acceptodds
Under review as a conference paper at ICLR 2027

Flow-Matching Based Refiner for Molecular Conformer Generation

Abstract

Low-energy molecular conformer generation (MCG) is a foundational yet challenging problem in drug discovery. Denoising-based generative models, including diffusion and flow matching, have become the dominant paradigm for MCG and achieved state-of-the-art performance by transporting samples from a simple base distribution to the molecular conformer distribution. However, sequential sampling can accumulate errors, especially in low-SNR regimes where inference states deviate from their timestep-conditioned training distributions. This train-inference mismatch makes vector-field prediction on off-trajectory states unreliable, further propagating errors through denoising. To address this issue, we propose a mismatch-aware flow-matching framework for molecular conformer refinement. Trained on data-centered perturbations, the refiner corrects upstream conformers without requiring explicit estimates of their residual error. Experiments on GEOM-QM9 and GEOM-Drugs show consistent quality improvements across strong denoising-based generators, while preserving diversity and reducing the total number of sampling steps.

Primary area:
generative models
Keywords:
Generative Model, Molecular Conformer Generation

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_nqRXOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This work introduces a mismatch-aware flow-matching framework to refine molecular conformer predictions. This model can be trained and then applied without per-model tuning. When the refiner is applied RMSD-based conformer ensemble metrics improve.

Critical strengths
  • The application of the refiner is well describes and does result in small reduction in steps
  • Extensive ablations on the impact on the refiner on the set benchmarks and evaluated mdoels.
Critical weaknesses
  • The paper discusses the goal of generating low energy conformers yet energy-based benchmarks are not performed [1, 2]. This work should also be benchmarked against and discussed as it related to low energy structures and energy-based benchmarks.
  • Given the focus of the work is to introduce a new flow matching theoretical framework why isn't this applied to other tasks like for example 3D molecule generation Protein design or images? With only the AMR benchmarks with known flaws and minor improvements its hard to understand if this is hacking the benchmark or there is generalizable theoretical value for structure-based diffusion or flow refiners.
  • Does not compare against AverageFlow[3] (both DiT non SE3 equivariant, and Nequip based)

[1] LoQI https://chemrxiv.org/doi/full/10.26434/chemrxiv-2025-k4h7v/v3 [2] https://pubs.rsc.org/dd/article/4/11/3282/889094 [3] AvgFlow https://arxiv.org/html/2507.09785v1

Questions for authors
  1. How does the refiner impact the energy metrics presented in [1]. In general its hard to assess the impact of the refiner as computational chemists do not use RMSD to know if the conformer is good or not as at test time you often do not have a reference as that is the whole point of the model.
  2. How do the Coverage results change as a function the delta? Do stricter cutoffs lessen or improve the impact of the refiner?
  3. How does the refiner work when applied to Average Flow at 1,2, and 100 steps?
  4. How does the refiner compared to xtb relaxation in both energy, rmsd, and time cost?
8 comments

W1/Q1: Energy-based evaluation

Thank you for raising this point. We agree that RMSD alone is insufficient to assess the physical quality of generated conformers. We already report an energy-related evaluation in Appendix D, “Chemical Property,” Table 6,

Following the protocol of Jing et al. (2022), we evaluated the same 100-molecule subset, selecting min⁡(2K,32)\min(2K,32) generated conformers for a molecule with KK reference conformers. After GFN2-xTB relaxation, we compare the generated and reference ensembles in terms of energy (EE), dipole moment (μ\mu), HOMO–LUMO gap (Δϵ\Delta\epsilon), and minimum energy (Emin⁡E_{\min}).

Median absolute errors in ensemble properties. Ensemble averages use Boltzmann weighting; Emin⁡E_{\min} compares the ensemble minima. EE, Δϵ\Delta\epsilon, and Emin⁡E_{\min} are in kcal/mol, and μ\mu is in Debye. Lower is better.

