{"id":"b5c34048-0677-4dfe-a429-83f4629898d3","arxiv_id":"1909.02695","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"In cancer gene networks, hub-centered communities with large degree discrepancies show Poisson eigenvalue spacing, while low-discrepancy subnetworks show Wigner-Dyson spacing.","lead":"This paper analyzes gene interaction networks from cancer cells using spectral methods and finds that networks localize around hub genes when hub degree exceeds a fitted critical value around 40. It also splits the networks by degree difference between connected genes and finds different random-matrix statistics in the two parts.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Wigner-Dyson/Poisson contrast is not yet shown to be a property of cancer-gene community structure; the delta=25 split is post hoc and the Poisson support is weak (76% KS), so a null-model test is needed.","rationale":"The reader's weakest assumption (delta=25 as a valid, pre-defined community boundary) is indeed load-bearing and is the same area I would flag. I mark partial rather than full agreement because the concern can be sharpened: the delta split is not merely an arbitrary threshold but a structural filter that would be expected to produce a Wigner/Poisson contrast in any strongly disassortative network, and the Poisson-side KS support (76%) is too weak to distinguish a real modular signal from a generic sparse-star artifact. The localization transition dc≈40 is also model-dependent, since Eq. (14) is fitted rather than predicted and the fit is acknowledged to break down in denser networks. I nevertheless keep the reader's CONDITIONAL verdict unchanged rather than moving to REJECT: the paper is exploratory, and a null-model comparison plus a threshold-stability analysis could validate or refute the central claim. The proposed concrete test directly targets the difference between a biological community signal and a generic artifact of the edge-filtering procedure.","tokens_in":9469,"tokens_out":4260,"duration_ms":47612,"concrete_test":"Run the same Figure 7 pipeline on degree-preserving randomized null graphs: generate 100 configuration-model versions of each of the 27 networks, apply the identical delta<25 and delta>=25 edge split, unfold eigenvalues, and compute KS acceptance rates for Wigner-Dyson and Poisson. If the randomized null graphs reproduce the same Wigner/Poisson contrast and similar acceptance rates, the observed dichotomy is a generic consequence of filtering by degree discrepancy, and the claim that it encodes hub-centered cancer-gene communities is not supported. If the contrast disappears in the null graphs, the concern is mitigated. As a secondary check, recompute the contrast for delta thresholds 10, 15, 20, 30, 40, and 50 to test whether delta=25 is a stable boundary rather than a post hoc cut.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that splitting a cancer gene network by the degree-discrepancy threshold delta=25 yields two spectrally distinct regimes: Wigner-Dyson spacing for delta<25 and Poisson spacing for delta>=25 (Section 3, Figure 7). This claim requires that the delta-based partition isolates biologically meaningful hub-centered communities. The load-bearing weakness is that delta=25 is introduced after inspecting Figure 5 and the dc≈40 discussion, with no derivation from a null model, no independent community-detection comparison, and no stability analysis with respect to the threshold. Because the cut removes all similar-degree edges, it mechanically separates a dense, assortative core from a disassortative, hub-anchored periphery; any scale-free network with broad degree distribution might show the same spectral contrast after such a filter. The Poisson claim is also not strongly supported: only 76% of 502 eigenvalue segments pass the KS test at alpha=0.05, well below the roughly 95% expected under a true Poisson null. In addition, the localization threshold dc is estimated by fitting Eq. (14), which is derived for a Poisson random graph plus one hub, to scale-free, disassortative data; the authors themselves note that the fit fails for dense networks with c>8. Thus both the critical degree and the community boundary are model-dependent or data-dependent, and the observed spectral dichotomy has not been shown to be a property of cancer-gene biology rather than of the filtering procedure.