acceptodds
Under review as a conference paper at ICLR 2027

OctMole: Multiscale 3D Molecular Generation with Hierarchical Spatial Flows

Abstract

Generating accurate 3D molecules requires jointly determining a variable number of atoms, their identities, and geometry across spatial scales: global shape emerges coarsely, whereas chemical bonds demand fine local precision. Existing atom-based diffusion and flow models operate largely at a single scale, while dense spatial fields become costly at high resolution. We introduce OctMole, a multiscale octree flow framework for 3D molecular generation. A hierarchical occupancy flow progressively resolves sparse molecular support from coarse to fine, and a joint atom flow assigns atom types and continuous in-cell offsets to occupied leaves. This decomposition combines efficient global structure with subcell geometric precision, while cell corruption improves robustness to generated support. On QM9 and GEOM-DRUG, OctMole achieves strong chemical, distributional, and geometric quality, with particularly clear gains over Uni-3DAR on GEOM-DRUG and substantial efficiency gains over dense 3D generation. As an additional capability, intermediate occupancy states and selected regions provide spatial conditioning without task-specific retraining, supporting multiscale shape conditioning, regional editing, linker generation, and shape interpolation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
3D molecular generation, multiscale generation, octree representations, flow matching, shape-conditioned generation

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_J56jOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents OctMole, a multiscale framework for 3D molecular generation based on hierarchical octree representations and flow matching. The proposed method first generates molecular spatial occupancy in a coarse-to-fine manner and then predicts atom types and continuous local coordinates. The authors evaluate the approach on QM9 and GEOM-DRUG and demonstrate promising results in molecular quality, generation efficiency, and spatial conditioning.

Critical strengths
  1. Interesting and relevant research topic. The paper addresses an important problem in 3D molecular generation. Exploring multiscale spatial representations is a promising direction.

  2. Clear presentation. The paper is generally well-written and easy to follow.

  3. Reasonable architectural design. The use of hierarchical octree grids is intuitive and well-motivated. Progressively refining spatial occupancy from coarse to fine provides a natural way to capture molecular geometry while avoiding the computational overhead of dense voxel representations. The separation between occupancy generation and atom-level generation is also a reasonable design choice.

  4. Promising empirical results. The proposed method achieves strong results on QM9 and GEOM-DRUG, particularly in molecular stability and connectedness on GEOM-DRUG. The experiments on spatial conditioning and the efficiency comparisons with dense voxel representations further demonstrate the potential of the framework.

Critical weaknesses
  1. Discussion of related works. The paper could provide a more comprehensive discussion of recent hierarchical molecular generation methods, such as [1]. Although the generation settings may differ, a discussion of the similarities and differences would help clarify the novelty of OctMole.

[1] Hu, Jiapeng, et al. Hierarchical Multi-Scale Molecular Conformer Generation. ICLR 2026.

  1. Ablation of the hierarchical occupancy flow. It is unclear how much of the improvement comes specifically from flow matching rather than the octree representation or the continuous atom-coordinate decoder. It would be helpful to compare the proposed flow-matching approach with an autoregressive approach under the same octree representation and atom decoder, such as [1].

[1] Shuqi Lu, Haowei Lin, Lin Yao, Zhifeng Gao, Xiaohong Ji, Yitao Liang, Weinan E, Linfeng Zhang, and Guolin Ke. Unified cross-scale 3D generation and understanding via autoregressive modeling.

Questions for authors
  1. Could the authors clarify the main methodological differences between OctMole and existing hierarchical molecular generation approaches, particularly the work by Hu et al. (ICLR 2026)?

  2. How much of the empirical improvement can be attributed specifically to hierarchical flow matching? Would alternative occupancy-generation methods achieve comparable results when using the same continuous atom decoder?

  3. What causes the relatively low uniqueness on QM9? Is there evidence that the model memorizes training molecules, or is the decrease primarily caused by candidate selection?

2 comments

We thank the reviewer for the constructive questions. We clarify the hierarchy and provide a new fixed-decoder comparison and molecule-level diversity analysis.

