{"id":"21a0d4b5-85b9-41b7-aaeb-a931df6f75cb","arxiv_id":"2505.23334","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"X2Graph converts tabular genetic and clinical samples into knowledge-graph-structured inputs for graph neural networks, reporting strong cancer subtyping accuracy on small TCGA datasets.","lead":"The paper introduces X2Graph, a method that turns each row of a biological dataset into a graph whose connections come from external gene-interaction knowledge bases, then classifies cancer subtypes with a graph neural network. The authors report improved accuracy over tree-based and deep learning baselines on three TCGA breast cancer datasets, though the gains are small and lack statistical testing.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"X2Graph's gains may come from using a smaller feature set (the KB intersection), not from graph message passing, because all baselines are trained on all features. A same-subset baseline comparison is missing.","rationale":"Good-faith reading: The mechanism proposed is that KB-defined edges act as inductive bias for message passing, reducing overfitting on small N/D data. That is plausible. But the experimental comparison does not control for the fact that X2Graph sees a much smaller feature space. In high-dimensional omics with N ~ 977, a simple feature-selection heuristic (keep only genes in a curated pathway/disease KB) could yield most of the observed gain; if so, the novelty is not the graph conversion but the feature filter. The paper even claims 'This demonstrates that the strong inductive bias introduced by the KBs is the key factor' while comparing against an MLP on all features, which is a non-sequitur. The reader's concern about KB completeness is real but secondary: even a perfect KB would not validate the graph mechanism unless the feature subset is controlled. This is why my stress-test centers on the missing control rather than on noisy KBs. I agree with the reader that the paper should be conditional; the missing same-subset comparison is a concrete additional condition.","tokens_in":9900,"tokens_out":4704,"duration_ms":44501,"concrete_test":"Retrain XGBoost, MLP, and TabPFN using only the features in S for each KB (and the union of the three S for the fusion model) with the same preprocessing, oversampling, and 10-fold CV used in the paper. Compare each per-KB and fusion result against Tables II-IV. Additionally, run an edge-less MLP on the same node embeddings (feature-index embedding + value linear layer, no message passing) to test whether the graph edges themselves contribute. If baseline accuracy/AUC on S equals X2Graph within noise, the graph architecture is not responsible for the SOTA claim; if X2Graph still dominates, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1 (line 1) constructs S = S_D ∩ S_B, discarding every column not present in the KB. For CNV, the full table has 23,286 features while MoNDO has 4,152 nodes; for RNA, 20,530 features vs 4,152 MoNDO nodes. Thus X2Graph trains on roughly 80% fewer features than the baselines in Tables II and III. The paper attributes the improvement to the 'strong inductive bias introduced by the KBs' (Section V-C), but no experiment isolates feature subsetting from graph-based message passing. Table V ablates node pruning and ID indexing only on MoNDO; it never runs MLP or XGBoost on the same S. If a plain MLP or XGBoost on S matches X2Graph's per-KB results, the central mechanism claim collapses to KB-guided feature selection. The Clinical KB likewise selects 50 of 138+ engineered columns, and Clinical results are excluded from the interpretability comparison because 'all features have a clear correlation with cancer,' further suggesting the KB mainly reduces noise. This is a correctness risk to the attribution, not just to KB completeness.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes X2Graph, a method that converts each row of a biological tabular dataset into a graph whose nodes are (feature index, cell value) pairs and whose edges are defined by external knowledge bases (MoNDO, Reactome, BioGrid, and a manually constructed Clinical KB). Node representations are learned by a GNN, and predictions from KB-specific models are combined with a late-fusion layer. The method is evaluated on TCGA BRCA CNV, RNA, and Clinical data using 10-fold cross-validation, comparing accuracy, AUC, F1, and Cohen's kappa against MLP, TabPFN, XGBoost, LASSO, and Ridge. The authors report state-of-the-art results on most metrics, and also present a GNNExplainer-based feature-importance analysis plus ablations of node pruning and feature-index encoding.","tokens_in":10190,"tokens_out":5176,"duration_ms":56181,"significance":"If the performance attribution were established, X2Graph would be a useful contribution: it injects domain knowledge into tabular deep learning without requiring a large auxiliary knowledge graph, and the row-to-graph conversion with feature-index coding is simple, interpretable, and clearly described. The use of public TCGA data and three distinct gene-interaction KBs is a strength, as is the explicit ablation of ID indexing and node pruning. However, the central mechanistic claim — that the KB-defined edges, rather than the KB-induced feature subset, drive the gains — is not tested, and the statistical support for the headline results is weak. The paper gives no code, and the MLP baseline relies on an unpublished 'under review' reference. With a same-subset