{"id":"dc571a75-38f6-4edd-91c5-45fac893fe93","arxiv_id":"2411.13887","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A structural similarity measure that compares molecules through the Gromov-Hausdorff ultrametric distance between spaces of harmonic cohomology generators.","lead":"The authors propose measuring molecular similarity by turning each molecule's loop structures, represented by harmonic cohomology generators, into a metric space and comparing molecules with the Gromov-Hausdorff ultrametric. Their clustering features separate halide perovskite structures by halide atom more cleanly than 3D coordinates, ECFP, or MACCS fingerprints in the reported test.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported clustering result is not yet shown to be stable: L1 distances among harmonic representatives depend on the arbitrary orthonormal basis of the zero eigenspace, and no sensitivity analysis is reported.","rationale":"The reader's weakest assumption identifies the non-uniqueness of harmonic representatives as the key fragility, and the manuscript itself concedes this limitation in the Discussion. My stress-test agrees: the L1 distance between coordinate vectors in Definition 1 is not invariant under orthogonal transformations of the zero eigenspace, and the uGH construction inherits this non-invariance. The article's numerical claim is an empirical one, with Table 1 reporting high ARI values, but no experiment is reported that controls for the arbitrary choice of basis. Consequently, the central claim that the method 'effectively clusters' molecular structures is not yet established in a reproducible, basis-independent sense. This is not a mathematical inconsistency in the paper's own terms, but it is a correctness risk for the headline result. I therefore do not call for rejection, since a basis-invariant formulation or a sensitivity analysis could rescue the claim; however, acceptance should be conditional on demonstrating that the reported results are stable under the acknowledged ambiguity. The concrete test proposed here would settle whether the basis ambiguity materially changes the clustering outcome.","tokens_in":13062,"tokens_out":3320,"duration_ms":37992,"concrete_test":"Take the 300 configurations in one phase (e.g., tetragonal) and the alpha-complex thresholds used in the paper. For each configuration, compute the null space of L1 and generate 100 random orthonormal bases by applying random orthogonal matrices to the returned zero eigenvectors. Recompute all pairwise L1 distance matrices, the resulting uGH matrices, and the K-means ARI for X-site clustering for each basis choice; report the mean, standard deviation, and range of the ARI. If the ARI varies materially (for example, drops from 1.000 to below 0.9) or the uGH matrices change by more than a small tolerance, the central clustering claim is not stable under the acknowledged basis ambiguity. A complementary check is to recompute the entire pipeline with a different eigensolver, such as dense SVD instead of sparse eigensolvers, and compare the resulting ARI values.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central construction in the Method section, specifically Definition 1, treats zero eigenvectors of the 1-dimensional Hodge Laplacian as a finite set of points with a coordinate-wise L1 distance. This set is not a well-defined object when dim ker L1 > 1: any orthonormal basis of the zero eigenspace is equally valid. If an orthogonal matrix is applied to the returned eigenvectors, the pairwise distances ||v_i - v_j||_1 change, so the ultrametric space H1(K) obtained by the dendrogram transformation changes, and therefore the uGH values between molecules change. The paper's Discussion explicitly acknowledges this: \"the choice of kernel vectors for the Hodge Laplacian is not unique. Therefore, for the same structure, we may have different choices for cohomology generators.\" However, the experimental section uses one arbitrary output of an eigensolver without quantifying how much the clustering outcomes vary under basis rotation. For alpha complexes of the OIHP structures at filtration thresholds 3.5, 4, 5, and 6 Å, the first Betti number is typically larger than one, so the ambiguity is not vacuous. Since the headline empirical claim is the effectiveness of the descriptor in clustering, and the descriptor itself is basis-dependent, the reported ARI values of 0.980–1.000 could be artifacts of a particular solver basis. No invariant reformulation is provided, and the sign-fixing convention does not resolve rotations within a multidimensional kernel.