{"id":"69cf5dc7-13f3-41b9-8a7d-3d0db792ace3","arxiv_id":"2502.07746","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"HiPoNet combines learned feature reweighting, Vietoris-Rips complexes, and simplicial scattering transforms to classify high-dimensional point clouds, reporting top accuracy on several single-cell and spatial transcriptomics tasks.","lead":"HiPoNet is a neural network that classifies and regresses whole high-dimensional point clouds, such as patient-level single-cell datasets, by building several differently-weighted graphs and extracting multiscale topological features with simplicial wavelets. The paper reports strong accuracy on melanoma immunotherapy response and other single-cell cohorts, though the headline claim of universal outperformance is not fully supported by its own tables.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Empirical superiority may rest on test-set-informed tuning: best test metric over epochs plus per-dataset epsilon/bandwidth selection (App. H, J) can inflate HiPoNet's margins.","rationale":"The Reader's weakest assumption correctly identifies per-dataset epsilon selection as fragile; Table A6 alone shows accuracy ranging from 27.27% to 90.90% on Melanoma as epsilon varies. I generalize this to a broader evaluation-protocol concern: Appendix H's use of best test-set metrics across training runs, combined with threshold and bandwidth choices that appear to be made after seeing test performance, means the reported numbers may be optimistically biased. This is load-bearing because the paper's central claim is the empirical superiority statement in the abstract. The paper has real strengths—an anonymous code release, multi-seed reporting, and extensive ablations—so I do not reject it; however, the current evidence does not establish superiority over KNN-GIN and TopoGNN under a clean validation protocol. The theoretical sections (especially Theorem 4.3 and Corollary 4.5) also need repair, but they are secondary to the empirical headline. A controlled re-run with validation-based selection for both HiPoNet and baselines would either confirm or refute the claim; hence the Reader's conditional verdict remains appropriate, with the condition being a corrected evaluation protocol.","tokens_in":21727,"tokens_out":6687,"duration_ms":59341,"concrete_test":"Reproduce the Melanoma and PDO classification experiments with a patient-level train/validation/test split (e.g., 60/20/20), and choose epsilon, bandwidth, number of views, simplex order, and training epochs on the validation split only (e.g., best validation accuracy/epoch). Apply the identical protocol to all baselines. Report the test-set accuracy for each configuration and also for a fixed default threshold (e.g., epsilon in {0.15, 0.50}). If HiPoNet no longer beats KNN-GIN on Melanoma or remains second to TopoGNN on PDO with comparable variance, the abstract claim should be revised to 'competitive' rather than 'outperforms'.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central empirical claim—HiPoNet outperforms graph/point-cloud baselines on single-cell data—is not yet supported by the reported protocol. Appendix H states: 'We record the best metric ... achieved on the test set in each training run.' That is test-set model selection, not validation. Appendix J (Tables A6, A7) shows epsilon and kernel bandwidth are tuned per dataset (Melanoma best at epsilon=0.50, PDO at 0.15; bandwidth 1.00 vs 0.75), with Melanoma accuracy spanning 27.27–90.90 across thresholds. The text says thresholds are chosen at a 'critical point' beyond which edge count grows rapidly, but no concrete algorithm is given, so the rule is unfalsifiable as stated. Indeed, the Appendix J text notes that PDO's critical point at 0.15 'leads to improved performance compared to what was reported in the main text,' suggesting post-hoc adjustment. If these hyperparameters were selected using the same test folds that later produce the reported scores, HiPoNet's margin (90.90±4.92 vs KNN-GIN 85.45±3.63 on Melanoma; 77.38±0.94 vs TopoGNN 79.90±16.15 on PDO) can reflect selection bias rather than architectural superiority. The baselines do not appear to receive analogous per-dataset test-set tuning. This is the most load-bearing threat because the headline contribution is the empirical win; the theoretical results (Theorems 4.3–4.5) are also under-derived, but even a corrected theory would not establish the empirical claim without a valid