{"id":"55ce128c-121c-49d8-afe7-fec460d08c76","arxiv_id":"2412.02601","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A hierarchical graph built from both spatial and image-feature clustering improves joint gene expression prediction from whole slide histology images over 1-hop GNNs and prior transformer/CNN baselines.","lead":"MERGE uses a graph neural network with two kinds of neighborhoods, spatial proximity and visual similarity, to predict gene expression across tissue slides from histology images. The paper also argues that gene-aware smoothing of the expression data produces biologically cleaner training targets than the spatial averaging used by earlier work.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation target is SPCS-smoothed expression, and MERGE's graph mirrors SPCS's two smoothing factors, so the reported gains may reflect reproducing the smoothing field rather than true gene expression.","rationale":"The reader's weakest assumption identifies the most load-bearing point, and I agree. The paper's central quantitative claim is a PCC/MSE improvement on a ground truth that is itself produced by a global two-factor smoothing operator. The graph construction is not an independent architectural choice: it uses spatial proximity and image-feature similarity, and the image-feature similarity is learned against the SPCS-smoothed labels. This creates a structural match between model and target that is specific to the smoothing, as Tab. 4's differential ranking change suggests. The baseline-reproduction issue is also real, but it is secondary: even a perfect reproduction of prior methods would not resolve whether the smoothed target distorts biology, and the internal 1-hop ablation is subject to the same target-matching concern. The correct remedy is not rejection — the method is coherent and the ablations are internally consistent — but a conditional acceptance that requires an evaluation on un-smoothed data (or at minimum a thorough sensitivity analysis across smoothing methods) plus a demonstration that the reported gains persist. I therefore keep the reader's CONDITIONAL verdict.","tokens_in":14990,"tokens_out":7089,"duration_ms":72643,"concrete_test":"Run the same eight-fold cross-validation protocol on the raw, un-smoothed log-normalized UMI counts as ground truth for MERGE, the 1-hop GNN, ResNet+FCN, and TRIPLEX (with its native preprocessing), reporting PCC, MSE, and MAE. If MERGE's ST-Net PCC advantage over the 1-hop baseline (0.6795 vs 0.5789) and over TRIPLEX (0.6795 vs 0.2320) shrinks substantially or reverses, the hierarchical graph's gains are an artifact of matching the SPCS-smoothed target rather than predicting real expression. As a secondary check, build feature-space clusters using an ImageNet-pretrained encoder not fine-tuned on expression; if the gain over 1-hop vanishes, the graph's benefit depends on target-tuned embeddings.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that the SPCS-smoothed expression used as ground truth (Sec. 5, Pre-processing: \"We smooth the gene expression data using SPCS...\") is a valid proxy for true gene expression. SPCS (Sec. 4) replaces each spot's expression with a weighted mixture of all other spots' expressions, with weights based on spatial distance and gene-expression pattern similarity (PCA + Pearson correlation). MERGE's graph (Sec. 3.1) draws edges between spots that are spatially close and between spots whose image features are similar, and the image features come from an encoder fine-tuned on the same SPCS-smoothed targets (Sec. 3.3). The graph is therefore structurally matched to the smoothing operator: its two clustering facets mirror SPCS's two weighting factors. Tab. 4 shows the target choice is not neutral: switching from 8n to SPCS drops TRIPLEX's PCC from 0.3014 to 0.2320 while raising MERGE's from 0.6645 to 0.679, i.e., the method whose graph best matches SPCS benefits most. If the smoothing distorts the underlying biology (e.g., by propagating global expression patterns), the internal gain of the hierarchical graph over 1-hop edges (0.5789 to 0.6795) and the claimed superiority over prior methods may reflect reproducing the smoothing field rather than predicting true gene expression. The biological-interpretation claims (Sec. 5.2) are all made against this smoothed target.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MERGE, a graph neural network for predicting per-spot gene expression from whole-slide histopathology images. The method builds a sparse multi-faceted hierarchical graph from two clusterings (spatial coordinates and image-feature embeddings), connects each cluster to a centroid spot, adds complete shortcut edges among centroids, and retains the 8-neighbor spatial edges; a four-layer GAT then predicts 250 gene expression values per spot. The authors also advocate replacing the commonly used 8-neighbor spatial smoothing with SPCS, a two-factor smoother combining spatial distance and gene-expression pattern similarity. Experiments on ST-Net, Her2ST, and SCC with eight-fold cross-validation report PCC, MSE, and MAE against six baselines, with