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a molecular similarity descriptor built from cohomology. For each molecular structure, an Alpha complex is constructed at several filtration thresholds; the 1-dimensional Hodge Laplacian is formed; its zero eigenvectors, called cohomology generators, are treated as points in a metric space under one of three pairwise distances (L1, cocycle, Wasserstein). The resulting distance matrix is converted into an ultrametric space via a dendrogram construction, and the Gromov-Hausdorff ultrametric (uGH) between two such ultrametric spaces is used as a dissimilarity measure. The method is applied to 900 molecular-dynamics configurations of nine organic-inorganic halide perovskite structures, and K-means clustering by halide atom is compared against clustering from 3D coordinates, ECFP, and MACCS fingerprints. The paper reports near-perfect Adjusted Rand Index values for the uGH features (0.980-1.000) and claims the approach captures loops and cavities better than traditional persistent homology.","tokens_in":13300,"tokens_out":5572,"duration_ms":58271,"significance":"If the construction is well defined and stable, the paper offers a novel, parameter-free topological descriptor for molecular similarity that is attractive in several ways: it uses Hodge-theoretic harmonic representatives rather than only Betti numbers, it avoids fitted constants and label-dependent tuning, and the authors provide code and data in a public repository. The experimental comparison is meaningful in design because the uGH features are computed directly from atomic coordinates without training on cluster labels. However, the central object of the method, the set of cohomology generators of the 1-dimensional Hodge Laplacian, is not basis invariant when the first Betti number exceeds one, and the paper's own Discussion acknowledges this non-uniqueness. The empirical claims therefore currently rest on an arbitrary eigensolver basis, which is a load-bearing issue that must be resolved before the reported clustering results can be interpreted as properties of the method.","major_comments":[{"comment":"Definition 1 defines distances between individual zero eigenvectors of L1. When β1 > 1, the zero eigenspace has dimension greater than one, and the eigenvectors returned by an eigensolver form an arbitrary orthonormal basis. An orthogonal change of basis generally changes the vectors themselves, the pairwise L1 distances ||v - w||_1, and hence the dendrogram, the ultrametric space H1(K), and the resulting uGH values. The sign-fixing convention described in the Method fixes only the sign of each eigenvector and does not resolve rotations within a multidimensional kernel. The Discussion acknowledges this non-uniqueness, but the Results use a single solver basis with no sensitivity analysis. Since the ARI values in Table 1 (0.980-1.000) are the paper's main empirical claim, the current evidence does not establish that the clustering is a property of the method rather than of a particular eigensolver output. Please either reformulate the construction using a basis-invariant object (for example, distances between the harmonic subspaces themselves, or spectral invariants of the Hodge Laplacian) or add a systematic sensitivity analysis over random orthogonal bases of ker L1 and report the range and robustness of the resulting ARI values.","section":"Method, Definition 1; Discussion"},{"comment":"The claim that the method 'incorporates geometric information' is only partially supported. With the L1 distance (Definition 1) and the cocycle distance (Definition 2), the pairwise distances between cohomology generators are computed solely from the entries of the harmonic representatives, which depend only on the simplicial complex combinatorics; atomic coordinates enter only through the filtration threshold used to build the Alpha complex. The Wasserstein distance (Definition 3) does use simplex coordinates, but the OIHP experiments reported in the Results state that only the L1 distance was used. The reported separation of halide atoms may therefore be driven by the multiscale Alpha filtration rather than by geometric information within the cohomology-generator metric. Please clarify this point, and if a geometric contribution is claimed, provide a comparison against a purely combinatorial analogue or report experiments using the Wasserstein distance.","section":"Abstract; Results; Definition 1/Definition 3"},{"comment":"The baseline comparison is under-specified. It is not clear how '3D coordinates' are converted into a fixed-length feature vector for K-means clustering, and the statement that fingerprints are compared 'without including additional information such as atomic number and weight' is difficult to interpret because ECFP and MACCS keys are defined in terms of atom invariants and structural keys; modifying those inputs changes the