evaluation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"HiPoNet models each point cloud as several Vietoris-Rips simplicial complexes built from learned feature-reweighted views, applies simplicial wavelet/scattering transforms to extract multiscale features, and feeds the aggregated representation to an MLP for point-cloud-level classification, regression, and representation learning. The paper presents theoretical results linking heat diffusion on simplicial complexes to 0-homology, to an associated simplicial graph, and to manifold geodesics, and reports experiments on melanoma, patient-derived organoid, and spatial transcriptomics cohorts, including a persistence-feature regression task.","tokens_in":22127,"tokens_out":7536,"duration_ms":67881,"significance":"High-dimensional point-cloud cohorts are an important and growing problem, and a scalable architecture that preserves geometric and topological structure would be valuable. The multi-view reweighting mechanism and the use of simplicial scattering are sensible ideas, and the comparison against a broad set of graph, topological, and point-cloud baselines is a useful contribution. The paper is also transparent in providing ablations, complexity analysis, and an explicit limitations section. However, the headline empirical claim is not supported by the current evaluation protocol: the reported numbers are selected on the test set, and key hyperparameters are tuned per dataset with an unspecified rule. In addition, the central theoretical equivalence in Theorem 4.3 is stated incorrectly as a block-diagonal identity. As presented, the evidence does not establish that HiPoNet outperforms existing methods, and the theory needs substantial correction.","major_comments":[{"comment":"The construction of the Vietoris-Rips complex is internally inconsistent. The text defines the kernelized distance as d^(v)_{i,j} = exp(-||x_i - x_j||^2 / (2σ^2)), which is a similarity in [0,1], and then declares that a simplex is included when d^(v)_{ip,iq} ≤ ε. Under this rule, small thresholds connect dissimilar (far) points and large thresholds connect similar points; with ε values in Table A6 ranging from 0.15 to 1.25, the resulting complexes are not Vietoris-Rips complexes in the usual distance sense. The manuscript needs a precise, correct definition of the filtration threshold used to construct S^(v).","section":"§3 and Algorithm 1"},{"comment":"The evaluation protocol uses test-set model selection. Appendix H states that the authors 'record the best metric ... achieved on the test set in each training run,' so the reported mean and standard deviation are maxima over epochs evaluated on the same test folds that are later used to report scores. This is not a valid held-out estimate and can only support an upper-bound comparison. Even if the baselines follow the same protocol, the reported numbers cannot establish the abstract's claim that HiPoNet 'outperforms' other models.","section":"Appendix H"},{"comment":"The hyperparameter sensitivity analyses in Tables A6 and A7 show that results depend strongly on the Vietoris-Rips threshold ε and kernel bandwidth σ, with Melanoma accuracy ranging from 27.27% to 90.90% across thresholds and the optimal values differing by dataset. The stated selection rule, tuning 'until reaching a critical point—beyond which the number of edges grows rapidly,' is not specified as a concrete algorithm, and the remark that PDO has a critical point at 0.15 'which leads to improved performance compared to what was reported in the main text' indicates post-hoc adjustment. Without a prespecified criterion and without analogous per-dataset tuning for the baselines, the reported margin over baselines can reflect selection bias rather than architectural superiority.","section":"Appendix J"},{"comment":"The empirical claims overstate the results. In Table 4, TopoGNN has a higher mean accuracy than HiPoNet on PDO (79.90 vs 77.38), and in Table 3, KNN-GCN has a higher AUC than HiPoNet on UPMC recurrence (0.668 vs 0.6044). The main text acknowledges the PDO result but then says HiPoNet 'outperforms all the graph-based, topological, and point cloud methods in Melanoma' and later claims 'top results in most settings' for spatial transcriptomics. The abstract's unqualified 'outperforms' claim and these text statements should be replaced with a dataset-by-dataset characterization.","section":"Tables 3 and 4"},{"comment":"The claimed block-diagonal identity is