internal ablations supporting the contribution of each graph component.","tokens_in":15240,"tokens_out":7512,"duration_ms":77158,"significance":"The multi-faceted graph construction is a sensible and potentially useful inductive bias for spatial transcriptomics: it explicitly creates short- and long-range edges between spots that are spatially close or morphologically similar, and the internal ablation in Table 2 is coherent (PCC rises from 0.5789 with 1-hop edges to 0.6795 with the full hierarchical graph). If the evaluation concerns below are resolved, the method would be a competitive contribution and the SPCS analysis could be a useful practical note. The paper does not provide code, and the current evaluation makes the claimed superiority over prior work not yet convincing because of the smoothed evaluation target, the modified TRIPLEX baseline, and the absence of uncertainty quantification.","major_comments":[{"comment":"The evaluation target is the SPCS-smoothed gene expression matrix, and SPCS constructs each spot's smoothed value as a weighted combination over essentially all other spots using two factors: spatial distance and gene-expression pattern similarity (PCA plus Pearson correlation). MERGE's graph and features are built from the same two factors: spatial clustering, feature-space clustering, and a patch encoder fine-tuned on the same SPCS-smoothed targets. The reported gains over 1-hop graphs (Table 2) and over baselines may therefore capture the model's ability to reproduce the smoothing field rather than to predict true gene expression. This is a correctness-risk concern, not a statement that the method is invalid. Concrete tests: report results with raw (or 8n-smoothed) expression as the training and evaluation target, and report results with a patch encoder that has not been fine-tuned on gene-expression targets. The 8n row of Table 4 partially mitigates the concern (MERGE PCC 0.6645 vs TRIPLEX 0.3014), but the internal graph-construction gain is not shown on non-SPCS targets, and the biological-interpretation claims in Sec. 5.2 are all made against the smoothed field.","section":"Sec. 5 (Pre-processing), Sec. 4, Sec. 3.3"},{"comment":"The TRIPLEX baseline appears to be a modified variant. TRIPLEX is described in the related work and in Sec. 3.3 as a multi-resolution method that extracts features from the target patch, its spatial neighborhood, and the whole slide, and the paper says it uses 'the same ResNet18 architecture as our own patch encoder' in TRIPLEX for fair comparison. Replacing TRIPLEX's multi-resolution encoders with a single-scale ResNet18 removes the mechanism that defines the method. The reported TRIPLEX numbers (e.g., PCC 0.2320 on ST-Net vs 0.579 for 1-hop MERGE in Table 1) should therefore be labeled as a re-implementation with a different backbone, or the original multi-resolution TRIPLEX should be run on the same smoothed data. Without this, the SOTA comparison is not established.","section":"Sec. 5 (Baselines), Table 1, Sec. 3.3"},{"comment":"No uncertainty quantification is reported. The text mentions eight-fold cross-validation and five reruns, but the supplementary material states that a constant seed is set and 'the best model is picked for each experiment'. Choosing the best of five runs removes any measure of run-to-run variability and can inflate reported values through selection. The key comparisons in Table 2 (PCC 0.6719 for 'w/o feature' vs 0.6795 for full MERGE) and across datasets in Table 1 are small in some cases, and without standard deviations or paired significance tests the superiority claims are not statistically grounded. Please report mean and standard deviation over folds and reruns under a fixed protocol, or clearly motivate the best-of-five selection.","section":"Supplementary Sec. 7.1, Tables 1-4"},{"comment":"The abstract's claim that the method 'outperforms state-of-the-art techniques across multiple metrics' is contradicted by the SCC results in Table 1: TRIPLEX has lower MSE (0.4891 vs 0.5353) and lower MAE (0.5356 vs 0.5838), while MERGE has only a slightly higher PCC (0.5512 vs 0.5416). The claim should be qualified to specific datasets and metrics, or the SCC discussion should directly address why MSE and MAE are worse for MERGE.","section":"Abstract, Table 1 (SCC row)"},{"comment":"The paper presents SPCS smoothing as a separate contribution and advocates adopting it, but Table 4 shows that SPCS is not uniformly beneficial: for TRIPLEX, switching from 8n to SPCS decreases PCC from 0.3014 to 0.2320 and increases MSE from 0.0760 to 0.1472. The recommendation to use SPCS therefore depends on the model, and the paper should discuss the conditions under which gene-aware smoothing helps or hurts rather than presenting it as a general improvement.","section":"Table 4, Sec. 4"}],"minor_comments":[{"comment":"In the sentence introducing shortcut edges, 'facility long-range communication' should be 'facilitate long-range communication'.","section":"Sec. 3.1"},{"comment":"The supplementary text refers to the 'FNAS' gene in several places; the gene name is FASN, and this typo should be