{"id":"4491d130-3855-4017-98e0-db5c1858e45c","arxiv_id":"2507.04891","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MurreNet decouples pathology and genomic features into shared and modality-specific components, then fuses them orthogonally, reporting state-of-the-art C-indices on six TCGA cancer cohorts.","lead":"This paper presents MurreNet, a neural network that splits tumor image and genomic data into shared and unique parts and then recombines them to predict cancer survival. On six public cancer datasets it reports better concordance indices than 15 prior methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim rests on C-index gaps that are never tested for significance, and the reported difference-loss mechanism is internally inverted.","rationale":"The reader's weakest assumption identifies the misdescribed KL loss in Eq. 8, and that is indeed a genuine internal inconsistency: under Eq. 6 with positive beta, minimizing D_KL(h_c||h_s) drives common and specific representations toward each other, contradicting the stated disentanglement purpose. This is a real mechanical flaw and worth flagging. However, I do not consider it the single most load-bearing concern about the paper's central claim, which is the comparative SOTA assertion in Section 3.3. That claim is an empirical statement about C-indices, and its support is statistically fragile: only 5 Monte-Carlo folds, no model-versus-model significance test, and several headline margins are smaller than the reported standard deviations. The KL issue attacks the mechanism explanation and the interpretability of the ablation, but the headline numeric claim could survive even if Eq. 8 is a sign typo; conversely, a wrong-sign loss does not by itself invalidate the C-index table if the engineering still works. The reader's verdict is CONDITIONAL, and my analysis supports keeping that verdict, but for a more load-bearing reason: the absence of paired significance testing and code-level reproducibility leaves the SOTA claim unverified. My concrete test therefore targets both: a paired significance check of the headline numbers, and an ablation sign-flip test that would settle whether the KL mechanism is a red herring. Thus my agreement with the reader is partial: the mechanistic critique is correct and should be fixed, but the decisive threat to the primary claim is the missing significance evidence.","tokens_in":9273,"tokens_out":5527,"duration_ms":44237,"concrete_test":"Re-run LUAD and UCEC with the author's data splits (or released code if provided) and compute paired per-fold C-index differences between MurreNet and SurvPath/CMTA/MCAT, testing significance with a paired permutation test at p<0.05. Separately, run the ablation with the sign of the L_diff term flipped (maximize D_KL in Eq. 8 instead of minimize, keeping all other hyperparameters fixed) on BLCA and UCEC; if C-index does not drop, the reported benefit of the difference loss is not caused by the stated objective.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is the Section 3.3 statement that MurreNet 'achieves superior performance on all cancer types,' with C-index improvements over MCAT of 2.37% to 14.29% across six TCGA cohorts. This claim requires that the reported gains are real and not noise. The paper reports only mean C-index and standard deviation from 5-fold Monte-Carlo cross-validation; it performs no statistical test comparing MurreNet's C-index to any baseline. The 'T-test analyses' mentioned in Section 3.2 are applied to high- versus low-risk group survival (the log-rank test in Section 3.5), not to model-versus-model C-index differences. With typical fold-level stds of 0.03-0.06 and only 5 folds, several reported advantages are within or near one pooled standard error. On LUAD, MurreNet (0.691±0.040) is only 0.015 above SurvPath (0.676±0.036); on UCEC, MurreNet (0.752±0.075) overlaps CMTA (0.740±0.066), SurvPath (0.737±0.049), and PORPOISE (0.737±0.097) within one standard deviation. Second, the ablation's largest single jump (model D to E in Table 2, adding the difference loss) is attributed to a loss term whose effect, by the equations in Section 2.4, contradicts its stated purpose. Eq. 6 minimizes L_total with positive beta, and Eq. 8 defines L_diff = D_KL(h_c || h_s), so minimizing this term reduces the divergence between modality-common and modality-specific features. The text in Section 2.4 claims L_diff 'ensure[s] that the modality-common and -specific representations capture distinct aspects,' but no mechanism in the loss enforces distinctness. Thus the reported disentanglement benefit and its ablation gain are not explained by the stated mechanism; either the loss should be maximized or redefined (e.g., as