1. Relation to Hu et al. (ICLR 2026). Both methods use coarse-to-fine generation. Hu et al. generate conformers conditioned on a given molecular graph, with a hierarchy of chemical substructures such as scaffolds. OctMole performs de novo generation of atom counts, types and coordinates. Our hierarchy is a spatial partition: progressively finer octree occupancy defines the support for continuous atom generation. We have added this comparison to the related-work discussion.

2. Autoregressive occupancy with the same atom decoder. Following your suggestion, we feed depth-6 (64^3) support from the official Uni-3DAR checkpoint into our default GEOM-DRUG atom decoder. (Both use the same physical box and cell size; aligning their coordinate systems requires only a translation.) This hybrid method uses OctMole's atom-confidence reranking, retaining approximately 20% of candidates, comparable to Uni-3DAR's native selection rate (20%).

Below, each method has 10,000 final samples evaluated identically, without clash-atom removal or coordinate optimization; validity requires whole-graph sanitization.

MethodMean atomsAtom stableMol stableValidUniqueConnectedValid & connected
Uni-3DAR30.07099.06478.58096.63098.68689.89086.860
Hybrid31.15299.21381.00097.30099.34294.02091.480
OctMole43.67599.63687.27098.18099.98096.99095.220

Distribution distances (lower is better; length in Å, angle in degrees):

MethodNodes W1Atom TVBond TVValency W1Length W1Angle W1
Uni-3DAR16.5808150.0412730.0221110.0229450.0017121.026799
Hybrid15.4991290.0431870.0206240.0235680.0027711.654627
OctMole2.9797200.0143850.0032310.0082700.0006320.893857

Molecule stability (%) by atom count:

Method≤2526–3536–4546–55>55
Uni-3DAR85.8878.0172.3865.6050.00
Hybrid86.4680.1078.4678.6250.00
OctMole93.6287.5487.0387.3486.97

(The AR groups have very few samples above 55 atoms, so conclusions rely on the better-populated bins.)

The hybrid improves molecule stability over native Uni-3DAR by 2.42 points. OctMole further improves molecule stability by 6.27 percentage points and improves all six distribution distances over the hybrid. Its stability advantage persists within size bins. The hybrid retains the AR support's small-molecule bias because its atom count is determined by Uni-3DAR.

These results demonstrate a contribution from our occupancy generator beyond the atom decoder.

3. QM9 uniqueness and training-set overlap. Both overfitting and candidate selection contribute. On equal-size 10k QM9 sample sets, OctMole's uniqueness drops from 98.54% before selection to 80.80% after selection. We also compared generated molecules against the training set using canonical SMILES. Both generated samples and the full training sets are evaluated without clash-atom removal or coordinate optimization; validity requires whole-graph sanitization. Training overlap counts generated samples matching the training set, with valid samples as the percentage denominator.

QM9:

MethodSamplesTotalValidTraining overlapOverlap %
OctMoleRaw98,46480,76126,26332.52
OctMoleReranked10,0009,9729,12291.48
Uni-3DARRaw98,46483,2857,4918.99
Uni-3DARReranked10,0009,9802,09721.01

GEOM-DRUG:

MethodSamplesTotalValidTraining overlapOverlap %
OctMoleRaw49,70946,576300.064
OctMoleReranked10,0009,818280.285
Uni-3DARRaw49,70945,9221050.229
Uni-3DARReranked10,0009,663890.921

We acknowledge molecular-identity memorization on QM9: reranking strongly favors training molecules and amplifies overfitting, plausibly related to its small training set (~100k). On GEOM-DRUG (~5.54M training conformations), OctMole retains 99.98% uniqueness and 99.715% novelty among valid reranked samples, with lower training overlap than Uni-3DAR both before and after selection. Thus, the QM9 limitation does not persist on the larger dataset. We have replaced the original speculation with these quantitative findings.

Thanks for the response from the authors. I will keep my positive rating.