fingerprints themselves. The very low ECFP ARI for the orthorhombic phase (0.030) suggests the baseline features may not be standard implementations. Please specify the exact fingerprint generation parameters and the coordinate-feature representation so that the comparison in Table 1 is reproducible and fair.","section":"Results, Table 1"}],"minor_comments":[{"comment":"The barycentric-coordinate definition of a p-simplex writes the condition as sum λ_i = 0; it should be sum λ_i = 1, since the current condition is inconsistent with 0 ≤ λ_i ≤ 1 except for the zero vector.","section":"Background, Simplicial complexes"},{"comment":"The fingerprint name is written as 'MACC' in the Results and Table 1; it should be 'MACCS' (as in the Introduction and reference [29]).","section":"Results, Table 1 and text"},{"comment":"The sign convention 'we enforce that the first element of each cohomology generator is non-negative' is incomplete when the first entry is zero; a tie-breaking rule, such as the first non-zero entry, should be specified.","section":"Method, L1 Distance"},{"comment":"The paper advertises analysis of 1-dimensional and higher-dimensional (co)homology, but all numerical experiments use only the 1-dimensional Hodge Laplacian. The Discussion correctly defers higher-order Laplacians to future work, but the abstract and introduction should be aligned with what is actually demonstrated.","section":"Abstract and Introduction"},{"comment":"The manuscript refers to 'Algorithm 1 in [48]' for the transformation from a distance matrix to an ultrametric space, but does not state the algorithm. Since the resulting uGH values depend on this transformation, a brief description or restatement would improve reproducibility.","section":"Method, ultrametric transformation"}],"recommendation":"major_revision","confidential_remarks":"The central basis-invariance concern is genuine and is acknowledged by the authors in the Discussion; it is not an artifact of the review process. The paper's core contribution is defensible if the authors can either make the construction invariant or show that the reported clustering results are stable under arbitrary choices of harmonic representatives. The availability of code and data is a strength. I would not reject outright, but the revision must directly address the rotation ambiguity, not merely list it as future work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the specific pipeline: take harmonic representatives of 1-dimensional cohomology from a Hodge Laplacian, treat them as points in a metric space, dendrogram-transform to an ultrametric, then compare molecules via the Gromov-Hausdorff ultrametric. That combination is not in the cited brain-network literature, and it is a sensible way to inject geometric content into loop-level topology. The paper is also honest about its own limitations, which counts for something: the Discussion explicitly flags the non-uniqueness of kernel vectors for the Hodge Laplacian.\n\nWhat works: the background is clearly written, the workflow is reproducible in principle (code is promised in a repo, and the data are MD snapshots of OIHPs, not fitted to labels), and the clustering comparison against 3D coordinates, ECFP, and MACCS keys is a reasonable baseline for a methods paper. There are no fitted constants, so circularity is not a real concern.\n\nWhere it gets soft: the central construction in Definition 1 is not basis invariant when dim ker L1 > 1, and for the alpha complexes used here it almost always is greater than one. The authors acknowledge this but do not quantify it. The reported ARIs of 0.980–1.000 could, in principle, be an artifact of one arbitrary eigensolver output. A sensitivity analysis under random orthogonal transformations of the kernel basis is the obvious fix, and its absence is a real gap, not a nitpick. Also, only the L1 distance is tested; the cocycle and Wasserstein variants are defined but never used, so the title's generality is not backed by experiments. The abstract's claim about higher-dimensional cohomology is not supported by anything in the results. And despite the text saying the method provides deeper insights than persistent homology, there is no persistent homology baseline at all. The tetragonal ARI of 0.980, while high, comes with no variance estimate across the 100 MD snapshots, and K-means is run once, presumably.