negative KLD or an orthogonality penalty), or the ablation gains must be attributed to another component. Both issues are concrete and verifiable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes MurreNet, a multimodal survival analysis model that combines histopathology whole-slide images and genomic profiles. The model contains a Multimodal Representation Decomposition (MRD) module that splits each modality into specific and common representations, a Deep Holistic Orthogonal Fusion (DHOF) module that recombines them, and a composite training objective with similarity, difference, reconstruction, and survival losses. Experiments on six TCGA cohorts against 15 baseline methods report mean C-indices, with the authors claiming state-of-the-art performance on all cohorts. An ablation study and Kaplan-Meier risk-stratification analysis are also presented. The main contribution is the specific decoupling/fusion architecture and its reported empirical performance.","tokens_in":9743,"tokens_out":7057,"duration_ms":71416,"significance":"If the reported C-index gains are statistically robust, MurreNet would be a useful advance in multimodal survival prediction and would provide evidence that explicit decoupling of shared and specific representations helps integrative pathology-genomics modeling. The paper compares with a broad set of 15 baselines on six public TCGA benchmarks, which is a strength. However, the headline claim is not yet supported: no significance test is applied to model-vs-model C-index differences, and the difference-loss mechanism in the training objective is defined in a way that contradicts its stated purpose. The manuscript would need corrected loss definitions and additional statistical evidence before the central claim can be accepted.","major_comments":[{"comment":"The statement that MurreNet 'achieves superior performance on all cancer types' is not backed by any statistical comparison of C-index values. The table reports only mean±std over five Monte-Carlo CV folds, and several improvements over the strongest baselines are within one standard deviation (e.g., LUAD: 0.691±0.040 vs. SurvPath 0.676±0.036; UCEC: 0.752±0.075 vs. CMTA 0.740±0.066 and PORPOISE 0.737±0.097). The 'T-test' and log-rank analyses in Sections 3.2 and 3.5 test high- vs. low-risk group separation, not model-vs-model differences. The authors should report paired significance tests (e.g., paired permutation or bootstrap over folds, or DeLong tests on pooled risk scores) and confidence intervals for the C-index differences, and should restrict the SOTA claim accordingly.","section":"Sec. 3.3, Table 1"},{"comment":"The difference loss is internally inverted. Eq. (6) minimizes L_total with positive beta, and Eq. (8) defines L_diff as D_KL(h_c || h_s). Minimizing a positive-weighted KL divergence drives the modality-common and modality-specific representations to become more similar, which is the opposite of the stated goal in Sec. 2.4 ('ensure that the modality-common and -specific representations capture distinct aspects') and of the 'repel' annotation in Fig. 1. The ablation gain from model D to model E in Table 2 therefore cannot be attributed to the mechanism described. In addition, KL divergence is only defined for probability distributions; applying it to raw feature vectors that may contain negative values and are not normalized is mathematically invalid. The authors need to either reverse the objective (e.g., maximize the divergence or use a negative weight), use a proper distance measure, and re-run the affected experiments, or clearly explain why the current formulation is intended.","section":"Sec. 2.4, Eqs. (6)-(8)"},{"comment":"The similarity loss formula does not match its description. The text states that L_sim reduces the discrepancy between 'the two modality-common representations' (presumably h_c_p and h_c_g), but Eq. (7) computes L1(h_o_p, h_c_p), which aligns the raw pathology embedding to the pathology-common representation. As written, the loss cannot perform the claimed cross-modal common-space alignment, and it may also be invalid if h_o_p (768-d) and h_c_p (d-d) have different dimensions. The formula should be corrected or the description revised.","section":"Sec. 2.4, Eq. (7)"},{"comment":"The co-attention matrix A is not well-defined. Eq. (2) computes A as the scalar dot product Linear(h_o_p)^T · Linear(h_o_g), but Eq. (1) then uses A^T * h_o_p and MLP(A) * h_o_g as if A were a matrix. If A is intended to be a co-attention map, it should be constructed as a matrix (e.g., an outer product or a softmax of QK^T), and the dimensions should be stated. As written, the MRD module cannot be reproduced from the equations.","section":"Sec. 2.2, Eqs. (1)-(2)"},{"comment":"The