issue, and it is indeed the most load-bearing concern. The paper frames the task as predicting expression from H&E and claims clinical decision-support, but the evaluation protocol supplies test-expression-derived edge types that are not available at inference. The authors' own ablation provides the decisive internal evidence: morphology-derived typing, the only test-time-feasible alternative, actively hurts performance relative to no typing. Therefore the claimed architectural benefit of DAEWConv is contingent on a target-derived input. The multi-seed reproducibility, negative control, and final holdout do not resolve this because they all inherit the same test-time expression usage. This is not a matter of disagreement with community consensus; it is an internal inconsistency between the problem statement and the evaluation. While the method may be useful in transductive settings where expression is available for the target section, the preprint's central claim for H&E-only prediction is not demonstrated. The reader's REJECT verdict is justified; my analysis does not change it.","tokens_in":39208,"tokens_out":5572,"duration_ms":58547,"concrete_test":"Re-run the primary breast multi-section benchmark (Table 3) and single-section benchmark (Table 4) with HierarchicalDAEW modified so that test-spot edge types are computed without expression: either (a) use no edge typing (all edges type 0) at test time, or (b) use morphology-derived typing from UNI embeddings for all spots. Keep all other training and evaluation details identical, and compare PCC against the same baselines under the same condition. If HierarchicalDAEW's PCC falls to or below the no-typing level (≈0.651) or no longer exceeds SEPAL/THItoGene, the SOTA claim is an artifact of test-expression leakage.","verdict_should_be":"REJECT","load_bearing_attack":"In the evaluation, domain labels for held-out test spots are assigned by nearest-centroid matching in expression-PC space (Eq. 4). This makes test expression an input to the edge-typing mechanism, so the model is not predicting purely from H&E. The paper's own ablation (Sec. 6.1.5, Table 12) shows that the benefit of DAEWConv disappears when edge types are derived from morphology: morphology typing gives PCC 0.598, below the no-typing control (0.651), while expression-derived typing gives 0.683. Thus the reported advantage over thirteen baselines (e.g., 0.696 vs 0.638 in multi-section training) depends on test-expression leakage that cannot be replicated on a new H&E slide, where expression is unknown. The negative control (Sec. 6.2.4) does not address this: edge types are kept at their real-expression values while only labels are shuffled, so it only rules out positional shortcuts, not target-derived inputs. This is a load-bearing flaw for the paper's central claim of predicting gene expression from H&E for clinical deployment.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HierarchicalDAEW, a dual-graph architecture for predicting spatial gene expression from H&E histology. The spot-level encoder, DAEWConv, types spatial edges as intra-domain, inter-domain, or boundary based on Leiden domains computed from expression data, and learns separate projections per edge type. A domain-level GCN with cross-scale gating adds hierarchical context, and a gene-graph decoder propagates landmark-gene predictions through STRING-DB and co-expression edges. Uncertainty is modeled with a Normal-Inverse-Gamma evidential loss. The authors report state-of-the-art Pearson correlation against thirteen baselines across six Visium sections, with ablations, negative controls, multi-seed checks, and calibration studies.","tokens_in":39522,"tokens_out":4055,"duration_ms":39614,"significance":"If the central claim were valid, HierarchicalDAEW would be a substantial advance toward H&E-only spatial expression prediction with calibrated uncertainty. The manuscript is unusually thorough in its evaluation: multi-seed reproducibility, bootstrap confidence intervals, per-gene win rates, an untouched holdout, and conformal calibration are all included. However, the core edge-typing mechanism requires expression data at test time through Eq. (4), which is not available for a new slide; the paper's own ablation (Sec. 6.1.5) shows that the morphology-derived proxy is actively harmful. This undermines the stated practical significance and the validity of the reported benchmark gains.","major_comments":[{"comment":"Test-spot domain labels are assigned by nearest-centroid matching in expression-PC space using the test spot's expression vector z_i. These labels determine edge types in Eq. (5), which are consumed by DAEWConv in Eqs. (6)-(7). Thus, at evaluation time, the model's edge-typing input contains information from the very expression values being predicted. This contradicts the