{"id":"220e6b78-07b4-4a73-aeda-a3f24058caca","arxiv_id":"2608.09773","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"Graph measures computed on reservoir-computed connectivity maps correlate with electrode firing rates in neuronal cultures, but the correlations may be inflated because the maps are inferred from the same activity.","lead":"The authors turn connectivity maps inferred from lab-grown neuron recordings into graphs and check whether graph measures predict how active each electrode site is. They find correlations, but the maps are derived from the same activity data, which complicates the interpretation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Experimental correlations do not yet validate the ICM: T0 is trained on the same traces that supply the dependent variables, so Fig. 6 associations may be recoded rate statistics; a surrogate-data null control is required before accepting Eq. (13).","rationale":"The reader's weakest assumption coincides with the concern I find most load-bearing. The paper is best read as making two separable claims: (i) the RC-derived ICM can recover structural connectivity in controlled simulations, and (ii) the experimentally observed correlations between ICM-based centralities and node activity independently validate the inference. Claim (i) has genuine support in Section 3.2 (Fig. 2, Table 2), with the caveat that thresholds and F1 operating points are tuned on these simulations. Claim (ii) is insecure. The ICM is a function of the training data; the activity observables are summaries of the same data. Correlating a quantity inferred from a signal with another summary of that signal is not an out-of-sample validation. The fact that ICM centralities outperform ground-truth centralities in predicting node activity is precisely what a circular estimator would do; it therefore weighs against, not for, the claim that the ICM is a better structural map. The proposed surrogate test is decisive: if scrambled data with matched rates still produce the correlations, the experimental validation fails; if not, the concern is retired. Secondary issues (selection of 49 of 170 recordings by q, threshold tuning, no multiple-comparison correction, non-independent node samples) should also be addressed, but the surrogate null is the single decisive check. Because this is a condition that can be tested and reported, I do not change the reader's CONDITIONAL verdict: the manuscript should be accepted only after the null control is reported or the experimental validation claim is weakened.","tokens_in":20187,"tokens_out":7600,"duration_ms":74832,"concrete_test":"Run a surrogate null on all 170 MEA recordings: for each recording, independently shuffle spike times within each electrode (or generate Poisson trains matched to per-electrode rates and ISI statistics), destroying inter-electrode temporal structure while preserving marginal firing statistics. Train the identical RC pipeline with identical hyperparameters on the surrogate traces, infer T0, apply Eqs. (14)-(15), recompute the Fig. 6 centrality-activity correlations, and compare against a null distribution built from multiple surrogates. If correlations of comparable magnitude persist, the experimental validation is circular; if they drop near zero, the concern is resolved in the authors' favor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is that T0, obtained by linearizing the RC model trained on the spike-rate traces, contains structural information beyond a recoding of those traces. Eq. (13) treats T0 as an effective adjacency matrix, and Section 3.4/Fig. 6 then uses centrality measures computed from T0 to validate the inference against node-level observables extracted from the same recordings. Because the training data and the dependent variables coincide, nonzero centrality-activity correlations are expected whenever the learned operator encodes per-electrode firing statistics; such correlations do not test whether T0 approximates the structural adjacency matrix. The observation in Fig. 6A that ICM-based centralities correlate with firing rate more strongly than ground-truth-based centralities is the expected signature of this circularity, and the offered explanation (linearization redistributes inhibitory pathways, inferred self-loops) is post hoc. The in silico benchmark in Section 3.2/Table 2 is independent and valuable, but it cannot validate the experimental claim. Section 4.6 acknowledges that some correlations 'may appear mathematically expected from an analytical perspective,' which is precisely the unresolved point. A rate-matched surrogate control would settle whether the reported associations reflect genuine cross-electrode structure or training-data imprinting.