acceptodds
Under review as a conference paper at ICLR 2027

Hypergraph-guided Complex Structure Enhancement for Accurate 3D Molecule Generation

Abstract

To accelerate the development of drug discovery, de novo 3D molecule generation has achieved many advances, aiming to synthesise accurate and reasonable 3D molecules from the given noise. Since molecules usually involve complex and diverse structures, it is challenging to precisely construct atoms and corresponding relations. Currently, most methods mainly employ flow- or diffusion-based techniques to effectively improve the quality of 3D molecules. However, they usually rely on atom-level and pairwise relation modelling, which may be insufficient for explicitly capturing higher-order chemical structures, e.g., aromatic rings, resulting in unstable patterns, unrealistic scaffolds, and degraded chemical validity in the generated molecules. To this end, inspired by the powerful relationship modelling capabilities of hypergraphs, we explore exploiting hypergraphs to perform higher-order structure enhancement. And a method of HSA-Mol3D is proposed. Concretely, HSA-Mol3D introduces a dynamic hypergraph branch into the generative backbone to model higher-order chemical structures. We further construct a staged and stabilised hypergraph fusion strategy in conjunction to make such higher-order information stable and effective throughout generation. Our proposed method enhances the stability of higher-order structures by progressively integrating hypergraph information into the generation process of the primary flow field model. Experiments show that our method achieves competitive performance on drug-like molecular generation, with particularly strong results in scaffold quality and the investigation of novel, higher-order skeletal medicinal molecules with superior properties.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
3D Molecular Generation, Hypergraph, Flow Matching, Higher-Order Chemical Structures, De Novo Molecular Design, Molecular Graph Generation, AI for Drug Discovery

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_aQPuOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary
  • The paper introduces HSA-Mol3D, a co-generation model for molecular graphs and 3D coordinates. HSA-Mol3D generates atom types, formal charges, bonds, and coordinates using a flow-based backbone augmented with a dynamic typed hypergraph branch for rings, aromatic systems, and conjugated structures.
  • HSA-Mol3D learns and refines hyperedges from intermediate molecular states and uses staged fusion to incorporate higher-order structural information into the generative backbone. Experiments on QM9 and ablation studies on GEOM-Drugs evaluate molecular validity, geometry, ring and scaffold distributions, and molecular properties, with component ablations assessing refinement, bond-aware construction, and fusion.
Critical strengths

None

Critical weaknesses
  • Regarding methodological novelty: previous graph or molecular generative models have shown that computing cycle, spectral, molecular, or RRWP features from noisy molecules and providing them as additional inputs to the denoiser improves generative fidelity [1, 2]. Thus, adding information that a neural network may not adequately capture is already a well-known strategy for molecular generation. In HSA-Mol3D, learned higher-order structural information represented by dynamic typed hypergraphs is added during generation. However, the paper does not sufficiently establish what methodological contribution the proposed hypergraph mechanism offers beyond existing structural feature augmentation strategies. I agree with the authors that the reported ablations show performance gains, but these gains alone do not establish methodological novelty beyond existing structural feature augmentation strategies.
  • Limited baseline comparison: QM9 contains molecules with at most nine heavy atoms, making evaluations on larger molecules important for assessing the proposed method’s gains. Although the paper reports results on GEOM-Drugs, broader comparisons with recent models (e.g., SemlaFlow) on this more challenging dataset would provide a more informative assessment of its performance gains.

[1] DiGress: Discrete Denoising Diffusion for Graph Generation (ICLR, 2023)

[2] DeFoG: Discrete Flow Matching for Graph Generation (ICML, 2025)

Questions for authors
  • Regarding conventional feature augmentation: What additional benefits does the proposed hypergraph representation offer over conventional graph feature augmentation, such as RRWP [2], cycle/ring statistics [1], and spectral features? Could the authors provide a controlled comparison with these alternatives using the same generative backbone and matched training and sampling settings?
Reviewer_C7WMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces HSA-Mol3D, a joint 3D molecular generation framework that augments an equivariant flow-based backbone with dynamic hypergraph representations of rings, aromatic systems, and conjugated components. The method constructs candidate hyperedges from predicted node, pair, and bond information, refines their representations with set-level supervision, and injects their features into the backbone through confidence-controlled, staged fusion. Experiments on QM9 and GEOM-Drugs evaluate chemical validity, distributional similarity, and geometric quality. Additional ablations investigate the effects of fusion timing, bond-aware construction, and learned proposal refinement. The reported results are promising, particularly for validity and several structural metrics on GEOM-Drugs, although aspects of the evaluation protocol and higher-order recovery analysis require clarification.

Critical strengths

A well-motivated integration of dynamic chemical structure into joint 3D generation. The method addresses the practical difficulty that rings, aromatic systems, and conjugated components are uncertain during intermediate generation steps. Inferring these structures from the evolving molecular state and controlling their contribution through confidence and time is a sensible design. The resulting framework offers an explicit mechanism for incorporating chemically meaningful set-level context into atom- and pair-level predictions.

Useful ablations supporting the role of delayed structural feedback. The appendix includes an auxiliary-only baseline, early fusion, rule-based proposals, and removal of bond-aware modulation. These comparisons go beyond evaluating the full model against external methods. For example, relative to auxiliary-only training, staged fusion improves scaffold similarity from 0.233 to 0.542, ring-size TV from 0.170 to 0.092, and bond-angle Wasserstein distance from 3.506 to 1.650. The deterioration of several metrics under early fusion also supports the motivation for delaying intervention.

