acceptodds
Under review as a conference paper at ICLR 2027

ScaleMol: Defining and Leveraging Scales for 3D Molecular Generation

Abstract

Small molecules lack a natural definition of scale, making it difficult to organize their generation from coarse structure to fine detail. We address this problem by introducing resolution-consistent molecular mass grids (MMGs), which define molecular scales through the spatial resolution of a molecular field. Coarse MMGs capture regional element content, while finer MMGs progressively resolve its spatial distribution, with exact consistency across scales. Based on this scale definition, we introduce ScaleMol, which explicitly predicts and supervises molecular fields from coarse to fine. On QM9 and GEOM-Drugs, ScaleMol improves atom and molecule stability over existing field-based generators, while controlled ablations verify the benefits of both MMGs and coarse-to-fine prediction. The same scale hierarchy also enables conditioning on a reference molecule at different resolutions without additional training. Together, our results provide a simple definition of scale for small molecules and demonstrate its utility for 3D molecular generation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI for Science, Molecule Generation, Generative Models, Drug Discovery

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_cBDhOpenReview ↗
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
  1. 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.

  2. 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
  1. 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.

  2. 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.

  3. 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
  1. 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.

  2. 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?

  3. 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?

  4. 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
AuthorsOpenReview ↗

Response to Reviewer cBDh (1/2)

We thank the reviewer for the thoughtful questions about hierarchical generation, chemical connectivity, and distribution fidelity.

Structural guidance from coarse predictions (W3, Q1, Q2)

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 x0x_0 parameterization: the score is computed as sθ(Y)=(X^θ(Y)−Y)/σn2s_\theta(Y)=(\widehat X_\theta(Y)-Y)/\sigma_n^2, where X^θ(Y)\widehat X_\theta(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.

ModelMol. stab. (%) ↑Atom stab. (%) ↑Valid (%) ↑Unique (%) ↑Valency W₁ ↓Atom TV ↓Bond TV ↓Length W₁ ↓Angle W₁ ↓
ScaleMol w/ ind93.1399.5398.7095.480.0140.0260.0090.0022.810

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.

Advances over VoxMol and computational cost (W2)

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...)

AuthorsOpenReview ↗

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.

AuthorsOpenReview ↗

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.

Reviewer_z2zBOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ScaleMol, a framework for 3D molecular generation based on resolution-consistent molecular mass grids. The grids integrate additive atomic Gaussian densities over nested spatial cells, enabling hierarchical prediction and multiscale supervision within a shared denoiser. Walk–jump sampling generates molecular fields, from which atoms are recovered. The same hierarchy supports reference conditioning without additional training. Experiments on QM9 and GEOM-Drugs demonstrate improvements in molecular stability, particularly on GEOM-Drugs.

Critical strengths
  1. The MMG formulation provides an interpretable scale hierarchy with exact aggregation and clear element-count semantics.

  2. The hierarchical readout and reference-conditioning mechanism are well motivated, with useful ablations and detailed supplementary diagnostics.

Critical weaknesses
  1. Stability improvements do not consistently extend to other quality metrics. QM9 bond-angle W1 is 2.659 versus 1.960 for VoxMol, while GEOM-Drugs atom-type TV is 0.111 versus 0.033. Validity is also evaluated only on the largest connected fragment.

  2. The representation ablation changes both the field representation and atom-recovery procedure, preventing clean attribution of the MMG gains. Component ablations are limited to QM9 despite the strongest results occurring on GEOM-Drugs.

  3. Reference control is evaluated on only 32 QM9 molecules. Coarse conditioning yields 57.03% molecule stability, and Table 12 shows substantial disagreement after atom recovery despite exact preservation of the predicted coarse field.

Questions for authors
  1. Can the authors report connected-output rates and atom-count distributions from existing samples to help explain the stability gains alongside weaker distribution matching?
  2. What evidence separates the contribution of MMGs from that of the changed atom-recovery procedure in Table 3?
  3. What primarily causes the large gap between grid preservation and recovered molecular agreement under coarse conditioning?
3 comments
AuthorsOpenReview ↗

Response to Reviewer z2zB

We appreciate the reviewer’s careful assessment of our evaluation and the questions about distribution matching, atom recovery, and reference control.

