acceptodds
Under review as a conference paper at ICLR 2027

AnchorSet: Structure-Anchored Set-Valued Canonicalization for Non-Equivariant 3D Molecular Generation

Abstract

Equivariant diffusion models have achieved strong performance in three-dimensional molecular generation by encoding Euclidean symmetries directly into the denoising architecture. Their specialized geometric operations, however, can increase computational cost and constrain the scalability and flexibility of the generative backbone. Existing canonicalization can support non-equivariant generation, but it can become unstable when molecular geometry does not define a unique orientation, replacing rotational redundancy with an arbitrary or perturbation-sensitive frame. We introduce AnchorSet, a structure-anchored, set-valued canonicalization framework that combines chemically informed pose estimation with explicit preservation of unresolved frame ambiguity. AnchorSet estimates molecular orientation from structure-conditioned anchors that combine a chemical-scaffold prior with learnable atom-type weights, biasing pose estimation toward stable molecular cores rather than flexible peripheral atoms. When the geometry does not identify a unique orientation, it preserves a compact set of symmetry-equivalent frames related by proper signed permutations, enabling a standard non-equivariant generator to operate in canonical coordinates. Across QM9 and GEOM-Drugs, AnchorSet outperforms prior non-equivariant approaches and matches or surpasses equivariant diffusion models in sample quality. Its stability under global rotations and non-anchor perturbations demonstrates that structure-aware set-valued canonicalization is a practical, scalable alternative to built-in equivariance.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
3D molecular generation, molecular diffusion model, structure-anchored canonicalization, set-valued alignment, non-equivariant generative

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

AnchorSet generates 3D molecules unconditionally using a fixed EGNN encoder and standard decoder with a latent diffusion Transformer. It fits a weighted PCA frame with learned weights per atom type on QM9 Bemis-Murcko scaffolds and GEOM-Drugs heavy atoms. When the third moment has low sign confidence, it forms 24 proper signed permutation frames, sampling one in training and picking one by a geometric fingerprint at test time. On QM9, the authors also run rotation and perturbation checks alongside reconstruction audits.

Critical strengths

The canonicalizer is simple and easy to follow. Section 3.2 separates anchor selection from the learned weights per atom type, while Section 3.3 gives a clear rule for discrete frame ties. Restricting candidates to proper rotations preserves input chirality, and test sampling needs no source scaffold.

To be fair, the pipeline beats several baselines. In Table 2, QM9 molecule stability hits 91.3% and validity reaches 95.5%, ahead of RADM DiT-B (87.3% and 94.1%) and GeoLDM (89.4% and 93.8%). On GEOM-Drugs, atom stability rises from RADM's 85.0% to 86.4% with validity near 99.1%, and Tables 5 and 6 show low reconstruction error.

Critical weaknesses
  1. My main worry is that the ablations never isolate the candidate set. Table 3 drops Murcko-Single even though it appears in the frame and reconstruction audits. There is no learnable single frame baseline either. I couldn't tell if sampling from the candidate set beats one deterministic frame on the same scaffold features, since moving from All-Atom and Heavy-Atom to Murcko-Set changes atom support and alignment along with the candidate policy without matched PCA controls.
  2. I wasn't convinced by the explanation for the main gain, and a key baseline is missing, so I lean toward a weak reject. From Murcko-Set to Learnable-Set, Table 3 stability jumps from 83.39% to 91.30%. Yet Table 9 frame error worsens from 0.192∘0.192^\circ to 0.235∘0.235^\circ and flips rise from 1.3 to 1.7. Section 4.4 triggered likelihood drift is also worse at 1.189 versus 0.457. In Table 4, perturbing outside the anchor mask gives zero drift by construction since those atoms never enter the covariance or fingerprint. Also, Section 2.2 cites InertialAR, but Table 2 omits it despite its 94.7% QM9 stability and 97.4% validity beating AnchorSet's 91.3% and 95.5%.
  3. The evaluation setup shifts across tables. Appendix A sets τc=10−3\tau_c=10^{-3}, where Table 8 lists candidate fractions of 11.25% for Murcko-Set and 11.02% for Learnable-Set with 6 to 8 flips, matching Table 7. Then Table 9 reports 7.344% and 7.083% with 1.3 and 1.7 flips, and I couldn't tell what changed.
  4. The baselines and likelihood checks leave out too much. Section 4 never maps All-Proc and Heavy-Proc to All-Atom and Heavy-Atom in Table 3. In Table 3, All-Atom hits 78.91% stability versus 87.3% for RADM DiT-B in Table 2, likely because RADM trains its encoder while Section 3.4 freezes it. Section 4.4 also leaves its NLL estimator undefined, and the unweighted Section 3.6 loss defines no likelihood.
Questions for authors
  1. Under the same backbone and budget, how do Murcko-Single and Learnable-Single generate, and why does Learnable-Set generate better despite worse frame error and likelihood drift?
  2. What are All-Proc and Heavy-Proc, and why does All-Atom drop to 78.91% stability in Table 3 versus 87.3% for RADM DiT-B?
  3. What changed between Tables 7 and 8 and Tables 1 and 9, and what does the Section 4.4 NLL estimate?
  4. How does AnchorSet compare to InertialAR under matched evaluation?
4 comments
AuthorsOpenReview ↗

Response to Reviewer iNgi

Thank you for the questions. AnchorSet combines structural anchoring, candidate frames and learned weights for reconstruction.

