{"id":"e704a465-fcde-4593-a318-00ece90eed75","arxiv_id":"2412.14999","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Embedding fMRI-derived functional sub-circuits, particularly the default mode network, into echo-state reservoirs improves performance over null models on memory and decision tasks.","lead":"This paper tests whether wiring human brain functional networks into echo-state networks improves performance on memory and decision tasks. It reports that brain-derived sub-circuits usually beat random or simpler wiring, and that the default mode network stands out.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing premise is the one-to-one alignment of 463/1015 structural nodes with Yeo-7 labels from a different parcellation (Schaefer-500, HCP-YA); §5.3 concedes these parcellations may not be aligned, so FN-specific rankings could be mislabel artifacts.","rationale":"The paper is a credible empirical study: structural reservoirs are compared with degree- and weight-randomized nulls at matched node counts, and the DMN advantage appears consistently across tasks and many nulls. The central scientific claim, however, is that the neuro-physiological identity of the sub-circuit, not merely reservoir size or spectral radius, drives performance. That claim depends entirely on the assignment of Yeo functional labels to structural connectome nodes. The reader's weakest assumption identifies exactly this dependency, and I agree. The omission is not merely cosmetic: the manuscript uses structural connectomes from one cohort/parcellation (463/1015 nodes) and functional labels from a different cohort/parcellation (Schaefer-500), and §5.3 itself states that these parcellations are not information-theoretically aligned. A label-permutation null is mentioned, which would be the decisive control, but its construction and results are not available in the text under review because the appendix pointer is missing. This does not warrant rejection: the empirical machinery is standard, and the permutation control could resolve the concern if properly reported. It does, however, keep the correctness risk at medium and supports conditional acceptance pending the explicit mapping/permutation check. Therefore the reader's CONDITIONAL verdict should be unchanged.","tokens_in":12449,"tokens_out":9015,"duration_ms":78424,"concrete_test":"Recompute the α=0.95 read-out comparisons (Figures 8-9) and the statistic-performance regressions (Figure 11) using Yeo labels assigned by direct anatomical overlap between the 463/1015-node structural parcellation and the Schaefer-500 parcellation in a common coordinate space, and repeat with an independent atlas (Schaefer-400 or Glasser/HCP-MMP). If the DMN/SM ranking and Figure 11 correlations do not survive in the concordant subset, the reported FN-specific effects are artifacts of the unverified node-mapping assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 combines structural consensus connectomes parcellated into 463 or 1015 nodes [2] with Yeo-7 functional-network labels obtained from a Schaefer-500 resting-state fMRI parcellation [49]. The structural and functional node sets differ in size, come from different cohorts, and the text never states how a structural node receives a Yeo label (nearest-parcel assignment, atlas overlap, or other). If this mapping is wrong, the induced subgraphs of §3.3 and the best-performance counts by functional network in Figures 8-9 are not evaluating the named circuits. Figure 11's communicability and group-betweenness correlations with PDMDR performance inherit the same flaw. The manuscript itself, in §5.3, states that Yeo networks derived from the Lausanne parcellation [2] and the Schaefer parcellation [49] are not information-theoretically aligned, an explicit admission that the central mapping is not secure. A label-permutation null is listed in §4.4 and Figures 8-9, which would be the right control, but its construction and results are not reported in the available text, and the pointer to the appendix is missing ('Appendix ??'). Without either a principled node-mapping description or a fully reported label-permutation null, the abstract's claim that performance optimums depend on the neuro-physiological characteristics of the sub-circuits remains unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a pipeline for embedding human brain functional subcircuits, derived from fMRI, into echo-state networks (ESNs) whose reservoir connectivity is given by structural connectomes. The authors evaluate several reservoir topologies—structural connectome, vertex-induced subgraph, Maslov–Sneppen rewired, uniform-weight, complete bipartite, configuration model, and label-permutation null—on four synthetic tasks (PDM, PDMDR, CDM, MemCap) across a range of spectral radius values. They find that structural and subgraph reservoirs outperform degree-preserving nulls and bipartite