{"id":"ce488f53-cb73-4333-abc2-91185de65e9b","arxiv_id":"2608.00405","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"MSGR's Gene Ontology-guided hierarchical decoder improves spatial gene expression prediction from histology images, with the biological structure adding a +0.027 gain over an equivalent random hierarchy.","lead":"This paper introduces MSGR, a gene-decoding module that organizes genes by their known biological roles and predicts spatial gene expression from tissue images in a coarse-to-fine way. It reports small but consistent average gains over flat-decoding baselines on nine public datasets, and shows the biological structure matters by beating a random group structure.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Plug-in evidence is confounded: Table 2 uses a lightweight MLP decoder and, for STFlow, a different loss, not the full GO decoder, so plug-in gains are not cleanly attributable to GO structure.","rationale":"The paper's strongest central claim has two parts: GO-structured decoding beats flat decoding, and the gain is due to biological structure. The second part is well supported for the standalone MSGR by the random-hierarchy ablation (0.490 vs 0.517) and the SSGT control. The reader's weakest_assumption about GO DAG projection is a genuine limitation but not the most load-bearing issue: the random-hierarchy control already shows that even the tree-projected GO structure outperforms an arbitrary tree, and the multi-parent averaging experiment does not invalidate that. The more decisive gap is in the 'seamless plug-in replacement' claim, which is prominent in the abstract and Section 4.3. Appendix E reveals that the plug-in experiments use a different, lightweight MLP decoder rather than the full GO-guided decoder, and for STFlow the loss is changed from flow matching to MSE. Without a capacity-matched flat-head control and with the objective change, the plug-in gains cannot be attributed to the GO hierarchy. This does not overturn the main framework, but it does mean the model-agnostic plug-in claim is currently under-supported. The reader's CONDITIONAL verdict remains appropriate; no verdict change is needed, but the revision should add the matched flat-head control and clarify the plug-in decoder variant.","tokens_in":26725,"tokens_out":11179,"duration_ms":122945,"concrete_test":"Retrain the Table 2 plug-in variants exactly as reported, but replace the GO hierarchy in the lightweight decoder with a flat MLP head matched in parameter count, layer count, and hidden width, trained with the same loss as the +MS variant (for STFlow, use the same MSE loss rather than flow matching). Compare average PCC-200 and per-dataset deltas. If the flat matched-MLP head reproduces the +MS gains within 0.005 average PCC-200, then the plug-in improvements are capacity effects, not GO effects; if it does not, the plug-in claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.3 presents Table 2 as evidence that 'the linear output heads of STFlow, EGN, and ST-Net are replaced with the proposed GO-guided decoder,' with modifications 'strictly limited' to the head and the loss. But Appendix E says the plug-in actually uses 'a lightweight variant of the GO-guided decoder based on MLP layers rather than the full Transformer backbone used in MSGR.' For STFlow, the base objective is flow matching, not the 'standard MSE loss' that is substituted. Thus each plug-in comparison changes three things at once: head architecture (linear to deeper multi-scale MLP), capacity (more parameters/layers), and, for STFlow, the training objective. No flat-head control with matched capacity and the same loss is reported. The observed average gains (+0.019/+0.017/+0.018, with several negative per-dataset deltas) could reflect extra head capacity or the objective switch rather than GO structure. The MSGR-Random ablation is a good control for the standalone model, but it does not validate the lightweight plug-in variant used in Table 2.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MSGR, a hierarchical decoder for predicting spatial gene expression from histopathology images. It projects the Gene Ontology DAG into a four-level tree (root, domains, terms, genes) and uses a shared Transformer backbone to produce residual corrections from coarse to fine scales, supervised by multi-scale MSE targets derived from descendant-mean aggregation. The authors claim two central results: (1) GO-structured decoding outperforms flat decoding, including a state-of-the-art generative baseline (STFlow), across nine HEST-1k datasets; and (2) the gain comes from biological ontology structure rather than hierarchy per se, based on a +0.027 average PCC-200 margin over MSGR-Random, a structurally equivalent random hierarchy. They also claim the decoder is a model-agnostic plug-in that improves STFlow, EGN, and ST-Net by average gains of +0.019, +0.017, and +0.018. The paper includes ablations for the latent highway, residual inheritance, loss weights, alternative biological priors, and single-parent versus multi-parent projections.","tokens_in":27014,"tokens_out":2884,"duration_ms":31867,"significance":"If the claims hold, the paper makes a useful contribution by introducing a biologically structured decoding prior into a task where existing methods treat the gene output space as flat. The random-hierarchy control and the comparison against alternative biological priors (pathway, PPI, coexpression) are well-designed and provide credible evidence that the structured decoding interface itself is not