problem formulation in Eq. (2) and the claimed H&E-only input. The negative control in Sec. 6.2.4 shuffles labels but keeps the real expression-derived edge types, so it does not test this form of leakage. A valid control would recompute or permute edge types from shuffled expression.","section":"Sec. 3.4, Eq. (4)"},{"comment":"The paper shows that expression-derived typing (PCC 0.683) outperforms no typing (0.651), while morphology-derived typing performs worse than no typing (0.598). Since expression is unavailable for an unseen slide, the deployment-time variant of the model would have to use morphology-derived typing, which would be worse than a standard GCN. The reported multi-section advantage (Table 3: 0.696 vs 0.638) therefore does not carry over to the clinical setting claimed in the abstract and introduction. This is acknowledged in Sec. 6.1.5 as 'an important limitation', but the conclusion and abstract still frame the method as an H&E-only predictor.","section":"Sec. 6.1.5, Table 12"},{"comment":"All baselines are trained with a single learning rate (5e-4) with no per-model tuning, while HierarchicalDAEW's hyperparameters (depth, hidden dimension, neighborhood size, Leiden resolution, loss weights, temperature) are selected via grid search in Sec. 6.1. This asymmetry can inflate the reported margins. At minimum, a learning-rate sensitivity analysis for the strongest baselines (SEPAL, THItoGene, MERGE) is needed to support the claim that HierarchicalDAEW is the strongest method, especially given the non-significant single-section gap to SEPAL.","section":"Sec. 4.3, baselines"},{"comment":"In the single-section benchmark, HierarchicalDAEW's PCC (0.704) is not statistically distinguishable from SEPAL (0.699). The paper's headline claim of outperforming thirteen published baselines is therefore supported only in the multi-section setting (Table 3), where the expression-derived edge-typing leakage is present. The per-gene win rate in Sec. 6.2.1 does not resolve this aggregate non-significance for the closest competitor; a method that is not significantly better than a baseline on the standard single-section protocol cannot be described as unequivocally strongest.","section":"Sec. 5.2, Table 4"}],"minor_comments":[{"comment":"The heading 'Leakage-Free Leiden Domain Assignment' overstates what the procedure achieves. The nearest-centroid matching avoids training-set contamination, but it still injects test-side expression information into the model inputs via the edge typing. Please reword to reflect that the procedure only prevents fold-level contamination, not target-derived inference leakage.","section":"Sec. 3.4"},{"comment":"TRIPLEX is approximated with single-scale UNI embeddings because multi-resolution inputs are unavailable. This should be stated as a limitation in the main experimental section, not only in the implementation details, since it may materially change the baseline's performance.","section":"Sec. 4.3, TRIPLEX"},{"comment":"The metric sets differ between Table 3 and Table 4: the former reports PCC, Spearman, R2; the latter adds CCC and drops some entries. Please use a consistent set of metrics across tables to make comparisons easier.","section":"Tables 3 and 4"},{"comment":"The Spearman/Pearson correlations in the calibration summary (Table 19: 0.176/0.119) differ from those in Figure 30 (0.249/0.249). The text explains that these come from separately trained instances, but the discrepancy should be made explicit in the table or figure caption to avoid confusion.","section":"Sec. 6.4.2"},{"comment":"Code is listed as 'available upon acceptance'. For a benchmark paper with strong empirical claims, a public repository at submission time would strengthen reproducibility and is customary for this venue.","section":"Data and code availability"}],"recommendation":"reject","confidential_remarks":"The paper is well-written and the experimental appendix is unusually extensive, but the central claim is not valid for the stated task. The edge-typing mechanism requires expression data at test time, and the authors' own ablation proves that the only inference-time-compatible alternative (morphology) is harmful. This is not a local fix; it requires redesigning the method's core input. I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe punchline: this is a serious, well-run paper that overclaims its central result. The architecture is novel — DAEWConv with expression-derived edge types (intra/inter/boundary), hierarchical aggregation via DomainGCN and CrossScaleGate, a gene-level decoder that mixes STRING-DB with co-expression, and evidential NIG uncertainty. The empirical work is unusually thorough: thirteen baselines under a shared protocol, multi-section joint training, ablations, multi-seed checks, a negative control, conformal prediction, and a final untouched holdout. The uncertainty calibration results (NIG 0.903 coverage at 90% vs MC dropout 0.323) are interesting and plausibly a real contribution.