acceptodds
Under review as a conference paper at ICLR 2027

FLDWM: World Model for learning approximate SO(3) equivariance for scalable and generalizable molecule generation

Abstract

Equivariant neural networks provide physically accurate representations for 3D molecular generation but introduce computational bottlenecks that limit scalability, while scalable voxel-based methods lack explicit rotational symmetries. To bridge this gap, we introduce the Forced Latent Denoising World Model (FLDWM), the first world-model-based framework for 3D molecule generation. Operating on compressed voxel representations, FLDWM formulates symmetry learning as a latent dynamics problem: the model is trained to invert random SO(3) rotation augmentations, conditioned on the applied rotation via Feature-wise Linear Modulation (FiLM), which we show theoretically forces the denoiser to learn approximate SO(3) equivariance, bounded only by voxel discretization error. On GEOM-Drugs, FLDWM significantly outperforms state-of-the-art non-equivariant (VoxMol, NEBULA, RADM) and equivariant (EDM, MiDi) baselines on stability and validity while retaining fast inference, and demonstrates zero-shot generalization to the out-of-distribution PCQM dataset without fine-tuning. As a practical illustration, seed-conditioned generation from five FDA-approved drugs yields novel lead-optimization-range analogs (ECFP4 ), showing FLDWM's potential to propose alternative drug candidates.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
World Model, Molecules

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_JaoMOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes FLDWM, a rotation-conditioned latent denoiser for 3D molecular generation. Molecules are voxelized and encoded with a VQ-VAE. During training, the model receives a noisy latent code of a randomly rotated molecule and the applied rotation matrix, and predicts the latent code in the molecule’s original pose. At inference, generation starts from a seed molecule and uses Walk–Jump Sampling in latent space. The paper reports results on GEOM-Drugs and PCQM and claims a bound on approximate SO(3) equivariance.

