{"id":"aefd9fe6-ec9a-4104-9c3c-32930e43c589","arxiv_id":"2412.17404","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"BrainMAP adaptively sequences brain-region graphs, routes them through a Mixture of Experts of Mamba models, and achieves state-of-the-art results on five HCP fMRI benchmarks with saliency-based explanations.","lead":"BrainMAP is a new deep learning method that turns brain connectivity graphs into ordered sequences and uses several sequence-model experts to predict behavior and highlight important brain regions. It reports accuracy gains over existing graph models on five HCP fMRI datasets and claims to reveal the multiple activation pathways the brain uses during tasks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'activation pathway' claim is not established: the order learner is trained purely by task-loss contrast, and the paper never validates that learned orders correspond to neural pathways.","rationale":"The reader's weakest assumption correctly identifies the task-loss-to-pathway proxy as the core gap. I agree that this assumption is load-bearing: the central claim explicitly says performance 'demonstrates its ability to extract long-range dependencies between brain regions along multiple pathways.' Without validation that the learned node order corresponds to a neural activation pathway, that inference is unsupported even if the accuracy numbers hold. I add a further point: even if the proxy were valid in principle, the architecture does not explicitly identify graph paths; it creates one total order per order-learner and applies sequence models over that entire order. The MoE experts all see the same full sequence, so 'multiple pathways' are a post-hoc interpretation rather than an architectural guarantee. The explanation study measures saliency of regions, not ordered path structure, so it does not close the gap. The proposed concrete test would directly resolve whether learned orders have path-like structure and whether that structure is responsible for the reported gains. I do not see a reason to move beyond the reader's CONDITIONAL verdict: the mechanical errors (impossible MRR, swapped columns, missing validation split) plus the unvalidated pathway interpretation warrant conditional acceptance rather than rejection, because the core architecture is novel and the predictive results, if reproducible, are still of interest even without the pathway interpretation. The concern is not that the accuracy numbers are fraudulent; it is that the paper's biological interpretation overreaches what the experiments demonstrate. This keeps the reader's verdict unchanged while sharpening the required condition.","tokens_in":16262,"tokens_out":5356,"duration_ms":57283,"concrete_test":"On HCP-Task MOTOR, train BrainMAP and extract the learned order for each test graph. Measure: (i) the fraction of adjacent node pairs in the learned order that are connected by functional edges above the 90th percentile FC weight; (ii) whether known motor ROIs from HCP task contrasts or NeuroSynth maps appear as contiguous blocks in the learned order; and (iii) whether the task-loss improvement of the learned order over a random permutation is larger than the standard deviation across seeds reported in Table 2. If learned orders are not significantly more path-like than random permutations, or the improvement over random order is negligible, then the activation-pathway interpretation fails and the paper should be reframed as an adaptive-permutation sequence model rather than a pathway-extraction method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim has two parts: (1) BrainMAP achieves superior predictive performance, and (2) this demonstrates extraction of long-range dependencies along multiple activation pathways. Part (1) is plausible from Table 2, though table typos and the missing validation split reduce confidence. Part (2) is unsupported. In Section 4.1, Eq. 3 and Eq. 6 train the order-learning GNN by contrasting sampled orders with low vs. high downstream task loss; the only supervision signal is task performance, not any neural-pathway ground truth. Positive and negative orders are selected by the same model's loss, so the learned order is rewarded for being a convenient permutation for the Mamba sequence model, not for corresponding to an actual signal-propagation path. The architecture then feeds each expert a full linear sequence of all N nodes, so 'multiple pathways' are not enumerated or verified as graph paths. The explanation study in Section 5.4 only computes saliency scores for individual brain regions and never compares learned orders to known pathways. Therefore, even if the accuracy improvements are real, the pathway-level interpretation stated in the central claim is not justified; the contribution may reduce to an adaptive graph sequentialization trick that improves prediction without recovering meaningful neural pathways.