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript analyzes 254 gene co-expression networks of 8,000 nodes each from the TCNG database. It reports a power-law degree distribution with exponent gamma ~ 3.87, a spectral tail exponent mu ~ 6.5, eigenvector-centrality localization on high-degree hubs with a critical degree d_c ~ 40 for mean degree c = 6.6, and disassortative degree correlations k_nn(k) ~ k^-0.37 for k >~ 40. The authors then select 27 networks with very large hubs and split each network into two subgraphs according to the degree discrepancy Delta between linked nodes: Delta < 25 and Delta >= 25. They report that the nearest-neighbor eigenvalue spacing P(s) follows the Wigner-Dyson distribution in the Delta < 25 subgraphs (93% of 906 segments pass the KS test) and the Poisson distribution in the Delta >= 25 subgraphs (76% of 502 segments pass the KS test). They conclude that hub-centered communities in cancer gene interaction networks are modular and that spectral statistics can identify them.","tokens_in":9806,"tokens_out":7118,"duration_ms":73838,"significance":"If the reported spectral dichotomy were robust, it would provide a relatively simple spectral signature of hub-centered modular organization in gene networks. The paper's strengths are its use of a public, large dataset, explicit KS pass-rate statistics, and a falsifiable aggregate claim: the Delta-split P(s) contrast. However, the two load-bearing ingredients of the claim—the Delta = 25 cut and the critical degree d_c—are chosen or fitted after inspecting the data rather than derived from a null model, and the Poisson claim is statistically weak at a 76% pass rate. These issues leave the central claim plausible but not yet established.","major_comments":[{"comment":"The Delta = 25 threshold is introduced after observing the degree-discrepancy histogram in Fig. 5 and the d >~ 40 discussion, and it is never derived or tested. Because the split deletes every edge whose endpoints have similar degree, it mechanically separates a dense assortative residual graph from a disassortative hub-anchored subgraph. A scale-free network with a broad degree distribution could display the same spectral contrast after such a filter even with no biologically meaningful community structure. Please add a stability sweep over Delta (e.g., 10, 15, 20, 30), a configuration-model null with the same degree sequence subjected to the same Delta-split, and a comparison with an independent community-detection method (e.g., modularity optimization or Louvain).","section":"Section 3, Fig. 7 and Eq. (16)"},{"comment":"The localization transition d_c is not predicted from the data; it is obtained by fitting Eq. (14), a formula derived for a Poisson random graph plus a single hub, to scale-free, disassortative networks, with the mean degree c treated as a free fitting parameter (c-dagger = 13 for c = 4.7, and c-dagger = 20 for c = 6.6). The authors explicitly state that the fit fails for c > 8. Thus d_c ~ 40 is an empirical crossover value, not a theoretically derived critical point, and the abstract's 'd_c ~ 40' should be presented as a fitted value. A direct plot of v_1^2 and the inverse participation ratio against d, without the fitted curve, would support the claimed transition independently of Eq. (14).","section":"Section 3, Fig. 3 and Eq. (14)"},{"comment":"The Poisson claim is not strongly supported by the reported KS statistics. For a true Poisson null and alpha = 0.05, approximately 95% of independent segments should pass the KS test; the reported 76% pass rate (N = 502) is far below that expectation and indicates systematic deviations. The authors should report the histogram of KS p-values and compare it with the uniform distribution expected under the null, and provide confidence intervals on the averaged P(s) curve, rather than reporting only the pass rate.","section":"Section 3, Fig. 7(b)"},{"comment":"The fraction of edges with Delta >= 25 varies from 5.7% to 65.7% across the 27 networks, and the fraction of nodes varies from 15.3% to 92.7%. The averaged P(s) in Fig. 7 could therefore be dominated by a few networks with very large hub communities. Please report per-network P(s) or per-network KS pass rates and state the weighting used in the average, to confirm that the Poisson behavior is not driven by a small subset of networks.","section":"Table 1"}],"minor_comments":[{"comment":"Typographical errors should be corrected: 'dissasortative' (Sections 2.3 and 3), 'eacn' (Section 2.2), 'statisitics' (Table 1 caption), 'Kormogorov-Smirnov' (Section 3), and the placeholder 'greaterorsimilar'.","section":"Throughout"},{"comment":"The unfolding and segment construction are only described by reference to Ref. [17]; since P(s) is the central observable, the segment size, the number of eigenvalues per segment, and the unfolding algorithm should be stated explicitly in this manuscript.","section":"Section 