\n\nBottom line: the idea is worth taking seriously, but the headline empirical claim is not yet stable in the sense that matters. A serious referee should ask for the basis-rotation sensitivity analysis and a persistent homology feature comparison, plus at least one of the other two distances. If the stability holds, this could be a useful descriptor for cheminformatics and materials screening. I would not cite it yet for my own work, but I would send it to review rather than desk-reject, because the core construction is novel and the failure mode is explicitly identified and potentially fixable.","headline":"A genuinely new descriptor idea with an acknowledged but unquantified basis-dependence problem; worth refereeing, but the empirical claim needs a stability check before trusting it.","tokens_in":13876,"tokens_out":761,"would_cite":false,"duration_ms":9973,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["55N31"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper introduces a cohomology-based Gromov-Hausdorff ultrametric for molecular similarity and shows it clusters halide perovskites by halogen atom with near-perfect Adjusted Rand Index.","keywords":["cohomology","Gromov-Hausdorff ultrametric","molecular similarity","Hodge Laplacian","simplicial complex","halide perovskite","topological data analysis"],"falsifier":"Take one OIHP configuration, compute its L1-based uGH features, then recompute after applying a random orthogonal transformation to the zero-eigenvalue eigenvectors of the Hodge Laplacian; if the resulting distance matrix and the K-means clusters change substantially, the method's central similarity claim is not stable.","tokens_in":12823,"feed_emoji":"🧪","tokens_out":6390,"duration_ms":57268,"temperature":0.7,"pith_summary":"The paper introduces a way to measure how similar two molecules are by comparing the geometry of their loop and cavity structures, rather than relying only on atom coordinates or chemical fingerprints. Each molecule is turned into a simplicial complex, cohomology generators are extracted from the kernel of the Hodge Laplacian, and distances between these generators are summarized by a Gromov-Hausdorff ultrametric. On a test set of organic-inorganic halide perovskite configurations, this descriptor clusters molecules by which halogen atom they contain with Adjusted Rand Index values at or near 1, outperforming 3D coordinates, ECFP, and MACCS fingerprints. The proposed significance is a topological-geometric similarity measure that captures geometric information that ordinary persistent homology does not provide.","feed_headline":"Cohomology metric clusters halide perovskites by atom type","feed_subtitle":"A Gromov-Hausdorff ultrametric on loop structure beats 3D coordinates and chemical fingerprints in tests.","key_machinery":"The load-bearing object is the cohomology generator space H^p(K), the set of harmonic representatives of cohomology classes obtained as the kernel of the combinatorial Hodge Laplacian L_p. By the Hodge decomposition, each cohomology class has a unique harmonic representative, so these eigenvectors can serve as coordinate descriptions of a molecule's loop and cavity geometry. Pairwise distances among generators—L1 distance, cocycle distance, or Wasserstein distance—turn the generator set into a finite metric space. Hierarchical clustering converts that metric space into a dendrogram, which is an ultrametric space, and the Gromov-Hausdorff ultrametric uGH between two such dendrograms is the final molecular similarity score.","core_discovery":"The central claim is that the space of harmonic cohomology representatives of a molecular simplicial complex can be treated as a metric space, and that the Gromov-Hausdorff ultrametric between two such spaces quantifies molecular structural similarity. Because cohomology classes correspond to loops, voids, and higher cavities, the descriptors carry topological information that is localized and geometry-aware. The paper constructs cohomology generators as eigenvectors spanning the kernel of the p-th combinatorial Hodge Laplacian, equips the generator set with one of three pairwise distances, converts the resulting metric space into a dendrogram and hence an ultrametric space, and then computes the Gromov-Hausdorff ultrametric between molecules. The numerical experiments show that this approach clusters halide perovskites by X-site atom nearly perfectly, and the paper presents it as the first cohomology-based Gromov-Hausdorff ultrametric method for such molecular similarity questions.","pith_inferences":["A natural stress test is whether the reported clustering advantage survives when the kernel-basis ambiguity is controlled by averaging or optimizing over orthogonal rotations of the harmonic generators; the paper itself leaves this open.","The same pipeline could be applied to protein conformations or ligand binding pockets, where loop and cavity geometry is thought to drive function; the paper mentions this direction but does not test it.","Combining uGH features with machine-learning models may yield descriptors complementary to persistent-homology barcodes, since