MethodEE ↓μ\mu ↓Δϵ\Delta\epsilon ↓Emin⁡E_{\min} ↓
GeoDiff0.310.350.890.39
GeoMol0.420.340.590.40
Torsional Diffusion0.220.350.540.13
MCF-L0.680.280.630.04
ET-Flow0.230.190.380.02
MCF-L + Refiner0.200.230.380.02
ET-Flow + Refiner0.210.180.390.01

The Refiner reduces the energy error from 0.68 to 0.20 for MCF-L and from 0.23 to 0.21 for ET-Flow. The minimum-energy error also decreases from 0.04 to 0.02 and from 0.02 to 0.01, respectively. These results provide evidence of improved agreement with reference energy statistics beyond RMSD.

We further report chemical-property errors before xTB relaxation. As shown in following tables, our refiner (20+20 steps) consistently reduces all reported errors relative to baseline (50 steps) under both uniform and Boltzmann weighting, demonstrating improvements in the generated conformers before geometry optimization.

Median errors with uniform weighting, before xTB relax.

MethodEE ↓μ\mu ↓Δϵ\Delta\epsilon ↓Emin⁡E_{\min} ↓
MCF-L2981.98002.437517.1721145.2360
ET-Flow7.68540.57731.78491.8387
ET-Flow + Refiner7.36090.47881.41491.0336

Table 2. Median errors with Boltzmann weighting at 298.15 K, before xTB relax optimization.

MethodEE ↓μ\mu ↓Δϵ\Delta\epsilon ↓Emin⁡E_{\min} ↓
MCF-L144.78901.29264.1674145.2360
ET-Flow2.26490.39951.32721.8387
ET-Flow + Refiner1.18380.39381.11411.0336

Before xTB relaxation, our refiner consistently reduces all reported chemical-property errors relative to ET-Flow under both uniform and Boltzmann weighting, indicating improvements beyond geometric similarity.

W2 Study Scope We thank the reviewer for encouraging us to consider other tasks, which helps us clarify the rationale and scope of our method. Our explanation is based on the representations used to predict coordinate updates. With a fixed molecular graph, relative atomic geometry can remain similar to that encountered during training even when the assigned time does not match the actual coordinate-error scale. We have revised Section 3.3 to discuss this mechanism further and added a Monte Carlo analysis in Appendix O to provide supporting evidence. Across 20 groups of one million trials, the results for the error interval [−0.1, 0.1] are:

Two-pixel RGB descriptor (%)Atom-pair distance (%)Three-atom angle (%)
0.00440 (0.00160)5.63084 (0.02621)25.72525 (8.91133)

Values are mean(std) across groups. The interval is measured in normalized intensity units, Å, and radians, respectively. A pixel hit requires all six channel errors to fall within the interval simultaneously. Under these simulation settings, distances and angles satisfy the interval criterion approximately 1,280 and 5,850 times as often as the joint RGB descriptor. Many coordinate configurations share similar distances or angles, supporting the possibility of encountering familiar local geometry despite time mismatch. Image-restoration studies [1] show the initial noise scale to zero from degraded images produces blurry outputs high FID. why successful refinement without added noise cannot be assumed to transfer directly to images.

Extending the method to other tasks requires studying whether their representations remain robust under time mismatch. We view conformer refinement as a particular setting whose geometric representations can support this behavior. Identifying which tasks and representations permit such refinement, and where it fails, is an interesting direction for future research.

[1] Ohayon, Michaeli, and Elad. Posterior-Mean Rectified Flow: Towards Minimum MSE Photo-Realistic Image Restoration. ICLR 2025. Paper

Q2 the performance under standard delta cutoff

We thank the reviewer for suggesting an evaluation across coverage thresholds, which provides a more comprehensive assessment of the Refiner’s effectiveness. The conventional cutoff of δ = 0.5 Å is relatively permissive for QM9: median COV-R is already 100% for all baselines below, and median COV-P is 100% for most of them. This limits its ability to distinguish improvements in geometric accuracy. We therefore use δ = 0.05 Å to provide a more discriminative evaluation.

