{"id":"82e452cc-8379-4092-a7b6-edbc871aa0f8","arxiv_id":"2608.13256","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"MK-TGAN, a multi-kernel graph neural network-based GAN, produces synthetic transcriptomic data with higher realism and downstream utility than existing generative models on an EMT gene panel.","lead":"This paper compares several AI methods for creating fake gene expression data and proposes a new one, MK-TGAN, that organizes genes as a graph. The new method generates synthetic breast cancer data that look and behave more like real data than earlier methods, which could help researchers augment and share medical data without exposing patient information.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never isolates the biological content of the prior: because all graph-informed models use the same GTEx correlation matrix, the reported gains could come from graph architecture/regularization rather than from knowledge guidance.","rationale":"The reader identifies the healthy-tissue-versus-tumor relevance of the prior as the weakest assumption. The concern here is related but more fundamental: the paper never demonstrates that the biological content of A, rather than the graph architecture itself, is responsible for the improvements. A shuffled or randomized prior is the standard control that would settle this. This does not contradict the reader's conditional verdict; it reinforces it. The paper is internally consistent and reports reproducible-looking metrics, but the central attribution to biological knowledge guidance is under-supported without this ablation. The recommendation remains CONDITIONAL acceptance, with the random-prior control as an explicit condition.","tokens_in":9313,"tokens_out":4747,"duration_ms":51872,"concrete_test":"Retrain MK-TGAN, and ideally GR-GAN, with A replaced by two negative-control priors: (i) a random symmetric matrix obtained by permuting the off-diagonal entries of A, preserving the empirical distribution but destroying gene-pair semantics; and (ii) a shuffled-gene version of A where the same correlation values are assigned to randomly permuted gene pairs. Keep all hyperparameters, training data, and evaluation metrics identical to the main experiments. If the random/shuffled priors reproduce the Table 1 correlation score (~0.914), recall (~0.784), and TSTR LR/MLP accuracies within one standard deviation, then the biological content of the prior is not load-bearing and the 'knowledge-guided' claim should be weakened. If the random-prior variants fall significantly toward the WGAN-GP or BioGAN levels, the biological prior matters.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Section 5) is that integrating prior biological knowledge improves realism and utility, and that MK-TGAN's multi-kernel GNN is the reason. This assumes the specific gene-gene correlations stored in A matter. However, every graph-informed model uses the same GTEx-derived Pearson matrix A (Section 2.5), and the only comparisons are against models without a graph or with a single-stream GNN. There is no ablation in which A is replaced by a randomized or shuffled correlation matrix while keeping architecture, training, and evaluation fixed. In the message-passing update Eq. (3), A serves as a fixed weighted adjacency matrix: each gene's embedding is updated by a weighted average of all other genes, which strongly changes the inductive bias, couples the generator's outputs, and drastically reduces the effective parameter count (<19k vs >100k). The observed gains in Table 1 could therefore be produced by this architectural regularization rather than by the biological content of the prior. This concern is underscored by the paper's own related-work discussion, which notes that Lacan et al. found performance gains driven largely by architectural complexity rather than biological knowledge (Section 1.1); yet no analogous control is run here. The issue is compounded by the correlation metric in Eq. (7), which scores agreement between pairwise gene-gene correlation matrices: because A is itself a gene-gene correlation matrix injected into the generator, a high correlation score is partly aligned with the injected prior and cannot independently certify biological plausibility. A random-prior control is the minimal experiment needed to attribute the improvement to knowledge guidance rather than to the graph architecture itself.