feedforward-like nulls, and that the ranking of read-out functional networks differs across tasks, with DMN performing best in three of four tasks at near-criticality. The paper then correlates graph statistics (communicability, modularity, group betweenness) of the functional subcircuits with task performance to argue that neuro-physiological characteristics of the embedded subcircuits determine reservoir performance.","tokens_in":12678,"tokens_out":4447,"duration_ms":29692,"significance":"If the central finding holds, it would show that the specific mesoscale functional topology embedded in a reservoir—not just spectral radius or raw network size—shapes computational performance, offering a neuroscience-inspired design principle for reservoir computing. The experimental design has notable strengths: multiple null models ablate distinct structural properties, four tasks span different computation–memory balances, 1000 runs are used, and several spectral radius values are swept. However, the load-bearing identification of functional network labels on structural nodes is not established, and a critical control (the label-permutation null) is referenced but not described or numerically reported. Until these gaps are closed, the abstract's claim that performance optimums depend on the neuro-physiological characteristics of the subcircuits remains unverified. As a result, the paper's significance is conditional on a fixable but nontrivial methodological clarification.","major_comments":[{"comment":"The mapping between the 463/1015 structural nodes from [2] and the Yeo-7 functional labels obtained from the Schaefer-500 parcellation ([49]) is never specified. The two node sets differ in size, parcellation, and cohort, and the text does not state how a structural node receives a Yeo label (nearest-parcel assignment, atlas overlap, or other). Without this mapping, the induced subgraphs of §3.3 and all functional-network-specific analyses in Figures 8–11 do not necessarily correspond to the named brain circuits. The manuscript's own §5.3 hypothesizes that Yeo networks derived from the Lausanne and Schaefer parcellations are 'not information-theoretically aligned,' which is an explicit admission that the central node-label correspondence is insecure.","section":"§4.1, §3.2"},{"comment":"The label-permutation null is listed as a model and its results appear in Figures 8 and 9, but its construction is not described anywhere in the available text, and the pointer to the appendix is missing ('Appendix ??'). This null is the appropriate control for testing whether functional-network-specific rankings are artifacts of the node-label assignment. Without a description of how labels are permuted and what the outcomes are, the claim that performance differences reflect the neuro-physiological identity of the subcircuits is not supported by the reported evidence.","section":"§4.4, Figures 8–9"},{"comment":"The text cites the wrong figures when discussing the association between node count, betweenness, and best-performance counts: it refers to 'Figure 5' and 'Figure 4,' but those figures show performance versus spectral radius and reservoir configuration box plots, respectively. The relevant figures are Figures 10 and 8–9. The mis-citation obscures the basis for the claimed association between DMN's size, betweenness, and its top performance, and makes the analysis difficult to audit.","section":"§5.1"},{"comment":"The correlation between communicability and PDMDR performance is presented as a scatter plot with only seven data points (one per functional network), and the reported value ρ=2.245e-10 is stated to be the slope of the best linear fit, not a correlation coefficient. No p-value, confidence interval, or goodness-of-fit measure is given. The assertion that the betweenness panel shows 'four separate clusters' is made without a clustering criterion or any quantitative support. These analyses are too thin to sustain the strength of the conclusions drawn about which graph-theoretical properties determine ESN performance.","section":"§5.2, Figure 11"}],"minor_comments":[{"comment":"The COVID-19 prediction experiment is referenced as 'Table ??' but no results table is provided in the manuscript; this section should either be completed with the actual numerical results or removed.","section":"§4.5"},{"comment":"Several placeholders 'Appendix ??' appear in the text, indicating missing supplementary material. The experimental setup details and additional COVID-19 model descriptions promised in these references are not available for review.","section":"§4.2, §4.5"},{"comment":"The caption of Figure 7 states that F1-score performance is shown 'on three datasets,' but the displayed subplot is only labeled for ContextDecisionMaking; please clarify whether panels are omitted or add the missing subplots.","section":"Figure 7"},{"comment":"The formula Q = ∑_{i,j}(a_{ij} − p_{ij})(δ_{σ_i,u}, δ_{σ_j,v}) is not self-contained: the Kronecker delta notation and the meaning of the partition vector σ are not defined in the text.","section":"§5.3"},{"comment":"References [5] and [45] appear to cite the same Yeo et al. work with identical titles; please consolidate to avoid duplicate entries.