{"id":"f5fb0c8e-6f3e-4c65-9c52-f69aece5b0b4","arxiv_id":"2607.21833","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"RECON replaces GRADE's simple nonzero rule with a GMM/max-ratio threshold, removing nearly all spurious regulatory edges while keeping true ones across simulated ODE networks.","lead":"A new statistical method, RECON, uses a nonparametric ODE model plus a data-driven threshold to rebuild microbial regulatory networks from time-course abundances, cutting spurious edges from 239 to zero in simulations. It also handles sparse, irregular patient sampling and gives each inferred link a sign, strength, and direction.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Threshold selection assumes a single well-separated gap between true and spurious edge strengths; the headline simulation result may be an artifact of a design where each node has one dominant true edge.","rationale":"The reader's weakest_assumption correctly identifies the unstated separation between true and spurious edge strengths as the fragile premise. I agree that this is the most load-bearing concern because the entire novelty of RECON is its data-driven threshold; if the threshold cannot distinguish true from spurious edges, the central claim collapses. I sharpen the concern by pointing to the max-ratio criterion's reliance on observed cluster extrema and to the specific simulation design (one true edge per node) that guarantees the gap. A concrete test with heterogeneous effect sizes would settle whether the method generalizes. The residual-term issue raised by the reader is real but secondary, as the simulations suggest it does not affect edge selection in the tested settings. The paper's single-run tables and lack of code or data further undermine confidence in the reported zero-spurious-edge results, but these are not by themselves fatal; they reinforce the need for a sensitivity check. Overall, the appropriate verdict remains CONDITIONAL: the paper should either prove or validate the separation assumption, or temper its claims.","tokens_in":34095,"tokens_out":9045,"duration_ms":90516,"concrete_test":"Modify Simulation 1-II to make true regulatory coefficients heterogeneous: for each block q, draw a multiplier m_q ~ LogNormal(0, 1) and replace the ODE coefficient 2qπ in Eq. (26) with m_q * 2qπ. This creates weak and strong true edges, so some true edge strengths may overlap with spurious ones. Reduce the number of subjects from R=200 to R=50 to increase estimation noise. Run RECON and GRADE over 100 random seeds. If RECON's AUC-ROC drops below 0.9 or its median FP exceeds 10 while GRADE's AUC-ROC stays near 0.85, then the separation assumption is load-bearing and the headline claim fails outside the favorable setting. If RECON maintains AUC > 0.95, the concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of RECON is that its data-driven threshold, chosen via GMM clustering and the maximum-ratio criterion in Section 3.4 (Eq. 19), removes nearly all spurious edges while retaining true edges. This claim depends on the unstated assumption that true and spurious edge strengths are well separated and that the largest relative gap between adjacent GMM clusters corresponds to the true/spurious boundary. The paper provides no theoretical or empirical support for this separation. The max-ratio criterion is especially fragile: it uses observed cluster min/max, which are noisy order statistics, and it always selects some gap—even when true strengths are heterogeneous or when spurious strengths form multiple clusters. If GMM selects K=1, the method is undefined. The headline result, in which spurious edges drop from 239 to 0 (Simulation 1-II), occurs in a setting where each node has exactly one true incoming edge, so after per-node normalization the true edge is at strength 1 and all spurious edges near 0, guaranteeing a large gap. This structure is not representative of typical regulatory networks, and the single-run tables provide no evidence that the gap is stable across randomness. Thus the claim that RECON 'consistently' outperforms GRADE is supported only in a narrow, favorable simulation regime; outside of it the method has no defined behavior.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RECON, an integral-based additive nonparametric ODE approach for reconstructing directed, signed, dynamic regulatory networks from longitudinal data. Methodologically, RECON estimates regulatory functions through B-spline expansions and group LASSO, then applies a data-driven threshold constructed from Gaussian mixture modeling and a maximum-ratio criterion to remove spurious edges. The paper claims that across five simulation studies RECON consistently outperforms GRADE, reducing spurious edges to zero or near zero while retaining true edges, and it applies the method to a longitudinal gut microbiota dataset from allo-HCT patients, reporting pre- and post-transplant networks with candidate keystone taxa.","tokens_in":34455,"tokens_out":3904,"duration_ms":49206,"significance":"If the reported results hold, RECON would provide a practical improvement over GRADE by adding a principled-looking, data-driven edge-selection step to a well-established nonparametric ODE framework, while also extending applicability to sparse irregular longitudinal designs. The real-data analysis addresses an important clinical dataset and