{"id":"fabdc717-326c-4343-88da-ebfa4be7f25c","arxiv_id":"2505.05650","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A plug-and-play combination of an equivariant GNN encoder and a hypergraph network with conjugated-bond hyperedges improves property prediction on large molecules, with inconsistent gains on small molecules.","lead":"This paper tests a simple recipe: take embeddings from an equivariant neural network, feed them into a hypergraph model where hyperedges are conjugated bonds, and predict molecular properties. The mix helps on large molecules, but the abstract overstates the geometry gain on one dataset and the equivariance property is never proven.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The strongest empirical claim is contradicted by Table 5: on Molecule3D, adding geometry to the hypergraph increases MAE from 117.55 to 122.25 meV, so 'adding geometric features further improves performance' is unsupported.","rationale":"The reader's strongest_claim is empirical, so I prioritise the empirical contradiction over the unproven equivariance claim. The paper's own tables provide direct evidence against a universal 'geometric features further improve' statement: Molecule3D shows EGNN-MHNN worse than MHNN, and OPV polymer tasks show similar reversals. This does not destroy the entire paper, because high-order structure still beats 2D graphs on the large datasets and the framework is modular, but the abstract and conclusion overstate the findings. The reader's CONDITIONAL verdict remains appropriate, with required revisions to qualify the central claim and to evaluate on official splits. I also note the equivariance issue: §3.2 asserts that initializing node features with backbone embeddings 'naturally embeds equivariant information', but no group action is defined on the hypergraph module, and the AllSet update operates on scalar embeddings, so the full pipeline is at most invariant for these scalar regression tasks. That issue is real but secondary to the empirical contradiction.","tokens_in":12591,"tokens_out":7895,"duration_ms":91064,"concrete_test":"Run the Molecule3D HOMO-LUMO task on the official benchmark split with identical hyperedge construction (RDKit conjugated bonds) and at least five seeds, comparing MHNN and EGNN-MHNN under matched training budgets. If EGNN-MHNN does not significantly beat MHNN (or remains worse), the abstract's 'adding geometric features further improves' claim must be restricted to specific datasets/backbones; the same check can also verify whether the reported 80/10/10-split results reproduce.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 5 (§4.3.4) directly contradicts the abstract's claim that adding geometric features to high-order structures further improves performance. On Molecule3D, the geometry-free MHNN achieves 117.55 ± 0.28 meV, while EGNN-MHNN — the only equivariant variant run on this large-molecule benchmark — achieves 122.25 ± 0.24 meV, a degradation of ~4.7 meV that is well outside the error bars. Table 3 also shows multiple OPV polymer tasks where EGNN-MHNN and FAFormer-MHNN underperform the geometry-free MHNN (e.g., polymer εHOMO: 66.67/54.85 vs 49.93 meV; polymer εLUMO: 69.32/72.05 vs 60.71 meV). Thus the central claim, stated as a general design rule, is not supported by the paper's own results; the benefit is dataset- and backbone-dependent. The claim that high-order interactions give only limited benefits on small molecules is also strained by OPV small-molecule tasks, where MHNN clearly beats GIN/GAT (e.g., Δε 34.02 vs 50.45/55.8 meV). The equivariance claim in §3.2 is also asserted without proof, but the empirical contradiction alone is decisive for the strongest_claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EquiHGNN, a framework that combines an equivariant geometric backbone (EGNN, FAFormer, or Equiformer) with a hypergraph neural network (MHNN/AllSet) for molecular property prediction. The authors construct hyperedges from conjugated bonds and initialize hypergraph node features with backbone embeddings plus scalar features. They report experiments on QM9, OPV, PCQM4Mv2, and Molecule3D and claim that high-order interactions give limited benefits on small molecules but consistently outperform 2D graphs on larger molecules, and that adding geometric features to high-order structures further improves performance. The paper also makes a methodological claim that the resulting pipeline is rotationally equivariant.","tokens_in":12880,"tokens_out":3891,"duration_ms":40088,"significance":"If