2.4"},{"comment":"The axis label 'v2_1' is ambiguous; it should be typeset as v_1^2 or v-squared_1, with a clear definition in the caption.","section":"Section 3, Fig. 3"},{"comment":"The condition 'lambda_2max > 200' is confusing. If lambda_max is the largest eigenvalue, write lambda_max^2 > 200 or lambda_max > sqrt(200), and use consistent notation.","section":"Section 3, Fig. 4"},{"comment":"The formula as printed is hard to parse; the hub case and neighbor case should be written as separate lines with clear parentheses, and the condition d > 2c should be stated in relation to the fitted parameter c-dagger.","section":"Eq. (14)"},{"comment":"The NDEx network is said to be accessible via a supplementary file, but no accession or stable URL is given; please include the NDEx UUID or a persistent link.","section":"Supplementary material"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is short and uses public data, so the requested analyses (Delta-sweep, configuration-model null, KS p-value distribution, and per-network results) are feasible within a revision. The main risk is that the Delta >= 25 Poisson signal is an artifact of the degree-discrepancy filter; if a degree-matched null shows the same dichotomy, the community claim should be substantially softened. The paper's topic fits the q-bio/physics interface, but the statistical standard for the central claim needs to be raised before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing you should know: this paper does something new—it splits inferred cancer gene networks by the degree discrepancy of linked nodes and reports that the Δ<25 part shows Wigner-Dyson level spacing while the Δ≥25 part shows Poisson spacing. That is a concrete, checkable observation, and I don't think it appears in the earlier literature. But the threshold is chosen after looking at Figure 5, and the Poisson claim only holds in 76% of the eigenvalue segments, so the contrast is not yet nailed down.\n\nWhat the paper does well: it uses a large body of data—254 inferred networks, about nine million edges—and it averages carefully. The disassortativity for hub degrees k>40 is plausible and well documented. The eigenvector centrality fit to Eq. (14) is honest: the authors report that the fitted critical degree dc is larger than the theoretical 2c and that the fit fails for dense networks (c>8). The power-law exponents γ≈3.8 and μ≈6.5 are consistent with the known relation μ=2γ−1. The paper does not oversell the mediator genes; that's a speculation at the end, not a claim.\n\nThe soft spots are all in the central contrast. The Δ=25 cutoff is not derived from a null model; it's introduced after inspecting the degree-discrepancy histogram. A spectral statistic that depends on where you cut the graph is exactly the kind of thing that can be an artifact of the filter. Cutting out all similar-degree edges mechanically separates a dense assortative core from a disassortative hub-anchored periphery; any scale-free network with broad degree distribution might show Wigner-Dyson in the core and Poisson in the periphery under that filter. Second, the Poisson support is weak: 76% of 502 segments passing KS at α=0.05 is below the ~95% one would expect from a true Poisson null. That's a meaningful gap. Third, the localization threshold dc is a fit parameter, not a prediction, and the paper says the fit breaks down for c>8, so the model is not universal.\n\nNone of these are fatal to the paper as an exploratory analysis. But they mean the title-level claim—that the spacing distribution encodes hub-centered community structure—is not yet established. A serious referee should ask for a null-model comparison (e.g., degree-preserving randomized graphs) and a threshold-sensitivity analysis.\n\nThis paper is for network biologists who already work with spectral methods and want a concrete example on real gene networks. It deserves peer review because it is honest, reproducible, and the observation is worth checking. I'd recommend sending it, expecting a major revision.","headline":"A useful empirical application of known spectral tools to cancer gene networks, but the headline Wigner/Poisson contrast depends on a post hoc threshold and needs a null-model check.","tokens_in":10301,"tokens_out":2512,"would_cite":false,"duration_ms":25881,"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":"Once a hub gene exceeds about 40 links, a cancer gene network localizes on that hub, and its eigenvalue spacing separates into Wigner-Dyson and Poisson regimes by degree discrepancy.","keywords":["spectral analysis","eigenvector centrality","localization","hub genes","community structure","Wigner-Dyson