acceptodds
Under review as a conference paper at ICLR 2027

Make Diffusion Models Great Again in Structure-Based Drug Design

Abstract

Recent Bayesian and flow-based generators have challenged diffusion's leading position in structure-based drug design (SBDD). This progress motivates us to re-examine the modeling choices within molecular diffusion, particularly its forward trajectories and intermediate atom-type representations. We introduce MEGA, a unified molecular diffusion framework that redesigns both for joint generation of coordinates and atom types. MEGA directly specifies Gaussian coordinate marginals and categorical parameter trajectories on the simplex, enabling explicit control over both forward noising paths. It separates atom-type parameter evolution from network observations, using stochastic Concrete samples to provide graded classwise evidence beyond hard one-hot observations. To further support learning under high noise and the sampling cold start, we augment MEGA with GeoRank, a geometric prior that provides persistent slot identities without requiring reference ligand conformations at inference. On CrossDocked2020, the MEGA framework leads the compared methods in 12 of 15 reported benchmark statistics, including all Vina evaluations, PoseBusters validity, and strain energy. It achieves 89.5% PoseBusters validity and a mean Vina score of -6.96 kcal/mol. With GeoRank, MEGA achieves the lowest weighted JSD for bond-length and bond-angle distributions among the compared methods, improving over MEGA alone by 20.2% and 15.1%, respectively. GeoRank also improves agreement with the reference distributions of ring sizes and functional groups. These results show that redesigning molecular diffusion can restore its leading performance in SBDD.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Diffusion model, Structure-based drug design

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_VegROpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies molecular diffusion for structure-based drug design. MEGA defines Gaussian coordinate marginals and categorical atom-type parameter trajectories under a shared signal schedule, and uses continuous Concrete observations for joint denoising. The authors also introduce GeoRank, which assigns persistent atom ranks from clean ligand geometry during training and uses randomly assigned ranks during generation. Experiments on CrossDocked2020 evaluate binding scores, molecular quality, and geometric distribution fidelity.

Critical strengths
  1. The MEGA results are encouraging. Table 1 shows strong Vina scores together with improvements in PoseBusters validity and conformational strain over the compared methods. The ablations in Appendix H also provide useful evidence for the contribution of the chosen schedule and continuous atom-type observations.

  2. GeoRank provides an interesting way to introduce geometric organization during high-noise denoising. It achieves improvements in bond-length, bond-angle, functional-group, and ring distributions, with additional category-wise analyses supporting the geometric gains. The paper also clearly explains how geometric ranks are used differently during training and generation.

Critical weaknesses
  1. The large effect of the forward corruption schedule would benefit from further explanation. In Appendix H, replacing the proposed schedule with Linear reduces PB-valid from 89.5% to 48.6% and increases median strain energy from 84.67 to 1302.21. These results support the chosen configuration, but provide limited insight into why this replacement causes such substantial degradation or which properties of the proposed schedule improve molecular geometry.

  2. The specific contribution of geometric ordering in GeoRank could be more clearly established. Appendix J compares virtual-anchor and spectral orderings, but both use clean ligand geometry. A control trained with geometry-independent persistent ranks and the same adapter would help distinguish the contribution of geometric information from persistent slot identities and additional message passing. The current results support GeoRank's distributional gains more directly than this particular explanation of their source.

Questions for authors
  1. Could the authors explain the large performance gaps between the proposed schedule and the Linear/Cosine variants in Appendix H, and confirm whether their training budgets and evaluation settings are matched? An analysis of SNR trajectories or denoising errors across noise levels could help explain these differences.

  2. Can the authors compare GeoRank with ranks sampled independently of ligand geometry during training and held fixed throughout corruption, using the same adapter, time gating, and training budget?

  3. Does GeoRank's denoising advantage mainly occur at high noise levels? Appendix E.4 reports losses averaged over sampled diffusion times; a comparison across noise levels would more directly support this motivation.

  4. Could the authors further discuss the trade-off between GeoRank's improved geometric JSDs and its worse Vina scores, strain energy, and SA? In what settings would GeoRank be preferable to MEGA alone?

  5. Could the authors consider including MolPilot [1] in the comparison, given its relevance to scheduling for SBDD? A comparison under consistent dataset splits, post-processing, and evaluation settings would help contextualize MEGA's results.