Reviewer_m7CtOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper proposes OctMole, a framework for multiscale 3D molecular generation. The method first generates occupied octree cells through hierarchical occupancy flows, and then generates atom types and continuous local coordinates with a joint atom flow. Cell corruption is introduced to reduce the mismatch between ground-truth and generated spatial support. The same framework also supports spatial conditioning at different resolutions without task-specific training. The idea is reasonable, and the experiments show promising results on GEOM-DRUG. However, I have several concerns about the baseline comparison, the contribution of individual designs, and the evaluation protocol. In the current version, I am not fully convinced that the advantages of the proposed multiscale framework have been sufficiently established.

Critical strengths
  1. The decomposition of spatial support and atom-level generation is clear and well motivated. Intermediate octree states provide a natural interface for controlling the resolution of spatial conditions.

  2. The improvement over Uni-3DAR on GEOM-DRUG is substantial. The size-stratified evaluation is useful, as it supports that the improvement in molecular stability is not simply obtained by generating smaller molecules.

  3. Cell corruption addresses a practical problem in the two-stage pipeline. Its improvement from 71.15% to 86.82% molecular stability on GEOM-DRUG is convincing within the reported setting.

  4. The paper includes extensive analyses and explicitly discusses several limitations. Reporting unsuccessful editing outcomes and the selection bias in physical evaluation improves the transparency of the work.

Critical weaknesses
  1. The baselines are not enough. The paper already discusses SemlaFlow and TABASCO, but does not compare with them in experiments. Since this paper also uses flow matching, I think these comparisons are necessary. The improvement over Uni-3DAR is clear, but how about other strong flow-based methods? From the current results, this is still not clear.

  2. The ablation cannot fully explain where the improvement comes from. In the dense-versus-octree experiment, the architecture, parameter number, and batch size are different. Also, the dense baseline only gets 4.91% molecule stability on GEOM-DRUG. With such a low result, I am not sure whether this is a suitable baseline to show the benefit of the hierarchy. For continuous offsets, putting all atoms at cell centers will naturally introduce coordinate errors. This experiment shows that fine coordinates are needed, but does not explain why the proposed way is better than other ways of predicting fine coordinates.

  3. The influence of reranking is not sufficiently studied. The model only keeps around 10% candidates on QM9 and 20% on GEOM-DRUG, while several baselines do not use such selection. I would like to see the results before reranking. Another concern is that QM9 uniqueness is only 80.77%. The authors mention possible memorization, but do not further check it at the molecule level. Since the training-set occupancy matches also increase after selection, I think this issue needs more investigation.

Questions for authors
  1. Can the authors provide results without reranking, together with molecular novelty against the training set and time per valid, unique molecule?

  2. How frequently does clash-atom removal modify generated molecules, and how do the main quality metrics change without this removal?

  3. Why is the GEOM-DRUG bond-length W1 approximately 0.001962 Å in Table 9, but 0.017 Å before snapping in the offset ablation? Please clarify the sample sets and aggregation protocols.

5 comments

We thank the reviewer for the constructive feedback on the method and experiments.

1. Results without reranking, novelty and efficiency. We evaluated all raw candidates (98,464 QM9 / 49,709 GEOM-DRUG) and 10,000 reranked samples per dataset, without atom removal or coordinate optimization and with whole-graph sanitization. Novelty is the fraction of valid samples whose canonical SMILES is absent from the training set. Time per final molecule comes from a single A100.

Dataset / samplesValid %Unique %Novelty %ms / reranked moleculems / validms / valid & unique
QM9 / raw82.0289.6267.48———
QM9 / reranked99.7280.808.52327.61328.53406.62
GEOM-DRUG / raw93.7099.99699.936———
GEOM-DRUG / reranked98.1899.9899.715342.82349.17349.25

Reranking improves validity but amplifies QM9 overfitting: it favors training molecules. On equal-size 10k QM9 sets, uniqueness drops from 98.54% to 80.80%. We acknowledge this limitation on the small QM9 training set (~100k). On GEOM-DRUG (~5.54M), reranked outputs retain 99.98% uniqueness and 99.715% novelty, with little additional cost per valid, unique molecule.