Following previous work, we also report results at δ = 0.5 Å. Our refiner achieves the best performance

MethodCOV-R Median (%) ↑COV-R Mean (%) ↑COV-P Median (%) ↑COV-P Mean (%) ↑
OMEGA100.085.5100.082.9
GeoMol100.091.5100.086.7
GeoDiff100.076.533.550.0
Torsional Diffusion100.092.8100.092.7
MCF-B100.095.0100.093.7
DMT-B100.095.2100.093.8
Etflow100.095.4100.092.8
Etflow + Refiner100.095.8100.094.4
Reviewer_TFf7OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a refinement framework to address training-inference mismatch generated by existing flow-based models in molecular conformation sampling settings for small molecules. The refiner is trained on Gaussian-perturbed reference conformers and learns to correct generated structures without explicitly estimating their residual errors. Experiments on GEOM-QM9 and GEOM-Drugs show consistent improvements across strong denoising-based generators, including MCF, ET-Flow, and DMT. The method improves conformer quality while preserving diversity and reducing the total number of sampling steps.

Critical strengths
  1. I think one of the main strengths of this paper is that a relatively small refiner can improve conformers from several strong generators across different datasets. The improvements are encouraging, and I appreciate that the authors also compare against baselines with more sampling steps. This makes the results more convincing and suggests that the gains are not simply due to insufficient baseline computation. The ability to reuse the same refiner across different generators, including RDKit, is also particularly appealing from a practical perspective.

  2. The comparisons with alternative refiners and MMFF, along with the runtime and sensitivity analyses, help build confidence in the approach. I especially like that the paper examines cases where refinement makes conformers worse rather than focusing only on improvements. The additional ablations and clarifications are also helpful. Overall, I think the experimental evaluation provides good support for the practical value of the method.

Critical weaknesses
  1. Novelty and Refinement Mechanism: I find the refinement approach promising, but I am not fully convinced that its main insight is clearly distinguished from standard denoising methods. The proposed flow is closely related to score-based probability-flow dynamics, and I think this connection should be better clarified (and acknowledged). I am also not entirely convinced by the self-calibration argument, since matching the initial noise scale does not guarantee that the evolving distribution aligns with the training distribution. The differences between the stated training objective and the implementation also need clarification. While the empirical results are encouraging, I would appreciate a comparison with simpler denoising baselines and an analysis of complete refinement trajectories to better understand what makes this approach effective.

  2. Baseline Comparisons and Practical Claims: I appreciate the extensive experiments, but I have some concerns about the strength of the conclusions. The reproduced ET-Flow results appear weaker than the published baseline, and I would like to understand why. The DMT results are more convincing, but the broader claims about physical quality and diversity seem less established. In particular, chemical properties are evaluated after additional optimization, while reference coverage alone does not demonstrate preservation of conformational basins. I would also like to see efficiency comparisons based on end-to-end runtime, including training costs where relevant, rather than sampling steps alone.

Questions for authors
  1. Mechanism and Contribution: Could the authors compare the proposed refiner with a simpler denoising model and continued upstream training under matched training budgets? I would be quite interested in how these methods handle different types of conformer errors, starting times, and already-correct inputs. This would help understand whether the proposed approach offers a distinct advantage and how well the self-calibration explanation reflects its actual behaviour.

  2. Robustness of Training: Could the authors further investigate how sensitive the method is to the time cutoff, noise settings, and choice of loss? I appreciate the existing stochastic-term ablation, but I think testing these additional choices across multiple seeds would help establish whether the method is robust to differences between the theoretical objective and its implementation.

  3. Physical Quality and Diversity: Could the authors evaluate the physical quality of the refined conformers before additional optimization, including energy, molecular geometry, and chirality? I would also like to see whether refinement preserves distinct conformational families, particularly for more flexible molecules.

  4. Baseline Fidelity and Computational Cost: Could the authors reproduce the published ET-Flow settings and apply refinement to those same outputs? I think this would help resolve the baseline discrepancies and make the improvements easier to interpret. I would also appreciate paired uncertainty estimates and a comparison of end-to-end runtime, memory usage, and training costs.

6 comments

W1/Q1 Novelty and Mechanism.

