{"id":"5fd48079-00cb-4ea8-9042-7808acc8bfa2","arxiv_id":"1908.09117","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Heat kernel features of energy transport differ significantly between autism and controls in peripheral brain subnetworks, while hub subnetwork features remain similar.","lead":"This paper applied a heat diffusion model to brain scans from people with and without autism and found that differences concentrate in peripheral connections, not the central hubs. It offers a new way to describe how autism affects whole-brain communication, but the effects are small and may depend on scan-site differences.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Control-derived hub labels make the 'preserved hub' claim unfalsifiable and may manufacture the peripheral gradient.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: hub and subnetwork labels are derived from a control-only, binarized group-average network, so the central claim of a preserved hub is not independently testable. This concern is concrete and grounded in the methods: Wgroup is built from controls alone (Section 2.2), and subject networks are masked by Wgroup (Section 2.3), meaning any edge absent from the control mask is treated as a non-edge and contributes zero direct energy transport. If autism reorganizes hubs, those reorganizations are systematically relabeled as feeder or non-edge differences, which could create the appearance of a peripheral gradient. The paper also lacks site correction and reports uncorrected p<0.05 stars alongside Bonferroni-corrected results, but the stratification issue is more central because it threatens the main interpretation, not just the robustness of the p-values. A sensitivity analysis using independently derived hub labels for each group would settle the question. Because the reader already assigned CONDITIONAL and this concern is precisely why that conditionality is warranted, no verdict adjustment is needed.","tokens_in":7935,"tokens_out":4011,"duration_ms":46117,"concrete_test":"Compute a separate binarized group-average adjacency matrix for the autism group using the same 75% retention threshold, identify its top-10 strength hub nodes, and compare the hub lists; then re-run the 32 subnetwork t-tests on each group using its own hub labels (or using the autism-derived labels for both, to break the control-only circularity). If the hub subnetwork gains significant group differences or the peripheral gradient disappears, the preserved-hub/peripheral finding is an artifact of control-derived stratification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that autism leaves the hub subnetwork preserved while peripheral subnetworks differ (Discussion, final paragraph). But the subnetwork labels and the subject matrices are both derived exclusively from a control binarized group-average network: Section 2.2 defines Wgroup using only controls and labels hub/feeder/seeder/non-edge edges relative to the top-10 strength nodes of Wgroup; Section 2.3 then sets each subject's network to the subject covariance masked by Wgroup. Consequently, the 'hub' subnetwork is, by construction, the set of edges that are consistently present in controls; autism-specific connections are forced to zero and assigned to the non-edge subnetwork. The observed 'preserved hub' is therefore not evidence that the functional core is preserved in autism—it only says that edges chosen for being reliable control hubs do not differ much, which is partly guaranteed by the selection. If autism alters hub organization, this design misclassifies true hubs as feeder/non-edge and moves genuine hub differences into peripheral layers. The increasing significance toward peripheral subnetworks, especially non-edge, could reflect the masking of ASD-specific edges rather than atypical peripheral energy transport.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an edge-centric analysis of resting-state fMRI functional connectomes in autism using heat kernels. The brain network of each subject is derived from a covariance matrix masked by a control-derived binary group network, Wgroup, and edges are classified into hub, feeder, seeder, and non-edge subnetworks. For each edge, the authors compute heat-kernel features (intrinsic time constant tc, peak heat hpeak, peak time tpeak, maximal heat change h'peak, and its time t'peak) over a range of diffusion times, average these features within each subnetwork, and test for group differences between autism and control subjects from ABIDE. The reported results show no significant hub-subnetwork differences but increasing significance in