the empirical pattern were robust, the paper would provide a useful design rule: hypergraph structure matters most for large molecules and 3D geometry adds value on top of high-order representations. The plug-and-play architecture is simple and could be a practical baseline for combining equivariant backbones with hypergraph message passing. The manuscript also has strengths: it evaluates on multiple public benchmarks, reports error bars, and ships code at a public repository. However, the central empirical claim is contradicted by the paper's own Table 5 and partly by Table 3, and the equivariance claim is asserted without a group-action definition or proof. These issues are load-bearing for the title and abstract, so the current version overstates what is demonstrated.","major_comments":[{"comment":"The abstract and Section 1 state that adding geometric features to high-order structures further improves performance, but Table 5 contradicts this on Molecule3D: MHNN achieves 117.55 ± 0.28 meV while EGNN-MHNN achieves 122.25 ± 0.24 meV, a degradation of about 4.7 meV that is well outside the reported error bars. Since the conclusion is phrased as a general design rule, this single counterexample is decisive and the claim needs to be restricted to the datasets and backbones where the improvement actually holds.","section":"§4.3.4, Table 5"},{"comment":"The name \"Equivariant HyperGraph Neural Network\" and the statement that the design \"naturally embeds equivariant information into the model\" are not supported by the construction as written. The hypergraph module is the standard AllSet/MHNN with permutation-invariant set functions; no group action of SO(3) or E(3) is defined on the node or hyperedge feature spaces, and no theorem shows that the hypergraph message-passing updates commute with rotations or translations. If the backbone outputs invariant scalar embeddings, the composed model is invariant, not equivariant. The authors should either define the group action on all intermediate representations and prove equivariance preservation, or reframe the contribution as a hypergraph model with geometric feature initialization rather than a rotationally equivariant model.","section":"§3.2"},{"comment":"The claim that high-order interactions offer limited benefits for small molecules is contradicted by the OPV molecular tasks. MHNN achieves 34.02 meV on Δε versus 50.45 and 55.8 meV for GIN and GAT, and 24.46 meV on εLUMO versus 53.29 and 46.68 meV; these are substantial improvements, not limited benefits. Conversely, on polymer tasks MHNN's advantage over 2D graphs is mixed, since GAT achieves a lower error than MHNN on polymer Δε (47.91 vs 48.95 meV). The abstract and Section 4.3.2 should be rewritten to describe the actual dataset- and task-dependent pattern rather than a uniform small-molecule/large-molecule dichotomy.","section":"Table 3, §4.3.2"},{"comment":"The large-scale comparisons rest on nonstandard evaluation protocols. PCQM4Mv2 experiments use only the training subset where 3D structures are available, with a custom 80-10-10 split, rather than the official PCQM4Mv2 train/validation/test split; Molecule3D is likewise split 80-10-10 with only one equivariant variant (EGNN-MHNN) evaluated on both large datasets. Consequently, the claim that high-order interactions \"consistently outperform 2D graphs on larger ones\" is based on custom splits and a single backbone, and the results are not directly comparable to standard benchmark numbers. Please specify the exact comparison protocol and test whether the conclusion holds under standard splits and with more than one equivariant backbone.","section":"§4.1, §4.3.3, §4.3.4"}],"minor_comments":[{"comment":"There are typos that should be corrected: \"symetry\" in Section 2.3, \"prediciotn\" and \"hidene\" in Section 3.1, and \"PCMQM4Mv2\" in the Table 4 caption.","section":"§2.3, §3.1, §4.3.4"},{"comment":"In the paragraph before Table 2, the citation for GAT appears as \"[?]