2. Effect of clash-atom removal. We evaluated the same samples with and without the 0.8 Å removal step, keeping other evaluation settings unchanged. Each row below contains 10,000 samples; the raw rows use fixed 10k subsets of the candidate pools. Arrows indicate “without → with removal”.

Dataset / samplesMolecules modifiedAtoms removedMol. stableValidConnected
QM9 / raw800 (8.00%)85264.55 → 62.8782.19 → 82.3784.53 → 83.42
QM9 / reranked41 (0.41%)4199.55 → 99.2399.72 → 99.7299.94 → 99.81
GEOM-DRUG / raw384 (3.84%)39662.42 → 61.4993.52 → 93.6184.20 → 83.13
GEOM-DRUG / reranked47 (0.47%)4787.10 → 86.8298.19 → 98.2096.99 → 96.90

Only 0.41%/0.47% of the final QM9/GEOM-DRUG samples are modified. Thus, the reported stability gains are not produced by removing clashing atoms.

3. Bond-length W1 discrepancy. Thank you for identifying this inconsistency. We apologize for using different evaluation protocols in Table 9 and the offset ablation.

  • Section E.2: unbinned bond-length W1 grouped by element pair and bond order, weighted by reference frequencies normalized over actual bonds; no clash removal.
  • Table 9: grouping by bond order with 0.01 Å bins and the full validation reference. Each bond-order distance is weighted by its frequency among all reference atom pairs, including non-bonded pairs in the normalization. This pipeline also includes 0.8 Å clash removal.

We have rerun the ablation using the same 10,000 final samples and the Table 9 pipeline for both original and snapped coordinates. Corrected results are:

DatasetBond-length W1 (Å), original → snappedBond-angle W1 (°), original → snapped
QM90.000805 → 0.0063770.437003 → 2.780979
GEOM-DRUG0.001962 → 0.0038630.891154 → 6.913506

Under the unified protocol, snapping worsens both bond-length and bond-angle distributions on both datasets, supporting the importance of continuous coordinate refinement.

4. More baselines. We agree that comparisons with strong flow-based models are important. We evaluated SemlaFlow, TABASCO, and EQGAT-diff on GEOM-DRUG using 10,000 final samples per method with the metrics in our paper.

The training distributions and modeled outputs differ: SemlaFlow uses one lowest-energy conformer per molecule, TABASCO up to five, and OctMole up to 30. SemlaFlow and EQGAT-diff predict molecular graphs; OctMole predicts atom types and coordinates, while TABASCO generates heavy atoms only and infers bonds in post-processing; we add hydrogens using RDKit for evaluation. Predicted graphs are retained where available; bonds are inferred for coordinate-only outputs.

MethodAtom stableMol stableValidUniqueConnectedValid & connectedPoseBusters
OctMole (ours)99.63687.27098.19099.98096.99095.22090.31
SemlaFlow95.53793.75094.30099.98996.82091.44087.07
TABASCO99.93297.63097.90098.99999.90097.80090.69
EQGAT-diff99.34892.82085.830100.00096.54082.98077.88

Bold and italicized entries indicate the best and second-best results, respectively. PoseBusters (%) is evaluated using TABASCO’s official configuration where a sample must pass all chemical and geometric checks in this configuration.

OctMole achieves the highest validity (98.19%) and 95.22% valid-and-connected samples, while SemlaFlow and TABASCO achieve higher molecule stability.

5. Continuous offset prediction. We agree that snapping alone only shows the need for fine coordinates. We have added two experiments to compare alternative offset predictions.

First, we fixed the octree support and atom types and replaced only the offsets. On the same 2,000 reranked GEOM-DRUG samples, molecule stability was 87.5% with learned offsets, 35.4% with cell centers, and 35.5% with per-element training-mean offsets. Shuffling offsets among atoms of the same element reduced stability to 2.0%, despite preserving their marginal distribution. This shows that offsets must match the local molecular context.

Second, we trained a direct L2 regression decoder with the same backbone, data, and training budget. We evaluated both decoders on the same 2,500 GT supports and 2,500 generated supports from GEOM-DRUG, without reranking.