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BrainMAP, a framework for prediction on fMRI functional-connectivity graphs. BrainMAP first learns one or more node orderings via an order-learning GNN, then feeds each ordering to a small set of top- gated sequential models (Mamba or Transformer experts) and aggregates the resulting representations. The authors claim that this procedure extracts 'multiple activation pathways' in brain networks, yields state-of-the-art accuracy on five HCP datasets, and provides saliency-based explanations that rank task-relevant brain regions. The paper includes ablations, an efficiency study, and proofs for two theorems about the approximate-rank contrastive loss.","tokens_in":16519,"tokens_out":8357,"duration_ms":78320,"significance":"If the empirical results are sound, the paper offers a practical recipe for combining adaptive graph sequentialization with mixture-of-experts sequence models, and it addresses a real limitation of message-passing GNNs on long-range brain-network dependencies. The paper is also commendable for providing code, evaluating on five datasets, running ablations, and attempting an interpretation study. However, the central conceptual claim that the method learns 'activation pathways' is not validated: the order learner is optimized purely through the predictor's task loss, and the explanation experiments only measure node-level saliency. The predictive gains are also not supported by significance testing, and the reported MRR values are impossible as written. The contribution is therefore best viewed as an accuracy-oriented sequentialization technique whose pathway interpretation needs either direct validation or substantial softening.","major_comments":[{"comment":"The paper's central claim that BrainMAP learns 'activation pathways' is not supported by any validation of the learned orders against known neural pathways. The order learner is trained by contrasting sampled permutations that happen to give low downstream task loss (the 'positive' set) with permutations that give high task loss (the 'negative' set); the only supervision signal is the predictor's own loss, not any pathway ground truth. Moreover, each expert consumes a full linear ordering of all N nodes, so 'multiple pathways' are never enumerated or verified as graph paths. The explanation study in §5.4 only aggregates node saliency scores; it never compares the learned order to known anatomical or functional pathways. At minimum, the pathway-language claims should be softened, or the authors should provide a direct validation that learned orders correspond to plausible activation pathways (e.g., agreement with known task-evoked networks or effective-connectivity estimates).","section":"§4.1, Eqs. (3) and (6); §5.4"},{"comment":"The headline 'outperforms all baselines' claim is not statistically supported. All experiments used only three seeds and no significance tests are reported, and under the reported standard deviations the BrainMAP mean overlaps with the best baseline on every dataset (e.g., HCP-Age: 48.44±1.65 vs. 46.35±2.73; HCP-Gender: 78.92±0.49 vs. 77.16±3.13; HCP-WM: 3.81±0.03 vs. 3.94±0.14). The authors should report paired significance tests across seeds or across subjects, or otherwise demonstrate that the improvements are not within noise.","section":"§5.2, Table 2"},{"comment":"The MRR values in Table 4 (3.07, 3.13, 6.27, 9.26) are impossible for a Mean Reciprocal Rank, which is bounded above by 1. Either the metric is misnamed or the computation is incorrect, and since this table is the quantitative evidence for the explanation claim, the results need to be recomputed and the metric defined precisely. The text also does not specify which 'Mamba-specific explanation method' was used or how the ground-truth activated regions were obtained.","section":"§5.4, Table 4"},{"comment":"The experimental protocol does not describe how data are split into train, validation, and test sets. Appendix C states that hyperparameters are chosen by grid search and each configuration is run for 100 epochs, but it does not state whether the reported test numbers correspond to the best validation configuration or to the best test configuration. Without a clear split and selection protocol, the reported superiority could be optimistic. Please specify the split procedure, number of folds or seeds, and how early stopping or model selection was performed.","section":"§5.1, Appendix C"}],"minor_comments":[{"comment":"The text refers to 'Table 5' for the main results, but the main results are in Table 2; the reference should be corrected.","section":"§5.2"},{"comment":"The