\" and should be resolved to reference [37].","section":"§4.3"},{"comment":"The Equiformer-MHNN row is misformatted for the Δε and ⟨R²⟩ columns; the values \"230.77±1.8 102815±445.433\" need to be split clearly so the table can be parsed.","section":"Table 2"},{"comment":"The phrase \"slightly higher error\" understates the EGNN-MHNN result on Molecule3D; the difference is about 4.7 meV and is outside the overlapping error bars, so the text should acknowledge that geometric features degraded accuracy on this benchmark.","section":"§4.3.4"},{"comment":"The limitations section reports that rings performed \"significantly worse\" than conjugated bonds but gives no numbers; adding this result would make the hyperedge-construction discussion more informative.","section":"§5, §6"}],"recommendation":"major_revision","confidential_remarks":"The paper's empirical contribution is potentially useful, but the title and abstract currently overclaim both equivariance and the benefit of geometric features. The Molecule3D result in Table 5 is a direct counterexample to the abstract's main design rule, and the equivariance claim in Section 3.2 lacks a mathematical formulation. These issues can be fixed by re-scoping the claims and either proving equivariance or renaming the framework, so I recommend major revision rather than rejection. The self-citations to the corresponding author's prior work are related-work citations and do not appear to be inappropriate. The custom 80-10-10 splits on PCQM4Mv2 and Molecule3D should be flagged prominently in any revision, since they limit comparability with published leaderboard results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a straightforward engineering study: take an equivariant GNN (EGNN, FAFormer, Equiformer), use its embeddings as node features for an AllSet/MHNN hypergraph module built on conjugated-bond hyperedges, and measure MAE on QM9, OPV, PCQM4Mv2, Molecule3D. That combination is new as a published benchmark, the datasets are public, and the authors report error bars. The code link is given. For a practitioner who wants to know whether this recipe helps, the tables are informative.\n\nThat said, the central claims need reworking. The abstract says adding geometric features to high-order structures further improves performance. Table 5 shows the opposite on Molecule3D: MHNN without geometry gets 117.55 meV, EGNN-MHNN gets 122.25 meV, a ~4.7 meV degradation outside the error bars. Table 3 also has polymer tasks where EGNN-MHNN and FAFormer-MHNN lose to geometry-free MHNN (polymer εHOMO 66.67 and 54.85 vs 49.93 meV). So the benefit is dataset- and backbone-dependent. The body actually acknowledges this in Section 4.3.4, but the abstract and conclusion state it as a general rule. That mismatch should be fixed.\n\nThe equivariance claim is also loose. The paper says the pipeline 'naturally embeds equivariant information' because the backbone is equivariant, but the hypergraph module is AllSet, which is permutation-invariant and operates on node features. If those features are invariant scalars, the whole model is invariant, not equivariant. No group action is defined on the hypergraph side. The title promise of a rotationally equivariant hypergraph network is unsupported. Either prove the property or describe the model as invariant with equivariant feature initialization.\n\nMethodology: PCQM4Mv2 and Molecule3D use an 80-10-10 split of the training subset because 3D geometries are only available there. That is reasonable for a self-consistent comparison, but it means the numbers are not comparable to published baselines that use the official split. The authors should say so clearly and ideally also evaluate on the standard split. The QM9 table has a likely typo: Equiformer-MHNN reports 230.77 meV for Δε and 102815 for ⟨R²⟩, far above all other rows; those entries need checking. Minor: the GAT citation in Section 4.3 is a missing placeholder.\n\nThe self-citations (refs 19, 40, 48) are to actual related work in equivariant and higher-order molecular networks, so I don't see a problem there.\n\nVerdict: this is a reasonable empirical study with useful benchmark data, but the paper overclaims in the abstract and hasn't demonstrated the equivariance property. With a revised abstract, a clear statement of invariance vs equivariance, and a corrected QM9 table, it could be a decent contribution. I'd send it to a serious referee, but the referee should be instructed to push on those specific points.","headline":"Useful empirical study on equivariant backbones plus hypergraph pooling, but the abstract's design rule is contradicted by its own Molecule3D table and the equivariance claim is asserted, not proven.","tokens_in":13429,"tokens_out":3085,"would_cite":false,"duration_ms":31410,"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":"EquiHGNN shows hypergraph structure plus rotation-aware geometric embeddings outperforms 2D graph models on large molecules.","keywords":["hypergraph neural networks","rotation