corrected throughout.","section":"Supplementary Sec. 8.1, Figs. 14-19"},{"comment":"The comparison in Fig. 5 is stated to be 'directly taken from the original SPCS paper', but the caption does not make this provenance clear to the reader; if this is a reproduction rather than an original experiment, the caption should say so explicitly.","section":"Fig. 5, Sec. 4"},{"comment":"The cluster-size ablation is performed only on ST-Net, and the text does not state whether the chosen cluster size (100) is then used for Her2ST and SCC without further tuning; this should be clarified so the reader can judge whether the reported Her2ST and SCC results involve any dataset-specific selection.","section":"Sec. 5.3, Table 3"},{"comment":"The text does not describe the hyperparameter search or training budget for the baselines; reporting these details would help the reader assess whether the comparisons are balanced.","section":"Sec. 5 (Baselines)"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about SPCS smoothing is legitimate and should be decisive: the evaluation target is a globally smoothed field and MERGE's graph construction mirrors the two smoothing factors, so the core superiority claim needs a non-smoothed or independently validated target to be convincing. I note, however, that the 8n row of Table 4 shows MERGE also outperforms TRIPLEX on 8n-smoothed data, so this is not an obvious reject; it is a fixable evaluation issue. The strongest additional problem is the modified TRIPLEX baseline, since replacing TRIPLEX's multi-resolution encoders with a single ResNet18 may invalidate the comparison. No code is provided, which makes the statistical issues harder to investigate. I would be comfortable seeing a revised version that reports raw/8n targets, original TRIPLEX, and error bars."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: the graph construction is a real improvement over 1-hop GNNs, and the ablation story is coherent. But the evaluation target is SPCS-smoothed expression, and MERGE's graph is structurally matched to that smoothing operator, so the headline gains may partly be an artifact of reproducing the smoothed field. The baseline reproductions also need scrutiny.\n\nWhat's genuinely new: the multi-faceted hierarchical graph. Merging spatial clustering and image-feature clustering, using centroid representatives and a complete shortcut graph among centroids, is a simple and effective way to get long-range message passing while keeping the graph sparse. The ablations show each piece adds a few points of PCC (1-hop 0.579, +spatial 0.636, +feature 0.672, full 0.680), and the monotonic pattern is believable. The comparison of SPCS against 8-neighbour smoothing is also a useful practical point, even if it's a recommendation rather than a new method.\n\nThe soft spots are mostly about the evaluation. First, the target itself: SPCS smooths each spot using spatial distance and gene-expression pattern similarity, and MERGE's graph draws edges based on exactly those two factors. The feature clustering uses an encoder fine-tuned on the same smoothed targets, so the clusters are partly target-informed. Table 4 shows the smoothing choice changes model rankings: TRIPLEX drops from 0.301 to 0.232 PCC when you switch from 8n to SPCS, while MERGE rises from 0.665 to 0.679. That doesn't kill the paper, but it means the reported superiority over TRIPLEX is entangled with the smoothing choice. Reporting raw and smoothed targets, or at least showing the 8n results for all baselines, would settle it.\n\nSecond, the reproduced baselines. TRIPLEX at 0.23 PCC on ST-Net is far below what the original paper reports, and BLEEP at 0.08 looks broken. The authors explain some of this in the supplementary (Harmony on SPCS data), but it still means the SOTA comparison isn't trustworthy as-is. Third, no error bars or significance tests; the supplementary says they run five times and pick the best model, which is the wrong way to report variance.\n\nFourth, the claim of \"outperforms across multiple metrics\" is overstated: on SCC, TRIPLEX beats MERGE on MSE and MAE, though MERGE wins on PCC.\n\nWho gets value: anyone working on spatial transcriptomics prediction or GNNs for WSIs. The graph construction is worth borrowing, but the evaluation needs a redo before I'd trust the magnitude of the improvement. No code is released, which is a problem for a method paper.