rows for HCP-FI and HCP-WM appear to be switched relative to Table 2: Table 2 reports HCP-WM as 3.81 and HCP-FI as 10.75, while Table 3 reports HCP-FI as 3.81 and HCP-WM as 10.75. The column labels and values should be checked.","section":"Table 3"},{"comment":"The proof of Theorem 4.2 states that Spearman's r satisfies 0 ≤ rs ≤ 1; this is false, since Spearman's rank correlation can be negative. The subsequent argument only needs rs < 1, so the proof can be repaired, but the claim should be corrected.","section":"Appendix B, after Eq. (26)"},{"comment":"The contrastive objective in Eq. (3) includes a weight λ that does not appear in the actual loss in Eq. (6). The relationship between these two formulations should be explained.","section":"§4.1, Eqs. (3) and (6)"},{"comment":"The abbreviations 'w/o LR' and 'w/o LB' are used without definition; the text should state that LR refers to the order learner and LB to the load-balancing loss.","section":"§5.3, Table 3"},{"comment":"The abstract gives the code URL as https://github.com/LzyFischer/Graph-Mamba, while the body gives https://github.com/LzyFischer/BrainMAP; the correct link should be used consistently.","section":"Abstract and body"}],"recommendation":"major_revision","confidential_remarks":"For the editor: this is a borderline paper. The predictive-accuracy contribution is plausible but not yet statistically established, and the central 'activation pathway' interpretation is currently an unsupported gloss on a task-loss-driven ordering procedure. The MRR and split-protocol problems are concrete and fixable. If the authors add significance testing, correct the explanation metrics, and either validate the pathway interpretation or substantially soften that claim, the paper could become publishable. I would not recommend rejection solely for disagreeing with the pathway metaphor, but the current manuscript overclaims relative to what is actually measured."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Good paper to know about, but don't take the 'activation pathway' language at face value. The architecture is a genuine novelty — adaptive graph sequentialization with learned permutations, fused with a mixture of Mamba experts — and the empirical work is mostly honest: five HCP datasets, ablations, and a correct derivation showing the contrastive loss matches Spearman rank correlation. The accuracy gains are modest but consistent, and the method beats Graph-Mamba on every reported dataset.\n\nThe soft spot is exactly where the stress-test points. The order learner is trained entirely by the model's own task loss: sample orders, keep the low-loss ones as positives and high-loss ones as negatives, and pull the learned order toward the positives. That is a way to find convenient permutations for the sequence model, but nothing in the training ties those permutations to neural activation pathways. No ground truth, no check that consecutive nodes in a learned order are connected by significant edges, no comparison with known anatomical pathways. So the central claim — that BrainMAP 'extracts long-range dependencies along multiple activation pathways' — is not established. What is established is that multiple learned orders, each processed by a gated expert, improve prediction. That's a useful engineering result, but it is not a neuroscience result.\n\nMinor issues are easy to list: MRR values in Table 4 are greater than 1 and the metric is undefined; Table 3 swaps the HCP-FI and HCP-WM columns; the abstract in the arXiv version points to the Graph-Mamba repo instead of the BrainMAP repo; there are no significance tests, and on HCP-Age and HCP-Gender the standard deviations overlap. None of these are load-bearing if fixed, but together they make the reported numbers less trustworthy than they should be.\n\nThe theoretical part is solid: the standardization trick in Eq. 4 and the equivalence proof in Theorem 4.2 are correct and genuinely useful for anyone working on differentiable sorting. The ablation shows both the learned order and the MoE matter, which supports the architecture's internal logic.