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of a computational neuroscience or neuromorphic computing journal. The main obstacle to acceptance is the unverified node-label mapping between structural and functional parcellations and the unreported label-permutation null. These issues are potentially fixable within the manuscript's scope, so I recommend major revision rather than rejection. I would also urge the editor to ensure that the missing appendices and tables are complete in any revised version, since several load-bearing results appear to reside in material that was not provided."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know about this paper: it takes the connectome-ESN idea a step further by using fMRI-derived Yeo functional sub-networks not just as labels, but as the actual read-in and read-out node sets on structural-connectome reservoirs. If the mapping between the structural nodes and the functional labels is trustworthy, the result—DMN and somatomotor readouts win on three of four tasks, while a dorsal-attention/frontoparietal preference appears on the perceptual task—is a genuinely new and potentially useful heuristic for reservoir design. The machine is standard ESN with ridge regression; the novelty is in the embedding and the null-model design.\n\nWhat it does well: the experimental scaffolding is solid. They compare the structural control against degree-preserving rewiring, configuration model, uniform-weight randomization, a bipartite null, and a subgraph-only variant, across four tasks and a spectral-radius sweep, with 1000 instantiations. The pattern that structural topology beats degree-preserving nulls is visible and likely robust. The graph-statistics analysis (betweenness, communicability) is a reasonable way to connect physiology to performance, and the paper is honest enough to state in §5.3 that the Lausanne and Schaefer parcellations may not be information-theoretically aligned.\n\nThe soft spots are in the load-bearing parts. The paper never says how a structural node, parcellated into 463 or 1015 Lausanne regions, receives a Yeo-7 label from a Schaefer-500 functional parcellation. Different cohorts, different parcellations, different node counts—if the assignment is nearest-parcel or overlap-based, small errors could reshuffle the FN-specific rankings. The label-permutation null listed in §4.4 is the right control, but its construction and results are missing (the pointer is 'Appendix ??'). That is not a small omission; it is the experiment that would validate the central claim. The also-missing code and data hamper verification. The 'bipartite as MLP' analogy is fine as a null, but describing it as 'feed-forward properties commonly seen in MLP' overreaches—a single random layer is not an MLP. The correlation plots in Figure 11 have no error bars, which would matter more if the mapping issue were resolved.\n\nNone of this refutes the main pattern; it means the paper is not yet at the stage where the FM-specific conclusions can be accepted. The structural-versus-null comparison likely survives the mapping issue, but the abstract's stronger claim about neuro-physiological characteristics is unverified without the permutation control and a written mapping procedure.\n\nThis paper deserves a serious referee. I would send it to review, but with a clear request: fix the mapping description, report the label-permutation null in full, and make code/data available. For my own work, I would not cite the FN-specific rankings yet. Bring it to reading group, though—the null-model strategy and the DMN question will spark useful discussion.","headline":"The new idea is embedding Yeo functional networks as read-in/read-out modules on structural reservoirs, and it deserves a careful referee; but the node-mapping between parcellations is a load-bearing gap that needs explicit testing.","tokens_in":13278,"tokens_out":2977,"would_cite":false,"duration_ms":24897,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Echo-state networks whose reservoirs embed fMRI-derived brain functional sub-circuits outperform bipartite and degree-preserving null models, with the best readout network varying by task.","keywords":["echo-state networks","reservoir computing","connectome-based reservoir computing","functional sub-circuits","structural connectome","functional MRI","network topology","memory capacity"],"falsifier":"Permute the assignment of the seven functional-network labels to the structural nodes and rerun the full ESN pipeline across many seeds: if the default-mode advantage on delayed-response, memory, and