We thank the reviewer for this suggestion. We cite the related work(Song et al., 2021) in the introduction. Our perturbation, x0=x1+σϵx_0=x_1+\sigma\epsilon, shares the additive Gaussian form used in VE-SDEs, but differs in its purpose and inference initialization. Standard VE-SDE generation uses a sufficiently large noise scale to reach a noise-dominated distribution, enabling approximately Gaussian initialization. Our perturbations instead model positional errors while retaining useful molecular structure. At inference, we start directly from an upstream-generated conformer, without knowing the effective flow time corresponding to its residual errors. Also, unlike SDEdit, our method does not add noise before refinement. Appendix F (revised during rebuttal) compares our method with two SDEdit-style baselines. SDEdit (ET-Flow) and SDEdit (Refiner) add noise to completed upstream outputs and then denoise using ET-Flow and our trained Refiner, respectively.

MethodStepsAMR-R Mean ↓AMR-R Median ↓AMR-P Mean ↓AMR-P Median ↓
ET-Flow500.0830.0310.1140.053
ET-Flow + SD-XL-style refiner20+200.0830.0290.1130.051
ET-Flow + SDEdit (ET-Flow)20+200.0790.0290.1160.051
ET-Flow + SDEdit (Refiner)20+200.0800.0300.1150.048
ET-Flow + Ours20+200.0760.0280.1030.031

Our method achieves the best performance across all metrics. Mechanism. We revised Section 3.3 to analyze geometric representation coverage under time mismatch. Although two atoms have six coordinate components, their pairwise distance is invariant to global translation and rotation, and many coordinate configurations share the same distance. Thus, substantial coordinate perturbations can still produce familiar local geometric features without requiring all six coordinates to remain close to their original values. Appendix O quantifies this effect through Monte Carlo experiments. For the error interval ±0.1\pm0.1 in the respective units, the hit probabilities are 5.63% for atom-pair distances and 25.73% for three-atom angles, compared with 0.0044% for simultaneously matching all six RGB channels of two pixels, under the specified perturbation settings.

W2a/Q4a ET-Flow baseline fidelity.

We thank the reviewer for this suggestion. We used the official public implementation and evaluated the released pretrained checkpoints with their corresponding configurations, selecting the best-performing checkpoint among those evaluated.

The maintainer explicitly states in the ET-Flow official repository:

I have changed some parts of the data preprocessing scripts to make it more efficient. However, these changes might mean that the configs might not lead to the same results as the one reported in the paper. I am working on reproducing the results with the new preprocessed data format. Thanks for your patience.

Despite our efforts using the available public implementation, checkpoints, and configurations, we could not fully recover the published performance. The preprocessing changes acknowledged by the maintainer may contribute to this discrepancy.

For our paired evaluation, we apply the Refiner directly to the same saved ET-Flow conformers evaluated before refinement, using identical reference conformers and evaluation procedures. Thus, the reported paired improvements measure the effect of refinement on the same upstream outputs.

W2b/Q3 Chemical Properties before xTB and conformational basin

We thank the reviewer for this suggestion. To separate the effect of refinement from subsequent geometry optimization, we evaluate chemical properties, molecular geometry, chirality, and conformational diversity directly on the generated and refined coordinates,

1. Chemical properties before xTB relax

We use GFN2-xTB single-point calculations to evaluate energy (EE), dipole moment (μ\mu), and HOMO–LUMO gap (Δϵ\Delta\epsilon). For each molecule, we compute the absolute difference between the generated and reference ensemble properties, and report the median error across molecules. We report both uniform and Boltzmann-weighted ensemble averages, together with the absolute difference between the minimum energies of the two ensembles (Emin⁡E_{\min}).

Median errors with uniform weighting, before xTB relax. Units: kcal/mol for EE, Δϵ\Delta\epsilon, and Emin⁡E_{\min}; D for μ\mu. Lower is better.

MethodEE ↓μ\mu ↓Δϵ\Delta\epsilon ↓Emin⁡E_{\min} ↓
MCF-L2981.98002.437517.1721145.2360
ET-Flow7.68540.57731.78491.8387
ET-Flow + Refiner7.36090.47881.41491.0336

Median errors with Boltzmann weighting at 298.15 K, before xTB relax optimization.

MethodEE ↓μ\mu ↓Δϵ\Delta\epsilon ↓Emin⁡E_{\min} ↓
MCF-L144.78901.29264.1674145.2360
ET-Flow2.26490.39951.32721.8387
ET-Flow + Refiner1.18380.39381.11411.0336

The Emin⁡E_{\min} errors are identical in the two tables because the minimum does not depend on ensemble weighting. Before xTB relaxation, our refiner consistently reduces all reported chemical-property errors relative to ET-Flow under both uniform and Boltzmann weighting, indicating improvements not from xtb optimization.