\n\nFor a reader: if you work on graph sequence models or connectomics, this is worth a slot in a reading group, mainly as an example of how to make permutation learning work with task-loss feedback. It should be cited for the method, not for the pathway claim. I would send it to peer review — a good reviewer will ask for a validation of the pathway interpretation, or a softening of the language, plus the mechanical fixes. The core idea is strong enough to deserve that round.","headline":"The architecture is a genuine novelty and the accuracy gains are believable, but the 'activation pathway' interpretation is not established—the learned orders are shaped by task loss, not validated against any neural pathway ground truth.","tokens_in":17072,"tokens_out":4216,"would_cite":true,"duration_ms":38886,"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":"BrainMAP models brain graphs as ordered activation pathways and beats prior fMRI predictors.","keywords":["fMRI","functional connectivity graphs","activation pathways","graph sequentialization","mixture of experts","long-range dependencies","brain network analysis","Mamba"],"falsifier":"Take an fMRI dataset with independently measured activation order (for example, from BOLD latency or effective connectivity) for a set of tasks, train BrainMAP, and compare its learned orders with the gold-standard orders under Spearman correlation. If the learned orders are no more aligned than random orders of the same regions, the pathway interpretation is falsified. A second check: randomize the learned orders while approximately preserving task loss; if accuracy does not drop, the sequentialization itself is not carrying the claimed signal.","tokens_in":16028,"feed_emoji":"🧠","tokens_out":9856,"duration_ms":83405,"temperature":0.7,"pith_summary":"BrainMAP is a framework for predicting behavior and demographics from fMRI-derived functional-connectivity graphs. Instead of relying on message passing over static edges, it learns an ordering of brain regions that approximates sequential information flow, feeds the ordered sequence to a mixture of sequential models, and aggregates the results. The paper argues that this captures multiple long-range activation pathways—chains of brain regions that co-activate during a task—which conventional graph models miss. On five real-world fMRI benchmarks it outperforms all baselines, and its saliency maps rank known task-relevant brain regions higher than prior models. The payoff is both better prediction and a more interpretable account of which brain regions work together for a task.","feed_headline":"BrainMAP beats prior fMRI models by reading brain graphs as pathways","feed_subtitle":"On five fMRI benchmarks it improves accuracy by up to 4.09% and locates known task-relevant regions better.","key_machinery":"Adaptive Graph Sequentialization: an order-learning graph neural network produces per-node scores; sorting these scores yields a candidate activation order. Because real ranks break differentiability, the paper standardizes scores to share the first two moments of a true rank variable (Theorem 4.1) and trains the order with a loss that is equivalent to maximizing Spearman rank correlation with sampled low-loss orders and minimizing it with high-loss orders (Theorem 4.2). Hierarchical Pathway Integration: for each learned order, a top-k gated mixture of P expert sequence models (selective state-space models such as Mamba) captures diverse pathways; the gating function uses self-attention and positional encoding so similar pathways consistently route to the same expert, and a weighted sum over M orders produces the final graph embedding.","core_discovery":"The central discovery the paper tries to establish is that a functional-connectivity graph can be re-expressed as several node sequences whose order encodes the direction of neural signal flow, and that this re-expression makes long-range multi-pathway structure learnable. The order is not fixed: an order-learning graph neural network assigns each brain region a score, and a contrastive objective pulls the resulting order toward sampled good orders (those with low task loss) and away from bad orders. A hierarchical mixture of sequence models then extracts multiple pathways within each order and merges representations across orders. The paper reports that this design outperforms all baselines on five HCP prediction tasks, that the mixture-of-experts component is the largest single contributor in ablations, and that the model's saliency maps identify expert-defined motor-task regions better than comparison models.","pith_inferences":["If the learned order truly approximates neural activation order, the same sequentialization-plus-mixture recipe could transfer to other domains where signal propagation follows ordered paths, such as gene regulatory cascades, epidemic spread, or circuit traces.","The order learner uses task loss as a proxy for pathway structure, so a stronger test would compare learned orders against gold-standard effective-connectivity or BOLD-latency orderings; if they diverge, the accuracy gains might come from a sequential inductive bias rather than faithful pathway discovery.","The interpretation results suggest BrainMAP could generate candidate brain regions for tasks where expert ground truth is unknown, and those candidates could then be validated with targeted fMRI experiments.","Because expert count saturates around four, the gating distribution