Single controls and candidate sets. All five AEs completed training. Their final checkpoints were evaluated on the same 1,280 validation molecules, with identical structural support and weight type within each Single/Set pair.

Final AECandidate useFrame errors >90°, evaluation seeds 0 / 1 / 2Coordinate RMSEDistance MAE
Murcko-Single0%10 / 11 / 130.0019890.002499
Murcko-Set11.25%7 / 8 / 60.0019530.002421
Learnable-Single0%16 / 12 / 100.0019140.002365
Learnable-Set11.02%7 / 7 / 60.0018980.002370
Murcko-Always24100%7 / 8 / 60.0019820.002420

Set reduces severe frame errors in every evaluation seed under both weighting choices. Murcko-Set matches Always24's error counts with candidates on only 11.25% of inputs.

Reconstruction uses latent means without encoder noise on the original inputs. Coordinate RMSE recenters the reconstruction, takes the square root of squared error divided by three times the active-atom count per molecule, then averages molecules, without extra rigid alignment. Distance MAE averages non-diagonal active-atom pairs within each molecule, then averages molecules. All five AEs achieve 100% active-atom type accuracy.

Generation metrics, including both Single configurations, are atom stability, molecule stability, validity, V×U and V×U×N, all in percent.

LDM AE canonicalizerAtom stabilityMolecule stabilityValidityV×UV×U×N
All-Atom97.78 ± 0.3878.91 ± 0.2990.90 ± 0.2388.91 ± 0.3145.60 ± 0.33
Heavy-Atom97.68 ± 0.4377.88 ± 0.5790.30 ± 0.4988.51 ± 0.3745.48 ± 0.28
Murcko-Single97.54 ± 0.1779.58 ± 0.5289.95 ± 0.5389.63 ± 0.4148.78 ± 0.37
Murcko-Set98.67 ± 0.3583.39 ± 0.3691.53 ± 0.2790.40 ± 0.3564.52 ± 0.36
Murcko-Always2497.85 ± 0.2179.50 ± 0.4090.40 ± 0.2590.27 ± 0.4553.50 ± 0.32
Learnable-Single98.26 ± 0.0780.72 ± 0.2791.60 ± 0.3891.32 ± 0.3250.75 ± 0.41
Learnable-Set99.21 ± 0.0991.29 ± 0.2395.50 ± 0.1693.62 ± 0.2369.48 ± 0.24

Set improves all five metrics under both weighting choices. Murcko-Set gains 3.81, 1.58, 0.77 and 15.74 percentage points over Murcko-Single in molecule stability, validity, V×U and V×U×N. Learnable-Set gains 10.57, 3.90, 2.30 and 18.73 points over Learnable-Single. Learnable exceeds Murcko in stability by 1.14 points for Single and 7.90 for Set, supporting learned weighting with candidate frames.

The interim Always24 result achieves 79.50% molecule stability and 53.50% V×U×N, versus Murcko-Set's 83.39% and 64.52%, despite identical severe-frame-error counts. Always24 randomly selects one of 24 frames for every training molecule; adaptive gating uses candidates for about 11% and a single frame for the rest. We interpret selective gating as handling ambiguous poses while reducing extra coordinate variation for high-confidence inputs, concentrating the non-equivariant DiT's training representation. This is consistent with the results; generation checkpoints and training budgets differ.

Frame consistency and generation quality. Frame error and flip count measure consistency across rotated inputs; RMSE measures reconstruction of the original input. Near-equal PCA eigenvalues can make discrete frame selection sensitive to small perturbations, without necessarily increasing reconstruction error on the original input. Learned weights optimize anchor covariance through reconstruction, so higher frame error can coexist with lower RMSE.

Table 9 reports higher operational-frame error for Learnable-Set than Murcko-Set (0.235° versus 0.192°), more flips (1.7 versus 1.3), and higher triggered-subset mean drift (1.189 versus 0.457). Section 4.4 reports median paired drifts of 6.1 × 10⁻⁵ and 4.6 × 10⁻⁵, with both p95 values below 7.1 × 10⁻⁴. Paired scores are stable under these rotations for most inputs; rare extremes dominate the triggered means. Each gate can select different molecules, so these means describe model-specific subsets.

Learnable-Set improves coordinate RMSE over Murcko-Set (0.001898 versus 0.001953) alongside generation metrics. Anchors stabilize global pose, candidates handle ambiguity, and weights adapt to reconstruction, improving the complete AE–LDM.