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a pipeline that takes the Intrinsic Connectivity Matrix (ICM) inferred by a reservoir-computing (RC) model from multichannel MEA recordings, treats the ICM as an effective adjacency matrix (T0 ≈ A, Eq. 13), and computes graph-theoretic centralities whose relationships with node- and culture-level activity measures are then examined. The authors first benchmark the RC inference against known ground-truth adjacency matrices in NEST simulations, studying how inhibitory balance and spectral radius affect inference accuracy, and then report correlations between centrality measures and firing/burst/ISI statistics in both simulations and experimental MEA recordings. The claimed contribution is a scalable graph-theoretic characterization of neuronal cultures and, at the same time, a validation of the RC-based connectivity inference.","tokens_in":20382,"tokens_out":4919,"duration_ms":43366,"significance":"The in silico benchmarking against ground-truth adjacency matrices (Section 3.2, Table 2, Fig. 2) is a genuine strength: it provides controlled evidence about when the linearized ICM approximates the structural connectivity, and it is independent of the experimental circularity discussed below. The application of established centrality measures to MEA-derived connectivity is also potentially useful for moving beyond electrode-wise activity summaries. However, the paper's central validation claim—that the experimental centrality–activity correlations support the RC inference and the T0 ≈ A assumption—is not yet established because the ICM is trained on the same spike-rate traces that supply the dependent variables. A surrogate-data control is needed before the experimental correlations can be interpreted as evidence for the inferred connectivity. With that control added, this would be a valuable contribution to functional network analysis of in vitro cultures.","major_comments":[{"comment":"The central experimental claim is not supported as it stands because of circularity. The ICM T0 is obtained by training the RC model on the instantaneous spike-rate sequences of each recording, and the dependent variables (AFR, ABR, μISI, σISI) are computed from the same recordings. Any learned operator that reproduces the training traces must encode per-electrode firing statistics, so centrality measures derived from T0 will be correlated with those statistics even if T0 carries no genuine cross-electrode structural information. The observation in Fig. 6A that ICM-based centralities correlate with firing rate more strongly than ground-truth-based centralities is the expected signature of this imprinting, not evidence that T0 ≈ A. The authors themselves note in Section 4.6 that some correlations 'may appear mathematically expected from an analytical perspective,' which is precisely the unresolved issue. To validate the experimental claim, the authors should add a surrogate-data null control: for example, train the RC model on temporally shuffled or phase-randomized ISR traces that preserve each electrode's rate and spectral statistics but destroy cross-electrode structure, and show that the Fig. 6B correlations disappear or become significantly weaker. Alternatively, they could compare against a null model that preserves node-strength and degree distributions while randomizing edge placement. Without such a control, the experimental correlations in Fig. 6 do not validate the RC inference.","section":"§2.6, Eq. (13); §3.4, Fig. 6"},{"comment":"The threshold w'_th = 0.2 is selected via F1-score analysis on simulations (Section 3.2) and then applied unchanged to experimental ICMs. Simulations use Izhikevich/NEST networks with known ground truth, while experimental recordings are noisier and may have a different effective connectivity scale, so a threshold optimized in silico need not be appropriate for the experimental data. Because the threshold directly determines which edges survive and therefore affects all downstream centrality measures, the authors should (i) justify the transferability of the threshold to experimental data, and (ii) report a sensitivity analysis of the Fig. 6 correlations over a range of thresholds (for example, 0.1–0.3) to show that the conclusions do not hinge on this specific choice.","section":"§2.6.1, Eq. (15)"},{"comment":"The experimental dataset used for Fig. 6B is a post-hoc selection of 49 out of 170 recordings based on the data-richness parameter q, and the reported correlations are not corrected for multiple comparisons across six centrality measures and several observables. The selection criterion is itself correlated with RC model performance, and the paper states that weaker correlations were associated with low q; retaining only high-q recordings therefore inflates the apparent strength of the effect. The authors should report the full 170-recording analysis or explicitly show how the correlations vary with q rather than filtering, justify the selection