2. Geometry, chirality, and conformational diversity

To assess whether refinement preserves conformational diversity, we assign each ET-Flow conformer to its nearest ground-truth reference before refinement and keep this reference fixed for the paired RMSD evaluation. Thus, the reported improvement reflects movement toward the originally assigned reference rather than rematching to a different reference after refinement. We additionally measure how often the original nearest-reference assignment is retained, providing an indication of whether refinement preserves the represented conformational modes.

All molecules

MetricET-FlowET-Flow + Refiner
RMSD to the fixed reference conformer (Å) ↓0.55840.5301
Chiral mismatch rate (%) ↓10.210510.2038

Flexible molecules (≥8 rotatable bonds)

MetricET-FlowET-Flow + Refiner
RMSD to the fixed reference conformer (Å) ↓0.76100.7323
Chiral mismatch rate (%) ↓13.800813.7995

The Refiner brings conformers closer to their originally assigned references in both groups, while slightly reducing chiral mismatch rates.

W2c\Q4b Training Cost and Performance under the Same Training Budget We thank the reviewer for this suggestion, which helps ensure a fair comparison. The Refiner was trained before the rebuttal period. Given our limited computational budget during rebuttal, we conducted a separate five-epoch timing run, measuring approximately 500 seconds per epoch on four NVIDIA A100 GPUs. This yields an estimated wall-clock time of 138.8 GPU hours for 250 epochs, comparable to the cost of continuing ET-Flow training for 250 additional epochs (140 GPU hours). GPU memory usage exceeded 70 GB per GPU and varied across batches with molecular sizes and graph structures. To examine whether the improvements arise simply from additional training, we continued training the pretrained ET-Flow model for another 250 epochs and compared it with ET-Flow + Refiner, which uses the same additional training budget of 250 epochs. AMR is reported in Å; lower is better. Relative reductions compare the 20 + 20-step Refiner pipeline with the 50-step continued-training baseline.

QM9

MethodInference stepsAMR-R mean ↓AMR-R median ↓AMR-P mean ↓AMR-P median ↓
ET-Flow (+250 epochs)100.0830.0340.1280.064
ET-Flow (+250 epochs)200.0800.0310.1210.055
ET-Flow (+250 epochs)500.0800.0300.1180.050
ET-Flow + Refiner10 + 100.0750.0270.1060.037
ET-Flow + Refiner20 + 200.0760.0280.1030.031
Relative reduction vs. 50-step baseline—5.00%6.67%12.71%38.00%

DRUGS

MethodInference stepsAMR-R mean ↓AMR-R median ↓AMR-P mean ↓AMR-P median ↓
ET-Flow (+250 epochs)100.4660.4370.5670.504
ET-Flow (+250 epochs)200.4640.4340.5530.488
ET-Flow (+250 epochs)500.4550.4260.5470.483
ET-Flow + Refiner10 + 100.4450.4110.5470.477
ET-Flow + Refiner20 + 200.4390.4100.5300.467
Relative reduction vs. 50-step baseline—3.52%3.76%3.11%3.31%

With the same additional training epochs, the Refiner achieves lower AMR across all four metrics on both datasets using 40 total inference steps versus 50, supporting its benefit beyond simply extending upstream training.

Reviewer_Huq7OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper deals with the problem of sequential error in flow matching, specifically for conformer generation task. The authors train an additional refiner network, using the pretrained ET-Flow network and finetuning it. The training data for it is created by adding Gaussian noise on reference conformations from the dataset. The authors argue that the resulting refinement trajectory can correct residual geometric errors without the need for estimating their scale. Experiments on GEOM-QM9 and GEOM-Drugs show improvements in geometric quality with fewer total sampling steps.

Critical strengths

1-) The experiments show meaningful improvements in some of the given metrics. The additional increased step control experiment also provides useful evidence that the gains are not simply explained by performing more sampling updates. In the end, the pipeline uses 20 generation steps and 20 refinement steps, compared with 50 steps for the main generator-only baselines.

