{"id":"4c9e30d9-7353-4ad0-a548-421925b335ea","arxiv_id":"2601.01337","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":11,"one_line_summary":"HyperNetWalk unifies personalized and cohort-level driver-gene identification by combining random-walk diffusion over PPI, regulatory, and dynamic co-expression networks with cross-patient hypergraph refinement, achieving competitive CGC Tier 1 benchmark performance across 12 TCGA cancer types.","lead":"This paper presents HyperNetWalk, an unsupervised method that ranks mutated genes as likely cancer drivers for individual patients and for whole patient cohorts at once, by diffusing scores across protein-interaction, gene-regulation, and gene co-expression networks. On 12 TCGA cancer types it reports competitive or superior performance against eight existing driver-gene tools, with the clearest advantage on cancer-type-specific drivers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No control for hub bias: CGC Tier 1 drivers are hubs, and Stage I RWR rewards hubs, so the claimed reverse-inference mechanism may not drive reported performance.","rationale":"The reader's weakest assumption identifies the same issue: the gold standard (CGC Tier 1) and the method's inductive bias overlap, because known drivers are hubs and Stage I explicitly rewards topological prominence. I agree this is the most load-bearing concern. The central claim includes both a performance claim and a mechanism claim ('reverse inference on layered signaling-regulatory network'). The performance claim may survive, but the mechanism claim is untested without a control that removes hub-ness. A degree-only baseline is a simple, decisive addition. I also note the abstract asserts ablation analyses that do not appear in the full text or listed supplementary; this is a separate missing-support issue that reinforces the conditional status. However, the hub-bias concern is the deepest because it affects interpretation of all benchmark results, not just the completeness of the manuscript. The verdict should remain CONDITIONAL: acceptance should require demonstration that the method outperforms degree-based controls, or a substantial softening of the mechanistic claims.","tokens_in":20659,"tokens_out":6988,"duration_ms":81649,"concrete_test":"Re-run the evaluation with a 'Degree' baseline: for each patient, rank mutated genes by their total degree in the induced PPI/GRN subnetwork (or an equally weighted combined degree), then apply the same personalized and cohort-level REA/Top-200 evaluation as in §3.5. Compare AUPRC/pAUPRC in Table S2. If Degree matches or exceeds HyperNetWalk, the reverse-inference mechanism is not load-bearing; if HyperNetWalk clearly outperforms Degree, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is that the evaluation design does not separate HyperNetWalk's claimed reverse-inference mechanism from the known topological property that CGC Tier 1 drivers are network hubs. §2.2 (Fig. S1a,b) shows drivers have higher PPI degree and GRN out-degree than non-drivers. Stage I's random walk on the induced PPI/GRN subnetworks (Eqs. 1–2) rewards high-degree nodes, and the final Stage I score is dominated by the static networks (γ = 0.8 in Eq. 16). Consequently, the reported Precision@K/AUPRC/pAUPRC values in Table S2 may largely reflect hub detection rather than the paper's claimed mechanism of tracing dysregulation back to upstream drivers. The paper includes no degree-only or degree-randomization control. If a degree-based baseline matches HyperNetWalk, then the architecture's contribution (reverse inference, dynamic co-expression, hypergraph refinement) is not established, and the central mechanistic claim would be unsupported even though the ranking numbers may be reproducible.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HyperNetWalk, an unsupervised two-stage framework for cancer driver gene identification that combines static (PPI, GRN) and dynamic (co-expression) networks with hypergraph diffusion. In Stage I, sample-specific RWR on patient-specific subnetworks yields per-gene scores; in Stage II, hypergraph RWR across similar samples refines these scores for both personalized and cohort-level predictions. The method is benchmarked on 12 TCGA cancer types against eight existing methods, with claims of personalized performance comparable to DawnRank and cohort-level performance second only to DriverMP, plus an IntOGen-based analysis of cancer-type specificity.","tokens_in":20816,"tokens_out":8201,"duration_ms":74513,"significance":"If the reported performance is robust and the mechanism is genuinely 'reverse inference' rather than a proxy for node degree, the paper offers a useful unified framework for both personalized and population-level driver prioritization. The strengths are the wide benchmark (12 cancer types, 8 baselines, multiple metric families), the fully unsupervised architecture, and the explicit integration of multiple network types. However, the absence of a degree-controlled baseline, the lack of statistical testing across datasets, and the unverified ablation claims mean that the central mechanistic claim is not yet established. The paper is likely of interest to the computational