Zero drift under peripheral perturbations follows by construction when support masks stay fixed and displaced atoms enter neither covariance nor fingerprint. Peripheral conformations can thus change without redefining global pose; candidates handle residual core ambiguity. The perturbation test shows support isolation; Single/Set tests frame ambiguity.

AuthorsOpenReview ↗

Response to Reviewer iNgi

Frame-evaluation differences and NLL definition. Tables 1/9 summarize frame consistency; Table 7 tests gating on/off at fixed checkpoints; Table 8 varies the threshold. Candidate rates of about 7% and 11% come from separate records, both at threshold 0.001, rather than a recorded threshold change. Joint evaluation of final AEs uses the same 1,280 molecules, support construction and paired rotations. At 0.001, Murcko-Set and Learnable-Set use candidates on 11.25% and 11.02%, with severe-error counts of 7 / 8 / 6 and 7 / 7 / 6. The fixed-checkpoint scan gives identical severe-error counts at 0.001, 0.01 and Always24, supporting selective gating at 0.001.

Latent diffusion variational-loss estimator. Section 4.4 estimates a stochastic latent diffusion variational-loss score per molecule, using natural logarithms.

D̂ = terminal-prior KL + T × weighted denoising loss at t
     + zero-time loss + implemented Gaussian constants − log p(N)
t is uniform over 1, …, T.

The evaluator canonicalizes and encodes each molecule, adds Gaussian encoder noise of amplitude 0.00316, and computes terminal KL, SNR-weighted denoising and a separate zero-time term. Errors sum over active latent coordinates and features per molecule. This differs from Section 3.6's dimension-normalized, unweighted training MSE. The score is a stochastic latent variational loss, distinct from exact molecular data-space likelihood or a pointwise upper bound for every random draw.

The rotation diagnostic takes the absolute difference of original/rotated scores with matched encoder noise, diffusion time and diffusion noise. The atom-count term cancels. Three rotations of 1,280 molecules give 3,840 paired measurements grouped by molecule; tail statistics pool these pairs. We will clarify this definition and name in Section 4.4. Unconditional generation starts from an atom count and Gaussian latent noise, then applies reverse diffusion and decoding. It uses no source molecule, scaffold or input rotation, and does not repeat input canonicalization at each step.

Baseline definitions and encoder freezing. All-Proc/Heavy-Proc denote Procrustes pose alignment using all/heavy atoms. All-Atom/Heavy-Atom name the corresponding support choices in the generation ablation. The former describe alignment operations; the latter describe AE configurations supplied to the LDM. We will clarify terminology and map the reported configurations.

All-Atom's 78.91% stability is an internal ablation; 87.3% is RADM DiT-B's published result. They differ beyond encoder freezing. Our randomly initialized encoder stays frozen during AE training, while reconstruction gradients pass through its inputs to the canonicalizer. LDM training freezes the entire AE. This retains a reference latent interface around which the canonicalizer and decoder adapt. RADM Algorithm 1 describes updating the encoder, while its public QM9 AE code freezes it. With a standard DiT denoiser and GNN decoder, AnchorSet achieves 91.29% stability and 95.50% validity, versus RADM's 87.3% and 94.1%.

Comparison with InertialAR. InertialAR v2 reports 94.5% QM9 molecule stability and 97.3% validity, versus AnchorSet's 91.29% and 95.50%, leading by 3.21 and 1.80 points. On GEOM-Drugs, AnchorSet achieves 99.1% validity versus 98.0%, and 86.4% atom stability versus 87.0%. AnchorSet demonstrates generation with structural anchoring and standard DiT-B latent diffusion.

InertialAR defines an inertial frame and atom order to form a canonical token sequence. Its hierarchical autoregressive Transformer sequentially predicts atom types and coordinates, with geometric positional encoding and a diffusion loss for coordinates. AnchorSet handles pose at the representation stage, combining structural support, learned atom-type weights and compact candidate sets with a standard latent diffusion Transformer. InertialAR generates tokens along an atom sequence; AnchorSet denoises latent atom representations in parallel within each reverse-diffusion step. This supports parallel computation; inference time and throughput were not measured in this comparison.

AnchorSet explicitly handles residual frame ambiguity within a structurally anchored representation. Structural support separates pose estimation from peripheral conformational changes, learned weights adapt geometry to reconstruction, and 24 signed permutations with determinant +1 retain discrete frame alternatives while preserving chirality. Latent diffusion then jointly generates molecules without InertialAR's canonical atom sequence or atom-by-atom autoregressive factorization.

Thank you for the additional experiments and clarifications. The new Single/Set results address my concern about missing controls. However, I still have some concerns about whether the full design is needed and how broadly it helps.

First, all generation results use the same AE–latent-diffusion setup, with anchor weights learned through reconstruction. If I understand correctly, your construction could be applied to other setup e.g. ambient space diffusion, flow matching. Could you comment on this and, if possible, include the results in the final revised manuscript?

Second, on the experiment itself, my impression is that the frame diagnostics still do not clearly explain the large generation gains. Separately, the different training budgets also make the Always24 comparison inconclusive.