2-) As my understanding, this refiner network is reusable for different models in the same task, which is a practical design choice. The paired displacement analysis, recall results, and per-conformer improvement and downgrade statistics also give a more informative picture of the method’s behavior than aggregate precision improvements alone.

Critical weaknesses

1-) A really important question for the paper is whether these improvements are worth the additional training cost. I think the paper does not provide enough information to answer this question. Section 4.1 specifies pretrained initialization and a learning rate change, but not which parameters are updated, the number of optimization updates, or GPU-hours. Since some improvements over ET-Flow are modest, these details matter for assessing whether the gains justify the additional training and deployment complexity. The sampling-step comparisons are useful, but they do not resolve this training-cost question.

2-) The comparison with other possible refinement approaches is still not fully convincing to me. The SDXL-style experiment is useful, however that baseline takes an intermediate generator state, while the proposed method takes a completed conformer. It is unclear how much improvement the proposed refinement formulation provides compared to using an existing pretrained model to refine the same completed conformers, for example by adding noise and denoising again. A comparison under a similar sampling budget would help establish why we should use this method over such an alternative.

Questions for authors

1-) Training-cost clarification: Which parameters are updated during refiner fine-tuning? Does the statement that other hyperparameters follow ET-Flow also include the training duration? So in the end, do you finetune the whole second network, and how much additional training does this require compared to ET-Flow? Can the authors report the number of optimization updates or epochs, hardware, and additional GPU-hours, separately from the pretraining used to initialize the refiner?

2-) Can the authors compare with adding noise to the same completed conformers and denoising them using the pretrained ET-Flow model, without training a separate refiner? This should follow ET-Flow's own training noise construction, with the restart level selected on validation data and a comparable sampling budget. It would help clarify what we gain from the proposed refinement formulation.

3-) Figure 4 shows that the velocity magnitudes stay similar when the time input is randomized, but it is unclear how this affects the final conformer quality. Can the authors evaluate the same trained refiner with fixed or randomized time conditioning, using the same starting conformers and sampling budget? This would help clarify how important the proposed time conditioning is for the improvements, beyond keeping the updates small.

I would be willing to raise my score if the authors clarify the computational trade-off and provide convincing evidence addressing the refinement comparison.

5 comments

W1/Q1 Comparison with continued ET-Flow training.

Thank you for raising this question. To examine whether the improvements arise simply from additional training, we continued training the pretrained ET-Flow model for another 250 epochs . We compare this baseline with ET-Flow + Refiner below. AMR is reported in Å; lower is better.

QM9

MethodInference stepsAMR-R mean ↓AMR-R median ↓AMR-P mean ↓AMR-P median ↓
ET-Flow (+250 epochs)100.0830.0340.1280.064
ET-Flow (+250 epochs)200.0800.0310.1210.055
ET-Flow (+250 epochs)500.0800.0300.1180.050
ET-Flow + Refiner10 + 100.0750.0270.1060.037
ET-Flow + Refiner20 + 200.0760.0280.1030.031
Relative error reduction vs. 50-step baseline—5.00%6.67%12.71%38.00%

DRUGS

MethodInference stepsAMR-R mean ↓AMR-R median ↓AMR-P mean ↓AMR-P median ↓
ET-Flow (+250 epochs)100.4660.4370.5670.504
ET-Flow (+250 epochs)200.4640.4340.5530.488
ET-Flow (+250 epochs)500.4550.4260.5470.483
ET-Flow + Refiner10 + 100.4450.4110.5470.477
ET-Flow + Refiner20 + 200.4390.4100.5300.467
Relative error reduction vs. 50-step baseline—3.52%3.76%3.11%3.31%

On QM9, the Refiner improves all four AMR metrics even compared with ET-Flow trained for an additional 250 epochs and sampled with 50 steps. These improvements also hold on DRUGS, showing that continued training alone does not recover the gains achieved by refinement.