\n\nRecommendation: yes, send it to peer review, but with the expectation of major revisions: add variance, validate baseline reproductions, report both smoothed targets, release the code. The core idea is good enough to deserve referee time.","headline":"The hierarchical graph is a genuine step forward, but the SPCS-smoothed evaluation target likely inflates the reported gains, and the baseline reproductions need verification.","tokens_in":15905,"tokens_out":5428,"would_cite":true,"duration_ms":52258,"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":"MERGE predicts per-spot gene expression by clustering tissue spots by both location and visual morphology; on ST-Net its Pearson correlation reaches 0.6795, up from 0.5789 with one-hop edges and 0.2320 for TRIPLEX.","keywords":["spatial transcriptomics","gene expression prediction","whole slide images","graph neural networks","hierarchical graph construction","histopathology","SPCS smoothing","gene-aware data smoothing"],"falsifier":"Re-run the full comparison (MERGE, the one-hop GNN, and TRIPLEX) with raw unsmoothed transcript counts as ground truth instead of SPCS-smoothed values; if the PCC ordering and the size of the gaps stay the same, the smoothing is not the driver, while a sharp shrinkage would tie the claimed superiority to the smoothing step rather than to true expression.","tokens_in":14696,"feed_emoji":"🧬","tokens_out":12146,"duration_ms":110225,"temperature":0.7,"pith_summary":"The paper sets out to show that gene-expression profiles at every spot of a tissue slide can be predicted from the histology image alone, provided the predictor is allowed to pool information across spots that are either close in tissue space or similar in appearance. It presents MERGE, which clusters image patches twice—once by physical coordinates and once by image-feature embeddings—then wires each spot to its two cluster centroids and the centroids to one another, so a graph neural network can pass messages between distant but morphologically related regions within a few hops. The reported results on three cancer spatial-transcriptomics datasets (breast and skin) place MERGE ahead of all compared baselines on mean squared error, mean absolute error, and Pearson correlation, with ST-Net Pearson correlation at 0.6795 versus 0.5789 for the same GNN with one-hop edges and 0.2320 for the reproduced TRIPLEX baseline. The paper further argues that the preprocessing choice matters: a gene-aware, space-aware smoothing (SPCS) produces ground-truth expression maps that better match tissue morphology than the eight-neighbor average used by prior work, and the reported superiority is measured against those SPCS-smoothed targets.","feed_headline":"Spot graph lifts gene-expression prediction to 0.68 correlation","feed_subtitle":"Linking distant but visually similar tissue spots beats per-patch and one-hop baselines across three cancer datasets.","key_machinery":"The load-bearing object is the multi-faceted hierarchical graph, built from two parallel clusterings of the $n$ spots in a slide: one in physical space using spot coordinates, and one in feature space using 256-dimensional patch embeddings from a fine-tuned ResNet18 encoder. For each of the $c$ clusters in either clustering, the spot nearest the feature-space centroid serves as the cluster centroid; every spot is joined to both of its centroids (internal edges), the $2c$ centroids form a complete graph among themselves (shortcut edges), and the eight-neighbor spatial edges are kept as well. A four-layer Graph Attention Network with eight attention heads performs message passing on this graph and produces a 250-gene expression vector at each node; the guaranteed three-hop connectivity between any two nodes is what allows short-range spatial context and long-range morphological similarity to be fused in a shallow network.","core_discovery":"MERGE's central claim is that joint prediction of all spots' gene expressions from histology images is improved by a graph whose edges encode two distinct notions of biological relevance. Spots are clustered once by tissue coordinates and once by image-feature (morphology) embeddings; within each cluster every spot is connected to a chosen centroid spot, all centroids are connected among themselves by shortcut edges, and the usual eight-neighbor spatial edges are retained. The resulting graph stays sparse—$O(n)$ internal edges plus $O(c^2)$ shortcut edges—yet guarantees that any two spots can exchange messages in at most three hops, letting the GNN propagate morphology-informed information between distant locations. On this graph a four-layer Graph Attention Network outputs a 250-gene expression vector per node, and the authors report that it outperforms the per-spot ResNet+FCN, the one-hop GNN of Hist2ST, and the multi-resolution TRIPLEX model on MSE, MAE, and Pearson correlation across ST-Net, Her2ST, and SCC. The paper also credits part of this success to adopting SPCS smoothing, a two-factor, transcriptome-aware preprocessing, as the source of the ground-truth expression targets.","pith_inferences":["A testable extension the paper does not run: retraining MERGE and the baselines on raw, unsmoothed transcript counts, or on a single-cell-resolved spatial assay, would show whether the graph's advantage reflects genuine biological signal or the recovery of the SPCS smoothing field.","The same two-clustering, centroid, and shortcut-edge recipe transfers to other whole-slide prediction problems—survival, grade, or mutation-status classification—where spatially distant but morphologically similar regions are plausibly informative; that transfer is an open experiment.","Because the feature-space clustering consumes any patch embeddings, MERGE's graph construction can sit on top of newer foundation-model encoders, and whether the reported