{"id":"9606c041-4adb-4339-bf73-66c2b487a073","arxiv_id":"2509.16301","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"TF-DWGNet combines XGBoost-derived directed weighted graphs with low-rank tensor fusion to improve multi-omics cancer subtype classification.","lead":"TF-DWGNet is a graph neural network that builds directed, weighted molecular interaction graphs from tree splits and fuses three omics data types with a low-rank tensor to classify cancer subtypes. It reports accuracy gains over five baselines on breast, uterine, and kidney cancer datasets, but ships no code or data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Supervised XGBoost graph construction may see test labels: the paper never states that it is re-fit inside each training split.","rationale":"The reader's weakest assumption is exactly the load-bearing concern: supervised graph construction using labels, if performed on the full dataset, would leak test information into the feature set and graph topology that drive the reported gains. The paper's language is ambiguous on this point and no code or data are provided, so the empirical claim is not yet verifiable. Because the concern is real but not yet proven, the existing CONDITIONAL verdict is appropriate; no new independent concern emerged that would force rejection or acceptance.","tokens_in":13355,"tokens_out":2869,"duration_ms":27747,"concrete_test":"Ask the authors to release code or state explicitly where XGBoost graph construction sits relative to the split loop. Independently rerun the BRCA experiment under two conditions: (a) graph built once on all 875 samples, as the current text could imply, and (b) XGBoost refit on each training set only (60%) with the resulting selected features and edges applied to validation/test. A correct implementation must match (b). If condition (a) reproduces the reported numbers while (b) drops TF-DWGNet's accuracy/F1-macro by more than a few points or eliminates the significant advantage over GEDFN, the central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that TF-DWGNet's directed weighted graphs plus tensor fusion outperform baselines. Graph construction is explicitly supervised: XGBoost uses labels Y to select features and build directed weighted edges. If XGBoost is run once on the full dataset before the 20 train/validation/test splits, then test labels influence (i) the selected feature set p* and (ii) the adjacency matrices used by the GNN and by the GCN/GEDFN baselines. The methodology section says 'For each omics modality, we generate an ensemble of M trees' and that XGBoost 'performs feature selection and constructs task-specific directed weighted graphs'; the experimental protocol only says '20 independent stratified train-validation-test splits (60%:20%:20%)' and 'GCN and GEDFN utilize graphs constructed via XGBoost,' but never says the graph construction is rerun inside each training fold. The supplemental material (A.2) lists only hyperparameters. Without this guarantee, the reported 8-12% BRCA gains and 1-3% margins on UCEC/KIPAN could be inflated by label leakage rather than by the model architecture. Code is 'available upon request,' so the ambiguity cannot currently be resolved empirically.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"TF-DWGNet proposes a multi-omics cancer-subtype classifier that combines supervised, tree-based directed weighted graph construction (via XGBoost) with low-rank CP-decomposed tensor fusion over three omics modalities, followed by a residual deep network. The method is evaluated on BRCA, UCEC, and KIPAN against RF, XGBoost, DFN, GCN, and GEDFN baselines over 20 stratified splits, reporting accuracy, F1-weighted, and F1-macro plus Welch t-tests. The authors claim consistent state-of-the-art performance and additionally provide modality- and feature-level interpretability scores.","tokens_in":13697,"tokens_out":4302,"duration_ms":39834,"significance":"If the experimental protocol is leakage-free, the paper offers a useful integration of tree-based supervised graph construction with low-rank tensor fusion, and the interpretability component is a practical addition. The evaluation is extensive for a bioinformatics methods paper: 20 repeated splits, confidence intervals, and pairwise statistical tests on three cancer cohorts. The biological plausibility checks (e.g., FAM134B, FOXC1) strengthen the interpretability claim. However, the central performance claim is contingent on an unresolved question about whether the supervised graph construction is refit within each training split, and the KIPAN results are mostly not statistically significant; both issues must be addressed before the 'consistently outperforms' conclusion can be accepted.","major_comments":[{"comment":"The graph construction is explicitly supervised: XGBoost uses labels Y to select features and build directed weighted edges. The protocol only describes '20 independent stratified train-validation-test splits' and says GCN and GEDFN use graphs constructed via XGBoost, but it never states that the XGBoost graph construction is rerun inside each training fold. If the graphs are built once on the full dataset before splitting, then test labels influence both the selected feature set p_i^* and the adjacency matrices used by TF-DWGNet and by the GCN/GEDFN baselines, inflating the margins in Table 3. This is load-bearing for the central claim. Please state explicitly that graph construction is performed per training fold (or per training+validation split only), or release code/rerun experiments under a nested-CV protocol.","section":"Section 3 ('Directed Weighted Graph Construction') and Section 4 ('Baselines and Experimental Setting')"},{"comment":"The abstract and conclusion state that TF-DWGNet 'consistently outperforms state-of-the-art baselines across multiple metrics and statistical tests.' On KIPAN, however, the absolute gains over the strongest baseline GEDFN are small (accuracy 