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes and benchmarks three knowledge-guided GAN variants for synthetic transcriptomic data: GM-GAN, GR-GAN, and MK-TGAN, all conditioning on a GTEx-derived Pearson correlation matrix A of 197 EMT genes. Using TCGA breast cancer data, the authors compare these with a GAN, a WGAN-GP, and a BioGAN-inspired graph GAN, reporting precision/recall, a gene-pair correlation coefficient, detectability, and TSTR classification utility. The central claim is that integrating biological prior knowledge improves realism and utility, and that MK-TGAN—a multi-kernel GNN generator—achieves the best overall performance (Recall10=0.784, correlation=0.914, LR accuracy=0.805).","tokens_in":9569,"tokens_out":4430,"duration_ms":45740,"significance":"The problem is timely and the design space (implicit modulation, regularization, explicit GNN) is sensibly surveyed. If the claims were fully supported, MK-TGAN would be a useful addition to synthetic transcriptomic tools, and the parameter-efficiency observation is interesting. The manuscript uses public datasets, reports standard deviations over ten runs, and documents hyperparameters clearly, which aids reproducibility. However, the evaluation currently conflates the biological prior with the graph architecture, and the primary correlation metric is structurally aligned with the injected prior; the empirical evidence needs additional controls and significance testing before the main claims can be accepted.","major_comments":[{"comment":"The performance claim lacks statistical significance testing. Many pairwise differences are within one or two standard deviations across the ten runs; for example, Precision10 for GM-GAN (0.967 ± 0.008) and MK-TGAN (0.960 ± 0.015) overlap, and Recall10 for GR-GAN (0.707 ± 0.027) and MK-TGAN (0.784 ± 0.020) overlap under a 2-SD convention. Provide paired significance tests (e.g., Wilcoxon or bootstrap over the ten sampling runs, or per-subtype comparisons) and report effect sizes for all key metrics.","section":"Section 4, Table 1"},{"comment":"The correlation coefficient defined in Eq. (7) measures agreement between pairwise gene-gene correlation matrices. Since the graph-informed models inject A—itself a gene-gene correlation matrix—into the generator, this metric is aligned with the injected prior. The reported correlation advantage (0.914 for MK-TGAN vs. 0.696 for WGAN-GP) may therefore partly or wholly reflect the evaluation rewarding the prior rather than improved biological realism. The paper should report a control trained on a permuted or shuffled A, and should include metrics not aligned with A, such as marginal distribution distances, subtype separation, or pathway-level enrichment.","section":"Section 3.1, Eq. (7); Section 2.5"},{"comment":"There is no ablation that isolates the biological content of the prior from the architectural inductive bias. All graph-informed models use the same GTEx-derived A, so the comparison against non-graph or single-stream baselines cannot distinguish \"knowledge guidance\" from the regularizing effect of message-passing over a dense weighted graph, which also reduces the parameter count from over 100k to under 19k. Given that Section 1.1 cites Lacan et al. finding gains largely driven by architectural complexity, a control with a randomized or null graph is essential. Without it, the central claim that prior biological knowledge, rather than graph architecture, drives the improvement is not established.","section":"Sections 2.4–2.5 and 4"}],"minor_comments":[{"comment":"The header \"WPGAN\" is a typo and should read \"WGAN-GP\".","section":"Table 1"},{"comment":"The bias term b^I is declared in R^{len(v)} but the equation h0 = ReLU(W^I v + b^I) requires a bias in R^{|genes|}; this is likely a typo and should be corrected.","section":"Section 2.4, Eq. (2)"},{"comment":"The word \"computated\" should be \"computed\".","section":"Section 2.2"},{"comment":"Reference [8] is cited for the TCGA breast cancer dataset, but the title given corresponds to a glioblastoma paper; the reference does not support the dataset description and should be replaced with the correct TCGA BRCA citation.","section":"References"},{"comment":"The text states that GR-GAN achieves intermediate detectability, but Table 1 shows MLP accuracy = 1.000 for GR-GAN; the \"intermediate\" statement refers only to the LR accuracy and should be clarified.","section":"Section 4, Table 1"},{"comment":"The paper should include a data and code availability statement, in particular for the exact TCGA/GTEx preprocessing pipeline and the constructed GTEx correlation matrix, to support reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a clear structure and the architectural proposal is reasonable, but the main empirical claim is currently underdetermined by the experiments. The missing ablation with a shuffled prior and the absence of significance testing are load-bearing and should be addressed before the paper is considered further. I also noticed that reference [8] appears to be mis-cited; please verify all references against their intended sources."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nMK-TGAN is a multi-kernel GNN generator for synthetic transcriptomics, and the comparison against a single-stream GNN baseline (BioGAN) makes the delta explicit. The paper is clearly written, the experimental setup is shared across models, and the evaluation covers precision/recall, TSTR utility, and detectability. Credit where due: the multi-kernel idea is a reasonable architectural variation, and the results show consistent gains over the implicit-knowledge variants (GM-GAN, GR-GAN) and the single-stream BioGAN.