SupportDecoderMol. stable (%) ↑xTB relaxation (median) ↓
GTFlow (ours)93.00.370
GTRegression93.20.043
GeneratedFlow (ours)62.61.39
GeneratedRegression39.03.09

Regression performs well on GT supports. On generated supports, flow improves molecule stability by 23.6 percentage points and requires less xTB relaxation. These results support the benefit of our flow decoder over direct regression in the actual multimodal generation setting.

Thanks for the reply. I still have some small questions. In the dense vs. octree comparison, the architecture and batch size are also different. Should I read the reported speedup as a comparison of the two full systems? Besides, in the new GEOM-DRUG comparison, did SemlaFlow and TABASCO also use reranking? How many raw candidates did each method generate to get the final 10,000 samples? It's not a must to answer the rest of the questions, since I think your response is convincing enough for my judgement on this paper.

Reviewer_8UhROpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces OctMole, a hierarchical multi-scale flow-matching model for molecular design. Molecules are generated by two subsequent flows. First, an occupancy flow determines where atoms should be placed through binary occupancies inside cells in 3D space. Second, an atom flow determines the identity and and precise positon of atoms within the occupied cells. The framework is additionally capable of regional editing; for example, the model can generate a linker given two pre-defined molecular fragments. OctMole performs well in comparison to the included baselines on the reported metrics.

Critical strengths

OctMole's approach to molecular generation is, to my knowledge, novel in the context of molecular design. Hierarchical and multi-scale generative approaches have been successful in 3D generation within other domains, and OctMole is (again, to my knowledge) one of the first non-autoregressive approaches tackling molecular design in this manner. The proposed method's performance appears to be competitive based on the results that are presented (with VERY important caveats; see below). This, in principle, makes the work well worth sharing with the community.

Critical weaknesses

Firstly, the arguments for the potential advantage of OctMole's approach to 3D generation are unconvincing. The introduction claims that 3D molecular generation is a "geometric problem at a chemically unforgiving scale" and positions multi-scale hierarchical generation as a solution. This is undermined by the prevalence of capable 3D molecular generative models that don't approach the problem in this way, however. This shortcoming is further reinforced by the chosen baselines; EDM, GeoLDM, and UniGEM all underperform (in terms of e.g. molecular stability) on the GEOM-Drugs dataset compared to OctMole. Considering the SemlaFlow paper referenced in the "Related work" section revealed that both EQGat-Diff and SemlaFlow, neither of which approach molecular generation through multi-scale/hierarchical approaches, outperform OctMole on the GEOM-Drugs dataset in terms of molecular stability. This seems to indicate that hierarchical/multi-scale approaches are unnecessary, and that baselines may have been cherry-picked to make OctMole look better in comparison.

Second, several small oddities in the method and results (e.g. the statistical significance of the results, the surgical modification of unrealistic samples, the low uniqueness on GEOM-Drugs, the removal of atoms outside the arbitrarily chosen bounding box) has a negative impact on the apparent soundness of the work.

Questions for authors

-- Critical --

\textbf{Metrics.} More of a comment than a question: Section 4.1 "Metrics" does not mention how atom stability is defined (though it is defined in the appendix), while the other metrics appear to be briefly explained there.

\textbf{Statistical significance of results.} Were several runs conducted to produce the results presented in the paper? If so, that should be stated clearly; if not, the advantage of OctMole over the presented baselines could in principle be a fluke.

\textbf{Baselines.} What merited the choice of baselines, given that other 3D generators from literature (e.g. EQGat-diff and SemlaFlow as stated under "Critical weaknesses") exhibit higher molecular stability than OctMole on GEOM-Drugs?

\textbf{Framing.} As outlined in the Critical Weaknesses section above, I am not sold on the necessity of the hierarchical/multi-scale generation here, especially given the choice of seemingly weak baselines. Could you elaborate on why hierarchical/multi-scale generation is advantageous for molecules?