Third, I am still unconvinced by the 8.39-point gap between the internal All-Atom ablation and the published RADM. I understand the two use different alignment methods, so a matched RADM comparison would help establish the improvement over the closest existing approach.

Finally, what confuses me is that GEOM-Drugs uses all heavy atoms as opposed to Murcko scaffolds to determine the frame. These GEOM results do not directly test the proposed benefit of restricting pose estimation to the scaffold. As a result, I am unsure how useful scaffold-based anchoring is for larger molecules. I think this is the most important question about scalability.

AuthorsOpenReview ↗

Response to Reviewer iNgi

Thank you for recognizing the Single/Set controls.

Extensions to other generators. AnchorSet handles pose and discrete frame ambiguity at the representation stage, independently of the latent diffusion objective. The revision will add latent flow matching to compare AnchorSet with baseline representations, reusing each representation's frozen AE with matched latent dimensions and generator capacity. This tests transfer across objectives while retaining reconstruction-learned representations. Coordinate-space diffusion requires new generator training and remains future work. The revision will describe its interface: fixed scaffold support and uniform weights define canonical clean-coordinate targets; generation starts from noise.

Frame consistency and training budgets. Main Learnable-Set results are 91.29±0.23% molecular stability and 95.50±0.16% validity (SD across three independent AE–LDM training runs, ddof=0). In the original AE comparison, Murcko-Set matches Always24's severe-frame-error counts using candidates on only 11.25% of inputs, demonstrating selective frame-ambiguity handling.

We additionally completed an equal-budget comparison: Murcko-Set and Always24 both use epoch-4000 EMAs with 1,564,000 LDM updates, each sampled three times with 2,000 molecules per run. Both have 81.37% mean stability; mean validity is 91.12% and 91.42%, respectively. This fixed-stage experiment shows no generative advantage for the adaptive configuration; the earlier score difference does not establish a gating-only benefit. The revision will hold training budgets, candidate construction and within-set selection fixed, compare ambiguity-triggered versus always-enabled candidates, and include independent training repeats.

On the same 1,280 validation molecules under three paired rotations, Set has lower p99 AE output drift than Single under both weighting choices; 95% intervals for mean differences include zero. In one-step decoded x0 results at t=100/500/900, Always24 has lower mean drift than Murcko-Set throughout. We therefore distinguish frame consistency from generation quality: the former measures how rotation sensitivity propagates to network outputs; the latter directly reports stability, validity, V×U and V×U×N. The revision will report both to test when geometric improvements accompany generative gains.

Official RADM reevaluation and matched comparison. Internal All-Atom stability is 78.91%, versus the published RADM DiT-B's 87.3%, a gap of 8.39 percentage points. Both use a learned rotation network and Procrustes projection; this internal-ablation versus published-result gap does not measure the canonicalization module's contribution. We evaluated the official QM9 DiT-B EMA and its embedded AE with a common evaluator, using three sampling seeds and 2,000 molecules each: stability is 86.82±0.63% and validity 94.20±0.22% (SD across fixed-model sampling repeats, ddof=1). This recovers stability close to the published value, but training conditions remain unmatched (official training update counts are unreported). The revision will match data splits, AE/DiT settings, budgets and checkpoint selection to directly assess the canonicalization module's contribution.

Scaffold use on larger molecules. GEOM's 86.4% atom stability and 99.1% validity use learned heavy-atom support and do not directly test the generative benefit of scaffold restriction. We added geometric diagnostics, applying the same trained Learnable-Murcko AE weights to heavy/Murcko support on 1,280 validation parents. Murcko-to-heavy fallback is 3/1,280; neither support has a >90° frame error under three random rotations.

For 1,034 parents with eligible peripheral heavy atoms, Gaussian perturbations at sigma=0.1 in input coordinate units yield mean frame drift of 6.356° for heavy support and zero for Murcko. Zero drift follows from excluding peripheral atoms from covariance and fingerprint, isolating peripheral changes from the fixed core. Under core perturbations, Murcko drift is 24.963° versus 12.555° for heavy support. Across natural conformer pairs, Murcko has lower mean frame drift but a slightly higher median than heavy support.

Of 55 parents with at least 35 heavy atoms, 51 qualify for peripheral perturbations: at the same scale, mean drift is 12.298° for heavy support and zero for Murcko. Under core perturbations across all 55, mean drift is 6.595° for heavy support and 13.838° for Murcko. Thus scaffold anchoring's value on larger molecules depends on core stability: it isolates peripheral changes, while broader heavy-atom support may be more stable when the core changes substantially. The revision will add a GEOM heavy/Murcko generation comparison matching data, capacity and budgets, reporting scaffold coverage, fallback, and geometry and generation quality by size and flexibility to test whether peripheral isolation benefits larger-molecule generation.