[1] Qiu et al., Piloting Structure-Based Drug Design via Modality-Specific Optimal Schedule, ICML 2025.

Reviewer_JX4XOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the task of structure-based drug design (SBDD), aiming to generate ligand atom types and 3D coordinates conditioned on a given protein binding pocket, with chemical bonds inferred through post-processing. The authors revisit two key design choices in joint diffusion modeling of continuous atomic coordinates and discrete atom types: (1) the noise schedule of the forward diffusion process and (2) the representation and observation of discrete atom types during diffusion. Based on these considerations, they propose MEGA, a framework that combines Gaussian diffusion, categorical probability parameter evolution, and continuous Concrete observations. In addition, the authors introduce GeoRank, which assigns persistent structural identities to atom slots through geometric ordering to improve the geometric distribution fidelity of generated molecules. Experimental results demonstrate strong generative performance on the CrossDocked2020 benchmark, while ablation studies provide evidence supporting the effectiveness of several key design choices. However, the main methodological components are closely related to existing approaches, and the paper does not sufficiently establish its independent methodological novelty or fully clarify the underlying mechanisms of some core design choices.

Critical strengths
  1. MEGA demonstrates competitive molecular generation performance on the CrossDocked2020 benchmark, outperforming several representative existing methods across multiple metrics, including predicted binding scores, PoseBusters validity, conformational strain energy, and synthetic accessibility.

  2. This work revisits two key design choices in the joint diffusion modeling of continuous atomic coordinates and discrete atom types: the construction of the forward diffusion process and the representation of discrete atom types as inputs to the denoising network. By integrating Gaussian diffusion, categorical probability parameter evolution, and continuous Concrete observations, the authors propose a joint diffusion modeling framework that demonstrates strong empirical effectiveness.

  3. The paper provides relatively comprehensive ablation studies that demonstrate the effectiveness of the proposed forward noise schedule and Concrete-based atom-type observations compared with commonly used alternatives, offering convincing empirical support for the main design choices.

  4. GeoRank introduces a persistent atom-slot identity mechanism based on geometric ordering, providing additional structural information during high-noise denoising stages. Experimental results demonstrate that this design substantially improves the geometric distribution fidelity of generated molecules.

Critical weaknesses
  1. Limited methodological novelty in continuous coordinate diffusion. MEGA adopts an existing linear Gaussian diffusion framework for continuous atomic coordinates. Its marginal distribution parameterization and the derivation of the corresponding SDE coefficients are mathematically similar to those in prior works such as DPM-Solver. Therefore, the contribution of this component primarily lies in the specific choice of the forward noise schedule and its application to molecular generation, rather than the introduction of a fundamentally new continuous diffusion mechanism. Although the ablation studies demonstrate that the selected schedule outperforms the DDPM Linear and Cosine schedules, these results mainly validate the effectiveness of a particular noise schedule and do not sufficiently establish the independent methodological novelty of the marginal-first parameterization itself.

  2. Limited novelty and insufficient justification of the atom-type diffusion formulation. MEGA’s atom-type modeling largely builds upon existing probability parameterization and continuous relaxation techniques. Specifically, its categorical parameter trajectory, θt=α(t)θ0+(1−α(t))u\theta_t=\alpha(t)\theta_0+(1-\alpha(t))u, has the same mathematical form as the marginal categorical probabilities of uniform-replacement diffusion in D3PM. Moreover, BFN and MolCRAFT have already explored modeling discrete variables in continuous probability spaces, while Concrete/Gumbel-Softmax is a well-established continuous relaxation technique. Therefore, the contribution of this component primarily lies in a particular combination and integration of existing ideas. Although the experiments show that the proposed formulation outperforms the evaluated one-hot baseline, the authors have not sufficiently clarified what distinctive advantages this combination offers over other reasonable discrete diffusion or continuous categorical modeling approaches.

  3. Insufficient justification of the necessity and distinctive advantages of Concrete observations. The authors argue that Concrete observations preserve richer classwise information than hard one-hot observations. However, Concrete sampling essentially introduces additional Gumbel randomness on top of the categorical probability parameters θt\theta_t. It remains unclear whether the observed improvements arise from the continuous soft categorical representation, the stochastic perturbation mechanism, or the specific properties of the Concrete distribution itself. Although the one-hot ablation supports the effectiveness of Concrete observations within the proposed framework, this comparison is insufficient to establish their distinctive advantages over other continuous stochastic categorical observation mechanisms. The authors should further clarify the theoretical motivation behind this design and provide more targeted ablation studies comparing alternative observation mechanisms.