not correct as stated. In Definition 4.2, the simplicial graph G(S) has edges only between same-order simplices that are upper or lower adjacent, and its standard graph Laplacian has off-diagonal entries of -1 (or -A_{ij}) for each adjacency. In contrast, the Hodge Laplacian Δ_k = B_k^T B_k + B_{k+1}B_{k+1}^T couples k-simplices to their boundary and coboundary simplices and has off-diagonal entries equal to the number of shared faces or cofaces (with signs in the oriented case). Therefore Δ_G cannot in general be written as the block-diagonal matrix of Hodge Laplacians displayed in Eq. (8), and the heat-equation agreement asserted in Theorem 4.3 does not follow from the given definitions.","section":"Theorem 4.3 and Appendix D"},{"comment":"The geometric results are under-derived. Theorem 4.4 invokes Varadhan's formula and convergence of the graph heat kernel to the manifold heat kernel, but it does not establish that the object used in the model, namely the 0-th order Hodge Laplacian on the simplicial complex or the heat kernel on G(S), is the object to which those convergence results apply, especially given the issues with Theorem 4.3. Corollary 4.5 then claims equivalence of dimension, volume, and total scalar curvature from Weyl's law and eigenvalue comparison theorems; these are asymptotic statements for the Laplace-Beltrami operator on a manifold, not immediate consequences of equality of discrete heat kernels. The statements may be true in a suitable limit, but the proof as written is a non sequitur.","section":"Theorems 4.4 and Corollary 4.5"}],"minor_comments":[{"comment":"There is a typo in the column header: 'Nuumber of cells' should be 'Number of cells'.","section":"Table 2"},{"comment":"The number of melanoma cells is inconsistent: the main text and Table 2 state 61K cells, while Appendix G reports 489 to 1784 cells per patient with a total of 11,862 cells. Please clarify which count is correct and whether this discrepancy affects the reported preprocessing or the fold construction.","section":"Appendix G vs Table 2"},{"comment":"The limitations section honestly notes that validation is confined to biological datasets and that generalization to other high-dimensional domains remains unexplored; this should be reflected in the abstract and conclusion, which currently present HiPoNet as a general solution for high-dimensional data analysis.","section":"Appendix L"},{"comment":"The notation for second-order scattering coefficients omits a closing parenthesis; the formula should read S^v_k[j,j'] X^v_k = |Ψ^{v,j'}_k ( |Ψ^{v,j}_k X^v_k | ) | for full clarity.","section":"Algorithm 1, line 14"}],"recommendation":"major_revision","confidential_remarks":"The central empirical claim is not supported by the current protocol because of test-set metric selection in Appendix H and the per-dataset hyperparameter tuning in Appendix J. This is fixable by re-running the experiments with a validation split and fixed, prespecified hyperparameters, and by reporting baselines tuned under the same rule. The theoretical issue in Theorem 4.3 is more than a presentation problem, since the proof sketch relies on a false block-diagonal identity; the authors should either correct the statement to a weighted multigraph Laplacian or substantially reframe the result. Given the scope of these issues, I recommend a major revision rather than rejection, because the architectural ideas and the breadth of the evaluation are potentially valuable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"HiPoNet is a real engineering contribution. The combination of learned multi-view feature reweighting, Vietoris-Rips complexes, and simplicial scattering wavelets for cohort-level prediction is new as far as I know, and the motivation is sound: 3D point-cloud networks do not transfer to high-dimensional single-cell data, and a single nearest-neighbor graph discards geometry and topology. The ablations show each component matters, and the experiments span clinically relevant cohorts with code provided. That deserves credit.\n\nThe soft spots are real, though. The theory in Theorem 4.3 is not correctly derived: the graph Laplacian of the simplicial graph G(S) is not block-diagonal with Hodge Laplacians in general, because the incidence matrix of G(S) is not the boundary matrix, and the off-diagonal entries count adjacencies differently. The proof sketch asserts the block structure without justification. Corollary 4.5 overreaches. These are not fatal to the architecture, but they need correction or weakening.