baseline comparison and variance/significance reporting, the claims could become solid; in its current form the evidence is suggestive rather than conclusive.","major_comments":[{"comment":"Algorithm 1 restricts the feature set to S = S_D ∩ S_B, so X2Graph trains only on the KB-intersection features. For CNV the full table has 23,286 columns while the MoNDO KB has 4,152 nodes, and for RNA the numbers are 20,530 versus 4,152; the baselines in Tables II and III are trained on the full feature set. The reported gains over XGBoost (e.g., +0.010 accuracy on CNV, +0.005 on RNA) could therefore be due entirely to KB-guided feature subsetting. Section V-C attributes the gains to 'the strong inductive bias introduced by the KBs', but Table V ablates only node pruning and ID indexing on the MoNDO graph and never evaluates an MLP, XGBoost, or LASSO on the same feature subset S. Please add a same-subset baseline comparison, and ideally a control with random edges or a feature-count-matched model, to separate feature selection from message passing.","section":"§IV-A (Algorithm 1), §V-C, Tables II-III"},{"comment":"No measure of variance is reported: Tables II-IV give only point estimates from 10-fold cross-validation. Many differences are small (CNV accuracy 0.7410 vs 0.7312 for XGBoost; RNA accuracy 0.9242 vs 0.9194), and on Clinical data XGBoost's F1 (0.6800) exceeds X2Graph's (0.6523), so the abstract's claim of 'superior performance' across the three datasets is not accurate. Without standard deviations, confidence intervals, or a paired significance test across the folds, the state-of-the-art claim is under-supported. Please report fold-level variability, run a paired test (e.g., Wilcoxon or corrected t-test), and soften the conclusions accordingly.","section":"§V-C, Tables II-IV"},{"comment":"The Clinical KB is manually constructed from literature that is also used to interpret the Clinical features, and §V-D states that Clinical interpretability results are omitted because 'all features have a clear correlation with cancer'. This is an acknowledged limitation, but it has consequences for the paper's third contribution: for Clinical data the 'external knowledge' is not independent of the outcome, so the edge-structure inductive bias cannot be credited with discovering relevant features. The interpretability conclusions should be scoped to CNV/RNA, and the manual Clinical KB should be described in enough detail (nodes, edges, and evidence per edge) to allow readers to assess potential label leakage into the graph definition.","section":"§III-B, §V-D, §VI"},{"comment":"The paper motivates X2Graph against other GNN-based tabular methods (IGGNet, TABGLM, Plato in Section II), but the experiments compare only against MLP, TabPFN, XGBoost, LASSO, and Ridge. The claim of state-of-the-art performance is therefore limited to non-graph baselines; to substantiate the advantage of the graph conversion, the authors should include a GNN-based tabular baseline (e.g., IGGNet or a GNN on a fully connected graph) in the comparison, or explicitly restrict the claim to the evaluated baseline families.","section":"§II and §V-C"}],"minor_comments":[{"comment":"The text says 'Electrical Heath Records'; this should be 'Electronic Health Records'.","section":"§III-A"},{"comment":"The function nodeFeat(v, s, **kwargs) is never defined; please specify how the node feature vector is assembled for each modality and what the allowed kwargs are.","section":"§IV-A, Algorithm 1"},{"comment":"The 'fraction of top genes with links to breast cancer' comparison does not state how many top genes were selected for each method or how the baseline attributions (split counts for XGBoost, integrated gradients for MLP) were normalized; please make this explicit.","section":"§V-D, Fig. 4(d)"},{"comment":"The hyperparameter search ('up to 50 random search iterations') is not accompanied by search ranges or the selected hyperparameters for the baselines; this information is needed to rule out undertuned baselines and to make the comparison reproducible.","section":"§V-B"},{"comment":"Reference [15] is an anonymous 'under review' work used both as the MLP baseline and as inspiration for the fusion method; since the reader cannot verify this baseline, please use a published baseline or provide a detailed implementation description, and consider moving the citation to a public version if one exists.","section":"References and §V-C"}],"recommendation":"major_revision","confidential_remarks":"The central empirical claim is interesting but the manuscript currently lacks the key control that would separate KB-guided feature selection from graph message passing. In addition, the use of an unpublished anonymous reference [15] as both baseline and fusion inspiration is a reproducibility concern that the editor may want the authors to address before review is complete. The manually constructed Clinical KB also deserves scrutiny for possible label leakage, since the selected features (e.g., ER/PR/HER2 IHC status) are known subtype markers; making the KB available would help. Overall, a major revision with the suggested ablations and statistical tests would bring the paper in line with the claims made in the abstract."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look, but the headline claim has a hole. X2Graph converts each row into a graph with edges from external KBs. That is genuinely different from IGGNet/TABGLM, which infer edges from the data, and unlike Plato it does not require a huge KB. The write-up is clear, the ablation of the gene-ID positional encoding is a real contribution, and the interpretability check against the literature is a nice verification step.