{"id":"8700395e-3d2b-4622-9617-0cc6173947cb","arxiv_id":"2509.01550","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"OTMol formulates molecular superimposition as fused supervised Gromov-Wasserstein optimal transport, yielding atom-order-independent, chirality-preserving alignments with preserved bond connectivity across ligands, peptides, lipids, and water or noble gas clusters.","lead":"OTMol is a new computer method for comparing 3D molecular shapes that is independent of atom ordering and refuses reflected, wrong-handed matches. It uses optimal transport math to pair atoms while keeping bonds and chemical features intact, and it runs faster than a widely used alternative on large clusters.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"One-to-one mapping is not enforced by OTMol's optimization; the argmax rounding is only checked post hoc, so the claimed guarantee is unsupported and can fail.","rationale":"The central claim advertises three features: low RMSD, chemical preservation, and one-to-one mapping. The one-to-one guarantee is the only one that is a binary claim directly contradicted by the algorithm: the optimization problem (Eq. 4) is over doubly stochastic matrices with equal marginals; its solutions need not be permutations, and the argmax rounding can collapse multiple rows to the same column. The algorithm's check is a filter, not an enforcement. If this check fails on any input, the method returns no alignment, which would make the tool unusable for that input and contradict the abstract's unconditional 'maintains molecular integrity... enforcing one-to-one mappings.' The reader's weakest assumption centers on whether the fsGW proxy is chemically faithful; while that is also important, the one-to-one issue is more fundamental because it is an internal inconsistency between the claim and the method, independent of external data. My proposed test directly measures the failure rate of the permutation check, which is a concrete and inexpensive validation. I therefore keep the CONDITIONAL verdict: the method may be useful in practice, but the stated guarantee needs to be either proven (e.g., by showing the solver with label costs always yields a roundable plan) or removed from the abstract.","tokens_in":17484,"tokens_out":14982,"duration_ms":182016,"concrete_test":"Instrument the released code to run Algorithms 1-3 on every benchmark pair and record, for each alpha (and each perturbation l), whether the argmax-rounded plan P' is a permutation. Report the per-pair and per-alpha success rate. If any pair has zero successful alpha values (or L empty for Algorithm 3), the algorithm fails to produce an assignment and the one-to-one guarantee is falsified. Additionally, construct a small molecule with two identical substituents (e.g., 2,2-dimethylpropane) and a conformer with rotated groups to see if rounding ever yields duplicate assignments; if duplicates occur, the claimed enforcement is violated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1 (and similarly Algorithms 2-4) solves an fsGW/GW/OT problem to obtain a soft plan P, then converts it to a permutation P' by taking j = argmax_k P_ik per row, and 'always check[s] if P' is a permutation matrix.' If the check fails for a given alpha, that alpha is simply skipped. There is no procedure that enforces a permutation; the algorithm only filters. The abstract's claim that OTMol 'maintains molecular integrity by enforcing one-to-one mappings between entire molecules' is therefore not a property of the optimization. In fact, since the fsGW problem has no constraint preventing two rows from having the same argmax, many-to-one roundings are possible whenever two atoms share the same label and the plan is not sufficiently sharp. For clusters, Algorithm 4's perturbation search similarly appends P' only when it is a permutation; if no perturbation yields one, L is empty and Algorithm 3 has no result. The pseudocode's final 'Take the lowest RMSD_alpha' assumes at least one valid permutation exists, but this is not guaranteed. The paper reports no failure counts across its 101 alpha values or 100 perturbations, so the claimed guarantee is empirically unverified. This is an internal inconsistency: the stated guarantee is not delivered by the described algorithm.