before presenting the correlations, and provide multiple-comparison-corrected confidence intervals or p-values (for example, FDR or permutation-based) for any claim of 'statistically robust associations.'","section":"§3.4, Fig. 6B"}],"minor_comments":[{"comment":"The sentence 'We focus on the some common measures reported in many MEA measurements studies' contains a typo; 'the some' should read 'some'.","section":"§2.4"},{"comment":"The caption contains a typo: 'buttom right' should be 'bottom right.'","section":"Fig. 6 caption"},{"comment":"The reference to the DLP stimulation system appears as 'DLP system []' with an empty citation; this should either be completed or removed.","section":"§2.7.1"},{"comment":"The missing entries in the 'Full Excitatory' row for inhibitory metrics are explained only after the table; using an explicit 'N/A' or en-dash with a table note would be clearer.","section":"Table 2"},{"comment":"The gamma-like fit 'ax^p e^{-qx}' is introduced without defining the parameters or reporting the fit uncertainty; adding confidence intervals or the fitted parameter values would make the claimed optimal inhibition regime more quantitative.","section":"§3.2.1, Fig. 3C"}],"recommendation":"major_revision","confidential_remarks":"The manuscript relies heavily on the authors' own previous work (Ref. [26]) for the RC model and its validation; the present paper does not independently re-derive or re-verify that component. The experimental circularity is the central concern, but it is addressable with a surrogate-data control, so I do not recommend rejection. I would also note that the paper does not include a data or code availability statement, which would be valuable given the emphasis on a 'scalable, data-driven framework.'"},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: the simulation benchmark is solid and worth a look; the experimental validation is not there yet. The paper's central move—treating the RC-inferred ICM as an adjacency matrix and using centrality-activity correlations to validate the inference—has a load-bearing circularity problem. T0 is trained on the same spike-rate traces that later supply the dependent variables in Fig. 6, so the reported correlations may just reflect the learned operator recoding rate statistics. In the simulations, T0-based centralities outperform ground-truth-based ones; that is the signature of this inflation, not evidence that T0 approximates A.\n\nCredit where it's due: the in silico part is careful and genuinely useful. The NEST setup with population nodes allows clean benchmarking of T0 against a known adjacency. The ROC/PR AUC, F1, and NMWA analysis is thorough and shows a non-obvious result: intermediate inhibition is the sweet spot for linearized RC inference. That finding should survive. The comparison with the earlier GLIF-based study [36] is also fair.\n\nThe soft spots are in the experimental sections. The circularity is unresolved. A surrogate control—shuffle spike trains across electrodes, retrain the RC model, recompute centralities and correlations—would settle whether the associations reflect true cross-electrode structure or just training-data imprinting. Second, selecting 49 of 170 recordings by data-richness q is post hoc, and there's no multiple-comparison correction across the many centrality-DV pairs. The 'large sample size' argument is weak because node samples within a recording are not independent. Third, the threshold w'_th=0.2 is tuned on simulations and applied to experiments without re-justification. The paper's own Section 4.6 concedes some correlations 'may appear mathematically expected,' which is precisely the unresolved point.\n\nThe paper is clearly thinking and honestly written, but the experimental claim as stated does not hold. Who is this for? Computational neuroscientists and MEA experimentalists who want network-level descriptors. The simulation part is worth a serious referee; the experimental part needs those controls before it can validate the ICM. I'd send it to peer review, with a request for a surrogate null and more careful statistics.","headline":"The simulation benchmark is solid and worth a look; the experimental validation is compromised by circularity and post-hoc selection.","tokens_in":20982,"tokens_out":4280,"would_cite":true,"duration_ms":39818,"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 the reservoir-computed connectivity map of neuronal cultures can be read as an adjacency matrix, and that node centrality computed on it correlates with measured firing and bursting at each electrode.","keywords":["neuronal cultures","microelectrode arrays","reservoir computing","connectivity inference","graph theory","centrality measures","intrinsic connectivity matrix","network neuroscience"],"falsifier":"Simulate a ground-truth network whose adjacency matrix is known, train