Stability and distribution matching (W1, Q1)

We have added the requested connected-output rates and atom-count distributions to Appendix N, using 10,000 existing GEOM-Drugs outputs. ScaleMol's single-fragment rate is 84.41%, higher than the 82.6% and 70.5% reported for VoxMol and FuncMol [1], respectively. Its atom-count distribution is similar to those of the baselines. Our Figure 16 can be compared qualitatively with Figure 10 in [1] (image url: https://arxiv.org/html/2501.08508v1/histograms1.png); Appendix N additionally examines ring-size distributions and element-conditioned distribution errors.

Regarding the divergence between stability and distribution-matching metrics, our interpretation is that MMGs integrate mass over spatial cells rather than sample pointwise Gaussian-field values, which may favor regional structural consistency over fine geometric accuracy. VecMol [2] discusses a related trade-off between global structural accuracy and local geometry, noting that small coordinate deviations can affect bond-length and bond-angle statistics.

[1] Score-based 3D Molecule Generation with Neural Fields. 2024, NeurIPS.

[2] VecMol: Vector-Field Representations for 3D Molecule Generation. 2026, ICML.

Separating representation and recovery effects (W2, Q2)

The following evidences shows that coordinate refinement has no hidden contributions on metrics. First we recovered the same VoxMol w/ MMGs fields using L-BFGS with an MMG-matched renderer. The main metrics remain essentially unchanged, preserving the conclusions of Table 4 in the revised manuscript.

ModelMol. stab. (%) ↑Atom stab. (%) ↑Valid (%) ↑Unique (%) ↑Valency W₁ ↓Atom TV ↓Bond TV ↓Length W₁ ↓Angle W₁ ↓
VoxMol w/ MMGs, original recovery89.1999.2498.7698.780.025440.024250.010780.003573.292
VoxMol w/ MMGs, L-BFGS recovery89.1399.2498.7398.780.024930.024250.010770.003273.337

Table 11 also provides complementary evidence from paired GT-field reconstruction: both representation–recovery pipelines achieve comparable chemical quality on accurately rendered fields, including 99.9% atom stability and 98.8% molecule stability for interior placements.

To further clarify, our choice of non-iterative MMG recovery is motivated by efficiency: Table 10 reports approximately 4× higher recovery throughput and lower peak GPU memory than occupancy recovery with L-BFGS.

Coarse-prefix control and molecular agreement (W3, Q3)

The limited information in coarse conditioning and the sensitivity of atom recovery to field variations give rise to two forms of disagreement as following. The first is structural disagreement between the generated and reference molecules under coarse prefix conditioning. This is consistent with our design: coarse condition retains coarse information while leaving finer arrangements flexible, much as a blurred image can correspond to different objects. Finer prefixes retain more reference detail and should therefore reduce diversity. This hierarchy is intended to let researchers control reference fidelity and molecular diversity at multiple spatial granularities.

The second is disagreement between the reference coarse field and the field re-rendered from the recovered molecule. Although prefix protection fixes the total mass of each element in each controlled coarse cell, finer heads can still redistribute that mass among its subcells. Such changes can weaken, shift, or create local peaks, affecting which atoms are recovered and where they are placed. Peak-based recovery is a nonlinear conversion to discrete atoms and does not enforce the original coarse-cell mass constraints; re-rendering these atoms can therefore produce a different coarse field. Thus, exact prefix preservation in MMG space does not imply exact preservation after atom recovery: for the final molecule, the prefix acts as a soft constraint rather than a guaranteed structural constraint.

We also added output-space conditioning based on MMG, which improves molecule stability under coarse (434^3) conditioning on QM9 from 57.03% to 93.85% (Section 3.5.2). The paired QM9 results in Section 4.3 show that multiscale prefix control need not rely on ScaleMol's particular hierarchical readout, suggesting that the MMG-based conditioning interface can extend to other network designs.

Thank you for encouraging us to examine these distinctions more closely and provide additional evidence for our conclusions.

AuthorsOpenReview ↗

A Note to Reviewer z2zB

As the discussion phase is approaching its end, we kindly request the reviewer to let us know if the above clarifications and the previously added experiments have addressed the remaining questions. We would be happy to address any additional points the reviewer may have during the remaining time of the discussion phase. We thank the reviewer for engaging with us in the discussion.