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces OTMol, an optimal-transport-based method for RMSD calculation that is invariant to atom ordering. It casts the atom-assignment problem as a fused supervised Gromov-Wasserstein problem (Eq. 4) with label constraints and distance matrices combining Euclidean and graph-geodesic distances, solves it with the POT library, converts the soft transport plan to a permutation by row-wise argmax with a validity check, and then applies Kabsch alignment. For homogeneous atom clusters and molecular clusters, it uses GW/OT with Hadamard powers and perturbation-based sampling of molecule-level assignments. The method is benchmarked against ArbAlign on ATP, imatinib, cyclic peptides, lipids, sugars, DNA, water clusters, neon clusters, and FGG/hydrate systems using RMSD, bond-connectivity inconsistency (BCI), and atom-mapping inconsistency (AMI). The paper claims order-independent, chirality-preserving, one-to-one mapping with low RMSD and O(n^3) scaling.","tokens_in":17818,"tokens_out":7340,"duration_ms":82363,"significance":"OTMol addresses a practical need: order-independent RMSD with chemical fidelity. Strengths include public code and data, use of established OT solvers, the introduction of BCI/AMI as complementary quality metrics, and explicit attention to chirality and molecular-cluster integrity. The comparison to ArbAlign on external benchmark sets is valuable. However, the reported results are weakened by post hoc selection over a large hyperparameter grid and by the lack of a hard permutation constraint in the optimization. The central claims are plausible but require stronger validation before the method can be regarded as a general-purpose tool.","major_comments":[{"comment":"The abstract claims OTMol 'enforces one-to-one mappings', but the algorithms do not enforce this. After solving the fsGW/GW/OT problem, P′ is obtained by taking j = argmax_k P_ik and accepted only if it is a permutation. The optimization itself has no constraint preventing two rows from sharing the same argmax; the algorithm merely filters. Consequently, for some α no valid P′ may exist, and Algorithm 4 can return an empty list, making the final 'Take the lowest RMSD' undefined. No failure counts are reported for the 101 α values, 9 p-values, or 100 perturbations. The stated guarantee is therefore unsupported. Either the method should solve a constrained assignment problem (e.g., Hungarian rounding with proven permutation output) or the claim should be weakened and failure rates reported.","section":"§2.2, Algorithms 1–4"},{"comment":"The selection rule used in the experiments differs from the pseudocode. Algorithm 1 returns the lowest RMSD over α, while §3.1.1 and §2.5.1 state that the user chooses the lowest-RMSD assignment among the set of assignments with the lowest BCI values. This BCI filtering is central to the reported 0% bond-mismatch results, yet it is not part of the provided algorithm listing. The method as described is therefore not fully reproducible; the pseudocode should include the BCI filter or the text should explain how the filter is applied.","section":"§2.5.1 vs Algorithm 1"},{"comment":"The evaluation protocol selects the best result over L_α = {0, 0.01, ..., 1}, dataset-specific c, L_p = {0.5, ..., 4.5}, and l = 100 after BCI filtering. The reported RMSD for each pair is thus a minimum over a large hyperparameter grid, not the performance of a fixed model. ArbAlign has no analogous selection step, so the comparison in Figures 5D/E is not apples-to-apples. Moreover, a single α curve is shown for only one FGG pair (Fig. 4B) and one water-cluster pair (Fig. 5A), so the reader cannot assess sensitivity. To support the 'consistently achieves low RMSD' claim, the paper should report fixed-hyperparameter results, cross-validation, or the distribution of RMSD across the grid.","section":"§3.1.1, §3.2.1–3.2.3"},{"comment":"Several molecular classes in the biomedical dataset consist of a single conformer pair: DNA (215D), DLP, EIC, BGC, and BGCGLC. Table 2 reports ArbAlign BCI statistics across Boltz-2 models, but for OTMol only aggregate statements are given; there is no RMSD distribution with sample sizes. The broad claim that OTMol 'consistently' outperforms across 'a wide range' of systems is not supported by n=1 per class. Additional independent pairs per class, or a clearly qualified statement, are needed.","section":"§3.1.5, Table 2"}],"minor_comments":[{"comment":"The statement that OTMol 'eliminates the need for manually defined cost functions' overstates the contribution; the method still requires choosing atom-label schemes, α, c, p, and l, and uses a label cost matrix in Eq. (4). Suggest revising the wording.","section":"Abstract"},{"comment":"The 'bilevel optimization' formula introduces a threshold ρ and an outer maximization over s that are never defined or used in the algorithms; it is unclear how Eq. (6) relates to the