Reviewer_spYUOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AnchorSet, a canonicalization method for non-equivariant molecular generative frameworks, as an alternative to built-in equivariance. The method uses chemically meaningful scaffolds and learnable atom-type weights to estimate the orientation to infer a canonical frame. The transformed molecule is then fed to a latent diffusion model. Evaluation on QM9 and GEOM-Drug datasets show competitive generation quality with established, mostly equivariant models. Ablation studies demonstrate the usefulness of all-atom representation and learnable weighting.

Critical strengths
  1. The work offers an interesting perspective on molecular symmetry. It proposes a useful alternative to built-in equivariance, which may increase computational burden or impose addition constraints. The proposed method is theoretically sound and empirically competitive.
Critical weaknesses
  1. Some of the design choices need further clarification or justification.
  2. Despite competitive performance, the scalability and efficiency claims need to be further supported.
Questions for authors
  • An overall schematic figure, in addition to the canonicalization schematic, would help readers understand the framework better.
  • What is the motivation behind still using an EGNN encoder, since the canonicalization already resolves the orientation? How does it compare to a non-equivariant one? Also, why is it fixed after initialization and how is it initialized?
  • Since the scalability and efficiency are presented as major motivations for non-equivariant frameworks, what is the train time, memory consumption and sampling throughput of the proposed method?
  • What do the learned atomic weights look like? Do they carry any chemical significance or other meanings?
1 comment
AuthorsOpenReview ↗

Response to Reviewer spYU

Thank you for the questions. AnchorSet uses structural anchors and learned geometric weights to align molecular representations, enabling a standard DiT denoiser and GNN decoder to model them. Below we explain the design choices and present the training costs, sampling efficiency and learned weights.

Overall pipeline. The framework has three stages. During AE training, the canonicalizer estimates a frame from the structural support and atom-type weights, a one-layer EGNN encodes the aligned molecule into coordinate and feature latents, and a non-equivariant GNN decoder reconstructs it. The encoder remains frozen while the decoder and canonicalization weights are jointly optimized through the reconstruction objective. During LDM training, the entire AE is frozen and DiT learns the latent distribution. Generation samples a molecule size and Gaussian latent noise, then applies reverse diffusion and decoding. We will expand the framework description to make these stages and trainable modules explicit.

Encoder design. Retaining RADM's one-layer EGNN interface (Ding and Hofmann, 2025, Section 3.1) allows us to examine the contribution of canonicalization while preserving the reference latent representation. Its coordinate latents rotate with the input (Garcia Satorras et al., 2021), so canonicalization supplies a consistent orientation for the non-equivariant decoder and denoiser. In the current implementation for training from scratch, the encoder is randomly initialized and then frozen, providing a fixed mapping around which the canonicalizer and decoder learn together; reconstruction gradients pass through the encoder's inputs to optimize the anchor weights.

We have also started a GNN-encoder comparison under the same Learnable-Single setting. The GNN and EGNN have 726,907 and 727,044 encoder parameters, respectively, differing by approximately 0.019%, with the structural support, weighting rule, frame policy, decoder architecture and latent dimensions matched. Both follow the same 200-epoch AE and 6,000-epoch DiT-B training budgets, followed by a 10,000-sample generation evaluation. The GNN LDM is currently training, and we will report the reconstruction and generation comparison after evaluation.

Training time, memory and sampling throughput. Based on existing Murcko and Learnable runs on an RTX 4090, AE training takes approximately 17 hours for 200 epochs with batch size 64, and DiT-B LDM training takes approximately 150 hours for 6,000 epochs with batch size 256. The AE and LDM training processes use approximately 6.1 GiB and 11.4 GiB of GPU memory, respectively. An available Learnable-Set checkpoint generated 1,000 QM9 molecules in 258.11 seconds with batch size 100 and 1,000 diffusion steps, achieving 3.87 molecules per second with 2.24 GiB peak PyTorch allocated memory during sampling.

The comparison with equivariant generative frameworks highlights the sampling efficiency of canonicalized representations combined with a standard DiT. Ding and Hofmann (2025, Table 3) report 55 and 49 seconds per 100 molecules for EDM and GeoLDM, respectively, using an RTX 4090, batch size 100 and 1,000 sampling steps. Our corresponding time is approximately 25.81 seconds, giving throughput approximately 2.1 times that of EDM and 1.9 times that of GeoLDM relative to these published timings. By incorporating pose alignment into the training representation, AnchorSet enables unconditional generation with a standard denoiser and decoder while retaining its structural-anchor canonicalization design.

Learned weights. The effective weights wt=0.05+softplus⁡(αt)w_t=0.05+\operatorname{softplus}(\alpha_t) learned by the final Learnable-Set and Learnable-Single AEs with training seed 0 are as follows.