0.964 vs 0.957; F1-macro 0.956 vs 0.946) and, by the authors' own report, 'most differences are not significant (p-value > 0.05).' This overstates the evidence. The claim should be qualified to BRCA (and partially UCEC), or the KIPAN proof-of-concept framing should be made explicit in the abstract and conclusion.","section":"Section 4.1, Table 3 (KIPAN rows)"},{"comment":"The manuscript states that datasets and code are 'available upon request' and will be made public upon publication. Because the central ambiguity in the experimental protocol (Major Comment 1) is precisely a graph-construction protocol ambiguity, the lack of released code makes the issue currently irresolvable. For a methodological paper whose contribution is the graph construction and fusion mechanism, providing the code is necessary to verify the reported results.","section":"Data and Code Availability"}],"minor_comments":[{"comment":"Typographical inconsistencies: 'TG-DWGNet' appears in the experimental setup text, and 'TF-DMGNet' appears in the Table 2 caption. These should be 'TF-DWGNet.'","section":"Section 4, Baselines and Experimental Setting; Table 2 caption"},{"comment":"The layer uses the symmetrically normalized adjacency ~D^{-1/2}~A~D^{-1/2}. For a directed weighted graph, ~A is not symmetric and ~D is not uniquely defined; the paper should specify whether ~D is the in-degree or out-degree matrix, or otherwise define the normalization.","section":"Section 3, GEDFN layer equation"},{"comment":"With P_i ∈ R^{R×(p_i^*+1)} and Z_i^* ∈ R^{n×(p_i^*+1)}, the product P_i Z_i^* is R×n, not n×R. The element-wise product across modalities and the subsequent fused representation should specify the orientation (e.g., transpose to n×R) to avoid confusion.","section":"Section 3, Tensor Fusion"},{"comment":"The superscript explanation reads 'TF-DWGNet is statistically significant than the corresponding baseline'; it should be 'significantly better than' or equivalent.","section":"Table 3 note"}],"recommendation":"major_revision","confidential_remarks":"In my view, the paper is not acceptable in its current form because the leakage ambiguity is central and cannot be resolved from the manuscript text. If the authors confirm that XGBoost graph construction is refit inside each training split, and if they soften the KIPAN significance claims, the paper would be a credible methods contribution. I would also urge the editor to require code release as a condition for any further review, since the graph-construction protocol is the crux."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper to know: TF-DWGNet reuses two known ideas - forgeNet's XGBoost-derived directed weighted graphs and TFN's low-rank tensor fusion - and combines them for multi-omics cancer subtyping. That combination is new, and the engineering is coherent. The experiments are reasonably careful: 20 stratified splits, multiple metrics, Welch's t-tests on all baselines, and the improvements on BRCA are substantial. On UCEC and KIPAN the margins are 1-3% and often not significant, which the authors mostly admit. The built-in feature and modality importance scores are a genuine plus.\n\nThe soft spot that matters is the graph construction. XGBoost uses the labels to select features and build the adjacency matrices. The paper never states that this step is re-run inside each training fold. If the graphs are built once on the full dataset, then the test labels influence both the selected feature sets and the edges used by the GNN and the GCN/GEDFN baselines. That would inflate the reported improvements. The stress-test note lands: this is not a manufactured flaw, it is a missing methodological detail in exactly the place where it matters. Until the authors state explicitly that graph construction is per-fold, or release code, the reported gaps are not fully trustworthy.\n\nMinor issues: no code or data release beyond \"available upon request,\" and the comparison omits several recent multi-omics GNNs that the paper itself cites, such as MOGAT, SUPREME and DeepMoIC. That weakens the \"state-of-the-art\" claim, but it is not a fatal omission.\n\nThe paper shows clear thinking and honest reporting of non-significant results, so I would send it to a serious referee rather than desk reject it. The referee should ask for the leakage clarification and ideally code. I would not cite the empirical numbers until that is answered. For a reading group, it is a good case study in how supervised preprocessing can blur the train/test boundary.\n\nRecommendation: send to peer review with a request to clarify the graph-construction protocol and share code. That is the difference between a solid incremental contribution and a possibly inflated one.","headline":"A competent but unverified combination of supervised tree-based graphs and tensor fusion; the label-leakage ambiguity around graph construction needs resolving before trusting the reported margins.","tokens_in":14055,"tokens_out":1432,"would_cite":false,"duration_ms":13168,"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":"TF-DWGNet claims that tree-ensemble split patterns, turned into directed weighted graphs and fused by low-rank tensor products, improve cancer subtype classification and yield built-in biomarker rankings.","keywords":["graph neural networks","multi-omics integration","directed weighted graphs","cancer subtype classification","tensor fusion","low-rank decomposition","model interpretability","tree-based feature selection"],"falsifier":"Inspect the split pipeline: if the tree-ensemble graphs are fit on all samples before the train-validation-test split, rerun the entire evaluation with graph construction and feature selection confined to each training fold. If accuracy and F1 gaps to the baselines collapse, the reported advantage is label leakage; if they persist, the central claim is supported.","tokens_in":13301,"feed_emoji":"🧬","tokens_out":8986,"duration_ms":80007,"temperature":0.7,"pith_summary":"TF-DWGNet