contextual tasks disappears under label permutation while the untreated connectome keeps it, then functional identity, not node size or placement, is what drives performance. The paper's own permutation null (Figure 8) suggests exactly this check is needed.","tokens_in":12194,"feed_emoji":"🧠","tokens_out":7473,"duration_ms":55426,"temperature":0.7,"pith_summary":"By embedding human structural connectomes as echo-state network reservoirs and using fMRI-derived functional sub-circuits (Yeo's seven networks plus subcortex) as the read-in and read-out node sets, this paper asks whether biological functional topology carries computational benefit. The authors show that these functional-sub-circuit-embedded reservoirs generally beat bipartite and degree-preserving null models across perceptual decision, delayed response, context-dependent decision, and memory-capacity tasks near criticality. They also find that the best-performing readout network depends on the task: default mode and somatomotor networks lead on three tasks, while dorsal attention and frontoparietal lead on perceptual decision-making. The paper concludes that performance is shaped by the graph-theoretic properties of the embedded functional sub-circuits, such as betweenness and communicability, and that reservoir performance is not strictly determined by the spectral radius (echo-state property).","feed_headline":"Brain sub-circuits beat null topologies in echo-state networks","feed_subtitle":"Embedding Yeo fMRI networks as readouts changes which reservoir tasks echo-state models solve best.","key_machinery":"The pipeline is an echo-state network (ESN) whose reservoir adjacency matrix is a min-max-scaled, spectral-radius-normalized structural connectome, with input routed to subcortical nodes and the readout ridge-regression layer attached to nodes of one of the seven Yeo functional networks. The functional sub-circuits thus act as induced subgraphs defining read-in and read-out structure, and the comparison against null models (degree-preserving Maslov-Sneppen rewire, configuration model, uniform-weight randomization, and complete bipartite) ablates specific topological features. Communicability and group betweenness centrality quantify the dynamical and static properties of the readout subgraph that the paper links to performance.","core_discovery":"The central claim is that the topology of the embedded fMRI-induced sub-circuits, not merely the reservoir's spectral radius, determines echo-state network performance, and that different functional networks are optimal for different computational demands. Concretely, when the structural connectome is the reservoir and a functional network is the readout node set, the default mode network (DMN) and somatomotor network outperform other networks on delayed-response, memory, and contextual tasks, whereas dorsal attention and frontoparietal networks win on the perceptual decision-making task; these rankings are tied to betweenness centrality and communicability of the readout subgraph. The paper further claims that reservoirs preserving the structural topology (control and induced-subgraph models) outperform degree-sequence-preserving nulls (Maslov-Sneppen rewire and configuration model) and the bipartite feed-forward-like null, contradicting the notion that the echo-state property, as measured by spectral radius, alone sets performance.","pith_inferences":["If the functional-identity result is real, a natural extension is to design neuromorphic reservoirs by optimizing readout subgraphs for communicability to the input nodes, rather than by tuning spectral radius alone; the paper does not propose such an optimization.","A same-subject multimodal dataset (diffusion MRI plus fMRI with identical parcellation) would directly test whether the reported rankings survive when structural and functional data come from the same brains; until then, cross-cohort node correspondence remains the main confound.","The paper's own permutation null appears to disrupt the DMN advantage pattern in the figures; systematically comparing that null against the control across seeds and tasks would clarify whether functional identity or node size and placement drive the ranking.","Because communicability is a global path-based measure, the authors' correlational analysis could be extended to predict the performance of arbitrary subgraphs in connectome reservoirs, allowing synthetic testbeds that separate size, placement, and biological identity."],"forward_implications":["Task-specific readout choice becomes a design lever: choosing DMN or somatomotor readouts for memory-heavy tasks and dorsal attention or frontoparietal readouts for perceptual decisions can improve ESN performance near criticality.","Connectome-based reservoir performance cannot be predicted from spectral radius alone; topological features of the embedded subgraph must be considered.","Degree-preserving