Thank you for the additional analyses. The comparison on the same generated MMG fields largely resolves my concern about coordinate refinement. The single-fragment rate also addresses my concern about disconnected outputs on GEOM-Drugs. The rebuttal strengthens my confidence in the contribution, and I maintain my recommendation.

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

This paper proposes a novel coarse-to-fine strategy and resolution-consistent molecular mass grids (MMGs) for generating small molecules. The proposed ScaleMol improves the molecular stability of the field-based method on QM9 and GEOM-Drugs. The method can also use reference conditioning of different resolutions for generation without training.

Critical strengths

The proposed strategy appears to be novel, while existing approaches are based on diffusion and flow matching.

The experiments are sound, with ablations demonstrating the effectiveness of the approaches.

Critical weaknesses

First, I am not so convinced by the motivation. Why "A natural question is how scale-based generation can be extended to three-dimensional molecules"(line 39-40) a natural question? Can you provide more insights why we want to extend scale-based generation into molecular generation? I do not think being widely applied in other fields is a good motivation

My biggest concern is the gridding of molecules, which results in substantial waste of cubes for some molecules. Besides, the adopted strategy simply aggregates the neighbouring 8 grid cells, while the molecule exhibits some natural hierarchical structure, such as the ring structure. Why not use the natural hierarchical structure?

The gridding strategy makes a 3D U-Net a must as the backbone denoiser, which is poorly scalable, hindering its application for large-scale problems.

Missing common baselines such as GeoBFN, SemlaFlow, Flexiflow etc. Based on the reported results, SelmaFlow outperforms ScaleMol

Questions for authors

Can you provide more motivations why we want to study the coarse-to-fine strategy for generating molecules rather than it's just a 'natural question' that other fields have adopted something similar?

Why not use the natural hierarchical structure, such as the ring structure of molecules?

How's the training compute compared with the other methods? How's the scalability of the method? Is the 3D-Unet heavy to train?

Can you provide more baselines? If the authors deliberately omit those baselines, please provide reasons to justify this

For reference conditioned generation, is there any application of this with a reference low-resolution molecule as the prior?

4 comments
AuthorsOpenReview ↗

Response to Reviewer fdqv (1/2)

We thank the reviewer for the constructive feedback on our molecular-scale motivation, baseline comparisons, and potential applications.

Molecular scales and hierarchical generation (W1, W2, Q1, Q2)

Hierarchical generation methods already have precedents within small-molecule modeling (e.g., HierDiff [1], JT-VAE [2], and HierVAE [3]). The motivation is to preserve meaningful molecular substructures while simplifying the generation of complex molecules. For example, fragment-based hierarchies pursue this goal through chemically meaningful units such as rings. However, this hierarchy has two limitations. First, the fragment and atom levels represent different objects and require different processing, as in HierDiff. This leads to separate generative modules for the different levels, which hinders the integrated use of information across levels. Second, a fixed fragment–atom decomposition provides only one fragment level above the atomic level, which limits the readily available choices of granularity for hierarchical modeling and reference control.

The concept of scales is not foreign to the molecular domain: related concepts already exist in biological and chemical research, such as electron-density resolution, which describes the level of spatial detail that can be distinguished [4]. Our definition of scale shares this underlying idea and intuition, as visually illustrated in Figure 6 of the newly added Appendix A. Scales allow the model to observe the molecule as a whole at different granularities, offering multiple perspectives beyond a single view of the molecule as a collection of atoms. Thus, scales have yet to be well defined and effectively leveraged for molecular generation, rather than being inherently unsuitable for it.

Molecular scales defined by MMGs addresses the limitations discussed above: all scales represent regional element mass, allowing one network to process them and combine their information. Its inherent multilevel nature provides a wider range of control granularities. With MMG-defined scales and ScaleMol, we improve the stability of generated molecules and enable prefix-conditioned generation, providing an interface for balancing molecular diversity and consistency with a reference molecule beyond unconditional generation.

[1] Coarse-to-Fine: a Hierarchical Diffusion Model for Molecule Generation in 3D. 2023, ICML.