\textbf{Post-processing.} Appendix C.1 states the following: "Following VoxMol (Pinheiro et al., 2023), we first remove atoms involved in unrealistically close contacts (below 0.8 Å)." To me, this feels almost like cheating; surgically correcting unrealistic samples in this manner could hide significant shortcomings of OctMole's sampling statistics. While OctMole follows VoxMol here, I find the choice just as questionable in that work. How much of an impact does this post-processing step have on the reported metrics?

\textbf{Stability metric.} Appendix C.1 furthermore states: "Stability is evaluated on the inferred graph before sanitization." Does that imply that the stability metric includes potentially invalid molecules? If so, how large is the fraction of stable molecules that fail sanitization?

\textbf{GEOM-Drugs uniqueness.} OctMole exhibits poor uniqueness on GEOM-Drugs. The authors speculate that memorization of the relatively small training dataset could be the culprit. Why is this a bigger issue for OctMole than for other approaches to 3D generation, which do not exhibit this shortcoming?

\textbf{Box bounds.} Appendix B.1 states the following: "If it exceeds the domain, we recenter it on the atom nearest the centroid and retain only atoms inside the box." That sounds potentially problematic, and implies that some molecules simply cannot be represented using the proposed method, potentially impacting OctMole's capability of learning the underlying molecular distribution. This issue seems large enough to at the very least merit mention in the main manuscript rather than being relegated to the Appendix. Is this a problem in practice? How many molecules in the relevant datasets does this apply to?

\textbf{Training-time data augmentation.} Appendix B.4 mentions that random rotations were applied during atom flow training on both datasets, while cell corruption was applied only to GEOM-Drugs. Two questions on that:

  1. Why was cell corruption not applied to the QM9 dataset?
  2. Were different rotations/cell corruptions applied in each epoch?

\textbf{Rotation augmentation.} Random rotations to inputs were applied during training; this should mean that the resulting model would be more or less agnostic to inference-time rotations. Table 12 nevertheless shows a noticeable degradation of molecular stability under rotations, particularly noticeable for depth 4. (The numbers in Table 12 also seem to disagree with the numbers stated in Section 7 "Limitations and scope," which should be corrected.) Could the authors elaborate on why this degradation might occur despite training-time rotation randomization?

\textbf{Sampling.} Section 4.1 "Baselines and sample selection" mentions the following: "OctMole retains approximately 10% of its QM9 candidates and 20% of its GEOM-DRUG candidates by mean atom-type log probability, [...]" Does this mean that you generate approximately 100 000 QM9 candidates and 50 000 GEOM-Drugs candidates to produce the final 10 000 samples for each? That seems to undermine the efficiency claims made regarding OctMole.

-- Clarifications --

\textbf{Sparse transformer.} To my knowledge, the term "sparse transformer" usually refers to a class of transformer models utilizing subquadratic self-attention formulations. What does sparsity entail in this work?

\textbf{Register tokens.} Appendix B.3 mentions that you insert eight learned register tokens partway through the transformer, which are removed before the output. What is the purpose of these register tokens?

\textbf{Cell corruption.} Is my interpretation correct in that the cell corruption scheme may displace the atom outside of the originally assigned cell? Appendix B.2 mentions that "Shifts outside the domain or into occupied cells are skipped," but the term "domain" appears to refer to the B^3 bounding box.

-- Future work --

\textbf{Regional editing.} To me, one of the most interesting parts of the paper is OctMole's regional editing and linker generation capabilities; these molecular design strategies are prevalent in drug design, but most work on 3D generators focus entirely on de novo design. I think it is a shame that it was not expanded on further in the paper, but I suspect the rather poor performance on these tasks is the reason for the lack of emphasis. Could the authors elaborate on their thoughts on why performance here is poor and how it could be rectified in future work?

4 comments

We thank the reviewer for the comprehensive and constructive feedback on our paper. Within our available resources, we have conducted additional experiments and clarified several details.

-- Critical --

Metrics. We have added the definition of atom stability to Section 4.1.

Statistical significance of results. The original sampling results used seed 0. We added sampling seeds 1 and 2 and now report means and sample standard deviations across three sampling seeds with fixed checkpoints (Tables 1–2 and Figures 3 and 10).