Critical strengths
  1. Conditioning a latent denoiser on a known rotation and training it to recover the original pose is a concrete learning objective that may encourage approximate SO(3) equivariance without requiring an explicitly equivariant backbone. The formulation is interesting.
  2. On GEOM-Drugs at 10 steps, FLDWM reports higher validity than NEBULA at nearly identical seed similarity. The drug-seed examples further show that the pipeline can produce structurally distinct analogs across different starting molecules.
Critical weaknesses
  1. Presentation and positioning
    1. Some aspects of the paper’s framing are unclear and may mislead readers who are less familiar with this area. The paper seems to blur the distinction between seed-conditioned generation and the broader task of molecular generation. Its generation experiments require a seed molecule, but the abstract and introduction frame the contribution more broadly as 3D molecular generation. The paper should make the narrower task explicit earlier.
    2. The discussion of baselines also blurs this distinction. NEBULA is designed for seed-conditioned generation, whereas EDM and MiDi are not seed-conditioned in their original form. Although the authors adapt the latter methods through partial noising, these are not originally the same generation task, and the comparison should be presented with that qualification.
    3. The positioning relative to prior work is questionable. In terms of its overall pipeline, this work largely builds on NEBULA and adds a rotation-aware denoiser conditioned on R. However, the related-work section states: “Closest to our setting, RADM likewise pursues scalable non-equivariant generation by handling rotations explicitly, through rotational alignment.”
  2. Experiments
    1. The proposed denoiser is substantially larger than NEBULA’s. It is therefore unclear whether the improvement over NEBULA comes from scaling up the model or from the rotation-conditioned denoiser. A comparison is needed to isolate the proposed contribution.
    2. Relatedly, comparing the present seed-conditioned model with EDM and MiDi is difficult to interpret. The authors adapt these models through partial noising, but the noise is applied in different spaces. EDM and MiDi perturb molecular representations directly, whereas FLDWM perturbs a compressed VQ-VAE latent. Even when a latent code is substantially noised, the decoder may still produce a molecule similar to the seed. However, direct noising in molecular space may destroy seed information much more rapidly. Matching the input signal-to-noise ratio therefore does not establish comparable effective perturbations or comparable outputs. Moreover, EDM is trained to reverse a multi-step diffusion process, whereas FLDWM is trained to perform a single denoising jump after its latent-space walk. These differences make the current comparison difficult to interpret as a controlled comparison of seed-conditioned generation.
    3. In Table 3, the PCQM “0 (data)” row reports valency W1=2.98W_1=2.98, atom TV =0.90=0.90, bond TV =0.24=0.24, bond-length W1=0.01W_1=0.01, and bond-angle W1=3.40W_1=3.40. The corresponding 10-step FLDWM values are 0.010.01, 0.010.01, 0.000.00, 0.000.00, and 0.500.50. If these distances use PCQM as their reference distribution, it is difficult to understand why the original data are so much farther from that reference than the generated molecules. One possible explanation is that the reference is instead GEOM-Drugs. If so, smaller distances would indicate greater similarity to GEOM-Drugs, rather than showing that the outputs match the PCQM test distribution. The authors need to specify the reference dataset if they want to use this table to support the cross-dataset fidelity.
  3. Math
    1. The proposed latent rotation T~R(Φ(x)):=Φ(TRx)\widetilde T_R(\Phi(x)):=\Phi(T_Rx) is not shown to be a well-defined, invertible, linear group action. Because Φ\Phi includes lossy VQ-VAE encoding, invertibility of RR in molecular coordinate space does not establish the corresponding properties in latent space.
    2. Theorem 1 assumes the decomposition fθ(⋅,R)=T~R−1∘gθ(⋅)f_\theta(\cdot,R)=\widetilde T_{R^{-1}}\circ g_\theta(\cdot), with an RR-independent gθg_\theta. FiLM conditioning does not enforce this decomposition, and a low denoising loss does not establish that the FiLM layers implement the inverse latent rotation.
    3. Equation (5) defines an expected squared loss over molecules, rotations, and Gaussian noise. The proof then uses it as a bound for individual clean inputs in Eqs. (9) and (13). In particular, suppressing the noise term and moving from an average error to a pointwise error require justification. The resulting pointwise equivariance bound does not follow from the stated assumption.
    4. Equation (14) defines εh\varepsilon_h as a supremum over molecules and rotations, whereas Appendix A.4 reports its average over sampled pairs. That average cannot be substituted into the theorem to obtain the claimed numerical “certificate.” Moreover, the appendix acknowledges an encoder contribution that need not vanish as h→0h\to0, which is difficult to reconcile with the stated εh=O(h)\varepsilon_h=O(h).
Questions for authors
  1. Is FLDWM intended only for seed-conditioned generation, or can it also generate molecules without a seed? If unconditional generation is within the intended scope, could the authors describe how sampling is initialized and provide corresponding results?

  2. How much of the improvement remains when the only substantive difference is rotation conditioning?

  3. What reference dataset is used for each distributional distance in Table 3, and how exactly is the “0 (data)” row computed?

  4. What assumptions ensure that T~R\widetilde T_R is a well-defined linear group action and that the FiLM-conditioned model has the factorization required by Theorem 1? How does an expected loss on noisy inputs yield the stated pointwise bound on clean inputs, and can the numerical certificate be recomputed using a quantity consistent with the supremum in Eq. (14)?

Reviewer_DEgqOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces FLDWM, a rotation-conditioned latent denoising model for seeded 3D molecule generation. A frozen VQ-VAE encodes voxelized molecules, and a FiLM-conditioned U-Net is trained to recover the original latent from a rotated, noisy input. Generation uses walk-jump sampling. The paper analyzes approximate rotational equivariance and evaluates generation on GEOM-Drugs and transfer to PCQM.

Critical strengths
  1. The rotation-conditioned reconstruction objective offers a simple way to incorporate geometric supervision into a latent generator.
  2. At comparable high seed similarity on GEOM-Drugs, FLDWM improves validity over NEBULA (96.62% vs. 90.29%).
  3. The evaluation includes two datasets, multiple sampling lengths, and reconstruction diagnostics.
Critical weaknesses

1. The equivariance guarantee is not established: Theorem 1 assumes a well-defined linear group action in latent space and a factorization f(y,R)=T~R−1g(y)f(y,R)=\widetilde T_{R^{-1}}g(y), with gg independent of RR. Neither property is established for the VQ-VAE/FiLM architecture: invertibility of spatial rotations does not imply invertibility after lossy encoding, and FiLM conditioning does not enforce this decomposition. The proof also converts an expected reconstruction loss over noisy inputs into pointwise, noiseless bounds (Eqs. 9 and 13). These gaps undermine the claimed guarantee for the implemented model.