\text{Atom type} & \text{Learnable-Set} & \text{Learnable-Single} \\\\ \hline \text{H} & 1.012734 & 1.021320 \\\\ \text{C} & 1.000865 & 1.000913 \\\\ \text{N} & 1.000469 & 1.000455 \\\\ \text{O} & 0.989440 & 0.989193 \\\\ \text{F} & 1.000433 & 1.002110 \\\\ \end{array}$$ These values show modest element-specific adjustments around uniform weighting. Structural anchors define the pose-estimation support, while the learned relative weights adapt its covariance to the reconstruction objective. This combines a chemically motivated scaffold prior with geometric fine-tuning. The interpretation also depends on support participation. Among 1,280 validation molecules, 430 use heavy-atom fallback, including one that further uses all atoms; hydrogen participates only through two atoms in that last molecule. We therefore interpret the coefficients as learned geometric adjustments within the selected support, rather than chemical importance rankings.
Reviewer_rEpmOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method for generating 3D molecules using non-equivariant neural networks while handling rotational symmetry through a preprocessing step. The proposed method aligns molecules into a consistent orientation based on PCA of chemically meaningful structural scaffolds (anchors). In case a unique orientation is difficult to determine, the proposed method considers multiple possible coordinate frames and selects a representative one. The resulting molecular representations are used to train a standard latent diffusion model with a DiT backbone. Experiments on QM9 and GEOM-Drugs show improved rotational consistency, perturbation robustness, and sample quality.

Critical strengths
  • The proposed candidate-set canonicalization is a good extension of single-frame canonicalization methods (e.g., RADM), addressing their limitation that a single stable canonical frame may be difficult to determine.
  • The paper systematically compares different canonicalization strategies while keeping the generative backbone fixed, providing evidence that the proposed canonicalization improves generation quality
Critical weaknesses
  • It looks like the proposed method depends on quite some heuristic hyperparameter values: for example, the confidence threshold that determines whether a single frame or multiple candidate frames should be used. Others include those of the fingerprint rule that selects the representative. It would be better if the paper could provide any principles for determining these values. A related question is: can we just always use the whole candidate set since its size is not large? How would it affect the model performance?
  • I didn't quite understand the config for generation. Some descriptions are confusing, e.g., in exp setup "the LDM test applies three rotations per molecule with matched diffusion noise, and generation uses 10,000 samples per configuration". To me the primary goal of canonicalization is to shape the input distribution (more like a preprocessing step), so I don’t expect the LDM needs the learned rotation during sampling. Can you clarify/expand this part a bit more?
  • The performance should be reported as mean and std across multiple runs.
Questions for authors

Please see Weaknesses

2 comments
AuthorsOpenReview ↗

Response to Reviewer rEpm

Reply 1 of 2: Parameter selection, frame diagnostics and generation protocol

Thank you for your comments. Our completed threshold scan supports retaining τc=10−3\tau_c=10^{-3}: it matches the severe-frame-error counts obtained with a larger threshold or always enabling candidates, while using candidates on approximately 11% of inputs. We have also completed all five AE variants and their common validation evaluation, providing the results below alongside clarifications of generation and repeated runs.

Parameter selection and Always24. The confidence threshold uses signed moments of the anchor geometry to determine when ambiguous PCA-axis signs require alternative frames, while the fingerprint selects a consistent representative at evaluation. The sensitivity check varied only the candidate gate on fixed Murcko-Set and Learnable-Set checkpoints using the same 1,280 validation molecules, with error counts below following evaluation seeds 0/1/2:

\text{Gate} & \substack{\text{Murcko}\\\\\text{candidate use}} & \substack{\text{Learnable}\\\\\text{candidate use}} & \substack{\text{Murcko errors}\\\\>90^{\circ}} & \substack{\text{Learnable errors}\\\\>90^{\circ}} \\\\ \hline 10^{-4} & 7.34\\% & 7.34\\% & 6 / 5 / 8 & 3 / 6 / 7 \\\\ 10^{-3} & 11.25\\% & 11.02\\% & 6 / 5 / 7 & 3 / 5 / 6 \\\\ 10^{-2} & 24.53\\% & 24.53\\% & 6 / 5 / 7 & 3 / 5 / 6 \\\\ \text{Always24} & 100\\% & 100\\% & 6 / 5 / 7 & 3 / 5 / 6 \\\\ \end{array}$$ The default $10^{-3}$ gives the same counts as $10^{-2}$ and Always24 with substantially fewer candidate activations, while improving some counts over $10^{-4}$, supporting its use to balance severe frame errors and candidate use. The fingerprint compares centered, scale-normalized projections and sums over atoms, giving translation invariance and independence from atom ordering. Scale controls score contrast and clipping bounds numerical range, while fixed element-type coefficients distinguish elements separately from learned PCA weights. With these coefficients held fixed, clipping values of 5, 10 and 20 give identical severe-error counts, whereas scale and direction affect some tail errors. The 24 proper frames address discrete sign/permutation ambiguity; continuous repeated-eigenvalue ambiguity remains outside this finite construction (Dym et al., 2024). The Always24 AE has completed training; this variant makes all 24 candidate frames available for every molecule and encodes the frame selected uniformly at random during training or by the fingerprint at evaluation. We compared the trained final AEs on the same 1,280 validation molecules using each model's deterministic evaluation policy: $$\begin{array}{lrr} \text{Final AE} & \text{Candidate use} & \substack{\text{Errors}>90^{\circ}\\\\\text{seeds 0 / 1 / 2}} \\\\ \hline \text{Murcko-Single} & 0\\% & 10 / 11 / 13 \\\\ \text{Murcko-Set} & 11.25\\% & 7 / 8 / 6 \\\\ \text{Learnable-Single} & 0\\% & 16 / 12 / 10 \\\\ \text{Learnable-Set} & 11.02\\% & 7 / 7 / 6 \\\\ \text{Murcko-Always24} & 100\\% & 7 / 8 / 6 \\\\ \end{array}$$ Under both Murcko and Learnable weighting, Set has fewer severe frame errors than its corresponding Single variant in every evaluation seed. Murcko-Set matches Murcko-Always24's counts while activating candidates on only 11.25% of inputs. **Generation protocol.** As you suggest, canonicalization shapes the training representation, while unconditional generation draws an atom count, starts from Gaussian latent noise, and applies reverse diffusion followed by decoding, independently of any source molecule, scaffold or input rotation. The three rotated copies measure rotation sensitivity on held-out molecules through a paired diagnostic with matched noise within each original/rotated pair, while the 10,000 samples in the main generation benchmark measure unconditional generation quality; we will clarify this distinction in the protocol description.
AuthorsOpenReview ↗