feeder, seeder, and especially non-edge subnetworks, which the authors interpret as a preserved central hub and atypical energy transport in peripheral subnetworks.","tokens_in":8156,"tokens_out":7561,"duration_ms":79350,"significance":"If the central claim is valid, the paper makes a useful methodological contribution by applying heat-kernel energy-transport features to hub-stratified subnetworks, and it adds to the growing literature on peripheral region involvement in autism. Strengths include the use of a large publicly available dataset (ABIDE), explicit correction for 32 multiple comparisons, acknowledgment of some limitations, and a theoretically motivated diffusion framework that, for nonnegative edge weights, captures indirect pathways. However, the central claim of a preserved hub and peripheral differences is currently undermined by the control-derived subnetwork definition, and the validity of the heat-kernel computation on the subject networks is not established because the covariance matrices can contain negative values. The lack of site and motion covariates in a multi-site dataset further weakens the group comparison. Because these issues affect the core results, the manuscript requires substantial revision and reanalysis before the findings can be considered reliable.","major_comments":[{"comment":"The hub and subnetwork labels are derived exclusively from the control group: Section 2.2 defines Wgroup from controls only, identifies the top-ten strength hubs from Wgroup, and labels edges relative to those hubs; Section 2.3 then masks every subject's covariance matrix with Wgroup. As a result, the hub subnetwork is, by construction, the set of edges that are reliably present in controls, and any autism-specific connections are forced to zero and assigned to the non-edge subnetwork. The reported 'preserved central hub' is therefore not evidence that the functional core is preserved in autism; it only shows that edges selected for being reliable control hubs do not differ between groups. If autism alters hub organization, this design misclassifies true hubs as peripheral and may manufacture the observed gradient of increasing significance toward peripheral subnetworks. I request a sensitivity analysis deriving hubs from the full sample or from each group separately, and a direct test of whether hub membership differs between groups.","section":"§2.2–§2.3, Table 3, Discussion (final paragraph)"},{"comment":"The subject network W is computed by multiplying the subject's covariance matrix, which can contain negative values, with the binary mask Wgroup. For a graph Laplacian to produce a well-defined diffusion process, edge weights must be nonnegative. With signed edge weights, the normalized Laplacian \\hat L = D^{-1/2}LD^{-1/2} is not guaranteed to be positive semidefinite, and exp(-t\\hat L) may not be a contractive, positivity-preserving heat kernel. The 'energy transport' interpretation is therefore not justified without additional assumptions. The authors should either confirm that the absolute value of the covariance is used for W, or demonstrate that the heat kernel remains physically meaningful for signed weights; this is load-bearing because every reported feature is derived from this matrix exponential.","section":"§2.3, Eqs. (1)–(2)"},{"comment":"ABIDE is a multi-site dataset, yet the group comparisons in Section 2.4 do not include site or motion as covariates, and Table 1 reports only age. Differences in site composition or head motion between the autism and control groups could confound the reported subnetwork differences, especially given the modest effect sizes in Table 3 (e.g., non-edge tc at 2%: 1.207 vs 1.215). The authors should add site and motion parameters as covariates or perform within-site analyses to rule out these confounds.","section":"§2.1 and §2.4, Table 1"},{"comment":"The intrinsic time constant tc in Equation (3) is defined using a criterion that divides by H_{u,v}(t). For the non-edge subnetwork, heat kernel values are near zero because these pairs have no direct edge in Wgroup, as the authors note in Section 4 and Figure 1. The division makes tc numerically unstable for these edges, and the non-edge subnetwork is exactly where the most significant group differences are reported (tc p<0.0001). The authors should demonstrate numerical stability of tc, for example by imposing a minimum denominator threshold, and report the range of H_{u,v}(t) values for the non-edge subnetwork. Without this, the leading result may be an artifact of floating-point noise.","section":"Eq. (3), Table 3 (non-edge row)"}],"minor_comments":[{"comment":"The