\n\nMore load-bearing is the evaluation protocol. Appendix H says the best metric on the test set is recorded in each training run, which is test-set model selection, not validation. Appendix J shows the Vietoris-Rips threshold and kernel bandwidth are tuned per dataset, with the selection rule stated only as a \"critical point\" that is not concretely specified, and the PDO note that the critical point leads to improved performance over the main text suggests post-hoc adjustment. That can inflate HiPoNet's margins, and the baselines do not appear to receive analogous per-dataset test-set tuning. The abstract's \"outperforms\" claim is not supported as written; Table 4 shows TopoGNN has higher mean PDO accuracy, and Table 3 shows KNN-GCN beats HiPoNet on UPMC recurrence. The persistence prediction experiments are better framed as an expressivity check than as evidence of topology preservation, which the paper mostly acknowledges.\n\nIn short: the core idea is worth taking seriously, but the empirical claim needs a clean protocol before it can be believed. I would send this to peer review, with major revision: fix the evaluation to validation-based selection, report test metrics once hyperparameters are fixed, specify the threshold rule, rerun baselines under the same protocol, and correct or weaken Theorem 4.3 and Corollary 4.5. A reader in geometric deep learning for biomedical data would get value from the architecture and the ablation study, but should not cite the current numbers as evidence.","headline":"A genuinely useful architecture for whole-sample prediction on high-dimensional point clouds, but the empirical win is not yet established because the protocol reports test-set best epochs and per-dataset threshold tuning.","tokens_in":22646,"tokens_out":2910,"would_cite":true,"duration_ms":28880,"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":"HiPoNet claims that modeling a high-dimensional point cloud as several feature-reweighted simplicial complexes, read by multiscale simplicial wavelets, preserves geometric and topological information that single-graph and 3D point-cloud…","keywords":["simplicial complexes","high-dimensional point clouds","single-cell data","scattering transforms","multiscale wavelets","feature reweighting","spatial transcriptomics","Vietoris-Rips complex"],"falsifier":"Run HiPoNet and its strongest graph baselines on the Melanoma and PDO cohorts with $\\epsilon$ and the kernel bandwidth $\\sigma$ selected by one fixed, prespecified rule rather than per-dataset tuning, and compare across equal hyperparameter budgets. The central claim of consistent superiority would be falsified if baseline GNNs match or exceed HiPoNet under that threshold policy, or if HiPoNet's accuracy collapses away from its tuned operating point ($\\epsilon = 0.50$, $\\sigma = 1.00$ on Melanoma).","tokens_in":21556,"feed_emoji":"🧬","tokens_out":9733,"duration_ms":79167,"temperature":0.7,"pith_summary":"HiPoNet is a neural network for whole-point-cloud prediction in high-dimensional spaces, the regime of single-cell cohorts where each patient or condition contributes its own cloud of thousands of cells. The paper argues that collapsing each cloud into a single nearest-neighbor graph throws away geometric and topological structure, and that an ensemble of simplicial complexes, one per learnable feature-reweighted view, preserves it. Each complex is read by simplicial wavelet transforms whose band-pass filters yield multiscale scattering coefficients, aggregated across views into a representation for classification or regression. The authors prove that heat diffusion on these complexes respects connected components, agrees with heat diffusion on an associated simplicial graph, and can approximate geodesic distances on the underlying data manifold. Empirically, they report that HiPoNet outperforms graph, topological, and 3D point-cloud baselines on melanoma immunotherapy-response prediction, patient-derived organoid treatment classification, and spatial transcriptomics outcome prediction.","feed_headline":"Multi-view simplicial network tops graph models on single-cell data","feed_subtitle":"By reading each cell cloud through several reweighted views, HiPoNet keeps geometry that single-graph methods discard.","key_machinery":"The load-bearing construction is the multi-view simplicial complex ensemble. Each view $v$ is a reweighted point cloud $\\tilde{X}^{(v)} = \\alpha^{(v)} \\odot X$; from it a Vietoris–Rips complex $\\mathcal{S}^{(v)}$ is formed at a fixed distance threshold, so that $k$-simplices encode interactions among $k+1$ cells rather than only pairwise edges. The Hodge Laplacian $\\Delta_k = B_k^\\top B_k + B_{k+1}B_{k+1}^\\top$ gives a simplicial random-walk matrix $P_k = \\Delta_k D_k^{-1}$, and differences of its powers at consecutive diffusion scales define the simplicial wavelets $\\Psi^{(v,j)}_k$; applying absolute values and a second wavelet pass yields the scattering coefficients. Theorems 4.1, 4.3, and 4.4 carry the theoretical claim: heat diffusion cannot cross connected components, the heat equation on the complex coincides with that on the associated simplicial graph, and the resulting diffusion approximates geodesic distance on the manifold underlying the point cloud.","core_discovery":"The paper's central claim is that a high-dimensional point cloud is better modeled as an ensemble of simplicial complexes, one per learned feature-reweighted view, than as a single graph or as a 3D-oriented point cloud. For each view $v$, HiPoNet rescales every point by a learnable vector $\\alpha^{(v)}$, builds a Vietoris–Rips complex from kernelized distances at a fixed scale, forms Hodge Laplacians and simplicial random-walk matrices, and applies simplicial wavelet transforms whose differences of diffusion at consecutive scales produce first- and second-order scattering coefficients. Aggregated across views, these coefficients feed a multilayer perceptron that outputs the cohort-level prediction. The paper asserts that this construction preserves geometry and topology in a provable sense, heat diffusion respects the 0-homology of the complex, agrees with the heat equation on the associated simplicial graph, and approximates geodesic distances on the manifold, and that empirically it predicts persistence features and clinical outcomes better than the graph and point-cloud baselines tested, with HiPoNet taking first place on melanoma and on most spatial transcriptomics tasks and second place with far lower variance on the organoid data.","pith_inferences":["Editorial inference: the learned reweighting vectors could be tested as hypothesis generators, checking whether the highest-weight features in each view recover known gene programs, cell-cycle modules, or immune markers would validate the disentangling-processes claim independently of prediction accuracy.","Editorial inference: the sensitivity to $\\epsilon$ and $\\sigma$ suggests a natural extension in which each view learns its own threshold or aggregates over a range of thresholds, effectively making the Vietoris–Rips scale a learned parameter rather than a tuned hyperparameter.","Editorial inference: the theory guarantees geodesic information for the heat diffusion stage, not for the scattering coefficients after absolute-value nonlinearities; the geometric-preservation claim should be read as proven for the diffusion operator and empirically supported for the full network.","Editorial inference: a fairer comparison would give baselines the same per-dataset threshold and bandwidth tuning; until then, the reported margins over GNNs should be read as performance under the paper's configuration rather than as an architecture-independent advantage."],"forward_implications":["Cohort-level prediction from high-dimensional single-cell and spatial omics can be performed end to end, without separately computing persistence diagrams or other topological summaries.","The learned feature weights provide per-marker importance within each view, so the same model that predicts outcomes can also indicate which proteins or genes organize the relevant cellular processes.","The theoretical results imply that the network's diffusion operators carry manifold-level information, including dimension, volume, and total scalar curvature, so its embeddings are geometry-preserving representations rather than merely predictive features.","Because the construction is differentiable and view-agnostic, the architecture transfers to any high-dimensional point-cloud cohort task, including non-biological domains where whole sets of points are the unit of prediction.","Mixing heterogeneous views, as done with spatial coordinates plus expression markers in spatial transcriptomics, is a natural way to fuse multiple data modalities inside one simplicial-wavelet model."],"supporting_citations":[{"why":"Supplies the