2. The equivariance baseline measures the wrong quantity: Appendix A.4 identifies ∥Φ(TRx)−Φ(x)∥\lVert\Phi(T_Rx)-\Phi(x)\rVert as the identity-map baseline. However, the identity map has zero equivariance error: g(T~Rz)−T~Rg(z)=0g(\widetilde T_Rz)-\widetilde T_Rg(z)=0 for g=Idg=\mathrm{Id}. The reported baseline measures variation across poses, not failure to commute with rotations. Consequently, the reduction from 1.9516 to 0.1607 cannot be interpreted as a 12-fold improvement in equivariance. The evaluation should explicitly define the backbone gg, the action applied to its predictions, and a common equivariance metric for all comparisons.

3. The PCQM distributional results have an unresolved reference-distribution issue: Table 3 reports Atom TV = 0.90 for raw PCQM data but 0.01 for FLDWM, interpreting the latter as a better match to PCQM. Yet NEBULA's discussion of the inherited baseline values interprets decreasing distances as reversion toward GEOM (NEBULA, Table 2 and Sec. 5.3). If GEOM remains the reference, near-zero distances do not support the claimed match to PCQM; if FLDWM uses PCQM instead, its results are not directly comparable to those baselines. The reference distribution and preprocessing must be consistent across methods before these metrics can support the transfer claim.

4. The ablation does not establish an advantage over ordinary rotation augmentation: The controls remove RR while retaining the target of reconstructing the original pose, making the inverse-rotation task ambiguous, as Appendix A.5 acknowledges. Their failure therefore shows that this particular objective needs rotation information; it does not show that ordinary augmented denoising is insufficient. The missing control is a denoiser trained to reconstruct Φ(TRx)\Phi(T_Rx) from Φ(TRx)+ϵ\Phi(T_Rx)+\epsilon, using the same encoder, comparable capacity, and the same training budget. Generation quality and OOD performance should be compared against this control to isolate the benefit of the proposed objective.

Questions for authors

Please see the Weaknesses section above.

Reviewer_skfQOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a method for generating 3D molecules that aims to combine the computational efficiency of voxel-based models with the rotational symmetry properties of equivariant neural networks. The main idea is to train a model to recover the original molecular representation from a randomly rotated and noisy version, using the rotation information as additional input. The authors argue that this encourages the model to learn approximate rotational equivariance without requiring specialized equivariant architectures. Experiments on GEOM-Drugs show improvements over several existing methods in molecular validity and structural quality, while experiments on PCQM demonstrate promising generalization to a different dataset.

Critical strengths
  • Learning rotational symmetry without explicitly enforcing it through network architecture is an interesting direction. The paper attempts to support this idea with theoretical analysis and empirical experiments.
  • The method achieves competitive or better results than several equivariant and non-equivariant baselines
Critical weaknesses
  • The presentation of this paper is poor. The proposed approach seems to be a mixture of existing works (voxel, rotation learning, etc.), but the difference between it and existing works is not clear. For example, I found the rotation learner quite similar to that of RADM, which also uses a VAE to learn rotations. The novelty/contributions of the proposed approach should be further clarified. Besides, a reference is missing regarding approximate equivariance (especially its theoretical part): SymDiff: Equivariant Diffusion via Stochastic Symmetrisation (ICLR 2025)
  • There are not enough details in the main body about the model architectures. Do you simply follow previous works or choose a particular architecture intentionally? Rationales should be provided. For example, the use of the FiLM layer is not natural to me.
  • Experiment-wise, I didn't observe a big difference in Table 2 regarding sample quality. While efficiency is one advantage claimed by the proposed approach, it has a lower sampling speed.
Questions for authors

Please see Weaknesses

Comments on the paper

1 comment

We thank the reviewers and the area chair for their time and careful feedback. We agree with the main concerns raised. We will substantially revise the work and resubmit.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.