phrase 'it's hub' should be 'its hub'.","section":"Abstract"},{"comment":"The text states that the top ten nodes with greatest strength are selected as hubs, but Table 2 lists only eight regions. Please clarify the discrepancy or add the missing entries.","section":"§2.2, Table 2"},{"comment":"Equation (4) appears to contain a typo: 'max|Hu,t(t)|' should read 'max|H_{u,v}(t)|'.","section":"Eq. (4)"},{"comment":"In the seeder row for tc, 2%, the control standard deviation '0.023' is missing its opening parenthesis; it should be '(0.023)'.","section":"Table 3"},{"comment":"Reference [17] spells the author name as 'Mller'; this should be 'Müller'.","section":"References"},{"comment":"The caption says 'Plots of mean values in the heat kernel matrix by subnetwork'; it would be clearer to state that the plotted quantity is H_{u,v}(t) averaged over all edges in each subnetwork.","section":"Figure 1 caption"}],"recommendation":"major_revision","confidential_remarks":"The control-derived hub stratification issue is severe because it directly undermines the paper's central conclusion; however, it is addressable through sensitivity analyses, so I do not recommend rejection at this stage. The signed-weight issue in Section 2.3 is even more fundamental: if the subject networks use raw covariance values, the heat kernel may not be a legitimate diffusion process. The authors should be asked to clarify this point first, as it determines whether the entire computational framework is valid. I would also encourage the editor to require the site-covariate analysis given the multi-site ABIDE dataset."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a five-page extension of the authors' prior heat-kernel and subnetwork frameworks, applied to ABIDE autism data. The empirical result is new for this dataset and question: heat-kernel features differ in the feeder, seeder, and non-edge layers, while hub-layer features do not. The paper is honest about its limits, but the main finding rests on a stratification choice that biases it, and the multi-site structure of the data is not in the model.\n\nWhat is genuinely good: the heat kernel encodes all paths, so edge-level features carry whole-network information; the non-edge category is a sensible fourth bin; the sample is large; Bonferroni correction is used properly; and the authors state plainly that they do not yet know what the feature differences mean. The pipeline is described well enough to reproduce on public data. This is honest plumbing, not an overclaim.\n\nThe soft spots, in order of importance. First, hubs and the subnetwork skeleton are defined from a binarized control group average (edges present in at least 75% of controls), and each subject's matrix is the covariance multiplied by that control mask. Autism-specific connections are zeroed before the heat kernel is computed. So 'preserved hub' says: edges selected for being consistent control hubs have similar heat kernel values in both groups — partly guaranteed by selection. If autism rewires the hub organization, those changes land in the non-edge layer, which is exactly where the largest differences appear. The hub-to-periphery gradient could be manufactured by the mask. The stress-test note says 'unfalsifiable' — that is a bit strong, because the hub subnetwork could still show group differences under this design; it simply did not. But the worry is real, and the authors should report sensitivity runs with subject-specific or combined-group hubs and with the unmasked covariance. Second, ABIDE is multi-site and the age range is 6–56, yet the comparisons are t-tests with no site or motion covariates. That is a serious gap for a dataset known for site effects. Third, a technical point: the group connectome is absolute-valued, but the subject matrices are not described as such. If negative covariance entries survive, the normalized Laplacian is non-standard, and the diffusion reading of the heat kernel needs a defense. Also, tc in the non-edge layer is computed on values near zero and is numerically fragile; the reported differences are tiny — a fraction of a percent against standard deviations of a few percent. Effect sizes throughout are small.