the RC model to obtain an ICM, then recompute the centrality–activity correlations after randomly permuting the entries of the ICM while preserving the distribution of edge weights and node degrees; if the permuted matrices yield correlations as strong as the unpermuted ones, the structural content of the ICM is not what drives the reported associations.","tokens_in":19911,"feed_emoji":"🧠","tokens_out":8097,"duration_ms":63436,"temperature":0.7,"pith_summary":"This paper argues that the connectivity map inferred by a reservoir-computing model from microelectrode-array recordings—called the Intrinsic Connectivity Matrix (ICM)—can be treated as an effective adjacency matrix of the cultured neuronal network. The authors compute standard graph-theoretic centrality scores on that matrix and show that these scores correlate, at the level of individual electrodes, with measured spiking and bursting activity in both in-vitro cultures and simulated networks. They also show that agreement between the inferred ICM and the known ground-truth wiring of simulated networks is highest at an intermediate inhibition level and degrades as network complexity grows. If the interpretation holds, it gives experimentalists a way to measure network-level organization and node importance directly from spike recordings, without direct anatomical tracing.","feed_headline":"Node centrality in inferred maps predicts firing","feed_subtitle":"Graph-theory scores computed on reservoir-computed connectivity track spiking, bursting, and interval statistics electrode by electrode.","key_machinery":"The central object is the Intrinsic Connectivity Matrix (ICM), denoted $T_0$, obtained by training a reservoir-computing model on instantaneous spike-rate sequences and then linearizing the trained operator while removing the influence of history; its entries are directed, signed weights (positive excitatory, negative inhibitory). The paper's operative identity is $T_0 \\simeq A$, treating the ICM as an effective adjacency matrix for graph-theoretic analysis. The centrality measures computed on $T_0$ (in/out degree, Katz, eigenvector, PageRank) are the machinery that bridges structure and activity: they are the independent variables whose Pearson correlations with measured firing, bursting, and ISI statistics quantify topology–dynamics coupling. The preprocessing threshold $w'_{th}=0.2$ (Eq. 15) sparsifies noise-dominated weights before graph metrics are computed, and the confidence measure $\\Gamma$ (Eq. 3) quantifies stability across reservoir initializations.","core_discovery":"Treating the reservoir-computed ICM as the adjacency matrix, $T_0 \\simeq A$ (Eq. 13), the paper establishes that node centrality measures—effective and absolute in-degree, out-degree, Katz, eigenvector, and PageRank—computed on the inferred connectivity correlate with node-level observables such as average firing rate, bursting rate, and inter-spike-interval statistics, in both simulated and MEA-recorded cultures. Across the 40 simulated networks and 170 experimental recordings (49 used after filtering for data richness), the correlations are consistently positive and strongest for Katz centrality and in-degree measures, while burst duration shows only weak associations. The paper further reports that the ICM's fidelity to ground-truth connectivity is maximal at an intermediate inhibitory–excitatory ratio and that reservoir-computing model performance declines with increasing spectral radius. Consequently, the authors conclude that the graph-theoretic interpretation of the ICM both provides a scalable framework for functional network characterization and independently supports the validity of the reservoir-computing inference.","pith_inferences":["A natural null-model test would permute ICM edge weights or degrees and recompute centrality–activity correlations; if the null reproduces the observed correlations, the inferred matrix would be shown to carry no structure beyond the activity statistics used to train the model.","The framework is portable to higher-density recording systems and in-vivo large-scale electrophysiology, where the authors expect denser sampling to sharpen the centrality–activity link, but this is not tested in the paper.","The observation that evoked-response training yields smaller inferred networks suggests stimulus-driven activity highlights a different subgraph than spontaneous activity; an explicit comparison of graph metrics between the two training conditions could reveal whether the effective connectivity is state-dependent."],"forward_implications":["If $T_0 \\simeq A$ holds, graph-theoretic descriptors of cultured networks can be computed from ordinary MEA recordings, giving a scalable summary of culture-level organization without invasive tracing.","Node centrality becomes a candidate predictor of local activity: electrodes high