\n\nThe soft spot is the missing control. Algorithm 1 starts with S = S_D ∩ S_B, so for CNV and RNA the model sees roughly 4,152 MoNDO genes instead of 20k+ columns, and all baselines are trained on the full column set. The paper credits the KB's inductive bias for the gains, but nothing rules out that simply dropping the unannotated columns is what helps. A plain MLP or XGBoost on the same S would settle it. The stress-test note is right, and this is a correctness risk to the attribution, not just a completeness worry. The Clinical KB also selects 50 of 138 columns, so the same issue applies there.\n\nThe quantitative issues are smaller but real: no error bars or significance tests, and the margins over XGBoost are thin (CNV +0.010 accuracy, RNA +0.005). On Clinical, XGBoost actually has the higher F1, so 'superior performance across three datasets' overstates the case. The unpublished, anonymous baseline [15] is used as the MLP comparator and as the source of the fusion idea, which adds opacity to the comparison.\n\nThe method is plausible, the experiments are extensive, and the direction is worth pursuing for medical ML and tabular GNN researchers. But the central mechanism claim is unproven until we see the same-feature-subset baselines. A serious referee should ask for that, plus confidence intervals and release of the handcrafted clinical KB and code. I would send it out rather than desk-reject: it is a solid workshop-to-conference submission with one clear empirical gap.","headline":"X2Graph is a genuinely new way to inject biological knowledge into tabular GNNs, but the paper never rules out that its gains come from the knowledge-base feature subset alone.","tokens_in":10697,"tokens_out":3333,"would_cite":false,"duration_ms":31234,"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":"Converting each table row into a graph whose edges come from external gene knowledge lets a graph neural network outperform tree-based and deep-learning baselines on three cancer subtyping datasets.","keywords":["tabular data","graph neural networks","inductive bias","knowledge base","cancer subtyping","TCGA BRCA","small-sample learning","biological tabular data"],"falsifier":"Train X2Graph on the clinical dataset with the knowledge-base edges randomly rewired while preserving node degrees; if the model still matches or surpasses the strongest tree-based baseline, the paper's inductive-bias explanation would be unsupported. Alternatively, remove a clinically documented edge, such as the link between the HER2 immunohistochemistry score and HER2 fluorescence in situ hybridization status, and check whether AUC drops toward the unconstrained deep-learning baseline.","tokens_in":9747,"feed_emoji":"🧬","tokens_out":8480,"duration_ms":73413,"temperature":0.7,"pith_summary":"X2Graph claims that converting each row of a small biological table into a graph, with edges defined by external knowledge bases such as gene interaction networks, lets a standard graph neural network beat tree-based and deep-learning baselines on cancer subtyping. The method is tested on three TCGA BRCA datasets (copy number variation, RNA expression, and clinical records) with hundreds of samples and tens of thousands of features. The reported gains come from the inductive bias that only known-related features talk to each other, reducing overfitting in the high-dimensional, low-sample regime.","feed_headline":"External gene networks give GNNs an edge on small cancer data","feed_subtitle":"X2Graph converts tabular rows into knowledge-based graphs, beating tree-based baselines on three cancer datasets.","key_machinery":"The central mechanism is the row-to-graph conversion: a node vector encodes the pair (feature index, cell value), which makes the graph view-invariant and lets the original row be reconstructed from the graph; an edge is present when the knowledge base's relation matrix links the two features. For copy-number-variation data, nodes with zero values are pruned because they indicate no copy-number change, further strengthening the inductive bias. For two of the datasets, predictions from three separate knowledge-base graphs are combined by a learned weighted sum.","core_discovery":"The paper proposes X2Graph, a graph-based deep learning method for biological tabular data. Each sample row is converted into a graph where nodes are the cell values paired with the feature index, and edges come from an external knowledge base that connects related features. Message passing is then applied with standard graph-neural-network layers. On the three TCGA BRCA datasets, X2Graph achieves the highest accuracy, macro AUC, and Cohen's kappa among all baselines, with F1 second to the strongest tree-based model on clinical data. The paper attributes the gains to the inductive bias supplied by the knowledge base.","pith_inferences":["The same row-to-graph recipe may transfer to any tabular domain, medical or not, where a reliable graph of column relationships exists; a natural next experiment is to apply it to financial or sensor data with known dependencies.","The paper's claim predicts a specific degradation pattern: randomly rewiring knowledge-base edges while preserving node degree should erode performance back toward the unconstrained