Training cost: We use the ET-Flow architecture and initialize the refiner from the pretrained ET-Flow checkpoint, fine-tuning all model parameters without freezing any parameters. On GEOM-DRUGS, continuing ET-Flow training for 250 additional epochs spent 140 A100 GPU-hours. The Refiner had already been trained before the rebuttal period. Given our limited computational budget during rebuttal, we conducted a separate five-epoch timing run, measuring approximately 500 seconds per epoch on four NVIDIA A100 GPUs (138.8 A100 GPU-hours)

W2/Q2 Comparison with noise-and-denoise refinement (SDEdit style). We thank the reviewer for this suggestion. We agree that this comparison is important for evaluating the effectiveness of our refinement approach. We evaluate two noise-and-denoise configurations applied to completed conformers x^1\hat{x}_1, using a fixed restart time of t=0.5t=0.5:

  1. ET-Flow-based restart: We construct x0.5=0.5x0+0.5x^1x_{0.5}=0.5x_0+0.5\hat{x}_1, where x0x_0 is sampled from the ET-Flow prior, and denoise using the pretrained ET-Flow model without training a separate refiner.
  2. Refiner-based restart: We construct x0.5=x^1+0.5σϵx_{0.5}=\hat{x}_1+0.5\sigma\epsilon, where ϵ∼N(0,I)\epsilon\sim\mathcal{N}(0,I), and denoise using the trained refiner. Both configurations use 20 generation steps followed by 20 denoising steps from t=0.5t=0.5 to t=1t=1, matching the total sampling budget of our 20+20-step method.
MethodStepsAMR-R Mean ↓AMR-R Median ↓AMR-P Mean ↓AMR-P Median ↓
ET-Flow500.0830.0310.1140.053
ET-Flow + SD-XL-style refiner20+200.0830.0290.1130.051
ET-Flow + SDEdit (ET-Flow)20+200.0790.0290.1160.051
ET-Flow + SDEdit (Refiner)20+200.0800.0300.1150.048
ET-Flow + Ours20+200.0760.0280.1030.031

At this restart level, both noise-and-denoise configurations provide only modest improvements. Our proposed refinement achieves lower AMR across all four metrics under the same sampling budget. In particular, it achieves an AMR-P mean of 0.103, compared with 0.116 and 0.115 for the two noise-and-denoise configurations.

Q3 Effect of randomized time conditioning on final conformer quality. We thank the reviewer for this suggestion. We agree that evaluating final conformer quality is important for understanding the role of time conditioning beyond controlling update magnitudes. We compare the same trained refiner under the regular schedule and randomized time conditioning, t∼U(0,1)t\sim\mathcal{U}(0,1), using the same starting conformers and the same budget of 20 refinement steps.

MethodAMR-R mean ↓AMR-R median ↓AMR-P mean ↓AMR-P median ↓
ET-Flow (20)0.0800.0310.1210.055
ET-Flow + Refiner (20+20, randomized time)0.0790.0310.1120.043
ET-Flow + Refiner (20+20, regular schedule)0.0760.0280.1030.031

Randomized time conditioning still improves precision, but the regular schedule achieves better performance across all four metrics. These results are consistent with Figures 4 and 5: small predicted velocities help limit error growth under time mismatch, while more compatible state–time conditions allow effective error correction. Randomized conditioning can still produce useful updates, but may make this compatibility less consistent, explaining the smaller overall improvement compared with the regular schedule.

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

This paper proposes a flow-matching-based refinement framework for improving the quality of molecular conformers generated by existing generative models. The key idea is to train the refiner using data-centered perturbations around reference conformers, allowing the model to learn how to correct geometric errors while preserving the overall molecular structure. The experiments show consistent improvements in several conformer-generation metrics, including coverage and average minimum RMSD.

Critical strengths

Instead of learning a direct mapping from random noise to molecular conformations, the proposed method constructs perturbations around reference conformers. This formulation is intuitively suitable for local geometric refinement, as it allows the model to focus on correcting deviations from valid molecular structures.

Beyond the main conformer-generation metrics, the authors investigate bond-length, bond-angle, and torsion-angle errors, as well as the effects of perturbation scales and sampling steps.