in in-degree or Katz centrality are expected to show higher firing and bursting rates, offering a structure-based readout for experiments.","The inhibition-balance result implies reservoir-computing connectivity inference is most reliable at intermediate inhibitory–excitatory ratios, so cultures outside that regime should be interpreted with caution.","The consistent structure–activity correlations in both simulations and experiments give independent support to the reservoir-computing inference model, extending its validity to datasets without ground-truth connectivity."],"supporting_citations":[{"why":"Supplies the reservoir-computing model and the ICM inference procedure that the entire pipeline builds on, including prior validation against ground-truth simulations.","marker":"[26]"},{"why":"Provides the earlier in-silico result that Katz centrality best predicts firing rate, which the present study contrasts with population-level MEA findings.","marker":"[36]"},{"why":"Defines the spiking-neuron model used for all simulated networks in the benchmarking experiments.","marker":"[41]"},{"why":"Supplies the regular-spiking parameterization that configures the simulated neurons.","marker":"[42]"},{"why":"Provides the simulation environment that generates the in-silico activity used for controlled benchmarking of the inference.","marker":"[31]"},{"why":"Supports treating inferred connectivity as a graph for functional characterization of in-vitro neuronal assemblies.","marker":"[30]"},{"why":"Defines the standard MEA node-level measure of average firing rate used as a dependent variable.","marker":"[23]"}],"fun_headline_variants":["Centrality on reservoir-derived graphs tracks firing","Inferred network centrality predicts spiking in cultures","Graph scores from RC connectivity mirror neural activity","Reservoir-based graph metrics forecast firing rates","Centrality in inferred maps links to bursting and spikes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on the assumption that the inferred connectivity matrix reflects the culture's actual wiring rather than being an artifact of training on the very activity traces it is compared with.","fun_headline_variants_meta":{"raw":{"variants":["Centrality on reservoir-derived graphs tracks firing","Inferred network centrality predicts spiking in cultures","Graph scores from RC connectivity mirror neural activity","Reservoir-based graph metrics forecast firing rates","Centrality in inferred maps links to bursting and spikes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00028,"raw_usage":{"total_tokens":1667,"prompt_tokens":955,"completion_tokens":712,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":642}},"tokens_in":571,"tokens_out":712,"duration_ms":7129,"temperature":1.0,"reasoning_tokens":642,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:01:09.535494+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a ground-truth network whose adjacency matrix is known, train the RC model to obtain an ICM, then recompute the centrality–activity correlations after randomly permuting the entries of the ICM while preserving the distribution of edge weights and node degrees; if the permuted matrices yield correlations as strong as the unpermuted ones, the structural content of the ICM is not what drives the reported associations.","supporting_citations":[{"cited_title":"Decoding neuronal networks: A reservoir computing approach for predicting connectivity and functionality,","cited_arxiv_id":null,"evidence_quote":"Supplies the reservoir-computing model and the ICM inference procedure that the entire pipeline builds on, including prior validation against ground-truth simulations."},{"cited_title":"From structure to activity: Using centrality measures to predict neuronal activity,","cited_arxiv_id":null,"evidence_quote":"Provides the earlier in-silico result that Katz centrality best predicts firing rate, which the present study contrasts with population-level MEA findings."},{"cited_title":"Which model to use for cortical spiking neurons?","cited_arxiv_id":null,"evidence_quote":"Supplies the regular-spiking parameterization that configures the simulated neurons."},{"cited_title":"Nest (neural simulation tool),","cited_arxiv_id":null,"evidence_quote":"Provides the simulation environment that generates the in-silico activity used for controlled benchmarking of the inference."},{"cited_title":"Functional connectivity in in vitro neuronal assemblies,","cited_arxiv_id":null,"evidence_quote":"Supports treating inferred connectivity as a graph for functional characterization of in-vitro neuronal assemblies."},{"cited_title":"Revealing neuronal function through micro- electrode array recordings,","cited_arxiv_id":null,"evidence_quote":"Defines the standard MEA node-level measure of average firing rate used as a dependent variable."}],"review_version":1}