cancer genomics community, but the current evidence does not fully separate the method's novelty from known hub properties of driver genes.","major_comments":[{"comment":"The central concern is hub-bias. The paper's own Fig. S1(a,b) shows that known CGC Tier 1 drivers have significantly higher PPI degree and GRN out-degree than non-drivers. Stage I's RWR (Eqs. 1–2) uses row-normalized transition matrices on the induced PPI/GRN subnetworks, which rewards high-degree nodes, and the final Stage I score is dominated by the static networks because γ = 0.8 in Eq. (16). No degree-only baseline (e.g., ranking by PPI degree) or degree-preserving network null is included in the benchmark. Consequently, the reported Precision@K and pAUPRC values may largely reflect hub detection rather than the claimed reverse-inference mechanism. Please add a control where only topological prominence is used, and/or an ablation that removes the static network component, to demonstrate that the co-expression and hypergraph components contribute beyond degree.","section":"§2.2, §3.3.1, Eq. (16)"},{"comment":"Table S2 provides only point estimates, and the text makes claims such as 'consistently outperforms' and 'significantly higher' without any error bars, confidence intervals, or between-method statistical tests. In fact, reading the table, DawnRank has higher AUROC in 10 of 12 cancer types and higher pAUROC in 7 of 12, while DriverMP has higher pAUPRC in most cancer types. The hedged phrasing in the results is appropriate, but the abstract's 'outperformed representative existing methods' is not supported. Please perform paired significance tests across the 12 cancer types (e.g., Wilcoxon signed-rank test) and report which differences are statistically significant.","section":"§2.3, Table S2"},{"comment":"The manuscript states 'Ablation analyses supported the contributions of the reverse inference formulation and layered network architecture,' but no ablation results are presented in the main text or in the listed supplementary figures/tables (Figs. S1–S7, Tables S1–S2). Without these results, the contribution of each component (PPI, GRN, co-expression, hypergraph refinement, ME prior) cannot be assessed, and the claim that the architecture matters is unverified. Please include ablation tables or remove the claim.","section":"§4 (Discussion), data availability"},{"comment":"The method has many free parameters (λ, θ, γ, β, τ, top-500 DEGs, σ, α, the ME cutoff |V_m|≥200, and the personalized horizon N_c), all set to fixed values without sensitivity analysis. Given that the performance differences between HyperNetWalk and the best baselines are often small (a few percent in pAUPRC/pAUROC), the conclusions may depend on these choices. Please provide a sensitivity analysis over the most influential parameters (e.g., λ, γ, τ, σ, α) or otherwise justify the fixed values.","section":"§3.3–§3.5"}],"minor_comments":[{"comment":"The abstract claims HyperNetWalk 'outperformed representative existing methods,' while the results state that it is 'comparable to DawnRank' and 'second-best behind DriverMP.' Please harmonize the wording.","section":"Abstract vs. §2.3"},{"comment":"There is a typo: 'the 10 will not update significantly' should be 'the equation will not update significantly.' Also, the convergence criterion for the RWR iterations is not specified.","section":"Eq. (10)"},{"comment":"The transition formula does not define the convention for nodes with zero degree in one network; a 0/0 term may arise. Please clarify that such terms are treated as 0.","section":"Eq. (1)"},{"comment":"The neighbor set requires sharing at least two mutated genes with the target sample. Many samples, especially those with low mutation burden, may have no neighbors, effectively reducing the personalized hypergraph to a single hyperedge. Please state how such cases are handled.","section":"§3.4.1, Eq. (17)"},{"comment":"The 'dynamic co-expression network' is not sample-specific; the adjacency matrix is fixed across all samples of a cancer type, and only the DEG set varies per patient. The phrase 'sample-independent random walks on patient-specific subnetworks' is therefore slightly misleading. Please clarify this point in the text.","section":"§3.3.2, Eq. (11)"}],"recommendation":"major_revision","confidential_remarks":"The hub-bias concern is the most important issue. The authors already show that known drivers are network hubs, which is the very confound that needs to be controlled for. A degree-only baseline or a degree-preserving null is essential to support the 'reverse inference' claim. The statistical testing and ablation issues are also substantial but fixable. The paper is within the scope of the journal, but the current framing overstates the evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a genuine methods contribution. HyperNetWalk is the first framework I know of that returns both personalized and cohort-level driver rankings from one unsupervised two-stage architecture, and the benchmark is broad — 12 TCGA types, 8 baselines, five metric families. The RWR and hypergraph-RWR math is standard and correctly stated. The full text is also honest about where it stands: comparable to DawnRank at the individual level, second to DriverMP at the cohort level. That is a credible result, not an inflated one.