\n\nThe soft spots are real but not fatal. The biggest one is that the paper never isolates the biological content of the prior. Every graph-informed model uses the same GTEx correlation matrix as the adjacency, and the only comparisons are against non-graph or single-stream models. As the authors themselves note from Lacan et al., performance gains in this area are often architectural rather than knowledge-driven; yet they don't run the minimal control of replacing the prior with a random or shuffled matrix while keeping architecture and training fixed. Without that, the claim that \"prior knowledge integration strategies improve performance\" is not fully supported — what the experiments show is that graph-based generators help, which is not the same thing. The correlation metric makes this worse, since it measures agreement between pairwise correlation matrices, and the injected prior is exactly that.\n\nA few smaller issues: no significance tests, and several comparisons have overlapping standard deviations (e.g., BioGAN correlation 0.801±0.015 vs GM-GAN 0.813±0.023). The abstract overclaims \"Unlike other methods, MK-TGAN leverages prior knowledge graphs by exploiting graph neural networks,\" which is directly contradicted by BioGAN. There's also a minor dimension mismatch in Eq. 2: the bias b^I should be in R^{|genes|}, not R^{len(v)}. No code is released, and the evaluation is limited to 197 EMT genes in one cancer type — the authors acknowledge this.\n\nOn balance: the empirical comparison is useful and the architecture is worth reporting, but the central interpretive claim needs the random-prior ablation before the \"knowledge-guided\" framing is justified. I'd accept it for peer review and ask for that control plus significance tests and code release.\n\nBest.","headline":"MK-TGAN is a legitimate architectural contribution, but the paper's 'knowledge guidance' claim is unproven without a random-prior ablation.","tokens_in":10177,"tokens_out":2962,"would_cite":false,"duration_ms":28808,"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":"This paper claims that explicitly injecting gene co-expression structure through multiple parallel graph neural network kernels in a GAN generator produces synthetic transcriptomic data that is more realistic and more useful for…","keywords":["synthetic data generation","generative adversarial networks","graph neural networks","transcriptomics","biological knowledge graphs","breast cancer subtypes","data augmentation"],"falsifier":"Train MK-TGAN with the identical multi-kernel architecture but with gene identities in the prior correlation matrix randomly permuted; if recall, correlation, and TSTR accuracy fall to baseline levels, the reported gains stem from prior content rather than from the multi-kernel mechanism.","tokens_in":9084,"feed_emoji":"🧬","tokens_out":4743,"duration_ms":43333,"temperature":0.7,"pith_summary":"The paper argues that synthetic gene-expression data become more realistic and more useful for downstream tasks when prior biological knowledge, specifically the pairwise co-expression structure of a set of genes, is injected directly into the generator of a generative adversarial network. It introduces three knowledge-guided GAN variants and shows that the most explicit form, MK-TGAN, which processes a fixed gene graph through several parallel graph-neural-network kernels, outperforms unconditional GANs and a single-stream graph-based GAN on a breast-cancer EMT gene panel. The result matters because synthetic transcriptomic data could mitigate data scarcity, imbalance, and privacy constraints in biomedical machine learning if generated samples faithfully preserve gene-gene relationships.","feed_headline":"Multi-kernel GAN tops synthetic gene expression realism","feed_subtitle":"Gene co-expression prior improves realism and breast cancer subtype prediction.","key_machinery":"The central object is a generator made of K = 4 parallel graph convolutional kernels, each taking gene embeddings as node features and propagating them over a dense, signed gene-gene correlation matrix derived from healthy breast tissue. Each kernel produces a per-gene embedding, and a learnable weight matrix W_F fuses the kernels' outputs into the final synthetic expression profile. This multi-stream design is what the paper credits with increasing representational capacity while remaining parameter-efficient (fewer than 19,000 learnable parameters).","core_discovery":"On 197 EMT-related genes from TCGA breast cancer samples, MK-TGAN achieves the best recall (Recall10 = 0.784), the highest correlation of gene-gene relationships (0.914), very high precision (0.960), the lowest detectability (MLP accuracy = 0.760), and the best train-on-synthetic/test-on-real classification accuracy (LR accuracy = 0.805, MLP accuracy = 0.791). The paper interprets this as evidence that a multi-kernel GNN generator, by capturing complementary relational patterns in parallel, balances the precision-recall trade-off better than single-stream graph methods, which tend to reproduce a limited subset of the distribution.","pith_inferences":["Because