could be used to estimate the effective number of pathways per task and check whether that count aligns with known parallel pathway organization, such as the dorsal and ventral visual streams."],"forward_implications":["On five HCP fMRI prediction tasks, BrainMAP outperforms all baselines, with improvements up to 4.09% over the strongest prior model and gains of up to 12.13% over traditional graph neural networks on HCP-Task.","Ablation studies show that both the learned sequentialization and the mixture-of-experts aggregation contribute to performance, with the mixture-of-experts being the most critical component and the load-balancing loss also helping.","For the MOTOR task, BrainMAP's saliency regions achieve Hit@10 of 19.38 and MRR of 9.26, outperforming the strongest sequence baseline (Hit@10 of 15.00 and MRR of 6.27), indicating better alignment with expert-defined activated brain regions.","Performance improves as the number of experts increases up to three and roughly plateaus beyond four, suggesting a limited number of latent activation pathways in the brain.","BrainMAP's average training time per epoch is higher than all baselines, ranging from about 9.23 to 34.93 seconds across datasets on four A100 GPUs."],"supporting_citations":[{"why":"Supplies the selective state-space sequence model used as each expert, giving linear-time extraction of long-range dependencies.","marker":"Gu and Dao 2023"},{"why":"Provides the NeuroGraph benchmark, the processed HCP datasets, and baseline configurations that BrainMAP is evaluated against.","marker":"Said et al. 2023"},{"why":"The state-of-the-art graph sequence baseline that BrainMAP compares against and extends with multiple pathways.","marker":"Wang et al. 2024a"},{"why":"Defines the rank correlation coefficient that the order-learning loss optimizes through differentiable approximate ranks.","marker":"Spearman 1904"},{"why":"Introduces the sparsely-gated mixture-of-experts mechanism that motivates the pathway aggregation module.","marker":"Shazeer et al. 2017"},{"why":"Provides the self-attention mechanism and positional encoding used in the gating function to select experts.","marker":"Vaswani et al. 2017"},{"why":"The source of the Human Connectome Project fMRI data used in all five datasets.","marker":"Van Essen et al. 2013"}],"fun_headline_variants":["BrainMAP maps brain's multiple activation pathways, boosting fMRI accuracy","BrainMAP beats five fMRI benchmarks by modeling pathways as orders","BrainMAP finds hidden brain pathways in fMRI, outperforming baselines","BrainMAP learns multiple brain activation routes, revealing key task regions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that an order of brain regions which lowers the task-prediction loss is genuinely closer to the order in which neural signals travel; if task loss is not a faithful proxy for pathway structure, the activation pathway interpretation is unsupported even if prediction accuracy still improves.","fun_headline_variants_meta":{"raw":{"variants":["BrainMAP maps brain's multiple activation pathways, boosting fMRI accuracy","BrainMAP beats five fMRI benchmarks by modeling pathways as orders","BrainMAP finds hidden brain pathways in fMRI, outperforming baselines","BrainMAP learns multiple brain activation routes, revealing key task regions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001281,"raw_usage":{"total_tokens":5216,"prompt_tokens":909,"completion_tokens":4307,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":4236}},"tokens_in":525,"tokens_out":4307,"duration_ms":31521,"temperature":1.0,"reasoning_tokens":4236,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:26:51.854754+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take an fMRI dataset with independently measured activation order (for example, from BOLD latency or effective connectivity) for a set of tasks, train BrainMAP, and compare its learned orders with the gold-standard orders under Spearman correlation. If the learned orders are no more aligned than random orders of the same regions, the pathway interpretation is falsified. A second check: randomize the learned orders while approximately preserving task loss; if accuracy does not drop, the sequentialization itself is not carrying the claimed signal.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the NeuroGraph benchmark, the processed HCP datasets, and baseline configurations that BrainMAP is evaluated against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the rank correlation coefficient that the order-learning loss optimizes through differentiable approximate ranks."},{"cited_title":"C.; Smith, S","cited_arxiv_id":null,"evidence_quote":"The source of the Human Connectome Project fMRI data used in all five datasets."}],"review_version":1}