nulls (Maslov-Sneppen rewire and configuration model) underperform the true structural topology, implying that the specific wiring pattern of brain connectivity carries computational value beyond degree sequence.","The subgraph model performs almost identically to the full structural model on three of four tasks, suggesting information principally propagates through input-output pathways in this setup.","Uniform-weight nulls with preserved topology are competitive, indicating that edge weights matter less than topology for some tasks."],"supporting_citations":[{"why":"Supplies the structural consensus connectomes (463 and 1015 nodes) and the prior connectome-based ESN methodology that this work extends to fMRI sub-circuits.","marker":"[2]"},{"why":"The conn2res toolbox provides the ESN implementation (update rule and ridge-regression readout) used in all experiments.","marker":"[4]"},{"why":"Defines the seven intrinsic functional networks used as the a priori readout node-sets and the functional parcellation.","marker":"[5]"},{"why":"Source of the fMRI-derived Schaefer parcellation (500 parcels, Yeo 7-network schema) whose functional labels are mapped onto structural nodes.","marker":"[49]"},{"why":"The Maslov-Sneppen algorithm provides the degree-preserving edge-rewire null model that ablates topology while keeping degree sequence.","marker":"[46]"},{"why":"The configuration model supplies the degree-sequence-matching random graph null model.","marker":"[47]"},{"why":"Defines the echo-state network update rule and the echo-state property / spectral-radius condition that the paper examines.","marker":"[42]"},{"why":"Prior work claiming that modular structure of reservoirs shapes learning, which the paper extends and partially contrasts.","marker":"[40]"},{"why":"Distance-dependent consensus thresholds used in building the group-representative structural connectomes.","marker":"[48]"}],"fun_headline_variants":["Default mode network wins echo-state memory tasks","Dorsal attention and frontoparietal nets win perceptual ESN tasks","Sub-circuit topology, not just spectral radius, decides ESN tasks","fMRI sub-circuits trump null models in echo-state nets","Echo-state reservoirs select task-optimal brain networks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"That the nodes of the diffusion-MRI-derived structural connectome correspond one-to-one to the fMRI-derived Schaefer and Yeo parcels, despite coming from different cohorts and different parcellation schemes; if this mapping is wrong, every functional-network-specific performance ranking could be an artifact of mislabeled nodes.","fun_headline_variants_meta":{"raw":{"variants":["Default mode network wins echo-state memory tasks","Dorsal attention and frontoparietal nets win perceptual ESN tasks","Sub-circuit topology, not just spectral radius, decides ESN tasks","fMRI sub-circuits trump null models in echo-state nets","Echo-state reservoirs select task-optimal brain networks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0024,"raw_usage":{"total_tokens":9257,"prompt_tokens":994,"completion_tokens":8263,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":8179}},"tokens_in":610,"tokens_out":8263,"duration_ms":47033,"temperature":1.0,"reasoning_tokens":8179,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:43:36.981661+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Permute the assignment of the seven functional-network labels to the structural nodes and rerun the full ESN pipeline across many seeds: if the default-mode advantage on delayed-response, memory, and contextual tasks disappears under label permutation while the untreated connectome keeps it, then functional identity, not node size or placement, is what drives performance. The paper's own permutation null (Figure 8) suggests exactly this check is needed.","supporting_citations":[{"cited_title":"E., Richards, B","cited_arxiv_id":null,"evidence_quote":"Supplies the structural consensus connectomes (463 and 1015 nodes) and the prior connectome-based ESN methodology that this work extends to fMRI sub-circuits."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The conn2res toolbox provides the ESN implementation (update rule and ridge-regression readout) used in all experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the seven intrinsic functional networks used as the a priori readout node-sets and the functional parcellation."},{"cited_title":"Functional Connectome Fingerprint Gradients in Young Adults","cited_arxiv_id":"2011.05212","evidence_quote":"Source of the fMRI-derived Schaefer parcellation (500 parcels, Yeo 7-network schema) whose functional labels are mapped onto structural nodes."},{"cited_title":"F., Griffa, A., Hagmann, P","cited_arxiv_id":null,"evidence_quote":"Distance-dependent consensus thresholds used in building the group-representative structural connectomes."}],"review_version":1}