paper states in Sec. 3.2 that 'T-test analyses' are used to assess significance, but Sec. 3.5 reports a Log-rank test for risk-group differences. These are different procedures, and the p-values from the log-rank test only show that the model can separate high- and low-risk groups within each cohort; they do not establish that MurreNet outperforms any baseline. The text should be corrected and the distinction made explicit.","section":"Sec. 3.2, Sec. 3.5"}],"minor_comments":[{"comment":"The text says '14.29% on UCED' but the dataset is UCEC; this typo should be fixed.","section":"Sec. 3.3"},{"comment":"The word 'preform' in the Introduction should be 'perform'.","section":"Sec. 1"},{"comment":"The text states the key-value pair is (f_o_c, f_o_c), but the formula uses (f_o_s, f_o_s); these should be aligned.","section":"Sec. 2.3, Eq. (3)"},{"comment":"The text mentions enhanced representations f'_s and f'_c, but only the definition for f'_s is given; the analogous definition for f'_c should be provided.","section":"Sec. 2.3, Eq. (4)"},{"comment":"The hidden dimension d, number of attention heads, decoder architecture, and other architectural hyperparameters are not reported; these details are needed for reproducibility.","section":"Sec. 3.2"},{"comment":"The notation in Eq. (8) sums over modalities while the logarithm is applied to vectors; the element-wise summation over feature dimensions should be written explicitly to avoid ambiguity.","section":"Sec. 2.4, Eq. (8)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the scope of the journal, but the decision should hinge on whether the authors can correct the loss definitions and provide statistical evidence for the C-index comparisons. If the experiments already used a correctly signed difference loss, the paper should state this explicitly and explain the current formulation. I recommend inviting a major revision rather than rejecting, since the identified issues are fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: MurreNet is a standard-combination architecture for combining H&E slides and genomics in survival models. It does not break new conceptual ground, but the specific recipe (decoupling, orthogonality, multi-loss regularization) is not in the cited prior work and the ablation is consistent. The problem is that the paper's main claim—state-of-the-art on six TCGA cohorts—rests on mean C-index differences that are never tested for significance, and one of the regularizers is described backwards relative to its own equation.\n\nWhat is actually new: MRD + DHOF plus the similarity/difference/reconstruction losses is a new combination, and the authors do include a broad baseline table. Using CHIEF as frozen feature extractor and TCGA cohorts is standard and fine. The writing is clear, and the reported mean numbers are consistently positive.\n\nNow the soft spots. The 'difference loss' in Eq. 8 is D_KL(h_c || h_s), and Eq. 6 minimizes it with positive beta. That makes the common and specific representations more similar, not more distinct. The text says it 'ensure[s] that the modality-common and -specific representations capture distinct aspects,' which is the opposite. This is not a typo-level detail: the ablation gives the largest jump to model E, which adds exactly this term, so the stated mechanism cannot explain the gain. Similarly, Eq. 7 calls L_sim a similarity between the two modality-common representations, but the equation is L1(h_o^p, h_c^p)—raw input vs common. The evaluation also lacks any test of whether C-index gaps are real; with 5 folds and stds of 0.03–0.08, LUAD (0.691 vs 0.676 for SurvPath) and UCEC (0.752 vs 0.740 for CMTA) are within a standard error. Related decoupling baselines (CCL, MuGI, PIBD) are mentioned in the intro but not compared. No code or data is released.\n\nNone of this means the model is bad. The mean gains are consistent, and fixing the loss to actually enforce difference (negative KL, or an orthogonality penalty) could make the story work. But as written, the SOTA claim is not fully supported.\n\nWho is this for: researchers working on histology-genomics fusion who want a recipe for disentanglement-style losses. It is not a paper that changes the field.