equivariance","molecular property prediction","geometric deep learning","higher-order interactions","conjugated bonds","large-scale molecular datasets","QM9"],"falsifier":"Run the trained EquiHGNN on rotated copies of the same molecules and check whether the prediction changes by the corresponding rotation or stays identical; if it stays identical, the model is invariant and the equivariance claim is unsupported. A second check is to rerun PCQM4Mv2 with 3D coordinates available for the entire test set, since the current setup only has coordinates for the training subset.","tokens_in":12349,"feed_emoji":"🧪","tokens_out":7517,"duration_ms":81038,"temperature":0.7,"pith_summary":"This paper argues that the best way to combine geometric and high-order structure for molecular property prediction is simpler than it sounds: embed atoms with an equivariant geometric model, then feed those embeddings into a standard hypergraph neural network whose hyperedges are conjugated bonds. The experiments claim that hypergraph message passing alone gives little on small molecules such as QM9 and OPV monomers, but consistently beats 2D GNNs on large molecules such as PCQM4Mv2, Molecule3D, and OPV polymers. Adding 3D geometric features on top of hypergraphs improves accuracy further on most datasets. If the pattern holds, practitioners get a modular, plug-and-play recipe: reuse a strong equivariant backbone, add a hypergraph readout, and scale to large chemical datasets without designing custom equivariant topological layers.","feed_headline":"Why hypergraph + 3D geometry beats 2D graphs on large molecules","feed_subtitle":"EquiHGNN: rotation-aware embeddings plus conjugated-bond hyperedges beat 2D GNNs on large molecules.","key_machinery":"The core mechanism is a two-stage pipeline: a rotation-equivariant backbone first produces symmetry-aware atom embeddings from 3D coordinates, and those embeddings initialize the node features of an AllSet-style hypergraph neural network, which models conjugated bonds as hyperedges and updates representations through two permutation-invariant set functions, one from nodes to hyperedges and one back. This carries the argument by turning the hard problem of equivariant hypergraph message passing into a feature-initialization problem, so the topology module stays standard and scalable while the geometry enters through the input representation.","core_discovery":"EquiHGNN's central claim is that you can get most of the benefit of equivariant geometric deep learning in a hypergraph model without designing an equivariant hypergraph message-passing layer: take an equivariant backbone to produce symmetry-aware atom embeddings, combine them with invariant scalar features, and use those as initial node features in the AllSet/MHNN hypergraph framework with conjugated bonds as hyperedges. The paper reports that this hybrid consistently beats 2D GNNs and the non-geometric MHNN baseline on large molecules, and that adding geometry further improves accuracy on QM9, OPV, and PCQM4Mv2. The one reported exception is Molecule3D, where plain MHNN edges out the geometric EGNN-MHNN variant. The authors read the overall pattern as evidence that high-order interactions matter most at scale and that spatial information is a complementary signal.","pith_inferences":["Because the equivariant information enters only through initial node features, the same recipe could transfer to other invariant topological networks: precompute embeddings from a strong equivariant backbone, freeze or fine-tune them, and keep the topological layer unchanged.","The Molecule3D result (plain MHNN 117.55 meV versus EGNN-MHNN 122.25 meV) suggests a boundary condition worth testing: geometric features may hurt when coordinates are noisy or when the dataset's property depends mostly on topology.","A direct rotation test on the full pipeline would settle whether 'rotationally equivariant' applies to the whole model or only to the backbone; if the backbone outputs invariant scalar embeddings, the hybrid is invariant, not equivariant.","The hyperedge construction is limited to conjugated bonds, and the paper notes rings performed worse; a systematic comparison of chemically meaningful high-order units could map when hypergraph structure helps most."],"forward_implications":["On large molecules, hypergraph models beat 2D GNNs, so high-order interactions become more valuable as molecular size grows.","Adding geometric features on top of hypergraphs further lowers MAE on QM9, OPV, and PCQM4Mv2, so