Baselines. Recent flow models use different GEOM-DRUG training distributions: SemlaFlow uses the first, lowest-energy conformer per molecule, while TABASCO uses up to five; OctMole uses up to 30 conformers per molecule. In addition, TABASCO generates heavy atoms and infers bonds in post-processing; we add hydrogens using RDKit for evaluation.

These differences motivated our original baseline selection and must be accounted for when comparing results. We evaluate these methods on GEOM-DRUG using the evaluation metrics in our paper, with 10,000 final samples per method. Predicted graphs are retained where available; bonds are inferred for coordinate-only outputs.

MethodAtom stableMol stableValidUniqueConnectedValid & connectedPoseBusters
OctMole (ours)99.63687.27098.19099.98096.99095.22090.31
SemlaFlow95.53793.75094.30099.98996.82091.44087.07
TABASCO99.93297.63097.90098.99999.90097.80090.69
EQGAT-diff99.34892.82085.830100.00096.54082.98077.88

Bold and italicized entries indicate the best and second-best results, respectively. PoseBusters (%) is evaluated using TABASCO’s official configuration where a sample must pass all chemical and geometric checks in this configuration.

Framing. Hierarchy offers concrete benefits for our spatial generation approach, although our experiments do not establish it as necessary for all high-quality molecular generators. On drug-like molecules, our multiscale model substantially outperforms our dense baseline (molecule stability below 5% on GEOM-DRUG). Our mean molecule stability of 86.18% is also higher than the 75.0±1.0% reported for VoxMol. The same model supports shape-preserving chemical exploration and finer reconstruction through coarse and fine occupancy conditions.

  • Note: We could not find public VoxMol checkpoints or evaluation scripts. Its evaluation framework is broadly similar to ours, so we cite its published result.

Post-processing. We evaluated the same samples with and without the 0.8 Å removal step, keeping other evaluation settings unchanged. Each row below contains 10,000 samples; the raw rows use fixed 10k subsets of the candidate pools. Arrows indicate without → with removal.

Dataset / samplesMolecules modifiedAtoms removedMol. stableValidConnected
QM9 / raw800 (8.00%)85264.55 → 62.8782.19 → 82.3784.53 → 83.42
QM9 / reranked41 (0.41%)4199.55 → 99.2399.72 → 99.7299.94 → 99.81
GEOM-DRUG / raw384 (3.84%)39662.42 → 61.4993.52 → 93.6184.20 → 83.13
GEOM-DRUG / reranked47 (0.47%)4787.10 → 86.8298.19 → 98.2096.99 → 96.90

Only 0.41%/0.47% of the final QM9/GEOM-DRUG samples are modified. Thus, the reported stability gains are not produced by removing clashing atoms.

Stability metric. This is possible by definition. However, across all four baselines and OctMole on both datasets (10,000 samples per method and dataset), every molecule classified as stable passes the validity check.

GEOM-Drugs uniqueness. The low uniqueness actually occurs on QM9. Our additional analysis identifies both overfitting and candidate selection as contributing factors. On equal-size 10k QM9 sets, uniqueness decreases from 98.54% before reranking to 80.80% afterward. Training-set overlap among valid samples increases from 32.52% to 91.48%, compared with 8.99% to 21.01% for Uni-3DAR. We acknowledge that OctMole exhibits stronger molecular-identity memorization on QM9, which confidence-based selection amplifies. This may be related to the small training set (~100k).

On GEOM-DRUG (~5.54M training conformations), OctMole achieves 99.98% uniqueness after reranking, compared with 98.69% for Uni-3DAR. Training-set overlap is only 0.285%, versus 0.921% for Uni-3DAR.

Box bounds. Clipping is rare in practice. For QM9, all centered conformations fit within the box under any rotation, as verified by their maximum radial extent. For GEOM-DRUG, a radius-based conservative bound accounting for recentering identifies at most 114 training conformations (0.00206%) and 656 atoms that could be clipped. Simulating the training rotation procedure at epoch 0 with seeds 0, 1 and 2 clips 22/24/29 conformations and 83/111/113 atoms, respectively, out of 5,538,014 training conformations.