  4. Insufficient disentanglement of the contributions of coordinate and atom-type noise schedules. Tables 5–6 in Appendix H demonstrate that MEGA’s forward noise schedule substantially outperforms the DDPM Linear and Cosine schedules. However, these experiments do not separately isolate the contributions of continuous coordinate diffusion and atom-type parameter evolution, nor do they directly establish the necessity of using a shared signal schedule. Consequently, the current results do not clearly reveal whether the performance gains primarily originate from the coordinate noise trajectory, the categorical parameter trajectory, or the interaction between the two. The authors are encouraged to conduct factorial ablation studies that independently vary the coordinate and atom-type schedules, as well as compare shared and independent scheduling strategies.

  5. Insufficient validation of GeoRank’s underlying mechanism and unexplained performance trade-offs. GeoRank simultaneously introduces geometric ordering information and persistent atom-slot identities. However, the existing experiments mainly compare the model without ranking against Virtual-anchor Ordering and Spectral Ordering, without sufficiently disentangling the respective contributions of geometric ordering and persistent slot identities. In particular, the paper lacks a control experiment in which randomly assigned ranks are kept fixed throughout the diffusion process during both training and generation. Therefore, it remains possible that the observed improvements primarily arise from persistent slot identifiers rather than the geometric information itself. Furthermore, although GeoRank substantially improves geometric distribution fidelity, as measured by JSD, it also leads to worse results on certain Vina scoring and conformational strain metrics. The authors do not provide a sufficiently detailed explanation of these performance trade-offs.

Questions for authors
  1. What mechanisms account for the performance improvements brought by Concrete observations. The authors argue that Concrete observations preserve richer relative classwise information than hard one-hot observations. However, the existing ablation studies do not sufficiently disentangle whether these improvements arise from the continuous soft categorical representation, potential regularization effects introduced by stochastic observations, or specific interactions between Concrete observations, the denoising network, and the training objective. Could the authors further clarify the underlying mechanisms and provide more targeted experimental evidence?

  2. What is the independent contribution of the time-dependent KL training objective? The paper employs a time-dependent KL loss to train the atom-type denoising network, but its practical advantages over standard categorical cross-entropy remain unclear. Given that KL divergence and soft-label cross-entropy are equivalent under certain conditions, could the authors clarify whether the performance gains primarily stem from the choice of supervision target, the time-dependent weighting scheme, or their combined effects? Additional ablation studies comparing standard cross-entropy, soft-label cross-entropy, and different time-dependent weighting strategies would help establish the necessity of the proposed training objective.

  3. How can the performance trade-off between geometric distribution fidelity and molecule-level generation quality introduced by GeoRank be explained? Experimental results show that GeoRank substantially reduces the JSD of geometric distributions, such as bond lengths and bond angles, while simultaneously degrading certain Vina scores, conformational strain energy, and synthetic accessibility metrics. Could the authors explain the reasons behind this discrepancy? In particular, is it possible that the geometric ordering constraints introduced by GeoRank improve the overall geometric statistics of generated molecules without sufficiently ensuring the physical plausibility of individual generated structures?

Post Rebuttal updates:

The additional experiments address several of my concerns and strengthen the empirical evidence. While concerns about methodological novelty remain, I have raised my score accordingly.

2 comments
Reviewer_JX4XOpenReview ↗

Thanks for the Rebuttal

I appreciate the authors’ detailed rebuttal and the additional controlled experiments, which substantially strengthen the empirical justification for MEGA’s key design choices and address several of my concerns.

However, I remain concerned that the proposed method primarily integrates existing techniques, and its independent methodological novelty is still not sufficiently established.

Overall, I acknowledge the improved empirical evidence and will raise my score accordingly.

Dear Reviewer JX4X,

Thank you very much for carefully considering our rebuttal and additional experiments, and for raising your score. We sincerely appreciate your recognition that the new controlled experiments strengthen the empirical justification for MEGA’s key design choices.

