- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes ScaleMol, a novel hierarchical framework for de novo 3D molecular generation. The method introduces resolution-consistent Molecular Mass Grids (MMGs) to represent molecular structures at different spatial scales. Building on a shared 3D U-Net, it employs a coarse-to-fine hierarchical readout mechanism to progressively predict molecular mass distributions, incorporating mass correction and multi-scale supervision. The predicted molecular fields are then used for Walk–Jump sampling, followed by atom recovery to construct 3D molecules. Experiments on QM9 and GEOM-Drugs demonstrate improvements in molecular stability, particularly on the more challenging GEOM-Drugs dataset. Ablation studies further validate the contributions of the MMG representation and hierarchical prediction design.
Critical strengths
-
Clear Presentation and Well-Motivated Design. The paper is clearly written. The motivation behind the proposed Molecular Mass Grids (MMGs) and hierarchical prediction framework is easy to follow, and the methodological components are explained well.
-
Comprehensive Experimental Evaluation. The paper provides extensive experiments on QM9 and GEOM-Drugs, including comparisons with existing approaches and systematic ablation studies. These experiments help demonstrate the effectiveness of the proposed representation and hierarchical design.
Critical weaknesses
-
Limited Explicit Modeling of Molecular Connectivity. ScaleMol primarily generates molecular mass distributions and relies on post-generation atom recovery and bond inference to construct molecular graphs. Unlike approaches such as HierDiff[1] or EDM[2], which explicitly incorporate chemical connectivity into the generation process, ScaleMol does not directly model bond formation or molecular topology. While this design simplifies the generation task, it raises concerns about whether the proposed framework can adequately capture complex chemical structures, particularly for larger molecules.
-
Unclear Methodological Advancement over VoxMol[3]. ScaleMol builds upon the field-based molecular generation framework introduced by VoxMol, replacing occupancy grids with Molecular Mass Grids and incorporating hierarchical prediction and multi-scale supervision. It largely retains the original field-based generation paradigm. Moreover, the use of dense volumetric representations continues to impose substantial computational costs.
-
Unclear Distinction Between Hierarchical Prediction and Hierarchical Generation. Although the paper emphasizes hierarchical molecular generation, the proposed coarse-to-fine mechanism operates primarily within each denoising network evaluation. The model extracts multi-resolution features through a shared U-Net and progressively predicts molecular mass grids, while the actual Walk–Jump sampling process evolves only the finest-resolution field. Consequently, the method appears closer to a single-scale generative process enhanced by hierarchical feature prediction and supervision than to a genuinely hierarchical generative process.
[1] Qiang, Bo, et al. "Coarse-to-fine: a hierarchical diffusion model for molecule generation in 3d." International conference on machine learning. PMLR, 2023.
[2] Hoogeboom, Emiel, et al. "Equivariant diffusion for molecule generation in 3d." International conference on machine learning. PMLR, 2022.
[3] O Pinheiro, Pedro, et al. "3D molecule generation by denoising voxel grids." Advances in Neural Information Processing Systems 36 (2023): 69077-69097.
Questions for authors
-
Could the authors clarify how the proposed hierarchy provides meaningful structural guidance during molecular generation? Since Walk–Jump sampling evolves only a finest-resolution noisy state and uses the final fine-scale score, it is unclear how the intermediate coarse-scale predictions influence the generation of chemically meaningful global structures, rather than simply improving feature extraction and denoising accuracy.
-
Followed by the previous question, could the authors provide a controlled comparison against independent multi-scale prediction heads with identical supervision, and analyze whether the coarse-scale predictions capture and preserve meaningful molecular structures throughout sampling?
-
Could the authors clarify the rationale for not explicitly modeling chemical bonds during generation? Since ScaleMol targets de novo 3D molecular generation, why are bond connectivity and bond types inferred only after atom recovery rather than jointly predicted with atomic structures? Could the authors discuss the limitations of this design and investigate whether incorporating bond-aware supervision or an explicit edge-prediction head would improve chemical validity, molecular topology, and distribution fidelity, particularly on GEOM-Drugs?