melanoma MIBI cohort and immunotherapy-response labels used for the main classification comparison.","marker":"[41]"},{"why":"Supplies the patient-derived organoid cohort and treatment labels for the second classification task.","marker":"[44]"},{"why":"Supplies the spatial transcriptomics CODEX cohorts, tasks, and evaluation folds.","marker":"[60]"},{"why":"Introduces diffusion wavelets, the basis for the simplicial wavelet transform and scattering construction.","marker":"[12]"},{"why":"Provides the geometric scattering transform whose absolute-value nonlinearities and multiscale coefficients HiPoNet adapts to simplicial complexes.","marker":"[21]"},{"why":"Defines the simplicial scattering transform used to extract multiscale features from each complex.","marker":"[34]"},{"why":"Justifies using random-walk matrices as a discrete approximation of heat diffusion on the data manifold.","marker":"[13]"},{"why":"Supplies the spectral convergence of graph heat kernels to manifold heat kernels used in Theorem 4.4.","marker":"[14]"},{"why":"Supplies Varadhan's formula connecting heat kernels to geodesic distance, the anchor of Theorem 4.4.","marker":"[52]"}],"fun_headline_variants":["Simplicial complexes beat single graphs on cell data","HiPoNet: reweighted complexes grasp cell geometry","Multi-view simplicial nets outscore graph baselines","Topological views win on single-cell point clouds","Learnable complexes outperform graphs on cells"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline succeeds only if a single Vietoris–Rips threshold $\\epsilon$ chosen for each reweighted view yields simplicial complexes whose multiscale wavelets contain task-relevant structure; the paper's own sensitivity analysis shows the best $\\epsilon$ differs by dataset, with Melanoma accuracy ranging from $27.27\\%$ to $90.90\\%$ as $\\epsilon$ varies.","fun_headline_variants_meta":{"raw":{"variants":["Simplicial complexes beat single graphs on cell data","HiPoNet: reweighted complexes grasp cell geometry","Multi-view simplicial nets outscore graph baselines","Topological views win on single-cell point clouds","Learnable complexes outperform graphs on cells"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000213,"raw_usage":{"total_tokens":1496,"prompt_tokens":1094,"completion_tokens":402,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":710,"completion_tokens_details":{"reasoning_tokens":328}},"tokens_in":710,"tokens_out":402,"duration_ms":4663,"temperature":1.0,"reasoning_tokens":328,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T11:42:17.991357+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run HiPoNet and its strongest graph baselines on the Melanoma and PDO cohorts with $\\epsilon$ and the kernel bandwidth $\\sigma$ selected by one fixed, prespecified rule rather than per-dataset tuning, and compare across equal hyperparameter budgets. The central claim of consistent superiority would be falsified if baseline GNNs match or exceed HiPoNet under that threshold policy, or if HiPoNet's accuracy collapses away from its tuned operating point ($\\epsilon = 0.50$, $\\sigma = 1.00$ on Melanoma).","supporting_citations":[{"cited_title":"Trevino, Eric Wu, Kyle Swanson, Honesty J","cited_arxiv_id":null,"evidence_quote":"Supplies the spatial transcriptomics CODEX cohorts, tasks, and evaluation folds."},{"cited_title":"Coifman and M","cited_arxiv_id":null,"evidence_quote":"Introduces diffusion wavelets, the basis for the simplicial wavelet transform and scattering construction."},{"cited_title":"Geometric scattering for graph data analysis","cited_arxiv_id":null,"evidence_quote":"Provides the geometric scattering transform whose absolute-value nonlinearities and multiscale coefficients HiPoNet adapts to simplicial complexes."},{"cited_title":"Unsupervised parameter- free simplicial representation learning with scattering transforms","cited_arxiv_id":null,"evidence_quote":"Defines the simplicial scattering transform used to extract multiscale features from each complex."},{"cited_title":"Diffusion maps","cited_arxiv_id":null,"evidence_quote":"Justifies using random-walk matrices as a discrete approximation of heat diffusion on the data manifold."},{"cited_title":"Spectral convergence of graph laplacian and heat kernel reconstruction in l∞ from random samples","cited_arxiv_id":null,"evidence_quote":"Supplies the spectral convergence of graph heat kernels to manifold heat kernels used in Theorem 4.4."}],"review_version":1}