uGH preserves geometric placement of features rather than only their persistence.","The three distance measures may behave differently under basis rotation, so choosing between L1, cocycle, and Wasserstein distances is an empirical question not settled in the paper."],"forward_implications":["If the method is correct, molecular similarity can be quantified from a molecule's intrinsic loop and cavity geometry without needing atom-type labels or chemical fingerprints.","Clustering tasks such as distinguishing halide perovskites by halide atom become nearly perfect using only structural coordinates, indicating that topology-geometry descriptors carry chemically relevant signal.","Because the Gromov-Hausdorff ultrametric is computable in polynomial time, the approach can in principle scale beyond the small molecules tested here to larger structural datasets, at least for first cohomology.","The framework extends to higher-dimensional cohomology groups by the same recipe, so voids and higher-dimensional cavities can be incorporated into the same similarity measure.","The dendrogram representation gives a hierarchical view of molecular similarity rather than a single scalar distance, which may support multi-scale analysis of structurally related molecules."],"supporting_citations":[{"why":"Defines the Gromov-Hausdorff distance between metric spaces, the dissimilarity concept the method adapts.","marker":"[42,43]"},{"why":"Defines the Gromov-Hausdorff ultrametric and supplies the polynomial-time algorithm used to compute it.","marker":"[48,49]"},{"why":"Identifies cohomology classes with harmonic representatives in the kernel of the Hodge Laplacian, the paper's cohomology generators.","marker":"[53,63,64]"},{"why":"Defines the ECFP fingerprint baseline used in the clustering comparison.","marker":"[28]"},{"why":"Defines the MACCS keys baseline used in the clustering comparison.","marker":"[29]"},{"why":"Shows computing the Gromov-Hausdorff distance is NP-hard, motivating the ultrametric variant.","marker":"[69]"},{"why":"Links dendrograms to hierarchical clustering and ultrametric spaces, the representation step before uGH computation.","marker":"[37,40,48]"},{"why":"Provides the halide-perovskite structural context and dataset used in the numerical experiments.","marker":"[70]"}],"fun_headline_variants":["Cohomology-GH metric spots molecule shape differences","Cohomology ultrametric sorts perovskites by atom element","New cohomology distance separates molecule cavities","Gromov-Hausdorff on cohomology outperforms chemical fingerprints"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the particular cohomology generators computed for a molecule are meaningful descriptors, but whenever a molecule has multiple independent loops the harmonic representatives can be rotated by an orthogonal change of basis, and the pairwise distances—and therefore the uGH features—can change.","fun_headline_variants_meta":{"raw":{"variants":["Cohomology-GH metric spots molecule shape differences","Cohomology ultrametric sorts perovskites by atom element","New cohomology distance separates molecule cavities","Gromov-Hausdorff on cohomology outperforms chemical fingerprints"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00099,"raw_usage":{"total_tokens":4183,"prompt_tokens":915,"completion_tokens":3268,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":3198}},"tokens_in":531,"tokens_out":3268,"duration_ms":23211,"temperature":1.0,"reasoning_tokens":3198,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:46:22.573255+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take one OIHP configuration, compute its L1-based uGH features, then recompute after applying a random orthogonal transformation to the zero-eigenvalue eigenvectors of the Hodge Laplacian; if the resulting distance matrix and the K-means clusters change substantially, the method's central similarity claim is not stable.","supporting_citations":[{"cited_title":"& Hahn, M","cited_arxiv_id":null,"evidence_quote":"Defines the ECFP fingerprint baseline used in the clustering comparison."},{"cited_title":"Installation and operational experiences with maccs (molecular access system)","cited_arxiv_id":null,"evidence_quote":"Defines the MACCS keys baseline used in the clustering comparison."},{"cited_title":"Computational aspects of the Gromov–Hausdorff distance and its application in non-rigid shape matching","cited_arxiv_id":null,"evidence_quote":"Shows computing the Gromov-Hausdorff distance is NP-hard, motivating the ultrametric variant."},{"cited_title":"V., Xu, Q., Wee, J., Xia, K","cited_arxiv_id":null,"evidence_quote":"Provides the halide-perovskite structural context and dataset used in the numerical experiments."}],"review_version":1}