Critical weaknesses

Evaluation inconsistencies and incomplete comparisons weaken confidence in the reported advantages. In Table 5, the auxiliary-only baseline has V&U = 0.904 but V&U&N = 0.913. If these are fractions over the same generated sample set, adding the novelty requirement cannot increase the score. This may be a reporting error, but it affects a central controlled comparison and needs correction. More broadly, Table 2 evaluates geometric quality against a substantially different baseline set from Table 1, omitting strong comparators such as FlowMol3 and UDM-3D. Consequently, the paper does not establish whether its geometric improvements hold against the strongest methods included elsewhere. The manuscript also does not clearly report generation sample counts, uncertainty across runs, or whether baseline results were reproduced under identical preprocessing and evaluation settings. This is especially consequential for small differences such as scaffold similarity of 0.542 versus 0.540.

The higher-order recovery analysis does not yet fully validate accurate recovery of complex structures. Table 7 evaluates raw proposals before learned membership refinement and uses non-exclusive matching at an atom-set IoU threshold of 0.5. Under this protocol, merged or duplicate proposals can receive favorable coverage scores without recovering distinct structures correctly. This matters because the constructor extracts connected components and their 2-cores, which do not necessarily separate individual rings in fused or spiro systems. The exact ground-truth definitions of rings, aromatic systems, and conjugated components are also insufficiently specified. Moreover, Full HSA has lower raw ring F1 than Rule-only (0.9307 versus 0.9858), despite better downstream ring-distribution fidelity. The authors acknowledge this difference, but evaluation of the refined hyperedges is needed to clarify what the learned refinement actually recovers. These limitations leave the central claim of improved higher-order structural modeling only partially substantiated.

Questions for authors

Can you clarify the denominators and reference sets used for V&U and V&U&N, and resolve the Table 5 inconsistency? Why do test-set molecules in the “Data” rows have zero novelty, and which reference set defines novelty for generated molecules?

Which baseline results were reproduced and which were taken from prior work? Please specify generation sample counts, preprocessing, and evaluation settings, and provide matched geometric-quality results for the strongest relevant baselines, particularly FlowMol3 and UDM-3D.

How are ground-truth ring, aromatic, and conjugation hyperedges defined, particularly for fused, bridged, and spiro systems? Can you report recovery after membership refinement using one-to-one matching, alongside the current non-exclusive coverage metric?

Does Figure 4 evaluate corrupted held-out molecules along the prescribed training interpolation, or actual free-running generation trajectories? If it uses the former, what evidence shows that hyperedge reliability and the confidence gate remain effective on states encountered during sampling?

Reviewer_wqaQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces dynamic, chemistry-typed hypergraph states into a joint discrete–continuous flow model for 3D molecular generation. I find the motivation appealing, particularly the decision to infer hyperedges from the evolving state rather than assume access to fixed structural priors. The staged-fusion design is also reasonable: early structural predictions are noisy, and the reported early-fusion results support the authors’ concern about error amplification.

Critical strengths

The core idea is sensible: rings, aromatic systems, and conjugated components are natural set-level chemical structures, so adding a dynamic hypergraph branch to a joint discrete–continuous 3D flow model is a reasonable inductive bias. Building hyperedges from the model’s evolving state, rather than assuming oracle structural priors, makes the approach usable during unconditional generation, and the staged-fusion design is well motivated because early structural predictions are noisy. The ablation suite is also useful—especially the Aux-only, early-fusion, Rule-only, and no-bond-aware variants—and some GEOM-Drugs gains are large enough to warrant follow-up (e.g., ring-size TV and bond-angle W1W1​).

Critical weaknesses

The main problem is trust in the empirical claims. Table 2’s FuncMol row appears to mix FuncMol and VoxMol values, and the claimed best QED no longer survives comparison with the published FuncMol number. There are also internal inconsistencies: Aux-only has V&U =0.904 but V&U&N =0.913, and the Data rows are described as test-set samples yet have zero novelty. The evaluation protocol is underspecified—sample counts, independent runs, variance, baseline rerun details, reference sets, and the energy definition are all missing or unclear, so small differences like 0.542 vs. 0.540 cannot support strong claims.

Conceptually, the paper overstates what ordinary molecular graphs cannot represent; the real issue is inductive bias and explicit set-level supervision, not missing information. A parameter-matched direct-feature or generic set-encoder baseline would better isolate the architectural contribution. The hyperedge definitions and recovery metric also need tightening: ground-truth extraction is underdefined, matching is non-exclusive, and Rule-only actually beats Full HSA on ring recovery, which narrows the learned-constructor claim. Figure 4’s recovery trend is partly an activation-schedule artifact, so it does not by itself show improving inference. Finally, the paper should discuss trade-offs (weak Frag/Scaf results), omitted recent baselines, overstated “medicinal” claims, computational cost, incomplete training details, and the size-stratified Figure 5 analysis.

Questions for authors

Can you provide a corrected and fully audited version of Tables 1–8, together with the exact source or rerun protocol for every baseline entry?

How are validity, uniqueness, and novelty computed, and why do the Data rows have zero novelty despite being described as test samples?

How many generated molecules and independent runs underlie each result? Do the main improvements remain significant under repeated sampling?

Table 7 and Eqs. (12)–(15) allow multiple predictions to match the same ground-truth component, while each component contributes at most once to coverage recall. This asymmetry can inflate precision and F1. Could you repeat the evaluation using one-to-one Hungarian matching at the same IoU ≥ 0.5 threshold and report precision, recall, and F1 alongside the current results?

How does the method compare with parameter-matched direct chemical-feature or generic set-encoding baselines, rather than only the current Rule-only variant?

What are the training and inference costs introduced by the dynamic hypergraph branch?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.