gains persist with stronger features is untested.","A head-to-head comparison the paper does not perform is a graph transformer given the identical hierarchical edges; if its accuracy matches the GAT, the hierarchy rather than attention would carry the result."],"forward_implications":["On all three benchmarks (ST-Net, Her2ST, SCC), MERGE with the hierarchical graph reports the highest Pearson correlation among the methods compared, lifting ST-Net PCC from 0.5789 with one-hop edges to 0.6795.","Because any two spots can exchange information within three hops, a four-layer GAT is enough; very deep networks are not needed to propagate context across a slide.","At inference the pipeline is inductive: given any whole-slide image, patches are extracted, the same two clusterings and edge sets are constructed, and all spots receive joint gene-expression predictions without retraining.","The smoothing step is not neutral: SPCS targets improve MERGE's PCC from 0.6645 to 0.679 relative to eight-neighbor averaging, while the same switch lowers TRIPLEX's PCC from 0.3014 to 0.2320."],"supporting_citations":[{"why":"State-of-the-art multi-resolution baseline (TRIPLEX) that MERGE must beat; also supplies the multi-resolution encoder variant used in the experiments.","marker":"[7]"},{"why":"The prior GNN approach (Hist2ST) whose one-hop graph is the direct point of comparison showing the value of long-range hierarchical edges.","marker":"[41]"},{"why":"Defines the ST-Net dataset, the 250-gene targets, and the per-patch prediction task that MERGE extends to joint prediction.","marker":"[13]"},{"why":"SPCS, the two-factor smoothing method whose output serves as the ground truth for training and evaluation.","marker":"[26]"},{"why":"Graph Attention Network architecture used as the message-passing model on top of the proposed graph.","marker":"[35]"},{"why":"ResNet18 backbone whose fine-tuned penultimate-layer features are clustered in feature space and used as node inputs.","marker":"[14]"}],"fun_headline_variants":["Graph links distant spots to boost gene prediction","Hierarchical graph improves spatial gene expression forecasts","Multi-faceted GNN edges beat one-hop for gene mapping","Distant tissue spots connected for better gene readouts","Gene-aware smoothing plus graph lifts expression accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that SPCS-smoothed expression—each spot's value replaced by a weighted blend of nearby and expression-similar spots—is the correct ground truth for training and evaluation, so the reported gains may partly measure how well the graph reproduces that smoothing field rather than how well it captures the underlying biology.","fun_headline_variants_meta":{"raw":{"variants":["Graph links distant spots to boost gene prediction","Hierarchical graph improves spatial gene expression forecasts","Multi-faceted GNN edges beat one-hop for gene mapping","Distant tissue spots connected for better gene readouts","Gene-aware smoothing plus graph lifts expression accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000626,"raw_usage":{"total_tokens":2923,"prompt_tokens":997,"completion_tokens":1926,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":1854}},"tokens_in":613,"tokens_out":1926,"duration_ms":12706,"temperature":1.0,"reasoning_tokens":1854,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:18:24.980397+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the full comparison (MERGE, the one-hop GNN, and TRIPLEX) with raw unsmoothed transcript counts as ground truth instead of SPCS-smoothed values; if the PCC ordering and the size of the gaps stay the same, the smoothing is not the driver, while a sharp shrinkage would tie the claimed superiority to the smoothing step rather than to true expression.","supporting_citations":[{"cited_title":"Accurate spatial gene expression prediction by integrating multi-resolution features","cited_arxiv_id":null,"evidence_quote":"State-of-the-art multi-resolution baseline (TRIPLEX) that MERGE must beat; also supplies the multi-resolution encoder variant used in the experiments."},{"cited_title":"Spatial tran- scriptomics prediction from histology jointly through Trans- former and graph neural networks","cited_arxiv_id":null,"evidence_quote":"The prior GNN approach (Hist2ST) whose one-hop graph is the direct point of comparison showing the value of long-range hierarchical edges."},{"cited_title":"In- tegrating spatial gene expression and breast tumour morphol- ogy via deep learning","cited_arxiv_id":null,"evidence_quote":"Defines the ST-Net dataset, the 250-gene targets, and the per-patch prediction task that MERGE extends to joint prediction."},{"cited_title":"SPCS: a spatial and pattern combined smoothing method for spatial transcriptomic expression","cited_arxiv_id":null,"evidence_quote":"SPCS, the two-factor smoothing method whose output serves as the ground truth for training and evaluation."},{"cited_title":"Graph Attention Networks","cited_arxiv_id":null,"evidence_quote":"Graph Attention Network architecture used as the message-passing model on top of the proposed graph."}],"review_version":1}