the sole source of the gain for the full MSGR model. The authors also ship source code, which supports reproducibility. However, the plug-in evidence, a central advertised strength, is currently confounded by architectural and objective mismatches, and the main empirical comparisons lack statistical significance testing despite the small reported margins.","major_comments":[{"comment":"The plug-in claim is not cleanly attributable to GO structure. Section 4.3 says modifications are 'strictly limited' to replacing the linear output head and substituting the loss, but Appendix E states that the plug-in uses 'a lightweight variant of the GO-guided decoder based on MLP layers rather than the full Transformer backbone used in MSGR.' For STFlow, the base training objective is flow matching, not 'standard MSE loss,' so swapping to multi-scale MSE changes the training objective. Each plug-in comparison therefore varies head architecture, capacity, and (for STFlow) the learning objective simultaneously. There is no flat-head control with matched capacity and the same loss. The observed average gains of +0.019/+0.017/+0.018, with per-dataset negatives (e.g., LUNG -0.004 for STFlow+MS, PRAD -0.035 for EGN+MS), are thus consistent with a capacity or objective effect. Please report","section":"Section 4.3 vs. Appendix E (Table 2)"},{"comment":"The central empirical claims are supported only by point estimates. Most reported differences are small (average PCC-200 gains of 0.014–0.027, with several per-dataset deltas near zero or negative), and the number of cross-validation folds is only 2–6 per dataset. The paper reports mean±std across folds but provides no paired significance tests, confidence intervals, or effect-size/rules-of-three analysis. The phrase 'consistently outperforms' is stronger than what the evidence supports. Please add per-dataset significance tests (e.g., paired bootstrap or Wilcoxon across folds) or otherwise quantify uncertainty in the average margins.","section":"Tables 1–3"},{"comment":"The random-hierarchy control is convincing for the standalone MSGR model but does not validate the lightweight plug-in variant. Table 3 shows MSGR-Random trailing MSGR by 0.027 under the full Transformer backbone, but Table 2's plug-in variants use the Appendix E MLP-based decoder, for which no random-hierarchy ablation is reported. The claim that the plug-in gains are 'attributable to biological ontology structure rather than hierarchical decomposition per se' therefore lacks a control for the plug-in setting. Either add an MSGR-Random version of the lightweight decoder for the three base models, or state the plug-in results as compatibility evidence only, without the biological-attribution claim.","section":"Section 4.5, Table 3 vs. Section 4.3, Table 2"}],"minor_comments":[{"comment":"Typo: 'Spaital Transformer' should be 'Spatial Transformer.'","section":"Figure 5 caption"},{"comment":"The organ column lists 'KIDNEY' for both CCRCC and KIDNEY rows; please distinguish renal cell carcinoma versus normal kidney or clarify the dataset naming.","section":"Table 5"},{"comment":"Equation (20) is stated but not numbered; please number it for cross-referencing in the discussion of residual inheritance.","section":"Appendix G.1"},{"comment":"The claim that MSGR 'outperforms STFlow on eight of nine datasets' should be accompanied by the fold-level standard deviations for those specific comparisons; Table 1 already includes them, but a direct paired-error visualization would help.","section":"Section 4.2"},{"comment":"The loss-weight ablation is only reported on five datasets; specify whether the remaining four datasets were omitted for computational reasons and note the limitation in the text.","section":"Appendix C.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical study with a well-specified method and a good random-hierarchy control for the standalone model. The main weakness is that the plug-in claim, highlighted in the abstract and Section 4.3, is not supported by the current experimental setup because of the lightweight-MLP and objective changes. I would encourage the editor to send the paper back for a revision that either adds the missing matched controls or softens the plug-in attribution claim. The lack of significance testing also needs to be addressed before the paper can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know: this is a solid, well-engineered empirical paper. The core claim—GO-structured decoding beats flat decoding, and the gain comes from ontology structure rather than hierarchy per se—is supported by the random-hierarchy ablation (0.490 vs 0.517). That control is well done, the code is public, and the benchmark is standard. It deserves a serious referee.\n\nWhat’s actually new: using GO as an explicit structural prior for histology-to-expression decoding, with a systematic four-level tree projection and residual coarse-to-fine refinement. The ablations against random, pathway, PPI, and coexpression hierarchies are a useful contribution, as is the external evaluation of intermediate GO-term predictions against Hallmark activities. The paper is honest about limitations: the DAG-to-tree simplification, mean aggregation of sparse pathway signals, and the MSE trade-off versus STFlow are all acknowledged.