implemented fsGW of Eq. (4). Please clarify or remove.","section":"§4.3, Eq. (6)"},{"comment":"Several typos and notation issues appear: 'Calulate' in Algorithm 3, 'RMSD P′' instead of 'RMSD_P′', and reuse of P for different plans in Algorithm 2. Using distinct symbols would improve readability.","section":"Algorithms 2–3"},{"comment":"The BCI definition 'the same atom pair is found in the aligned structure' is ambiguous: does this refer to atom indices after permuting B? It should be specified whether BCI is computed on the mapping P′ or on the aligned coordinates.","section":"§2.5.1"},{"comment":"There is a typo in the cyclic peptides row: 'smller dihedral variation' should be 'smaller dihedral variation'. Similar typos occur elsewhere, e.g., 'decribed' in §3.1.5.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the journal's scope. The main technical concerns are fixable within the scope of the paper: the argmax-filtering step should be replaced with a genuine permutation constraint or accompanied by failure-rate reporting, and the evaluation should avoid post hoc selection over the hyperparameter grid. If these are addressed, the paper could be acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"OTMol is worth a look if you work on molecular superimposition or RMSD benchmarking. It brings the authors' fused supervised Gromov-Wasserstein formalism to single-molecule and cluster alignment, using graph geodesic distances plus hard label costs to respect chemistry, and it packages this with a hierarchical treatment of water clusters. The external checks—BCI, AMI, and comparison against ArbAlign—are the right kind of grounding, and the paper is honest that low RMSD only matters when connectivity and chirality are preserved. The chirality issue with ArbAlign's reflections is real, and OTMol's no-reflection design is sensible.\n\nThe soft spots are where the claims get ahead of the evidence. First, the one-to-one \"enforcement\" is actually a post hoc filter: solve the OT problem, take argmax per row, then check whether the result is a permutation; if not, skip that alpha. That does not guarantee a permutation exists, and the paper never reports how often the check fails across the 101 alpha values or 100 perturbations. Without failure counts, the abstract's \"enforcing one-to-one mappings\" is unsupported. Second, the evaluation picks the lowest RMSD among all alpha (and p) values after BCI filtering. That is a best-case selection, not a fixed-model evaluation, so the reported numbers are optimistic and hard to compare fairly with ArbAlign's defaults. Third, the cluster algorithm can return nothing when no perturbed instance yields a permutation; again, no failure analysis. These are fixable, but they need to be addressed.\n\nThe math itself is a straightforward application of existing sOT/sGW theory, not a new branch of science, and the paper cites the prior work properly. The contribution is the application and the practical safeguards. A serious referee should see this, but the authors should be pushed to report failure rates, fix the hyperparameter protocol (e.g., a held-out alpha or a default setting), and soften the guarantee language in the abstract.\n\nWho gets value: people building or benchmarking alignment tools, especially for flexible ligands and water clusters. It is not a theoretical breakthrough, but it is a credible engineering contribution with real evidence. I'd probably cite it as related work in a methods paper, and I'd want to discuss it in a reading group, mostly to debate how much the cherry-picked hyperparameters inflate the results.","headline":"A promising application of supervised Gromov-Wasserstein to molecular alignment, but the reported guarantees are softer than the abstract implies and the validation protocol cherry-picks hyperparameters.","tokens_in":18293,"tokens_out":2048,"would_cite":true,"duration_ms":26387,"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":"Atom mapping, not rotation, is the real RMSD problem—OTMol solves it with optimal transport.","keywords":["RMSD","supervised optimal transport","Gromov-Wasserstein","molecular superimposition","atom mapping","chirality preservation","bond connectivity consistency","water clusters"],"falsifier":"Take a reference molecule, randomly permute its atom indices and apply a random rotation, and ask OTMol to align the permuted copy to the original; a correct method must recover the identity assignment with near-zero RMSD and zero BCI. Repeating this over many permutations and flexible conformers settles the claim—any non-identity assignment with nonzero RMSD on