spatial information and hyperedges complement each other.","Equivariant embeddings can be injected at initialization rather than by redesigning hypergraph message passing, yielding a modular recipe that reuses existing public backbones.","On small molecules such as QM9, plain hypergraph models underperform 2D GNNs; geometric features are needed to realize gains.","The Molecule3D result shows the geometric benefit is not universal, so the design rule likely depends on dataset characteristics such as conformational flexibility or coordinate reliability."],"supporting_citations":[{"why":"Supplies the AllSet multiset framework that the hypergraph module builds on.","marker":"[13]"},{"why":"MHNN is the hypergraph baseline that EquiHGNN extends and compares against.","marker":"[14]"},{"why":"Provides the scalarization-based equivariant backbone used to initialize node features, especially for large-scale datasets.","marker":"[31]"},{"why":"Provides the Fourier-domain equivariant backbone that achieves the strongest results on QM9 and OPV.","marker":"[32]"},{"why":"Supplies the frame-averaging equivariant backbone used as another geometric feature extractor.","marker":"[24]"},{"why":"QM9 is the small-molecule benchmark where hypergraph-only models underperform.","marker":"[27]"},{"why":"OPV provides both small-molecule and polymer tasks that support the size-dependent conclusion.","marker":"[28]"},{"why":"PCQM4Mv2 is the large-scale benchmark where hypergraph modeling and geometry both improve results.","marker":"[29]"},{"why":"Molecule3D is the large-scale benchmark where the geometric variant performs worse than plain hypergraph modeling.","marker":"[30]"},{"why":"GIN is the 2D graph baseline that the hypergraph models must beat on large molecules.","marker":"[36]"}],"fun_headline_variants":["Equivariant features plus hyperedges beat 2D GNNs on large molecules","Equivariant embeddings as initial nodes boost hypergraph models","High-order interactions pay off mainly on large molecules","Equivariant features, not layers, drive hypergraph gains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline is only as equivariant as the backbone that produces the initial atom embeddings; the hypergraph message-passing stage is a standard invariant layer with no defined rotation action, so if the backbone emits invariant scalars, the model is invariant rather than rotationally equivariant.","fun_headline_variants_meta":{"raw":{"variants":["Equivariant features plus hyperedges beat 2D GNNs on large molecules","Equivariant embeddings as initial nodes boost hypergraph models","High-order interactions pay off mainly on large molecules","Equivariant features, not layers, drive hypergraph gains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001758,"raw_usage":{"total_tokens":6916,"prompt_tokens":898,"completion_tokens":6018,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":5955}},"tokens_in":514,"tokens_out":6018,"duration_ms":39178,"temperature":1.0,"reasoning_tokens":5955,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:00:26.490362+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained EquiHGNN on rotated copies of the same molecules and check whether the prediction changes by the corresponding rotation or stays identical; if it stays identical, the model is invariant and the equivariance claim is unsupported. A second check is to rerun PCQM4Mv2 with 3D coordinates available for the entire test set, since the current setup only has coordinates for the training subset.","supporting_citations":[{"cited_title":"Molecular hypergraph neural networks.The Journal of Chemical Physics, 160(14):144307, 2024","cited_arxiv_id":null,"evidence_quote":"MHNN is the hypergraph baseline that EquiHGNN extends and compares against."},{"cited_title":"Equiformer: Equivariant graph attention transformer for 3d atomistic graphs","cited_arxiv_id":null,"evidence_quote":"Provides the Fourier-domain equivariant backbone that achieves the strongest results on QM9 and OPV."},{"cited_title":"Protein-nucleic acid complex modeling with frame averaging transformer","cited_arxiv_id":null,"evidence_quote":"Supplies the frame-averaging equivariant backbone used as another geometric feature extractor."},{"cited_title":"Message-passing neural networks for high- throughput polymer screening.The Journal of chemical physics, 150(23), 2019","cited_arxiv_id":null,"evidence_quote":"OPV provides both small-molecule and polymer tasks that support the size-dependent conclusion."}],"review_version":1}