\n\nThe soft spots are real but manageable. The stress-test concern about Table 2 is correct. The main text says the plug-in replaces the linear head with the proposed GO-guided decoder and substitutes the loss, but Appendix E says the plug-in uses a lightweight MLP-based variant, not the full Transformer backbone. For STFlow, the base objective is flow matching, not MSE. So the plug-in gains (+0.019/+0.017/+0.018) could come from extra head capacity or the objective switch rather than GO structure. The MSGR-Random ablation does not validate this lightweight variant. That weakens the \"seamless plug-in replacement\" framing. A matched-capacity flat-head control with the same loss would settle it.\n\nAlso worth noting: the main comparisons lack significance tests, the average margins are small (0.014–0.027 PCC-200), several per-dataset plug-in deltas are negative, and hyperparameters appear to be selected on the evaluation datasets. These are standard empirical weaknesses, not fatal flaws.\n\nWho this is for: people working on spatial gene expression prediction from H&E. The standalone MSGR result is credible and the idea is worth building on; I would cite it. But I would not rely on the plug-in claim until the confound is addressed.\n\nRecommendation: send it to peer review, conditional on fixing the plug-in evidence. The central contribution stands on its own.","headline":"Solid empirical paper with a real new idea and a good random-hierarchy control; the GO-structure claim holds, but the plug-in compatibility evidence is confounded and needs a matched-capacity control.","tokens_in":27488,"tokens_out":1850,"would_cite":true,"duration_ms":20262,"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":"MSGR argues that spatial gene expression prediction improves when target genes are decoded coarse-to-fine along a Gene Ontology tree rather than as a flat vector, and that the gain comes from the biological structure itself, not from hierar","keywords":["spatial transcriptomics","histopathology images","Gene Ontology","hierarchical decoding","residual refinement","gene expression prediction","plug-in decoder","HEST-1k"],"falsifier":"Train the identical MSGR decoder on a hierarchy in which GO term assignments have been randomly permuted across genes while preserving tree shape and term sizes. If the PCC-200 gap over flat decoding does not collapse toward zero, the claimed biological-structure contribution is not doing the work.","tokens_in":26642,"feed_emoji":"🧬","tokens_out":10221,"duration_ms":100742,"temperature":0.7,"pith_summary":"The paper argues that predicting spatial gene expression from histology images should not treat the target genes as an unstructured 200-dimensional output. It organizes the genes into a four-level tree derived from Gene Ontology (virtual root, GO domains, GO terms, individual genes) and decodes coarse-to-fine, with each level adding a residual correction to the inherited parent prediction under scale-weighted supervision. On nine HEST-1k datasets, the GO-guided decoder reports an average PCC-200 of 0.517, ahead of flat decoders including the generative baseline STFlow (0.503). A structurally identical random hierarchy scores 0.490, a 0.027 gap that the paper attributes to the biological content of the ontology rather than to hierarchy depth or shape. Because the decoder only replaces the output head, it can be inserted into existing pipelines and the paper reports consistent improvements for STFlow, EGN, and ST-Net (+0.019, +0.017, +0.018 average PCC-200).","feed_headline":"Ontology-guided decoding lifts gene prediction from tissue images","feed_subtitle":"A four-level GO hierarchy plus residual corrections beats flat decoding on nine benchmarks—and random trees by 0.027.","key_machinery":"The central object is the four-level GO tree projection: virtual root, up to four level-1 nodes (three GO domains plus an unannotated proxy), a level-2 layer of GO terms or proxies, and level-3 leaf genes, with each gene assigned exactly once to its most specific annotation and each GO term connected to its largest parent. Coarse targets are means over descendant leaves, so supervision is consistent across scales. The mechanism that carries the argument is the shared multi-scale residual decoder: scale-0 outputs a root estimate; every finer scale broadcasts the parent prediction to its children and adds a residual from a shared AdaLN Transformer block, while a latent highway concatenates par","core_discovery":"MSGR's central claim is that a panel of hundreds of genes can be predicted from an H&E image by first forming a coarse functional picture and then refining it. The target genes are arranged offline into a four-level GO tree; the expression target of every non-leaf node is the mean expression of its descendant genes. A decoder shared across levels starts from the virtual root, and at each finer level the parent prediction is broadcast to children, a residual correction is learned from conditioning tokens, and latent context from the parent scale is propagated forward through a cross-attention highway. The paper's evidence that this structure matters: the same backbone without the hierarchy (S","pith_inferences":["If the biological-structure effect is genuine, then other functional gene groupings—pathway modules, protein interaction communities, even data-driven co-expression modules—should also help when used as the hierarchy; the paper's own alternative-prior ablation shows they help less but still beat a random tree, suggesting the benefit comes from coherent functional grouping rather than from GO's cur","A natural next test is a full DAG-aware decoder that retains gene-to-term multi-parent edges and term-to-term connectivity; the paper's multi-parent averaging variant is not that test and loses 0.017, so