\n\nRecommendation: send it to peer review, not desk reject, but with the expectation of major revision. A serious referee should ask for significance testing, corrected loss definition, and the missing baselines.","headline":"MurreNet is a plausible incremental architecture, but the SOTA claim is under-supported and the 'difference' regularizer is backwards relative to its own equation.","tokens_in":10265,"tokens_out":4780,"would_cite":false,"duration_ms":52559,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Disentangling images and genomics lifts cancer survival prediction","keywords":["multimodal survival prediction","whole slide images","genomic profiles","representation disentanglement","orthogonal fusion","co-attention","TCGA","deep learning"],"falsifier":"Train MurreNet on one TCGA cohort and record the KL divergence between the modality-common and modality-specific representations at the end of training; if the divergence has dropped toward zero, the difference loss is not separating the representations. A direct ablation that removes only the difference term and shows no change in C-index would further confirm that the stated disentanglement benefit is attributable to other components.","tokens_in":9027,"feed_emoji":"🧬","tokens_out":3580,"duration_ms":39487,"temperature":0.7,"pith_summary":"MurreNet is a deep learning model that combines whole-slide pathology images with genomic profiles to predict cancer survival. The paper argues that previous multimodal methods over-rely on shared information and under-use modality-specific signal, so it decomposes each patient's features into a common component and two private components, then recombines them with an orthogonal-fusion step. The authors report state-of-the-art concordance-index scores on all six TCGA cancer cohorts tested, with gains over the strongest prior model ranging from about 2.4% to 14.3%. If these results hold, the key lesson is that explicitly separating and re-integrating private and shared multimodal information improves prognosis.","feed_headline":"Disentangling images and genomics lifts cancer survival prediction","feed_subtitle":"Six TCGA cohorts show consistent C-index gains; the split-and-recombine design is the reason.","key_machinery":"The Multimodal Representation Decoupling (MRD) module produces pathology-specific, genomics-specific, and modality-common embeddings, with the common branch built from a co-attention matrix between the two modalities. The Deep Holistic Orthogonal Fusion (DHOF) then computes, for each specific feature, the residual left after projecting it onto the common representation; that orthogonal residual is concatenated and pooled with the common features to form the final multimodal embedding. The training objective is the weighted sum of an L1 similarity loss, a Kullback-Leibler divergence term (written as a difference loss), an MSE reconstruction loss, and a Cox-style negative log-likelihood survival loss.","core_discovery":"The central claim is that survival prediction improves when histopathology and genomic features are explicitly disentangled into modality-common and modality-specific representations and then fused through a step that removes the common projection from the specific ones. MurreNet uses three parallel encoders, a co-attention module for the common branch, a cross-attention transformer to refine the split features, and an orthogonal fusion that keeps only the component of the specific features orthogonal to the common representation. Trained with a loss that combines similarity, a stated difference term, reconstruction, and negative log-likelihood survival loss, the model achieves higher C-index values than all fifteen baselines on six TCGA datasets. The paper takes that as evidence that holistic modeling of both inter- and intra-modality interactions is what multimodal survival analysis needs.","pith_inferences":["As written, the KL 'difference' loss in Eq. 8 minimizes the divergence between modality-common and modality-specific features when added to the total loss with a positive weight, so the paper's stated mechanism for enforcing distinctiveness may be acting in the opposite direction; the observed gains could come from the reconstruction or similarity losses instead.","A sharper test of the disentanglement claim would be to measure the actual KL divergence after training; if it has decreased, the model is not separating the representations the way the prose describes.","The comparison against MCAT may be confounded by the stronger pathology encoder (CHIEF) used in MurreNet; a fairer attribution would require re-running MCAT with the same feature extractor.","The orthogonal-fusion step is a testable design: on synthetic paired data with known shared and private signal, one can verify whether the orthogonal residual indeed carries the survival-relevant private information."],"forward_implications":["If MurreNet's reported results hold, multimodal survival models should adopt explicit decoupling of common and private features rather than simple concatenation or attention-only fusion.","The per-cancer gains over MCAT (2.37% to 14.29% C-index) suggest the benefit of disentanglement varies with tumor type, with the largest gains