[2] Junction Tree Variational Autoencoder for Molecular Graph Generation. 2018, ICML.

[3] Hierarchical Generation of Molecular Graphs using Structural Motifs. 2020, ICML.

[4] Crystallographic Structure Factors and Electron Density, Resolution, and R-value. n.d., PDB-101.

Computational cost and scalability (W2, W3, Q3)

ScaleMol introduces multiscale modeling at a computational cost comparable to VoxMol while achieving better molecular stability with a smaller training budget. Table 8 confirms comparable training costs per update despite the added multiscale architecture and supervision. Our representation and design also support a lower overall training cost, for example, on QM9, ScaleMol achieves higher molecule stability using only 70% of VoxMol's reported sample-exposure budget[1]. Its successful application to GEOM-Drugs further demonstrates sufficient scalability for our target small-molecule generation tasks. For substantially larger systems such as proteins, FuncMol-style latent-space generation [2] offers a promising way to reduce volumetric computation, as discussed in the newly added Limitations section.

[1] 3D Molecule Generation by Denoising Voxel Grids. 2023, NeurIPS.

[2] Score-based 3D Molecule Generation with Neural Fields. 2024, NeurIPS.

(Please see next post for remaining replies...)

AuthorsOpenReview ↗

Response to Reviewer fdqv (2/2)

Baseline comparisons (W4, Q4)

Our initial baseline selection followed voxel- and field-based generators, including VecMol [1] (ICML, July 2026), because they recover atomic structures from spatial fields. We follow FuncMol's evaluation setup [2], with Open Babel bond inference and MiDi-defined metrics, and separately report ScaleMol_ref with VecMol-style chemical postprocessing.

To broaden the comparison, we reevaluated the official SemlaFlow GEOM-Drugs checkpoint on 10,000 molecules under our evaluation pipeline. SemlaFlow [3] evaluates its generated bond types even when they differ from those inferred by Open Babel from the generated geometry. This differs from our evaluation pipeline, so we also report results using Open Babel to infer bonds from the same generated molecular geometries.

ModelMol. stab. (%) ↑Atom stab. (%) ↑Valid (%) ↑Unique (%) ↑Valency W₁ ↓Atom TV ↓Bond TV ↓Length W₁ ↓Angle W₁ ↓
SemlaFlow\mathrm{SemlaFlow}97.499.893.7100.00.0360.0610.0250.0021.255
SemlaFlowobabel\mathrm{SemlaFlow}_{\mathrm{obabel}}82.199.492.8100.00.2530.0610.0240.0031.135
ScaleMol\mathrm{ScaleMol}88.399.595.397.70.2520.1110.0450.0031.470
ScaleMolref\mathrm{ScaleMol}_{\mathrm{ref}}90.499.792.897.60.2640.1220.0430.0042.458

The unsubscripted row uses generated bonds and formal charges; obabel\mathrm{obabel} denotes their inference from the same atom identities and coordinates using Open Babel, which aligns with our evaluation pipeline. Compared with SemlaFlow$_{\mathrm{obabel}}$, ScaleMol (Table 2) retains higher molecule stability (88.3% vs. 82.1%), atom stability (99.5% vs. 99.4%), and validity (95.3% vs. 92.8%).

[1] VecMol: Vector-Field Representations for 3D Molecule Generation. 2026, ICML.

[2] Score-based 3D Molecule Generation with Neural Fields. 2024, NeurIPS.

[3] SemlaFlow – Efficient 3D Molecular Generation with Latent Attention and Equivariant Flow Matching. 2025, AISTATS.

Applications of prefix conditioning (Q5)

The primary intended application is ligand-based drug design (LBDD): generating diverse analogues while retaining selected characteristics of a reference ligand. Existing approaches condition on molecular shape, as in ShapeMol [1], or pharmacophore features, as in PGMG [2]. MMG prefixes provide an additional control axis through spatial granularity, from regional elemental composition to finer arrangements, allowing reference consistency and diversity to be adjusted within one model. We have explicitly added this LBDD application scenario to the Introduction to help readers understand the intended use of prefix-conditioned generation.

[1] Shape-conditioned 3D Molecule Generation via Equivariant Diffusion Models. 2023, arXiv.