-
Could the authors explain why the substantial improvements in molecular stability do not consistently translate into better distribution matching? In particular, ScaleMol achieves 88.3% molecule stability on GEOM-Drugs but underperforms several baselines in atom-type TV, valency W1, and bond-angle W1. Does the proposed hierarchical mass-based representation introduce a bias toward certain molecular compositions or structurally stable configurations?
I will increase my score if authors can solve my concerns.
5 comments
Response to Reviewer cBDh (2/2)
Bond modeling and connectivity (W1, Q3)
Our decision not to model chemical bonds during generation reflects both the continued use of distance-based bond inference in many models and the practical challenges of representing bonds in field-based methods such as MMGs, rather than to simplify the task. Although some current models directly model bond types, inferring bonds from generated atomic geometry remains common in point-cloud models like GeoBFN [1] and UniGEM [2] (EDM likewise infers bond types from generated interatomic distances rather than explicitly incorporating chemical connectivity into generation [3]), and is standard practice in field-based generators such as VoxMol [4], FuncMol [5], and VecMol [6]. As a more fundamental, general-purpose computational chemistry method, density functional theory also uses electron density and nuclear positions without predefined bond types [7], supporting the chemical rationale for this choice.
For explicit joint atom–bond generation, the main difficulty for MMGs is representational. Each atom contributes unit mass centered at its coordinates, so every scale measures regional element content. Bonds, which are essentially a visual representation of the interactions between atoms, are not additional atom-count mass sources with independently specified positions. Thus, explicitly representing bonds requires an additional relational representation and a mechanism to associate it with atoms recovered from the fields. This is a nontrivial extension. Meanwhile, the newly added Appendix N reports a single-fragment rate of 84.41% for ScaleMol on GEOM-Drugs, exceeding the reported rates for VoxMol (82.6%) and FuncMol (70.5%) [5], further highlighting the potential of MMG-based generation in connectivity modeling.
Adding auxiliary bond- or connectivity-aware supervision remains a promising extension and a valuable research question. However, this requires more than simply adding channels to the representation. Without a directly applicable precedent, we have not yet identified a suitable formulation. We have therefore explicitly acknowledged the absence of such supervision in the newly added Limitations section.
[1] Unified Generative Modeling of 3D Molecules via Bayesian Flow Networks. 2024, ICLR.
[2] UniGEM: A Unified Approach to Generation and Property Prediction for Molecules. 2025, ICLR.
[3] Equivariant Diffusion for Molecule Generation in 3D. 2022, ICML.
[4] 3D Molecule Generation by Denoising Voxel Grids. 2023, NeurIPS.
[5] Score-based 3D Molecule Generation with Neural Fields. 2024, NeurIPS.
[6] VecMol: Vector-Field Representations for 3D Molecule Generation. 2026, ICML.
[7] Inhomogeneous Electron Gas. 1964, Physical Review.
Sources of distributional discrepancies (Q4)
We have added Appendix N to analyze the generated sample distributions and found no clear evidence of a distributional preference unique to ScaleMol. On GEOM-Drugs, excess hydrogen and deficient carbon together account for 87.24% of Atom TV. Carbon also contributes 90.87% of reference-weighted valency W₁ and 85.98% of bond-angle W₁, identifying the main contributors to these discrepancies.
To examine how these biases relate to model components, we apply the same analysis to all nine QM9 ablation models. We find no uniform composition-bias pattern across variants: replacing VoxMol's Gaussian occupancy representation with MMGs tends to slightly lower the carbon fraction, while the transition from single-scale to multiscale readout reverses the hydrogen excess in the incremental comparison. Valency and bond-angle errors show similar patterns across models, with carbon dominating their weighted contributions, consistent with GEOM-Drugs.