Training-time data augmentation.

  • QM9 configuration: Cell corruption provides a modest gain on QM9 (molecule stability: 99.23% → 99.73%), where metrics are already near saturation. Therefore, we left it disabled in the main configuration.
  • Augmentation sampling: The random seed depends on the global seed, epoch and sample index: rng_seed = global_seed + epoch * 1000003 + sample_index. Rotations and cell corruptions are resampled for each sample across epochs.

Rotation augmentation. Random rotations are applied only during atom generation (Stage 2) training. Occupancy generation (Stage 1) was trained without them because they slowed convergence. We evaluated the same 256 molecules with and without random rotation, voxelized them at different depths, and measured the resulting decrease in generated-molecule stability:

Clamp depthSettingDecrease (pp) [95% CI]
6stage 2 only1.6 [0.0, 5.2]
4stage 1 & 211.5 [7.1, 15.4]
4stage 1 & 2 (w/o cell corruption)31.7 [27.0, 36.0]
2stage 1 & 2−1.7 [−3.9, 0.5]

The depth-4 rotation sensitivity arises primarily along the Stage 1 path; the decrease is smaller when Stage 2 runs alone. An axis-aligned grid produces orientation-dependent occupancy conditions, which may lead Stage 1, trained without rotation augmentation, to generate completions of different quality. Cell corruption reduces the stability drop from 31.7 to 11.5 pp.

Sampling. Yes. Producing 10,000 retained samples requires 98,464 candidates on QM9 and 49,709 on GEOM-DRUG. Our reported generation time per retained molecule already includes the cost of generating all discarded candidates. In the single-A100 40GB benchmark with 500 retained samples, the default 50-step configuration takes 327.61 ms per retained QM9 molecule and 342.82 ms per retained GEOM-DRUG molecule.

-- Clarifications --

Sparse Transformer. Sparsity refers to the spatial representation, in contrast to dense voxel representations such as VoxMol. Stage 1 assigns tokens to occupied parent cells, and Stage 2 processes occupied leaves. Attention is global and non-causal over the existing tokens.

Register tokens. The eight learned tokens provide shared intermediate states for information aggregation and processing. They participate in attention but do not correspond to atoms or occupied cells, and are removed before the output heads.

Cell corruption. Yes—the atom can lie outside its reassigned cell, while its absolute position remains unchanged. In cell units, shifting the GT cell from c to c+δ changes the offset target from d to d−δ, preserving the physical coordinate. The domain is the entire B^3 box. We sample δ from the six unit axis directions and skip shifts that leave the box or enter an occupied cell.

-- Future work --

Regional editing. We agree that these tasks deserve further attention. Fixing occupancy does not fully preserve an existing molecular fragment: even when the same cells remain occupied, Stage 2 can change their atom types and coordinate offsets.

For linker generation (100 targets, up to 16 candidates each), 73% of candidates are disconnected; 71% of those failures are failures to bridge the fragments. Candidate selection is also limiting: oracle best-of-16 succeeds on 46% of targets, whereas the current score selects a successful top-1 candidate for only 16%.

For GEOM-DRUG plane editing, strict success is 6.0%. Retained-side positional deviation has a median of 0.37 Å, but only 16.5% of samples preserve its atom types because Stage 2 resamples them. These findings motivate fragment-conditioned Stage 1 fine-tuning, linker-aware reranking, and explicitly fixing retained atom types and coordinates.

Thank you the comprehensive reply! My main concerns with the work have been answered. While I am still not fully sold on the hierarchical approach in molecular generation, I appreciate that it is beneficial in this context. I believe the table with additional baselines should be added to the paper, with the accompanying discussion on training distribution differences. With those changes, I would be more than happy to raise my score to a 3.

Thank you for your thoughtful feedback and we appreciate your reconsideration of our work! We will add the additional baseline comparisons and the accompanying discussion of training distribution differences to the revised paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.