distribution","Poisson distribution","degree correlation"],"falsifier":"Recompute $P(s)$ for the same 27 networks with the splitting threshold varied over $\\Delta = 10, 15, 20, 30, 35, 40$: if the Wigner-Dyson/Poisson contrast does not persist across neighboring thresholds, the reported regimes are an artifact of the $\\Delta=25$ choice. Alternatively, rewire each network while preserving the degree sequence; if the contrast disappears under rewiring, the effect is purely structural rather than a signature of biological gene communities.","tokens_in":9273,"feed_emoji":"🧬","tokens_out":11595,"duration_ms":106711,"temperature":0.7,"pith_summary":"The paper sets out to show that the eigenvalue statistics of cancer gene interaction networks encode the network's community organization. It observes that when a hub gene's degree passes a critical value $d_c \\simeq 40$, the leading eigenvector localizes on that hub and its neighbors, and hubs attach preferentially to low-degree nodes. In 27 strongly localized networks whose hubs have degree $d>125$, the paper splits the graph by the degree discrepancy $\\Delta$ between linked nodes and finds two distinct spacing laws: edges with $\\Delta < 25$ give Wigner-Dyson nearest-neighbor spacing, while edges with $\\Delta \\ge 25$ give Poisson spacing. On the paper's interpretation, this spectral dichotomy is a signature of hub-centered modular sub-networks, and it opens a spectral route to extracting gene communities that may determine cancer-cell behavior.","feed_headline":"Cancer gene networks show two spectral regimes once hubs pass ~40 links","feed_subtitle":"Degree-discrepancy splitting makes the spectrum switch from Wigner-Dyson to Poisson, exposing hub-centered modules.","key_machinery":"The key objects are the adjacency matrix $A$ of each gene network, the eigenvector centrality $v_i$ (the $i$-th component of the eigenvector of the largest eigenvalue $\\lambda_1$), and the inverse participation ratio $\\Psi(\\lambda_i)=\\sum_j (v_i^{(j)})^4$, which detects localized eigenvectors. The paper uses a theoretical formula for $v_i^2$ in a random graph with an added hub, Eq. (14), to locate the localization transition by fitting $d_c = 2c^\\dagger$; it defines the degree discrepancy $\\Delta=|k_i-k_j|$ for edges and uses it to partition each network. The central diagnostic is the unfolded nearest-neighbor spacing distribution $P(s)$, compared with Wigner-Dyson and Poisson forms by goodness-of-fit tests. The machinery works by turning community structure into a spectral dichotomy: the $\\Delta<25$ sub-network is the delocalized 'hairball' showing correlated eigenvalues, while the $\\Delta\\ge25$ hub-centered sub-network shows uncorrelated Poisson eigenvalues.","core_discovery":"The central discovery is that a single spectral observable, the nearest-neighbor eigenvalue spacing distribution $P(s)$, cleanly separates two regimes in cancer gene interaction networks once the graph is split by the degree discrepancy $\\Delta = |k_i - k_j|$ of linked nodes. In the 27 strongly localized networks whose largest hub has degree $d>125$, the sub-network formed by edges with $\\Delta < 25$ yields a Wigner-Dyson $P(s)$, with 93% of 906 eigenvalue segments passing a goodness-of-fit test at $\\alpha=0.05$, while the sub-network of edges with $\\Delta \\ge 25$ yields a Poisson $P(s)$, passing in 76% of 502 segments. The paper links this to localization: eigenvector centralities take finite values only on hub nodes and their neighbors once the hub degree exceeds $d_c \\simeq 40$, the degree correlation function becomes disassortative with $k_{nn}(k) \\propto k^{-0.37}$ for $k \\gtrsim 40$, and the $\\Delta\\ge25$ component contains the hubs and their small-degree partners. The conclusion is that hub-centered gene communities can be read off the spectrum of the adjacency matrix.","pith_inferences":["A natural extension is to test whether the $\\Delta=25$ cutoff rescales with the network's mean degree or node count; if the qualitative Wigner-Dyson/Poisson contrast survives rescaling, the spectral signature is general, while a fixed cutoff would be an artifact of this particular dataset.","Applying the same splitting to single-cell expression networks or protein interaction networks would show whether the two-regime spacing law is a generic feature of scale-free biological networks or specific to these inferred cancer networks.","A degree-preserving random rewiring of the same networks would isolate whether the Poisson spacing in $\\Delta\\ge25$ parts comes from hub-localized topology alone or from biologically meaningful community organization.","The fraction of segments that