\n\nOverall, the pattern is plausible but underdetermined. I do not think the central claim is refuted; I think it is not yet tested. This paper is for people working on connectome subnetwork analysis: worth one read, and worth a serious referee, but the referee should ask for the sensitivity analyses above. My call: send it to peer review, expect a request for major revision.","headline":"A clean, small extension of the authors' own heat-kernel and subnetwork methods to ABIDE autism data; the peripheral-layer result is new and plausible, but the control-derived hub mask and missing site covariates leave the main claim underdetermined.","tokens_in":8660,"tokens_out":7361,"would_cite":false,"duration_ms":73617,"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":"In autism, heat-kernel energy-transport features differ significantly in peripheral subnetworks while the hub subnetwork remains indistinguishable from controls.","keywords":["connectome","heat kernel","energy transport","subnetworks","hubs","autism","resting-state fMRI","functional connectivity"],"falsifier":"Recompute the top-ten hub nodes from the autism group's own average connectome and re-run the heat-kernel subnetwork comparisons. If the hub list differs substantially between groups, or if the autism-defined hub subnetwork shows significant group differences, then the claimed preservation of the hub is an artifact of the control-derived stratification.","tokens_in":7747,"feed_emoji":"🧠","tokens_out":5229,"duration_ms":49286,"temperature":0.7,"pith_summary":"This paper argues that in autism the brain's functional core, its hub subnetwork, transports energy much as it does in controls, while the peripheral subnetworks (feeder, seeder, and non-edge connections) show measurable, statistically significant differences in how energy spreads over time. The authors model each subject's resting-state fMRI connectome as a graph and solve a diffusion process on it with heat kernels, which capture energy flow through all direct and indirect pathways rather than only single edges. They apply this to hundreds of controls and autism subjects from a large public multi-site dataset, stratifying edges by distance from hub nodes. The result matters because it shifts attention from hub regions to peripheral regions in autism and offers time-based features that could describe the disorder's distributed impact.","feed_headline":"Autism alters energy flow in brain's periphery, not its hub","feed_subtitle":"Heat-kernel analysis of fMRI connectomes finds significant transport changes in feeder, seeder and non-edge subnetworks.","key_machinery":"The heat kernel $H(t)=\\exp(-t\\hat L)$ is the fundamental solution to the diffusion equation $\\partial H/\\partial t = -\\hat L H$ on the graph; entry $H_{u,v}(t)$ is the amount of energy transferred between regions $u$ and $v$ through all pathways after time $t$. From a sweep of $t$ values it yields eight edge features per connection: the intrinsic time constant $t_c$ (at several thresholds), the peak energy $h_{\\mathrm{peak}}$, its time $t_{\\mathrm{peak}}$, the maximal change $h'_{\\mathrm{peak}}$, and its time $t'_{\\mathrm{peak}}$. These features are averaged within subnetworks defined by each edge's relation to the top-ten strength hubs of the control group's binarized group-average connectome.","core_discovery":"Using heat kernels $H(t)=\\exp(-t\\hat L)$ derived from the normalized Laplacian of each subject's functional connectivity matrix, the authors compute five types of edge-wise energy-transport features over time, average them within hub, feeder, seeder, and non-edge subnetworks, and compare groups. They find no significant group differences in the hub subnetwork, but significant differences in all three peripheral subnetworks, with the largest effects in the non-edge subnetwork. In autism, the time-related features $t_c$, $t_{\\mathrm{peak}}$, and $t'_{\\mathrm{peak}}$ are shifted later in peripheral subnetworks, indicating that peak energy transfer occurs later even though overall transport profiles look similar. The authors interpret this as a preserved functional core with atypical energy dispersion in peripheral regions.","pith_inferences":["Because hubs are defined from controls only, a separate analysis that recomputes hubs within the autism group would directly test whether the 'preserved hub' is real or a consequence of the stratification; this is my extension, not the paper's.","A natural next step would be to test whether the later timing of peak energy transfer in peripheral subnetworks correlates with clinical severity or age, which the paper does not do.","The same heat-kernel subnetwork pipeline could be applied to structural connectomes or to task-based fMRI to see whether the peripheral energy-transport signature is specific to resting-state functional organization."],"forward_implications":["If peripheral subnetworks carry the autism-related signal, studies of autism should weight feeder, seeder, and non-edge connections at least as heavily as hub