it does not settle whether richer topology would help or hurt.","Since the plug-in gains are similar across a CNN, a Transformer regressor, and a flow-matching model, the same gene-side hierarchy should transfer to other multi-gene prediction heads, such as super-resolution expression or cell-type deconvolution, without retraining the image-side modules.","A complementary experiment would hold the decoder fixed and build the hierarchy from expression data per tissue instead of from GO; if a learned hierarchy matches or exceeds GO, it would measure how much curated biological prior remains useful once data are abundant."],"forward_implications":["The GO-guided decoder can replace the final linear head of diverse models, with average PCC-200 gains of +0.019 for STFlow, +0.017 for EGN, and +0.018 for ST-Net across nine datasets.","GO-structured decoding improves the top-10, top-50, and top-200 gene correlations, with reported averages of 0.776, 0.710, and 0.517, respectively.","Intermediate GO-term predictions constitute interpretable outputs that align with independent pathway activity annotations, with a reported mean correlation of 0.896 on SKCM.","The gene-side decoder is orthogonal to image-side multi-magnification methods, so both enhancement strategies can be combined in the same architecture.","Both cross-scale communication (latent highway, -0.015) and residual inheritance (-0.013) contribute to the final average PCC-200, so the benefit is not from the tree alone but from the coarse-to-fine refinement protocol."],"supporting_citations":[{"why":"Defines the Gene Ontology hierarchy that is the paper's structural prior.","marker":"[1]"},{"why":"Supplies the nine spatial transcriptomics datasets and the benchmark protocol used in all comparisons.","marker":"[14]"},{"why":"Provides the generative flow-matching baseline and the spatial-conditioning architecture that MSGR extends; its flat head is replaced in plug-in experiments.","marker":"[13]"},{"why":"Provides the exemplar-guided regression baseline whose output head is replaced in plug-in experiments.","marker":"[42]"},{"why":"Provides the CNN regression baseline (ST-Net) and the DenseNet-121 backbone used in plug-in experiments.","marker":"[10]"},{"why":"Supplies the GO term-to-gene annotation query interface used to construct each dataset's four-level tree.","marker":"[16]"},{"why":"Supports the biological premise that genes sharing GO annotations tend to show correlated expression.","marker":"[35]"},{"why":"Supplies the frozen foundation-model image encoder used to extract patch embeddings across compared methods.","marker":"[4]"},{"why":"Supplies the pathway gene-set database used in the alternative-prior ablation.","marker":"[8]"},{"why":"Supplies the protein-protein interaction network used in the alternative-prior ablation.","marker":"[32]"}],"fun_headline_variants":["Ontology tree refines gene prediction from tissue images","GO-guided decoder lifts spatial gene prediction","Hierarchical GO beats flat decoding in gene prediction","MSGR uses Gene Ontology to sharpen gene predictions"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The projection of the GO DAG into a single-parent tree must retain enough functional coherence that the coarse-scale supervision targets are biologically meaningful; if that projection scrambles the signal, the method would reduce to an arbitrary hierarchy.","fun_headline_variants_meta":{"raw":{"variants":["Ontology tree refines gene prediction from tissue images","GO-guided decoder lifts spatial gene prediction","Hierarchical GO beats flat decoding in gene prediction","MSGR uses Gene Ontology to sharpen gene predictions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000149,"raw_usage":{"total_tokens":1028,"prompt_tokens":739,"completion_tokens":289,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":230}},"tokens_in":483,"tokens_out":289,"duration_ms":3921,"temperature":1.0,"reasoning_tokens":230,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T04:13:00.018460+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the identical MSGR decoder on a hierarchy in which GO term assignments have been randomly permuted across genes while preserving tree shape and term sizes. If the PCC-200 gap over flat decoding does not collapse toward zero, the claimed biological-structure contribution is not doing the work.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the nine spatial transcriptomics datasets and the benchmark protocol used in all comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the generative flow-matching baseline and the spatial-conditioning architecture that MSGR extends; its flat head is replaced in plug-in experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the exemplar-guided regression baseline whose output head is replaced in plug-in experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the CNN regression baseline (ST-Net) and the DenseNet-121 backbone used in plug-in experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the GO term-to-gene annotation query interface used to construct each dataset's four-level tree."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports the biological premise that genes sharing GO annotations tend to show correlated expression."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the pathway gene-set database used in the alternative-prior ablation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the protein-protein interaction network used in the alternative-prior ablation."}],"review_version":1}