We also appreciate your remaining concern regarding independent methodological novelty. Our intended contribution is the unified formulation of molecular evolution and intermediate-state representation for joint coordinate–type denoising. We will further sharpen this distinction in the revision, making clearer how this formulation relates to the established techniques it builds upon.

Your questions motivated valuable controlled experiments and helped us make the paper more precise and informative. Thank you again for your time, constructive feedback, and thoughtful reassessment.

Best regards,
The Authors

Reviewer_ppjpOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper revisits the forward process and atom-type representations of diffusion models for SBDD, proposing two components: MEGA and GeoRank. MEGA directly specifies coordinate Gaussian marginals and categorical parameter trajectories, and constructs continuous atom-type observations through Concrete sampling. GeoRank derives atom ordering from ligand geometry and introduces persistent geometric ranks and additional message passing to facilitate high-noise denoising.

On CrossDocked2020, MEGA outperforms major baselines in Vina, PoseBusters validity, and strain energy. GeoRank further improves geometric distribution fidelity, but some docking and molecular-quality metrics deteriorate.

Critical strengths
  • Well-motivated problem and meaningful method design. MEGA adopts a marginal-first formulation and introduces continuous atom-type observations, providing a unified framework for joint coordinate–type denoising.
  • Competitive experimental results. MEGA achieves strong Vina, PB-valid, and strain energy results on CrossDocked2020, outperforming several major baselines.
Critical weaknesses
  • Marginal-first Gaussian diffusion and categorical parameter interpolation have been explored in prior work, while Concrete/Gumbel-Softmax is also a well-established technique. MEGA mainly combines and applies existing techniques. The authors should further clarify the independent methodological contributions or new insights of this work.
  • Insufficient justification and validation of GeoRank's independent contribution. Rank proximity does not guarantee spatial proximity between atoms in 3D space. Traditional diffusion already preserves atom-index consistency, while GeoRank introduces additional message passing. Further ablations with alternative ordering strategies are needed.
Questions for authors

GeoRank appears primarily motivated by improving molecular geometry quality, but the experiments do not demonstrate a clear overall improvement. Have the authors explored other geometric priors or conditioning strategies? What insights or takeaways have been gained from these attempts?

3 comments
AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer ppjp,

Thank you again for your thoughtful review and for recognizing the motivation behind our method and MEGA’s strong empirical performance. Your comments helped us sharpen the presentation of our contribution and investigate the role of geometric conditioning more carefully.

In our rebuttal, we clarified MEGA’s central contribution: a unified formulation that jointly designs how molecular geometry and atom-type uncertainty evolve, and how their intermediate states are represented to the denoiser. For GeoRank, we distinguished geometry-informed conditioning from ordinary atom-index consistency and provided additional structural analyses and comparisons of alternative orderings. We also reported an updated nine-layer configuration that offers a better balance between geometric fidelity and molecular quality, while explicitly discussing the remaining trade-offs.

We hope these clarifications and additional results address your main concerns. If you find that they strengthen the paper’s contribution and supporting evidence, we would be very grateful if you would consider a more favorable assessment and score.

If any concern remains unresolved, please let us know. We would welcome the opportunity to clarify it or point to the relevant evidence before the discussion closes. Thank you again for your time and consideration.

Best regards,

The Authors

I have carefully read the authors’ rebuttal and appreciate the additional experiments. The updated GeoRank results and ordering ablations partially address my concerns about its effectiveness and geometric motivation.

However, unified modeling of continuous and discrete variables, as well as continuous representations of categorical uncertainty, have been extensively explored in prior work.

Overall, the rebuttal strengthens the empirical evaluation, the main claims still require more rigorous support. Considering your extensive results, I raise my score to 3.

Dear Reviewer ppjp,

Thank you for your careful consideration, constructive feedback, and positive reassessment. We greatly appreciate your recognition of our strengthened empirical evaluation and the additional evidence supporting GeoRank.

Our work identifies and addresses an underappreciated bottleneck in molecular diffusion: how the corruption process and intermediate molecular representations are jointly designed. Our controlled experiments indicate that these choices are a key contributor to the performance gap with competing generative approaches. Addressing them enables MEGA to bring diffusion models back to leading performance on CrossDocked2020 across binding and molecular-quality metrics.

Your feedback has helped us articulate this contribution more precisely. In the final revision, we will sharpen these statements and clarify the distinctions from prior work.

Thank you again for your thoughtful engagement.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.