offers falsifiable, network-level hypotheses about microbial regulation after transplantation. However, the central threshold assumption and the single-run simulation evidence are not yet sufficient to support the headline claim of consistent, near-perfect reconstruction. No circular reasoning is apparent: the GMM threshold is based on estimated edge strengths without using true adjacency labels, and the GRADE comparison uses identical model estimates with a different post-hoc threshold.","major_comments":[{"comment":"The threshold selection procedure assumes that normalized strengths of true and spurious edges are well separated and that the largest ratio min(C_l)/max(C_{l+1}) identifies the correct boundary. This assumption is load-bearing: the headline result of reducing spurious edges from 239 to 0 (Simulation 1-II) occurs in a design where each target node has exactly one true incoming edge, so after per-node normalization the true edge has strength 1 and spurious edges are near 0. The paper provides no theoretical or empirical support for the separation in other configurations, and the procedure is undefined when BIC selects K=1. The authors should either provide a theoretical characterization of the gap or demonstrate robustness through simulations with overlapping strength distributions, heterogeneous true degrees, and multiple true edges per node.","section":"§3.4, Eq. (19)"},{"comment":"All simulation results appear to be single runs with no standard errors, confidence intervals, or repeated-seed analysis. The claim that RECON 'consistently outperforms' GRADE cannot be evaluated from one realization per setting, particularly when AUC values are reported to four decimal places. Additionally, the reported 'AUC-ROC' is computed as (1 + TPR - FPR)/2, which is a linear transformation of a single operating point, not the area under an ROC curve. This metric is misleading as an AUC and should either be computed as a proper area over a range of thresholds or renamed. Repeated Monte Carlo simulation and a correctly defined AUC are needed before the comparative claim is supported.","section":"§4.3 and Tables 1–5"},{"comment":"The residual integral term sum_k ∫ δ_jk(X_k(u;θ)) du is dropped as negligible, but no justification or numerical check is provided. This is especially concerning for the nonlinear Brusselator system in Eq. (31), where the additive B-spline assumption may not hold with a small residual over the trajectory support. The authors should either provide a bound on the dropped term or report a simulation diagnostic comparing fitted and true regulatory functions to show that the omission does not materially bias the estimated edge strengths in any of the five settings.","section":"§3.3, Eqs. (12)–(14)"},{"comment":"The real-data networks are interpreted as 'causal' and used to identify keystone taxa and regulatory dynamics, but no sensitivity analysis is given for the threshold selection or for preprocessing choices (PACE tuning, GMM/BIC, the maximum-ratio boundary). Since the entire network topology depends on the threshold in Eq. (19), the biological conclusions in Section 5 could change under a different but equally reasonable threshold. The authors should report how edge counts, modularity, and keystone identifications vary with the threshold and with reasonable perturbations of the preprocessing steps.","section":"§5.2–5.3"}],"minor_comments":[{"comment":"The sentence 'smaller TP or larger FN indicates possible model misclassification' is confusing; smaller TP is not a sign of better performance, and the intended meaning should be rephrased.","section":"§4.3"},{"comment":"Since group LASSO estimates are rarely exactly zero in finite samples, the definition of G^GRADE using strict positivity is not numerically operational as stated; the paper should clarify how zero coefficients are identified in practice.","section":"§3.4, Eq. (18)"},{"comment":"Gray arrows for spurious edges are difficult to distinguish from black arrows in small print; adding a separate panel with FP-only edges or using dashed/solid styles would improve readability.","section":"Figures 1, 4–8"},{"comment":"Several reference entries are incomplete or inconsistently formatted, e.g., [11], [73], and [80] mix 'et al.' styles or omit author lists. A careful reference cleanup is needed.","section":"References"},{"comment":"The sentence describing the filtering result is slightly confusing: 29 families are retained in total, with 21 common and 4 unique to each period, giving p=25 per window. This should be stated more directly to avoid implying 29 per window.","section":"§5.1"},{"comment":"The paper does not state whether simulation code or seeds are available. The interactive figures are useful, but providing the underlying analysis code and simulation scripts would strengthen reproducibility, especially given the single-run simulation concern.