deep-learning baseline.","The manually built clinical knowledge base (about fifty nodes and two hundred edges) suggests that even small hand-curated relation graphs can steer deep learning, which could make the approach usable in domains lacking large public knowledge bases."],"forward_implications":["On the copy-number-variation and RNA datasets, the fused X2Graph model outperforms the strongest tree-based baseline on all four reported metrics.","The ablation study shows that including the feature index in node vectors is essential, improving AUC by at least 15% regardless of node pruning.","Late fusion of three knowledge-base graphs outperforms an early-fusion multi-edge version by about one point of AUC on the copy-number-variation data.","The interpretability analysis identifies top contributing features, and a larger fraction of these features have literature evidence for breast-cancer linkage than for the tree-based and deep-learning baselines."],"supporting_citations":[{"why":"Documents the dominance of tree-based methods on tabular data, motivating the baseline comparison.","marker":"[2]"},{"why":"The strongest tree-based baseline that X2Graph must beat.","marker":"[12]"},{"why":"The deep-learning baseline and the source of the late-fusion weighting idea.","marker":"[15]"},{"why":"A transformer-based baseline designed for small tabular classification problems.","marker":"[16]"},{"why":"A taxonomy of graph-based tabular learning methods that positions X2Graph's approach.","marker":"[23]"},{"why":"The closest prior work using a knowledge base for edges, which X2Graph distinguishes by not requiring a KB larger than the dataset.","marker":"[26]"},{"why":"Supplies the methodology for constructing the gene knowledge bases from public interaction databases.","marker":"[30]"},{"why":"The interpretability method used to identify important nodes and map them back to table columns.","marker":"[33]"}],"fun_headline_variants":["Small cancer data? X2Graph uses gene graphs to beat baselines","Gene graphs from X2Graph win on tiny cancer datasets","X2Graph maps tabular samples to gene graphs for sharper subtyping","Turning rows into gene graphs, X2Graph excels on small cancer data","X2Graph leverages gene interaction graphs for cancer subtyping on tiny data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's gain rests on the assumption that the external knowledge bases already contain the most crucial relations between features for the prediction task; if a key relation is missing or an edge is spurious, the graph's inductive bias can mislead the model.","fun_headline_variants_meta":{"raw":{"variants":["Small cancer data? X2Graph uses gene graphs to beat baselines","Gene graphs from X2Graph win on tiny cancer datasets","X2Graph maps tabular samples to gene graphs for sharper subtyping","Turning rows into gene graphs, X2Graph excels on small cancer data","X2Graph leverages gene interaction graphs for cancer subtyping on tiny data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001593,"raw_usage":{"total_tokens":6260,"prompt_tokens":766,"completion_tokens":5494,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":382,"completion_tokens_details":{"reasoning_tokens":5399}},"tokens_in":382,"tokens_out":5494,"duration_ms":42186,"temperature":1.0,"reasoning_tokens":5399,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:46:57.462220+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train X2Graph on the clinical dataset with the knowledge-base edges randomly rewired while preserving node degrees; if the model still matches or surpasses the strongest tree-based baseline, the paper's inductive-bias explanation would be unsupported. Alternatively, remove a clinically documented edge, such as the link between the HER2 immunohistochemistry score and HER2 fluorescence in situ hybridization status, and check whether AUC drops toward the unconstrained deep-learning baseline.","supporting_citations":[{"cited_title":"Graph neural networks for tabular data learning,","cited_arxiv_id":null,"evidence_quote":"A taxonomy of graph-based tabular learning methods that positions X2Graph's approach."},{"cited_title":"Why do tree-based models still outperform deep learning on typical tabular data?","cited_arxiv_id":null,"evidence_quote":"Documents the dominance of tree-based methods on tabular data, motivating the baseline comparison."},{"cited_title":"Xgboost: A scalable tree boosting system,","cited_arxiv_id":null,"evidence_quote":"The strongest tree-based baseline that X2Graph must beat."},{"cited_title":"Scalable and loosely-coupled multimodal deep learn- ing for breast cancer subtyping,","cited_arxiv_id":null,"evidence_quote":"The deep-learning baseline and the source of the late-fusion weighting idea."},{"cited_title":"High dimensional, tabular deep learning with an auxiliary knowledge graph,","cited_arxiv_id":null,"evidence_quote":"The closest prior work using a knowledge base for edges, which X2Graph distinguishes by not requiring a KB larger than the dataset."},{"cited_title":"The multilayer community structure of medulloblastoma,","cited_arxiv_id":null,"evidence_quote":"Supplies the methodology for constructing the gene knowledge bases from public interaction databases."},{"cited_title":"Gnnexplainer: Generating explanations for graph neural networks,","cited_arxiv_id":null,"evidence_quote":"The interpretability method used to identify important nodes and map them back to table columns."}],"review_version":1}