Response to Reviewer rEpm

Reply 2 of 2: Generation results and performance comparison

Table 3. QM9 generation ablation. The table includes the new Murcko-Always24 generation results. All metrics are percentages and reported as mean ±\pm standard deviation; higher is better.

LDM AE canonicalizerAtom stabilityMolecule stabilityValidityV×UV×U×N
All-Atom97.78 ± 0.3878.91 ± 0.2990.90 ± 0.2388.91 ± 0.3145.60 ± 0.33
Heavy-Atom97.68 ± 0.4377.88 ± 0.5790.30 ± 0.4988.51 ± 0.3745.48 ± 0.28
Murcko-Single97.54 ± 0.1779.58 ± 0.5289.95 ± 0.5389.63 ± 0.4148.78 ± 0.37
Murcko-Set98.67 ± 0.3583.39 ± 0.3691.53 ± 0.2790.40 ± 0.3564.52 ± 0.36
Murcko-Always2497.85 ± 0.2179.50 ± 0.4090.40 ± 0.2590.27 ± 0.4553.50 ± 0.32
Learnable-Single98.26 ± 0.0780.72 ± 0.2791.60 ± 0.3891.32 ± 0.3250.75 ± 0.41
Learnable-Set99.21 ± 0.0991.29 ± 0.2395.50 ± 0.1693.62 ± 0.2369.48 ± 0.24

Why retain adaptive gating? Always24 is a viable, simpler alternative. The adaptive gate restricts candidate-frame disambiguation to inputs flagged by the confidence test, activating candidates for 11.25% of Murcko inputs while matching Always24's severe-frame-error counts in the fixed-model threshold scan. In Table 3, Murcko-Set exceeds Murcko-Always24 by 3.89 percentage points in molecule stability, 1.13 in validity and 11.02 in V×U×NV\times U\times N. These results support retaining the adaptive configuration as our default, combining selective candidate activation with better observed molecule stability and valid, unique, novel yield.

Performance comparison and standard deviations. We report performance means and standard deviations for the methods in Table 2. We also include available literature error terms, retaining their original definitions and precision.