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The paper has a useful core idea and a relevant application, but the central threshold claim is currently supported only in a narrow simulation regime, and the simulation evidence lacks replication. I would encourage the editors to request major revisions that add proper Monte Carlo evaluation, a correct AUC, and a robustness check of the GMM/max-ratio threshold. If the authors can show stable performance outside the one-true-edge-per-node design, the contribution could become publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: RECON is an incremental but honest extension of GRADE that adds a data-driven threshold to group-LASSO edge strengths plus PACE for sparse irregular data. In the simulated settings it convincingly does what it claims — massive false-positive reduction with near-perfect AUC. The soft spots are the heuristic nature of the threshold, the single-run results, and the real-data interpretation. Nothing here invalidates the main claim, but the generalization beyond well-separated strengths is unproven.\n\nWhat's new: the GMM/max-ratio thresholding (Section 3.4) and the PACE interpolation for ODE-based network inference. The simulations span linear and nonlinear (Brusselator) ODEs, block structures, and the Design II latent-coefficient setting, which is genuinely harder. Credit where due: the GRADE comparison is fair — identical estimates, only the post-hoc threshold differs — and the reported numbers back the abstract's \"239 to 0\" claim. The real-data analysis is also a sensible demonstration of the pipeline.\n\nSoft spots, in order of severity:\n\n1. The threshold assumes true and spurious normalized strengths are well separated. The max-ratio criterion always picks a gap, and if GMM selects one cluster the method has no defined output. The paper never states or tests this assumption. The simulations are structured favorably: per-node min-max normalization puts the true edge(s) near 1 and spurious edges near 0, which makes the gap large. That is a real limitation, but not a fatal one — the method is explicit about what it does, and the stress-test note overstates the point slightly: Simulations 2 and 3 have two true edges per node, so it is not just a one-dominant-edge artifact.\n\n2. All simulation results are single-run. With AUCs of 1.00 across many settings, one or two bad seeds could change the picture. Error bars or a few replicates would materially strengthen the claim.\n\n3. No code or data released. For a method whose main novelty is a threshold heuristic, reproducibility matters. The interactive plots are nice but do not substitute for scripts.\n\n4. The real-data network inferences (keystone taxa, pre/post differences) are stated without any uncertainty quantification. That is common in this literature, but the language overclaims.\n\nThe math is standard GRADE machinery; the dropped residual integral is inherited from GRADE and acknowledged as an assumption. Citation pattern looks fine. This deserves a serious referee — I would send it out, but ask for multi-run simulations, a handling of the K=1 case, and code. It is a conditional accept in my book.","headline":"RECON is an honest, incremental extension of GRADE whose GMM threshold convincingly kills spurious edges in clean simulations, but the separation assumption is untested, results are single-run, and no code is released.","tokens_in":34878,"tokens_out":2588,"would_cite":false,"duration_ms":28820,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G05","62-08"],"pacs":[],"model":"deepseek-v4-flash","headline":"A new edge-selection procedure, RECON, turns noisy group-LASSO ODE estimates into nearly perfect directed regulatory networks, cutting spurious edges from 239 to zero in simulations and exposing distinct pre- and post-transplant microbial r","keywords":["regulatory network reconstruction","nonparametric ODE","integral-based estimation","group LASSO","Gaussian mixture model","maximum-ratio criterion","longitudinal data","microbiome dynamics"],"falsifier":"Run RECON on a simulation where the weakest true edge has an estimated strength comparable to the strongest spurious edge (for example, by adding a true edge with a very small coefficient while keeping noise at the same level). If the GMM maximum-ratio threshold then fails to recover that true edge without adding spurious ones, the separation assumption is violated and the paper's headline claim collapses.","tokens_in":33969,"feed_emoji":"🦠","tokens_out":3676,"duration_ms":36451,"temperature":0.7,"pith_summary":"RECON claims to solve a practical bottleneck in reconstructing regulatory networks from time-course data: standard ODE-based methods like group LASSO leave many spurious edges because they lack a principled cutoff. The paper shows that a data-driven threshold—built from per-node normalized edge strengths, Gaussian mixture clustering, and a maximum-ratio gap criterion—can strip away almost all false edges while keeping nearly all true ones. Across five simulation studies spanning linear and nonlinear systems and dense or sparse sampling, RECON reaches AUC-ROC 1.00 in most settings, and in the hardest case reduces spurious edges from 239 to 0. Applied to longitudinal gut microbiota from transplant patients, it yields pre- and post-transplant networks with different topologies and keystone taxa, suggesting regulatory dynamics that abundance trends alone do not show.","feed_headline":"Data-driven threshold cuts spurious network edges from 239 to 0","feed_subtitle":"RECON's adaptive edge selection keeps true regulatory links while deleting noise—tested on simulated and transplant microbiota data.","key_machinery":"The load-bearing object is the data-driven edge threshold built from normalized regulatory strengths. For each node, the estimated L2 norms of group-LASSO