\n\nThe soft spot is load-bearing. Test-time edge typing uses the held-out spot's expression: Eq. 4 assigns domain labels by nearest-centroid matching in expression-PC space. So the graph structure the model reasons about contains information derived from the very expression values it is asked to predict. For a new H&E slide, that input is not available. The authors know this — Section 6.1.5 shows that a morphology-derived substitute is worse than no typing (0.598 vs 0.651 PCC), while expression-derived typing gives 0.683. That means the reported SOTA gains (0.696 in multi-section, beating 0.638 SEPAL) are not achievable under the paper's stated deployment scenario. The negative control in 6.2.4 does not address this, since edge types are held at their real-expression values while labels are shuffled. So the central claim of predicting from H&E alone is unsupported. The paper could be reframed as a transductive or expression-informed method, which would remove the overclaim but shrink the significance. The single-section comparison also has a caveat: the advantage over SEPAL is not statistically significant (and the bootstrap CI crosses zero, as the paper honestly reports).\n\nMinor issues: baselines share a uniform learning rate rather than per-model tuning (the paper argues a sweep showed this is fair, but it is a known way to handicap stronger baselines); TRIPLEX is approximated with single-scale inputs; no code is released yet.\n\nThis is still a paper worth engaging with. It is honest about its limitations in the discussion, and the architecture and uncertainty pipeline could be useful in settings where expression-derived domains are available at inference. I would send it to peer review — a good referee could push the authors to either fix the evaluation or reframe the contribution. But I would not cite the H&E-only result as reported.","headline":"Expression-derived edge typing at test time undercuts the H&E-only claim; the paper is thorough but the central result is not supported.","tokens_in":39969,"tokens_out":4906,"would_cite":false,"duration_ms":45270,"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":"A dual-graph model with domain-aware edge typing and evidential uncertainty outperforms thirteen baselines in predicting spatial gene expression from H&E histology, and its confidence intervals achieve near-exact 90% coverage.","keywords":["spatial transcriptomics","gene expression prediction","H&E histology","graph convolutional networks","tissue domains","edge typing","evidential uncertainty","Visium"],"falsifier":"Re-run the edge-typing experiment on a fresh slide with no expression data, using only a morphology-based domain predictor, and check whether mean PCC falls below the untyped GCN control; if it does not, the paper's central premise fails. Also, an independent reproduction of the 0.696 mean PCC on three breast sections, or of the 0.903 vs 0.323 coverage gap versus MC dropout, would directly test the headline numbers.","tokens_in":39113,"feed_emoji":"🧬","tokens_out":5187,"duration_ms":46097,"temperature":0.7,"pith_summary":"The paper tries to establish that gene expression at tissue spots can be predicted from H&E histology better when the graph message passing respects tissue domain structure instead of treating all spatial edges alike, and that evidential deep learning gives trustworthy per-prediction confidence intervals. A sympathetic reader would care because if true, this gives pathologists a cheap, assay-free route to spatially resolved molecular profiling, with calibrated flags for when predictions should not be acted on. The central claim is the architecture's performance: it beats thirteen published methods across six human Visium sections and its 90% credible intervals contain the true expression 90.3% of the time, versus 32.3% for Monte Carlo dropout. The paper further claims the benefit is specifically due to expression-derived domain typing, not any typed edges, and reports that a morphology-derived typing is harmful.","feed_headline":"Edge-typed graph model beats 13 baselines in spatial gene prediction","feed_subtitle":"Typing edges by tissue domain plus evidential uncertainty gives PCC 0.696 and near-exact 90% confidence coverage.","key_machinery":"The load-bearing mechanism is DAEWConv, a relational-style convolution in which edges in the spot graph are typed by Leiden-derived tissue domains (intra-domain, inter-domain, boundary) and each type has its own learned projection matrix plus a learnable scalar gate; a DomainGCN over pooled domain representations and a per-spot CrossScaleGate fuse coarse tissue context back into spot features. A second gene-level graph