& \text{QM9} & \text{QM9} & \text{QM9} & \text{QM9} & \text{QM9} & \text{GEOM} & \text{GEOM} \\\\ \text{Method} & \text{Atom sta.} & \text{Mol. sta.} & \text{Validity} & V\times U & V\times U\times N & \text{Atom sta.} & \text{Validity} \\\\ \hline \text{ENF} & 85.0 & 4.9 & 40.2 & 39.4 & \text{--} & \text{--} & \text{--} \\\\ \text{G-SchNet} & 95.7 & 68.1 & 85.5 & 80.3 & \text{--} & \text{--} & \text{--} \\\\ \text{EDM} & 98.7\pm0.1 & 82.0\pm0.4 & 91.9\pm0.5 & 90.7\pm0.6 & 59.6 & 81.3 & 92.6 \\\\ \text{GeoLDM} & 98.9\pm0.1 & 89.4\pm0.5 & 93.8\pm0.4 & 92.7\pm0.5 & 53.9 & 84.4 & 99.3 \\\\ \text{GDM-aug} & 97.6 & 71.6 & 90.4 & 89.5 & 66.8 & 77.7 & 91.8 \\\\ \text{GraphLDM-aug} & 97.9 & 78.7 & 90.5 & 89.5 & \text{--} & 79.6 & 98.0 \\\\ \text{RADM DiT-B} & 98.5\pm0.0 & 87.3\pm0.2 & 94.1\pm0.1 & 91.7\pm0.1 & \text{--} & 85.0 & 99.3 \\\\ \text{UniGEM} & 99.0 & 89.8 & 95.0 & 93.2 & \text{--} & 85.1 & 98.4 \\\\ \text{EquiFM} & 98.9\pm0.1 & 88.3\pm0.3 & 94.7\pm0.4 & 93.5\pm0.3 & 53.7 & 84.1 & 98.9 \\\\ \text{GeoBFN} & 98.64\pm0.1 & 87.21\pm0.3 & 93.03\pm0.3 & 91.53\pm0.3 & 64.4 & 78.9 & 93.1 \\\\ \text{GOAT} & 99.2 & \text{--} & 92.9 & \text{--} & 72.3 & 84.8 & 96.2 \\\\ \text{MOLTD} & 99.40\pm0.1 & 92.53\pm0.2 & 96.04\pm0.4 & \text{--} & 67.03\pm0.6 & 86.9 & 95.3 \\\\ \text{AnchorSet} & 99.21\pm0.09 & 91.29\pm0.23 & 95.50\pm0.16 & 93.62\pm0.23 & 69.48\pm0.24 & 86.4 & 99.1 \\\\ \end{array}$$ ENF and G-SchNet were introduced by Garcia Satorras et al. (2021) and Gebauer et al. (2019); their benchmark values, together with the QM9 EDM and GDM-aug results, follow Hoogeboom et al. (2022), Tables 1–2. GeoLDM, GraphLDM-aug and the GEOM-Drugs EDM/GDM-aug values follow Xu et al. (2023), Table 1. RADM, UniGEM, EquiFM and GeoBFN results are taken from Ding and Hofmann (2025), Table 1, Feng et al. (2025), Table 2, Song et al. (2023), Table 1 and Song et al. (2024), Table 1, respectively. GOAT and the $V\times U\times N$ values for EDM, GeoLDM, GDM-aug, EquiFM and GeoBFN follow Hong et al. (2025), Tables 1–2; MOLTD follows Zhang et al. (2025), Table 1.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Key Findings

We thank the reviewers for their constructive feedback. The discussion has strengthened the empirical support for AnchorSet and clarified its advantages in frame consistency, selective candidate handling, generation quality, and practical efficiency.

  1. Structure-aware, set-valued canonicalization provides an effective representation for molecular generation. AnchorSet combines chemically motivated structural support, learned atom-type weights, and a compact set of candidate frames. Structural anchors define molecular pose, learned weights adapt the anchor geometry to reconstruction, and candidate frames preserve discrete orientation alternatives. The 24 proper signed permutations preserve chirality while handling residual sign and permutation ambiguity. This representation enables a standard DiT denoiser and GNN decoder to operate in canonical coordinates.

  2. Selective candidate activation improves frame consistency. Across the same 1,280 validation molecules, Set reduces severe frame errors relative to Single under both Murcko and learned weighting. Murcko-Set matches Always24’s severe-error counts while activating candidates on only 11.25% of inputs. The threshold scan supports the default threshold of 0.001, concentrating candidate handling on ambiguous molecular geometries. This selective policy preserves a single frame for high-confidence inputs and restricts candidate-frame randomization to cases requiring disambiguation.

  3. AnchorSet achieves strong generation quality with a standard latent diffusion backbone. The main Learnable-Set configuration achieves 91.29±0.23% molecular stability and 95.50±0.16% validity on QM9, together with 69.48±0.24% valid, unique, and novel yield. On GEOM-Drugs, AnchorSet uses a heavy-atom prior with learnable atom-type weights and achieves 86.4% atom stability and 99.1% validity, demonstrating applicability beyond small molecules. Unconditional sampling starts from a molecule size and Gaussian latent noise, requiring no source molecule or scaffold.

  4. Concrete computational measurements establish practical feasibility. On an RTX 4090, AnchorSet generates 1,000 QM9 molecules in 258.11 seconds, achieving 3.87 molecules per second with 2.24 GiB peak allocated memory. The reported AE and DiT-B training runs take approximately 17 and 150 hours, respectively. These measurements demonstrate a practical implementation that combines structure-aware canonicalization with standard neural generation components.

  5. Scaffold anchoring provides targeted robustness to peripheral molecular changes. Separate GEOM support diagnostics show that scaffold-based pose estimation isolates peripheral changes when the core remains stable. Under the tested peripheral perturbations, Murcko support maintains zero frame drift, compared with 6.356° for heavy-atom support. This behavior also holds in the larger-molecule subset, where the corresponding heavy-atom drift reaches 12.298°. When the core itself changes, broader heavy-atom support can provide greater stability, and the benefit varies across natural conformers. These findings identify molecules with stable cores and flexible peripheral groups as a particularly suitable setting for scaffold anchoring.

The central finding is that chemically informed anchoring and explicit frame-ambiguity handling provide an effective representation-level foundation for competitive, practical molecular generation. The revised manuscript will consolidate these findings with clearer descriptions of the framework, evaluation protocols, and component roles.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.