fail the Poisson test in the $\\Delta\\ge25$ part suggests the boundary may be fuzzy; a generative model with a smooth mixture of Wigner-Dyson and Poisson spacing could estimate the mixing fraction as a function of $\\Delta$."],"forward_implications":["If the spectral dichotomy is real, the spacing distribution $P(s)$ of an unlabeled gene sub-network reveals whether it is hub-centered: Poisson spacing flags a disassortative hub community, and Wigner-Dyson spacing flags the rest.","The localization threshold $d_c \\simeq 40$ and the disassortative exponent $-0.37$ become empirical constraints that any proposed generative model of cancer gene networks should reproduce.","The $\\Delta \\ge 25$ sub-networks of the 27 super-hub networks define a concrete candidate list of hub-centered gene communities, so the small-degree mediator genes that connect hubs can be prioritized as potential disease-relevant targets.","Because the method needs only the adjacency matrix, it can be applied without community-detection algorithms or gene annotations, giving a purely spectral community extraction procedure."],"supporting_citations":[{"why":"It supplies the Bayesian network inference method and the gene-network data from which every adjacency matrix in the study is built.","marker":"[16]"},{"why":"It provides the gene-expression samples that are the raw input for the Bayesian network inference.","marker":"[18]"},{"why":"It establishes the unfolding procedure and the prior observation of Wigner spacing in dense gene networks that this paper extends to community splitting.","marker":"[17]"},{"why":"It supplies the theoretical eigenvector-centrality localization formula, Eq. (14), and the relation $\\lambda_{\\max}\\simeq\\sqrt{d}$ used to locate the hub-degree transition.","marker":"[19,20]"},{"why":"It supplies the degree-correlation function $k_{nn}(k)$ and the assortativity framework used to show that hubs connect to low-degree nodes.","marker":"[21,22,23]"},{"why":"It gives the tail power-law relation $\\rho(\\lambda)\\propto\\lambda^{-\\mu}$ with $\\mu=2\\gamma-1$, used to compare the spectral tail with the degree distribution.","marker":"[7]"}],"fun_headline_variants":["Hub degree 40 splits gene network spectra","Degree discrepancy dictates cancer spectra regime","Gene communities read from spectral switch","Wigner-Dyson to Poisson at hub degree 40"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the degree-discrepancy cutoff $\\Delta=25$ is a real community boundary rather than a value chosen after inspecting the data, so if the cutoff is arbitrary or data-dependent the reported Wigner-Dyson/Poisson contrast could be created by the cut itself.","fun_headline_variants_meta":{"raw":{"variants":["Hub degree 40 splits gene network spectra","Degree discrepancy dictates cancer spectra regime","Gene communities read from spectral switch","Wigner-Dyson to Poisson at hub degree 40"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00044,"raw_usage":{"total_tokens":2224,"prompt_tokens":931,"completion_tokens":1293,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":1239}},"tokens_in":547,"tokens_out":1293,"duration_ms":11680,"temperature":1.0,"reasoning_tokens":1239,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:42:03.641921+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute $P(s)$ for the same 27 networks with the splitting threshold varied over $\\Delta = 10, 15, 20, 30, 35, 40$: if the Wigner-Dyson/Poisson contrast does not persist across neighboring thresholds, the reported regimes are an artifact of the $\\Delta=25$ choice. Alternatively, rewire each network while preserving the degree sequence; if the contrast disappears under rewiring, the effect is purely structural rather than a signature of biological gene communities.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the Bayesian network inference method and the gene-network data from which every adjacency matrix in the study is built."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It provides the gene-expression samples that are the raw input for the Bayesian network inference."},{"cited_title":"(2018) Random Matrix Analysis for Gene Interaction Net- works in Cancer Cells","cited_arxiv_id":null,"evidence_quote":"It establishes the unfolding procedure and the prior observation of Wigner spacing in dense gene networks that this paper extends to community splitting."},{"cited_title":"N., Goltsev, A","cited_arxiv_id":null,"evidence_quote":"It gives the tail power-law relation $\\rho(\\lambda)\\propto\\lambda^{-\\mu}$ with $\\mu=2\\gamma-1$, used to compare the spectral tail with the degree distribution."}],"review_version":1}