\n\nWhat is actually new is the two-stage pipeline: sample-specific subnetworks from PPI + GRN + dynamic co-expression, then a hypergraph random walk that shares information across similar patients. The individual pieces exist in the cited literature, but the unification and the cohort-level hypergraph refinement for rare drivers are a real step. The code and data statement is a plus, though the repo needs a commit pin.\n\nThe soft spots, in order. First, the two abstracts disagree. The arXiv abstract claims outperformance at both prediction levels and refers to ablation analyses; the full text's abstract says 'superior or competitive,' and I could not find any ablation analysis in the text or the listed supplement. That mismatch needs fixing before anything else. Second, the hub-bias control is missing. The paper itself shows known drivers are PPI hubs and GRN high-out-degree nodes, and Stage I's static random walk, weighted at γ=0.8, rewards exactly that topology. A degree-only baseline or a network-label randomization control would tell us whether the reported precision is really coming from the reverse-inference mechanism or just from the well-known fact that drivers are hubs. I don't think this sinks the paper — the unified framework stands on its own — but the mechanistic claim is currently under-supported. Third, the evaluation reporting is thin: Table S2 is point estimates with no error bars or significance tests, and the personalized evaluation only keeps samples with at least three CGC drivers; the authors should disclose the sample-size reduction and add sensitivity analysis for the hand-set hyperparameters.\n\nBottom line: this paper deserves a serious referee. It is a solid, well-benchmarked methods contribution with a genuine unification, and the equations are sound. But it needs a revision that reconciles the abstracts and adds a degree control. I would send it to review, and I would wait for the revised version before citing it.","headline":"Genuinely unified two-level driver-gene framework with honest benchmarking, but the reverse-inference claim lacks a hub-bias control and the two abstracts disagree.","tokens_in":21509,"tokens_out":4155,"would_cite":false,"duration_ms":43338,"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":"HyperNetWalk claims that one unsupervised two-stage random-walk pipeline—reverse-tracing transcriptional dysregulation through protein interaction and gene-regulatory networks, then sharing signal across patients via hypergraph walks—ranks","keywords":["cancer driver genes","random walk with restart","hypergraph diffusion","protein-protein interaction network","gene regulatory network","personalized medicine","cohort-level prediction","TCGA"],"falsifier":"Rerun HyperNetWalk on the same 12 cohorts with the protein-protein interaction and gene-regulatory subnetworks replaced by degree-preserving random networks, keeping the expression-based restart priors unchanged. If Precision@K against the known-driver list stays nearly the same, the claimed reverse-inference mechanism is not what carries the ranking. As a second check, if ranking mutated genes by network degree and regulatory out-degree alone matches the top-200 recall, then hub bias explains most of the result.","tokens_in":20365,"feed_emoji":"🧬","tokens_out":17961,"duration_ms":134447,"temperature":0.7,"pith_summary":"HyperNetWalk proposes to unify two tasks that are usually separate: finding driver genes for a single patient and finding driver genes for a whole cancer cohort. The central idea is to treat driver identification as reverse inference: instead of asking which mutated genes are most frequent, walk backward from the observed transcriptional disruption of a patient to the mutated genes that could plausibly cause it, using a layered network of protein-protein interactions (upstream signaling), transcription-factor-to-target regulation (downstream control), and tumor/normal co-expression. A first stage performs patient-specific random walks with restart on these induced subnetworks; a second stage builds hypergraphs whose hyperedges are patients and runs random walks that let samples borrow strength from similar samples, producing both personalized rankings and cohort rankings from the same machinery. The paper evaluates this on 12 TCGA cancer types and reports that the method is competitive with DawnRank for personalized predictions, is second only to DriverMP for cohort-level predictions, and best captures cancer-type-specific drivers when checked against IntOGen reference sets. If these results hold, the contribution is a single unsupervised pipeline that gives both a clinician and a researcher a ranked list of candidate drivers without needing labeled training data.","feed_headline":"One architecture ranks cancer driver genes per patient and per cohort","feed_subtitle":"Unsupervised network walks match the best per-patient methods and place second for cohort-level drivers across 12 tumor types.","key_machinery":"The load-bearing object is a layered signaling-regulatory network: an undirected PPI network approximating