connections.","Heat-kernel timing features, especially the later $t_c$ and $t_{\\mathrm{peak}}$ values in autism, could serve as candidate biomarkers in functional-connectivity analyses.","The non-edge subnetwork's strong group differences link energy-transport features to a network's small-world capacity, suggesting a direct test of small-world propensity in autism.","The preserved hub result supports the view that developmental or compensatory changes in autism spare the functional core while altering flexible peripheral communication."],"supporting_citations":[{"why":"Supplies the heat-kernel feature framework for connectomes that this paper adapts to subnetwork analysis.","marker":"[6]"},{"why":"Defines the hub/feeder/seeder subnetwork stratification based on connection to hub nodes.","marker":"[24]"},{"why":"Supplies the large multi-site autism imaging dataset analyzed in the study.","marker":"[10]"},{"why":"Source for using node strength to identify hub regions and rich-club-style grouping.","marker":"[8]"},{"why":"Prior subnetwork pathology study showing core preserved and peripheral regions more affected, which the authors compare with.","marker":"[30]"},{"why":"Supplies evidence of atypical functional hierarchy in autism and the recruitment of seeder regions in long indirect pathways.","marker":"[16]"},{"why":"Reports globally atypical network organization in the same cohort with overlapping hub findings across metrics.","marker":"[17]"}],"fun_headline_variants":["Autism shifts energy timing in periphery, hub preserved","Autism's energy flow deviates in peripheral subnetworks, not hub","Peak energy transfer delayed in autism's peripheral brain regions","Autism alters energy transport in non-hub brain subnetworks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central result assumes that hubs derived from the control group's average connectome are the correct hubs for both groups; if autism reorganizes the hub structure itself, then labeling both groups with control hubs could manufacture a 'preserved hub' and push all group differences to the periphery.","fun_headline_variants_meta":{"raw":{"variants":["Autism shifts energy timing in periphery, hub preserved","Autism's energy flow deviates in peripheral subnetworks, not hub","Peak energy transfer delayed in autism's peripheral brain regions","Autism alters energy transport in non-hub brain subnetworks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00074,"raw_usage":{"total_tokens":3316,"prompt_tokens":967,"completion_tokens":2349,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":2288}},"tokens_in":583,"tokens_out":2349,"duration_ms":18244,"temperature":1.0,"reasoning_tokens":2288,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:20:42.876407+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the top-ten hub nodes from the autism group's own average connectome and re-run the heat-kernel subnetwork comparisons. If the hub list differs substantially between groups, or if the autism-defined hub subnetwork shows significant group differences, then the claimed preservation of the hub is an artifact of the control-derived stratification.","supporting_citations":[{"cited_title":"Neuroimage 141, 490–501 (2016) 10 Schirmer and Chung","cited_arxiv_id":null,"evidence_quote":"Supplies the heat-kernel feature framework for connectomes that this paper adapts to subnetwork analysis."},{"cited_title":"In: International Workshop on Connectomics in Neuroimaging","cited_arxiv_id":null,"evidence_quote":"Defines the hub/feeder/seeder subnetwork stratification based on connection to hub nodes."},{"cited_title":"Molecular psychiatry 19(6), 659 (2014)","cited_arxiv_id":null,"evidence_quote":"Supplies the large multi-site autism imaging dataset analyzed in the study."},{"cited_title":"Schizophre- nia Bulletin 40(2), 438–448 (2013)","cited_arxiv_id":null,"evidence_quote":"Source for using node strength to identify hub regions and rich-club-style grouping."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior subnetwork pathology study showing core preserved and peripheral regions more affected, which the authors compare with."},{"cited_title":"Nature communications 10(1), 1022 (2019)","cited_arxiv_id":null,"evidence_quote":"Supplies evidence of atypical functional hierarchy in autism and the recruitment of seeder regions in long indirect pathways."},{"cited_title":"Biological Psychiatry: Cognitive Neuroscience and Neuroimaging 2(1), 66–75 (2017)","cited_arxiv_id":null,"evidence_quote":"Reports globally atypical network organization in the same cohort with overlapping hub findings across metrics."}],"review_version":1}