with protein–protein interaction priors and co-expression edges propagates landmark predictions via source-typed attention-gated message passing, and a Normal-Inverse-Gamma evidential loss outputs both an expression estimate and a calibrated variance.","core_discovery":"On its own terms, the discovery is that a graph convolutional operator which types each spatial edge as intra-domain, inter-domain, or boundary—with separate learned projections per type—plus a hierarchical domain-level pooling with gated fusion, produces state-of-the-art correlation between predicted and measured expression (mean PCC 0.696 under multi-section joint training; 0.704 single-section), and that a Normal-Inverse-Gamma evidential head yields confidence intervals with near-exact 90% empirical coverage that remain informative under conformal calibration. The authors also show the gene graph decoder propagates landmark-gene predictions to unsupervized genes (imputation PCC 0.831), an","pith_inferences":["Editorial inference: because test-time domain labels are assigned by nearest-centroid matching in expression-PC space, the method cannot yet be applied directly to an unmeasured slide; the paper's own ablation suggests a morphology-based domain predictor would not recover the gain.","Editorial inference: the near-tie against the strongest baseline on single-section breast (PCC 0.704 vs 0.699, not significant) means the claimed advantage is partly contingent on multi-section joint training; single-section gains may not generalize to all tissues.","Editorial inference: the calibrated uncertainty degrades sharply under tissue-type shift (coverage 0.333–0.499), so the 'trustworthy intervals' claim should be read as in-distribution only; conformal recalibration on the target tissue would be a cheap practical fix.","Editorial inference: the success of expression-derived typing suggests that learning tissue-domain structure directly from H&E in a supervised or self-supervised way could close the deployment gap, offering a concrete next step the paper leaves implicit."],"forward_implications":["If correct, this method becomes the strongest published approach for H&E-to-expression prediction under the reported evaluation protocol, with a meaningful edge when training spans multiple sections.","The near-exact 90% coverage means evidential NIG can replace Monte Carlo dropout for trustworthy per-spot confidence intervals without extra forward passes.","The gene graph decoder extends prediction to genes never directly supervised (imputation PCC 0.831), suggesting scalable expansion of gene panels.","The hierarchy ablation (+0.044 PCC) implies domain-level context is a genuine source of predictive signal worth incorporating in future models.","The benefit concentrates in multi-section training and depends on expression-derived domains; morphology-derived typing actively hurts."],"fun_headline_variants":["Edge typing by tissue domain beats 13 baselines in gene prediction","Evidential uncertainty yields calibrated confidence for spatial gene prediction","Domain-aware edge weighting lifts gene prediction above 13 baselines","Hierarchical graph with evidential uncertainty predicts gene expression"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The method needs expression-derived tissue domains to type edges, but for a genuinely new H&E slide no expression is available at inference time; the authors assign test domains by nearest-centroid matching in expression-PC space (Eq. 4), and their own ablation shows replacing this with morphology-based typing drops PCC below the untyped baseline (0.598 vs 0.651).","fun_headline_variants_meta":{"raw":{"variants":["Edge typing by tissue domain beats 13 baselines in gene prediction","Evidential uncertainty yields calibrated confidence for spatial gene prediction","Domain-aware edge weighting lifts gene prediction above 13 baselines","Hierarchical graph with evidential uncertainty predicts gene expression"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000972,"raw_usage":{"total_tokens":4001,"prompt_tokens":811,"completion_tokens":3190,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":3122}},"tokens_in":555,"tokens_out":3190,"duration_ms":21616,"temperature":1.0,"reasoning_tokens":3122,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T09:04:46.371562+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the edge-typing experiment on a fresh slide with no expression data, using only a morphology-based domain predictor, and check whether mean PCC falls below the untyped GCN control; if it does not, the paper's central premise fails. Also, an independent reproduction of the 0.696 mean PCC on three breast sections, or of the 0.903 vs 0.323 coverage gap versus MC dropout, would directly test the headline numbers.","supporting_citations":[],"review_version":1}