upstream signaling connected to a directed gene regulatory network (transcription factors as interface nodes) approximating downstream transcriptional control, with a dynamic co-expression network layered on top. On this layered representation the paper runs random walks with restart (RWR), a diffusion process in which a walker moves along network edges and periodically resets to a prior distribution; the reverse-inference mechanism is encoded by reversing GRN edges and by using differential-expression z-scores as the restart prior, so probability flows from observed downstream perturba","core_discovery":"HyperNetWalk claims driver identification can be unified at two scales. For each patient, it induces subnetworks from mutated genes: PPI and GRN supply topology (GRN reversed, so walks run upstream), and a co-expression network gives perturbation signal from z-scores. Random walks with restart convert these into scores. Stage II makes hypergraphs with patients as hyperedges and runs hypergraph random walks to diffuse scores across similar patients, producing personalized and cohort rankings. On 12 TCGA cancers, it reports personalized precision on par with DawnRank, cohort performance behind only DriverMP, and clear cancer-specific enrichment against IntOGen; ablations support the design.","pith_inferences":["A decisive test of the reverse-inference story would be ablating the topology terms: if rankings hold when PPI/GRN subnetworks are replaced with degree-matched random graphs and only the expression-perturbation prior remains, then the causal 'reverse walk' narrative is not the active ingredient; the paper's own ablation goes partway but a random-network control would isolate topology from diffusio","The hypergraph refinement stage is score-agnostic; feeding it scores from any per-patient ranker (e.g., a simple differential-expression summary) would show whether cross-sample diffusion, not Stage I's specific priors, drives the gain.","Because the benchmark uses the curated Cancer Gene Census list, a set itself enriched for well-studied hub genes, the reported precision may overstate performance on undiscovered drivers; checking against held-out recent additions to the census, or against functional screens, would test generalization.","The diagonal cancer-type specificity pattern suggests a possible second use the paper does not develop: using the method's cohort-level rankings to assign a tumor sample to its tissue of origin, or to flag predicted drivers whose tissue association is ambiguous."],"forward_implications":["A single pipeline can supply both patient-specific shortlists and cohort-level driver lists, so clinical and research use cases no longer require separate models.","Rare, low-frequency drivers can be recovered if they sit upstream of strongly perturbed downstream genes, rather than being filtered out by mutation-recurrence thresholds.","Cancer-type-specific driver sets can be recovered, which is useful for tissue-specific therapeutic targeting and for distinguishing general drivers like TP53 from context-dependent ones.","High precision at the top of the ranked list means fewer candidates need experimental validation, making downstream functional screens cheaper."],"fun_headline_variants":["Unified network walk finds cancer drivers per patient and cohort","Reverse walks on layered network ID cancer drivers at two scales","HyperNetWalk: one model, two driver rankings","Personalized and cohort driver genes via one walk","Layer network reverse walk ranks drivers per patient and cohort"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The evaluation assumes that the curated list of known cancer driver genes is a complete and unbiased truth label, even though the model is built to reward the same network-centrality properties that helped genes get onto that list.","fun_headline_variants_meta":{"raw":{"variants":["Unified network walk finds cancer drivers per patient and cohort","Reverse walks on layered network ID cancer drivers at two scales","HyperNetWalk: one model, two driver rankings","Personalized and cohort driver genes via one walk","Layer network reverse walk ranks drivers per patient and cohort"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001011,"raw_usage":{"total_tokens":4100,"prompt_tokens":725,"completion_tokens":3375,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":469,"completion_tokens_details":{"reasoning_tokens":3312}},"tokens_in":469,"tokens_out":3375,"duration_ms":20314,"temperature":1.0,"reasoning_tokens":3312,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T12:51:16.981261+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun HyperNetWalk on the same 12 cohorts with the protein-protein interaction and gene-regulatory subnetworks replaced by degree-preserving random networks, keeping the expression-based restart priors unchanged. If Precision@K against the known-driver list stays nearly the same, the claimed reverse-inference mechanism is not what carries the ranking. As a second check, if ranking mutated genes by network degree and regulatory out-degree alone matches the top-200 recall, then hub bias explains most of the result.","supporting_citations":[],"review_version":1}