coefficients are normalized to [0,1], pooled across nodes, clustered by a Gaussian Mixture Model, and the optimal cluster boundary is chosen by the maximum ratio of the smallest strength in a large-strength cluster to the largest strength in the adjacent small-strength cluster (Equation 19). This threshold converts the dense set of nonzero estimates into a sparse, signed, weighted omnidirectional adjacency matrix. The rest of the pipeline—local polynomial or PACE smoothing, basis expansion, integrated basis functions, and group-LASSO estim","core_discovery":"The paper establishes that an adaptive threshold placed at the largest relative gap between clusters of normalized regulatory strengths effectively separates true edges from estimation noise in an integral-based additive nonparametric ODE model. After fitting the model with group LASSO, RECON normalizes each node's estimated edge strengths, clusters the pooled normalized strengths with a Gaussian Mixture Model, and selects the cluster boundary with the maximum drop ratio between adjacent clusters. This boundary becomes the cutoff for edge inclusion. In simulations, the procedure retains all or nearly all true directed edges while driving false positives to zero or near zero, improving on the","pith_inferences":["The method's success depends on the empirical separation between true and spurious edge strengths; if that separation is absent or weak in other applications, the GMM threshold may either retain spurious edges or drop legitimate weak regulators. This is a testable limitation not addressed theoretically in the paper.","The real-data 'causal' interpretation carries the implicit assumption that the additive nonparametric ODE structure (including the dropped basis-expansion residual in Section 3.3) is a faithful description of gut microbial dynamics. A reader should treat the inferred pre/post-transplant regulatory differences as hypothesis-generating rather than confirmatory.","The threshold logic could be transferred to other sparse-estimation settings where group coefficients need hard-thresholding, though doing so would require re-establishing the separation property for each new problem.","The paper's five simulation studies all use block-structured ground truths; a natural next test would be a scale-free network or one with heterogeneous edge strengths, where the gap between true and spurious strengths might be less pronounced."],"forward_implications":["If correct, regulatory network reconstruction from discretely observed time courses can be made nearly spurious-free without sacrificing true edges, making ODE-based inference practical for noisy biological data.","The method extends ODE-based reconstruction to sparse, irregular longitudinal sampling designs, which are common in clinical microbiome studies, by interpolating subject-specific trajectories onto a common dense grid.","Modeling edge effects as time-varying functions allows networks to be studied dynamically: regulatory relationships can strengthen, weaken, or reverse over the observation window, which is invisible to constant-coefficient methods.","The signed and weighted omnidirectional output supports direct biological interpretation, including activatory/inhibitory roles, keystone-node identification, and modularity, as demonstrated on pre- and post-transplant gut microbiota networks.","The reported reduction from 239 to 0 spurious edges in the most challenging simulation is concrete evidence that the threshold can perform well even when the baseline method is very noisy."],"fun_headline_variants":["RECON removes 239 spurious edges in network reconstruction","Adaptive cutoff drives false regulatory edges to zero","Causal network method deletes all noise edges from data","From 239 false edges to 0 with data-driven threshold","New ODE approach reconstructs networks without spurious links"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The edge cutoff assumes that the estimated strengths of true and spurious edges form well-separated clusters, so that the largest gap between GMM clusters is the correct boundary; if the strength distributions overlap, the threshold will either keep spurious edges or drop true ones.","fun_headline_variants_meta":{"raw":{"variants":["RECON removes 239 spurious edges in network reconstruction","Adaptive cutoff drives false regulatory edges to zero","Causal network method deletes all noise edges from data","From 239 false edges to 0 with data-driven threshold","New ODE approach reconstructs networks without spurious links"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1229,"prompt_tokens":757,"completion_tokens":472,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":392}},"tokens_in":501,"tokens_out":472,"duration_ms":4942,"temperature":1.0,"reasoning_tokens":392,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T06:33:00.668246+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run RECON on a simulation where the weakest true edge has an estimated strength comparable to the strongest spurious edge (for example, by adding a true edge with a very small coefficient while keeping noise at the same level). If the GMM maximum-ratio threshold then fails to recover that true edge without adding spurious ones, the separation assumption is violated and the paper's headline claim collapses.","supporting_citations":[],"review_version":1}