such constructed examples falsifies the method.","tokens_in":17341,"feed_emoji":"🧪","tokens_out":8864,"duration_ms":94922,"temperature":0.7,"pith_summary":"OTMol claims that the molecular RMSD problem can be solved properly by treating atom matching as a fused supervised Gromov-Wasserstein optimal transport problem, rather than as a linear assignment with hand-built cost matrices. The method replaces heuristic costs with pairwise intra-molecular distance comparisons—Euclidean plus graph-geodesic distances—and hard atom-label constraints, then applies the Kabsch rotation without reflection. On benchmarks spanning ATP, Imatinib, cyclic peptides, lipids, sugars, DNA, water clusters, and neon clusters, OTMol reports zero bond-connectivity inconsistencies and chirality-preserving alignments, and lower RMSD than ArbAlign for most tested pairs, while running in O(n^3) time. The paper argues that RMSD comparisons are only meaningful when chemical integrity is preserved, so a slightly higher RMSD from OTMol can be the correct answer when the lower value from a baseline comes from an atom mapping that breaks bonds or inverts stereochemistry.","feed_headline":"OTMol maps atoms via optimal transport, preserving bonds and chirality","feed_subtitle":"Low-RMSD alignments with zero bond mismatches across drugs, peptides, sugars, and water clusters.","key_machinery":"Load-bearing object: the fused supervised Gromov-Wasserstein objective (Eq. 4), combining a label cost (Cij = 0 for matching atom labels, infinity otherwise) with a Gromov-Wasserstein term that penalizes mismatches between pairwise distances inside each molecule. Distance matrices are a convex combination of Euclidean distance and graph-geodesic distance, so the plan respects geometry and covalent topology. OTMol scans alpha in [0,1], hard-thresholds each plan into a permutation, filters to the lowest-BCI assignments, and applies Kabsch without reflection. For clusters, the mechanism is hierarchical: molecule-level assignment via perturbed GW on representative coordinates, then block-structu","core_discovery":"Central claim: atom-to-atom correspondence, not the rigid-body fit, is the real bottleneck in RMSD, and the fused supervised Gromov-Wasserstein problem finds that correspondence. OTMol solves Eq. (4) with hard atom-label costs and D = (1 - c)DE + cDG; for each alpha it converts the transport plan to a permutation, keeps assignments with the lowest bond-connectivity inconsistency, and returns the lowest-RMSD one. Reflection is never applied, so chirality is preserved. Clusters of identical atoms use a GW-plus-OT refinement; clusters of identical molecules use a two-level hierarchical match. Result: chemically meaningful RMSD for single molecules and clusters without manual cost matrices.","pith_inferences":["The alpha-grid plus BCI filter compensates for the nonconvexity of Gromov-Wasserstein; a direct optimization of BCI/RMSD could replace the grid and likely improve robustness.","The same distance-preserving matching principle may transfer to point-cloud, surface, or density-map alignment problems where correspondence is unknown.","Because the implementation relies on a generic GW solver, the practical ceiling is set by solver local optima; a permuted-input recovery test would quantify this ceiling and could guide solver choice."],"forward_implications":["Atom-order-independent RMSD: structures from different software or databases can be compared directly, since the same assignment is found regardless of input ordering.","Low RMSD becomes interpretable: a BCI-minimal, chirality-preserving mapping means a low RMSD reflects genuine similarity rather than broken bonds or a reflected mirror image.","Cluster comparisons stop splitting molecules: the hierarchical match ensures each water molecule maps as a whole, removing a failure mode of the baseline.","The fsGW formulation is reusable for molecular analog comparison and maximum-common-substructure search, as the paper notes.","O(n^3) scaling keeps the method practical for larger systems such as neon clusters up to 1000 atoms."],"supporting_citations":[{"why":"Supplies the ArbAlign baseline method and its benchmark datasets (water clusters, FGG, S1-MA-W1, neon clusters) against which OTMol is compared.","marker":"[16]"},{"why":"Kabsch algorithm provides the optimal rigid-body rotation used after each candidate atom assignment.","marker":"[17]"},{"why":"Gromov-Wasserstein distances establish the metric formulation that motivates matching atoms by pairwise intra-molecular distances.","marker":"[30]"},{"why":"Gromov-Wasserstein averaging supplies the GW