claims that cancer subtype classification from multi-omics data improves when each omics modality is represented by a directed, weighted graph learned from supervised tree-ensemble split patterns, and when the three learned embeddings are fused through a low-rank tensor product that explicitly models single-modality, pairwise, and three-way interactions. Tested on breast, endometrial, and kidney cancer cohorts under 20 train-validation-test splits, the model is reported to beat all compared baselines on accuracy, weighted F1, and macro F1, with the largest margins on the five-class breast-cancer task. Because the graph edges and the fusion weights come from the trained model itself, the same modelling routine yields per-feature importance ranks and per-modality contribution scores, which the authors present as a built-in interpretability mechanism rather than a post hoc analysis. The underlying bet is that directionality and interaction strength—information that undirected similarity graphs and plain concatenation discard—are what carry the improvement.","feed_headline":"TF-DWGNet outperforms all baselines on three cancer datasets","feed_subtitle":"Tree-split directed edges and tensor fusion lift accuracy and reveal which features matter.","key_machinery":"The central object is the directed weighted graph built from tree-ensemble co-splits. Each tree is treated as a directed graph where an edge a→b means feature a appears before feature b on a splitting path, and edge weights count how often that directional co-split occurs across all trees; this graph provides both the topology and the connection strengths used by the graph-embedded layer. The other load-bearing piece is the tensor fusion module: augmenting each unimodal representation with a row of ones and taking the third-order outer product yields all unimodal, bimodal, and trimodal interaction terms, while CP decomposition approximates the tensor as a sum of R rank-one factors, reducing","core_discovery":"On its own terms, TF-DWGNet establishes that supervised graph construction and higher-order tensor fusion can be combined into a single architecture that is both predictive and interpretable. For each omics matrix, a tree-boosting ensemble is trained on the classification task; each tree is read as a directed graph in which an edge a→b records that feature a was split before feature b along a path, and the aggregate edge weight is the frequency of that directional co-split across all trees. The resulting sparse graph simultaneously selects features and encodes directional dependencies. A graph-embedded neural network per modality embeds the reduced feature matrix together with the normalized","pith_inferences":["If tree-split edge direction encodes regulatory order, the learned directed edges could be tested against known transcription-factor/target or methylation-expression relations; a positive match would turn the graph from a predictive device into a hypothesis generator.","The method could be applied to single-cell or spatial omics where regulatory direction is less well charted; edge frequencies might serve as a proxy for temporal or causal ordering.","Because the kidney-cancer gains are small, the method's advantage may be concentrated on harder, more imbalanced tasks; a deliberate stress test on a low-separation cohort would clarify where the margin actually comes from."],"forward_implications":["Multi-omics classifiers can learn task-specific graphs without external pathway databases; the same split-pattern construction can be reused on any new cancer cohort or omics combination.","Explicit unimodal, bimodal, and trimodal terms make it possible to see which omics pair or triple carries classification signal, not just which single omics is most informative.","Because feature and modality importance come from the trained connection weights, biomarker discovery and model interpretation do not require additional retraining or ablation runs.","The rank-R CP decomposition keeps higher-order interaction modeling tractable at thousands of input features, so the approach scales beyond the three-omics setting used here."],"fun_headline_variants":["Graph net with tensor fusion beats cancer subtype baselines","Directed weighted graphs plus tensor fusion improve cancer classification","TF-DWGNet: Tree-built edges and tensor fusion for cancer subtypes","New GNN fuses tensors and tree-based directed graphs for cancer"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise, which the paper does not explicitly state, is that the supervised tree-ensemble graph construction and feature selection are fitted on the training portion of each of the 20 splits; if the graphs are built on the full dataset, test labels could leak through the edges and inflate the reported improvements.","fun_headline_variants_meta":{"raw":{"variants":["Graph net with tensor fusion beats cancer subtype baselines","Directed weighted graphs plus tensor fusion improve cancer classification","TF-DWGNet: Tree-built edges and tensor fusion for cancer subtypes","New GNN fuses tensors and tree-based directed graphs for cancer"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000137,"raw_usage":{"total_tokens":994,"prompt_tokens":762,"completion_tokens":232,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":162}},"tokens_in":506,"tokens_out":232,"duration_ms":2929,"temperature":1.0,"reasoning_tokens":162,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T16:03:45.156470+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the split pipeline: if the tree-ensemble graphs are fit on all samples before the train-validation-test split, rerun the entire evaluation with graph construction and feature selection confined to each training fold. If accuracy and F1 gaps to the baselines collapse, the reported advantage is label leakage; if they persist, the central claim is supported.","supporting_citations":[],"review_version":1}