the evaluation's correlation metric measures the same gene-gene relationships that were injected as the prior, part of MK-TGAN's advantage may reflect alignment between prior and metric; an independent biological validation would be a stronger test.","If the advantage persists with a tumor-matched or tissue-specific prior, the multi-kernel design itself would be the likely cause; if not, prior content is the driver, not architecture.","The same multi-kernel injection strategy could apply to other omics data, such as methylation or proteomics, wherever a reliable feature-relationship graph can be constructed."],"forward_implications":["Incorporating prior biological knowledge through graph-structured generators improves both realism and downstream utility of synthetic transcriptomic data compared with unconditional GANs.","A multi-kernel formulation (MK-TGAN) outperforms a single-stream graph generator (BioGAN-style) on recall and detectability, suggesting that multiple parallel message-passing pathways preserve more of the data distribution.","Synthetic samples generated by MK-TGAN support breast cancer subtype classifiers at roughly 80 percent accuracy, indicating practical potential for data augmentation in imbalanced settings.","The architecture's small parameter count suggests it can scale to larger gene panels at modest computational cost."],"supporting_citations":[{"why":"BioGAN, the single-stream graph-based GAN that MK-TGAN is compared against and whose architecture it extends with multiple kernels.","marker":"[10]"},{"why":"Viñas et al. established WGAN-GP as a baseline for transcriptomic generation, which all models in this paper build on.","marker":"[13]"},{"why":"Lacan et al. introduced the correlation-based similarity coefficient used to evaluate gene-gene relationship preservation.","marker":"[6]"},{"why":"Kynkäänniemi et al. provided the precision and recall metrics used to assess fidelity and diversity of generated samples.","marker":"[5]"},{"why":"GTEx is the source of healthy breast tissue samples used to compute the Pearson correlation prior graph.","marker":"[3]"},{"why":"Gulrajani et al. supplied the gradient-penalty WGAN training framework used by all models.","marker":"[4]"}],"fun_headline_variants":["Graph-guided GAN beats rivals at synthetic gene data","MK-TGAN: GNN-powered GAN for realistic transcriptomes","Gene graph priors boost synthetic transcriptomic realism","Multi-kernel GNN GAN wins on gene expression synthesis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the gene co-expression structure measured in healthy breast tissue is a valid and sufficient biological prior for generating tumor transcriptomes, and the paper never independently verifies that assumption.","fun_headline_variants_meta":{"raw":{"variants":["Graph-guided GAN beats rivals at synthetic gene data","MK-TGAN: GNN-powered GAN for realistic transcriptomes","Gene graph priors boost synthetic transcriptomic realism","Multi-kernel GNN GAN wins on gene expression synthesis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1329,"prompt_tokens":832,"completion_tokens":497,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":448,"completion_tokens_details":{"reasoning_tokens":430}},"tokens_in":448,"tokens_out":497,"duration_ms":5824,"temperature":1.0,"reasoning_tokens":430,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:15:31.725019+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train MK-TGAN with the identical multi-kernel architecture but with gene identities in the prior correlation matrix randomly permuted; if recall, correlation, and TSTR accuracy fall to baseline levels, the reported gains stem from prior content rather than from the multi-kernel mechanism.","supporting_citations":[{"cited_title":"Bioengineering12(6), 658 (2025)","cited_arxiv_id":null,"evidence_quote":"BioGAN, the single-stream graph-based GAN that MK-TGAN is compared against and whose architecture it extends with multiple kernels."},{"cited_title":"Bioinformatics38(3), 730–737 (2022)","cited_arxiv_id":null,"evidence_quote":"Viñas et al. established WGAN-GP as a baseline for transcriptomic generation, which all models in this paper build on."},{"cited_title":"Bioinformatics39(Supplement_1), i111–i120 (2023)","cited_arxiv_id":null,"evidence_quote":"Lacan et al. introduced the correlation-based similarity coefficient used to evaluate gene-gene relationship preservation."},{"cited_title":"In: Advances in Neural Information Processing Systems","cited_arxiv_id":null,"evidence_quote":"Kynkäänniemi et al. provided the precision and recall metrics used to assess fidelity and diversity of generated samples."},{"cited_title":"Nature583(7818), 452–460 (2020)","cited_arxiv_id":null,"evidence_quote":"GTEx is the source of healthy breast tissue samples used to compute the Pearson correlation prior graph."},{"cited_title":"Advances in neural information processing systems30 (2017) 14 F.P","cited_arxiv_id":null,"evidence_quote":"Gulrajani et al. supplied the gradient-penalty WGAN training framework used by all models."}],"review_version":1}