on COADREAD, UCEC, and STAD.","Ablation results indicate that the similarity, difference, and reconstruction losses each add predictive value on top of the architectural modules, so the training regularization is a real contributor.","Risk stratification with the model yields statistically significant separation of high- and low-risk groups across all six cohorts, supporting potential clinical use in prognosis.","The framework extends naturally to other paired biomedical modalities, such as radiology images with lab values or methylation arrays with expression data."],"supporting_citations":[{"why":"CHIEF is the frozen pathology foundation model that embeds each WSI patch into instance-level features, the visual input of the whole pipeline.","marker":"[23]"},{"why":"CLAM is used to tile each whole slide image into non-overlapping patches, defining the input preprocessing for pathology.","marker":"[19]"},{"why":"MCAT is the current SOTA multimodal baseline whose C-index the paper claims to beat by 2.37% to 14.29% across the six cohorts.","marker":"[5]"},{"why":"CMTA serves both as a strong multimodal baseline and as the source of the Nyström attention used inside the transformer decoders.","marker":"[30]"},{"why":"SurvPath is a dense multimodal interaction baseline that MurreNet outperforms, supporting the claim of comprehensive interaction modeling.","marker":"[9]"},{"why":"The negative log-likelihood survival loss is the task objective that links the learned representations to censored survival data.","marker":"[27]"},{"why":"Kullback-Leibler divergence is the basis of the 'difference' regularization term that is supposed to separate modality-specific from modality-common features.","marker":"[12]"}],"fun_headline_variants":["Decoupling shared and unique features improves survival prediction","MurreNet separates modality signals to sharpen cancer prognosis","Orthogonal fusion of pathology and genomics lifts survival C-index","Disentangled histology and genomics predict survival more accurately","Splitting common and specific representations boosts survival models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes the divergence term in Eq. 8 forces modality-specific and modality-common features to become distinct, but as written, minimizing that term with a positive weight actually reduces the divergence between the two feature sets, so the claimed disentanglement mechanism may not be doing the work.","fun_headline_variants_meta":{"raw":{"variants":["Decoupling shared and unique features improves survival prediction","MurreNet separates modality signals to sharpen cancer prognosis","Orthogonal fusion of pathology and genomics lifts survival C-index","Disentangled histology and genomics predict survival more accurately","Splitting common and specific representations boosts survival models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00038,"raw_usage":{"total_tokens":1998,"prompt_tokens":907,"completion_tokens":1091,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":1014}},"tokens_in":523,"tokens_out":1091,"duration_ms":11794,"temperature":1.0,"reasoning_tokens":1014,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:36:12.468292+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train MurreNet on one TCGA cohort and record the KL divergence between the modality-common and modality-specific representations at the end of training; if the divergence has dropped toward zero, the difference loss is not separating the representations. A direct ablation that removes only the difference term and shows no change in C-index would further confirm that the stated disentanglement benefit is attributable to other components.","supporting_citations":[{"cited_title":"Nature634(8035), 970–978 (2024) Modeling Holistic Multimodal Interactions for Survival Prediction 11","cited_arxiv_id":null,"evidence_quote":"CHIEF is the frozen pathology foundation model that embeds each WSI patch into instance-level features, the visual input of the whole pipeline."},{"cited_title":"In: Proceedings of the IEEE/CVF interna- tional conference on computer vision","cited_arxiv_id":null,"evidence_quote":"MCAT is the current SOTA multimodal baseline whose C-index the paper claims to beat by 2.37% to 14.29% across the six cohorts."},{"cited_title":"In: Proceedings of the IEEE/CVF International Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"CMTA serves both as a strong multimodal baseline and as the source of the Nyström attention used inside the transformer decoders."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"SurvPath is a dense multimodal interaction baseline that MurreNet outperforms, supporting the claim of comprehensive interaction modeling."}],"review_version":1}