formulation and algorithmic basis OTMol adapts for cluster alignment.","marker":"[31]"},{"why":"Supervised optimal transport supplies the hard label-cost mechanism that prevents mismatched atom labels in the fused problem.","marker":"[32]"},{"why":"Supervised Gromov-Wasserstein with metric-preserving constraints supplies the sGW framework used as the Gromov term in Eq. (4).","marker":"[33]"},{"why":"Optimal transport for structured data on graphs supplies the fused-GW solver used in OTMol's practical implementation.","marker":"[35]"},{"why":"Water-cluster minima database is the source of the two newly curated water-cluster benchmark sets.","marker":"[39]"},{"why":"Supplies the EMD and fused-GW numerical solvers used to compute OTMol alignments.","marker":"[40, 41]"}],"fun_headline_variants":["OTMol replaces heuristics with optimal transport for RMSD","Chirality-safe molecular alignment via fused Gromov-Wasserstein","Data-driven atom matching beats hand-crafted cost matrices","OTMol: low-RMSD alignments that respect bonds and chirality","Optimal transport solves atom-order ambiguity in RMSD"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The pipeline assumes that the optimal-transport score built from atom labels and pairwise graph distances ranks chemically correct atom matchings above incorrect ones, so that the lowest-RMSD assignment surviving the bond-connectivity filter is the right one.","fun_headline_variants_meta":{"raw":{"variants":["OTMol replaces heuristics with optimal transport for RMSD","Chirality-safe molecular alignment via fused Gromov-Wasserstein","Data-driven atom matching beats hand-crafted cost matrices","OTMol: low-RMSD alignments that respect bonds and chirality","Optimal transport solves atom-order ambiguity in RMSD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000807,"raw_usage":{"total_tokens":3420,"prompt_tokens":824,"completion_tokens":2596,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":2520}},"tokens_in":568,"tokens_out":2596,"duration_ms":21672,"temperature":1.0,"reasoning_tokens":2520,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:25:58.248264+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a reference molecule, randomly permute its atom indices and apply a random rotation, and ask OTMol to align the permuted copy to the original; a correct method must recover the identity assignment with near-zero RMSD and zero BCI. Repeating this over many permutations and flexible conformers settles the claim—any non-identity assignment with nonzero RMSD on such constructed examples falsifies the method.","supporting_citations":[{"cited_title":"ArbAlign: a tool for optimal alignment of arbitrarily ordered isomers using the Kuhn–Munkres algorithm","cited_arxiv_id":null,"evidence_quote":"Supplies the ArbAlign baseline method and its benchmark datasets (water clusters, FGG, S1-MA-W1, neon clusters) against which OTMol is compared."},{"cited_title":"A solution for the best rotation to relate two sets of vectors","cited_arxiv_id":null,"evidence_quote":"Kabsch algorithm provides the optimal rigid-body rotation used after each candidate atom assignment."},{"cited_title":"Gromov–Wasserstein distances and the metric approach to object matching","cited_arxiv_id":null,"evidence_quote":"Gromov-Wasserstein distances establish the metric formulation that motivates matching atoms by pairwise intra-molecular distances."},{"cited_title":"Gromov-wasserstein averaging of kernel and distance matrices","cited_arxiv_id":null,"evidence_quote":"Gromov-Wasserstein averaging supplies the GW formulation and algorithmic basis OTMol adapts for cluster alignment."},{"cited_title":"Supervised optimal transport","cited_arxiv_id":null,"evidence_quote":"Supervised optimal transport supplies the hard label-cost mechanism that prevents mismatched atom labels in the fused problem."},{"cited_title":"Supervised Gromov–Wasserstein Optimal Trans- port with Metric-Preserving Constraints","cited_arxiv_id":null,"evidence_quote":"Supervised Gromov-Wasserstein with metric-preserving constraints supplies the sGW framework used as the Gromov term in Eq. (4)."},{"cited_title":"Optimal transport for structured data with application on graphs","cited_arxiv_id":null,"evidence_quote":"Optimal transport for structured data on graphs supplies the fused-GW solver used in OTMol's practical implementation."},{"cited_title":"Atlas of putative minima and low-lying energy networks of water clusters n= 3–25","cited_arxiv_id":null,"evidence_quote":"Water-cluster minima database is the source of the two newly curated water-cluster benchmark sets."}],"review_version":1}