[2] A pharmacophore-guided deep learning approach for bioactive molecular generation. 2023, Nature Communications.

We appreciate these suggestions, which have helped us better explain the rationale and practical scope of our approach.

Thanks for the reply.

Based on the results in Tables 1 and 2 from [1], ScaleMol underperforms SelmaFlow and FlexiFlow on Geom-drugs in molecular stability. Since the idea of coarse-to-fine is not new, I will keep my recommendations.

[1] FlexiFlow: decomposable flow matching for generation of flexible molecular ensemble, ICML2026.

AuthorsOpenReview ↗

Follow-up Response to Reviewer fdqv

Thank you for the follow-up. We would like to clarify two points central to our contribution and the performance comparison.

Methodological novelty

Coarse-to-fine generation is one motivation for our work, but does not capture the full scope of our contributions. Our contribution is to define molecular scales through MMGs and use this shared multilevel representation for both 3D generation and prefix-conditioned control. Among the precedents cited in our earlier response, JT-VAE [1] and HierVAE [2] generate molecular graphs rather than 3D structures. HierDiff [3] is a relevant 3D precedent, but its fragment–atom hierarchy uses different representations at different levels and offers a fixed fragment-based granularity, as discussed in our previous comment. MMGs instead represent regional element mass consistently across spatial resolutions, enabling joint multiscale modeling and graded control over similarity to a reference molecule.

[1] Junction Tree Variational Autoencoder for Molecular Graph Generation. 2018, ICML.

[2] Hierarchical Generation of Molecular Graphs using Structural Motifs. 2020, ICML.

[3] Coarse-to-Fine: a Hierarchical Diffusion Model for Molecule Generation in 3D. 2023, ICML.

Comparison with other baselines

The published stability ranking alone does not establish the ranking under a shared coordinate-based evaluation. As reported in our previous comment, we reevaluated the official SemlaFlow checkpoint on 10,000 GEOM-Drugs outputs. On the same samples, molecule stability is 97.4% using generated bonds but 82.1% when bonds are inferred from coordinates with Open Babel, as in our pipeline. ScaleMol achieves 88.3%, alongside higher atom stability (99.5% vs. 99.4%) and validity (95.3% vs. 92.8%). This distinction also matters when interpreting FlexiFlow: its Table 2 explicitly separates methods that infer bonds from coordinates from those that generate bonds directly [1].

[1] FlexiFlow: decomposable flow matching for generation of flexible molecular ensemble. 2026, ICML.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We thank reviewers fdqv, z2zB, and cBDh for their thoughtful feedback, which has helped us clarify the motivation, strengthen the evaluation, and better understand the limitations of our work.

  • Additional ablation: We added ScaleMol w/ ind, an independent multiscale-head control with the same supervision, to Table 4 and discussed it in Section 4.4, following reviewer cBDh's suggestion.
  • Molecular-scale motivation: We added Appendix A to explain the intuition behind defining and using molecular scales through an analogy with electron-density resolution, and thank reviewer fdqv for encouraging a clearer molecular motivation.
  • Output-space conditioning: We added a final-field projection method that applies prefix guidance without conditioning intermediate multiscale heads (Section 3.5.2), with paired QM9 comparisons in Section 4.3, and thank reviewer z2zB for prompting a closer examination of prefix control.
  • Generation dynamics: We added Appendix M to examine multiscale predictions along complete Walk trajectories through endpoint-deviation curves and molecular-field visualizations, and thank reviewer cBDh for prompting this analysis.
  • Sample distributions: We added Appendix N to analyze element-conditioned distribution errors, their patterns across QM9 ablation controls, and GEOM-Drugs fragment counts, ring sizes, and molecular sizes, and thank reviewers z2zB and cBDh for motivating these investigations.
  • Limitations: We added an unnumbered Limitations section covering the absence of bond-aware supervision, voxel/3D U-Net computational costs, the current scope of prefix-conditioned LBDD, and future pocket-conditioned generation, and thank reviewers fdqv, z2zB, and cBDh for helping us clarify these directions.

We also improved the manuscript’s organization, including paragraph structure and consistent naming of ablation variants. The revisions are highlighted in blue in the updated manuscript for ease of review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.