Regarding the divergence between stability and distribution-matching metrics, one possible explanation is that MMGs encode mass integrated over spatial cells rather than pointwise values of a Gaussian field. This may favor regional structural consistency without equally improving fine geometric accuracy. A related trade-off is discussed in VecMol [1, Section 4.2]: global structural accuracy can coexist with weaker local geometry, since small coordinate deviations can noticeably affect bond-length and bond-angle statistics.
[1] VecMol: Vector-Field Representations for 3D Molecule Generation. 2026, ICML.
We are grateful for these questions, which motivated new analyses and sharpened our understanding of the model’s behavior and limitations.
A Note to Reviewer cBDh
We hope the above clarifications and the additional experiments in the revised draft sufficiently addressed your concerns. If you are satisfied, we kindly request you to consider updating the score to reflect the newly added results and discussion. We remain committed to addressing any remaining points you may have during the discussion phase.
Response to Reviewer cBDh (1/2)
We thank the reviewer for the thoughtful questions about hierarchical generation, chemical connectivity, and distribution fidelity.
Coarse-scale predictions support the role of our method beyond hierarchical supervision or feature alignment alone in three ways. Architecturally, they directly shape the low-resolution, non-zero-sum component of the clean-field prediction used to compute the score. Empirically, our controlled ablation shows that hierarchical prediction improves generation quality, particularly molecule stability. Dynamically, visualizations show structural consistency between coarse and fine predictions within individual denoiser evaluations, while trajectory analysis suggests a possible coarse-to-fine tendency across the full sampling process.
In terms of model architecture, as shown in Figure 7, each finer-scale prediction is both conditioned on the predicted coarse field and constructed by adding detail to its mass-preserving upsampling. The detail branch redistributes mass within parent cells, while a separate correction branch allows their totals to be revised. Crucially, we use an x0 parameterization: the score is computed as sθ(Y)=(Xθ(Y)−Y)/σn2, where Xθ(Y) is the normalized finest clean-field prediction produced by this hierarchy. Consequently, coarse predictions influence the fine prediction that determines the Walk drift. Thus, coarse predictions supply regional elemental composition as an explicit basis for finer spatial arrangements. This shares the principle of building finer representations on coarser predictions with scale-wise autoregressive models such as VAR [1] and PAR [2].
In terms of generation quality, following your suggestion, we have added ScaleMol w/ ind to Table 4 in the revised manuscript. This control replaces the hierarchical readout with independent multi-scale prediction heads while retaining the same supervision and encoder fusion. Under the same training budget and sampling protocol, molecule stability decreases from 94.43% to 93.13%, supporting the benefit of using coarse-scale predictions to guide the fine-field prediction underlying the final score estimate.
In terms of generation dynamics, we examine both individual denoiser evaluations and the full sampling process. Figure 4 shows that, within a single evaluation, the finest-resolution field retains and refines the structure captured by coarse-scale predictions. The new trajectory analysis in Appendix M further shows that the mean relative deviation from each scale's own endpoint prediction generally increases with resolution: coarse fields stay closer to their endpoint predictions, and their deviations decrease more gradually. This suggests a possible implicit coarse-to-fine tendency across sampling, even without an explicit temporal scale schedule. Trajectory renderings in Appendix M also illustrate how high-resolution predictions use coarse-scale structure across the displayed sampling stages, showing that this behavior is not confined to the end of sampling.
[1] Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction. 2024, NeurIPS. [2] Protein Autoregressive Modeling via Multiscale Structure Generation. 2026, ICML.
Compared with VoxMol, our contribution is a new molecular representation, MMGs, with an inherent multilevel scale structure, together with a generative architecture designed to use this structure. Scale-prefix conditioning further extends the utility of multiscale information beyond unconditional generation. In terms of computation, Table 8 shows that training costs per update remain close to VoxMol despite the added multiscale architecture and supervision. This framework also supports a lower overall training budget, for example, on QM9, ScaleMol achieves higher molecule stability after only 70% of VoxMol's reported 500-epoch budget [1].
[1] 3D Molecule Generation by Denoising Voxel Grids. 2023, NeurIPS.
(Please see next post for remaining replies...)