Critical weaknesses

The derivation of t^* establishes a correspondence between the initial error scale of the input conformer and the perturbation scale used during training. However, the error distribution of the input conformer is expected to evolve throughout the refinement process. Therefore, matching the initial error scale to the training perturbation scale at a particular flow time does not necessarily guarantee that the predicted velocity field provides the correct refinement direction at that time. The authors should further clarify the theoretical justification of this assumption.

Although the proposed refinement strategy achieves consistent improvements when applied to several existing conformer generation models, its effectiveness has not been sufficiently compared with other alternative refinement strategies. For example, staged denoising approaches or other post-generation optimization methods may also improve the quality of generated conformers.

Questions for authors

See weaknesses.

3 comments

W1 Explanation of the mechanism

We thank the reviewer for pointing out the distinction between matching perturbation scales and obtaining an effective refinement direction. The derivation of t∗t^* establishes a correspondence at the level of the initial error scale. To explain why refinement can remain effective as the error distribution evolves, we examine how the model’s geometric representations can remain within the range encountered during training despite a mismatch in perturbation scale.

We have revised Section 3.3 to clarify this scope and discuss robustness to time mismatch at the representation level. With the molecular graph fixed, we examine the local geometric features used by the model. Under the Gaussian approximation, a mismatch in perturbation scale does not necessarily place pairwise geometry outside the range encountered during training. For example, with independent per-atom perturbations of $1$Å, a C–C bond of length $1.54$Å has approximately a 5.11% probability of retaining its distance within an error interval of [−0.1,0.1][-0.1,0.1]Å. Recurring atom-type combinations and repeated perturbations therefore provide opportunities to learn similar local geometry despite the scale mismatch.

Although the coordinates of two atoms also form a six-dimensional input, their pairwise distance is a scalar geometric feature that can remain similar across different coordinate configurations. In comparison, for six independent components: such as the six RGB channel values of two pixels in our experimental setting, the probability that all six errors simultaneously fall within a small interval is close to zero.

Appendix O quantifies this difference through a Monte Carlo comparison of atom-pair distances, three-atom angles, and two pixels. For the error interval [-0.1,0.1] in the respective units, the mean hit probabilities are 5.63%, 25.73%, and 0.0044%, respectively. The geometric features thus satisfy the interval criterion approximately 1,200 and 5,800 times more frequently under the specified experimental settings. These results support a representation-level explanation for robustness to time mismatch, rather than a guarantee that initial scale matching ensures correct refinement directions throughout the trajectory.

W2 Comparison with SXDL refinement and post-generation:

We thank the reviewer for this suggestion. We compare our method with both staged denoising strategies and classical post-generation optimization.

1. Staged denoising and noise–denoise baselines (Appendix F, Table 8). We evaluate an SD-XL-style refiner that takes over the upstream generation trajectory at t=0.8, and two SDEdit-style baselines that add noise to generated conformers before denoising with either ET-Flow or our trained Refiner. All refinement pipelines use 20 upstream steps and 20 second-stage steps.

MethodStepsAMR-R Mean ↓AMR-R Median ↓AMR-P Mean ↓AMR-P Median ↓
ET-Flow500.0830.0310.1140.053
ET-Flow + SD-XL-style refiner20+200.0830.0290.1130.051
ET-Flow + SDEdit (ET-Flow)20+200.0790.0290.1160.051
ET-Flow + SDEdit (Refiner)20+200.0800.0300.1150.048
ET-Flow + Ours20+200.0760.0280.1030.031

Our method achieves the best performance across all reported recall and precision AMR metrics. These comparisons support the benefit of our refinement design beyond introducing a second-stage model or applying an additional noise–denoise procedure.

2. Classical post-generation optimization (Appendix G, Table 9). In a separate experiment, we compare MMFF optimization and our Refiner on the same ET-Flow outputs on GEOM-QM9.

MethodAMR-R Mean ↓AMR-R Median ↓AMR-P Mean ↓AMR-P Median ↓
ET-Flow0.0820.0340.1260.066
ET-Flow + MMFF0.1820.1350.1870.147
ET-Flow + Ours0.0760.0270.1070.038

MMFF optimization degrades all reported conformer recovery metrics in this experiment, whereas our Refiner improves them. This comparison concerns agreement with the reference conformers; MMFF optimizes a force-field energy objective, which need not align with these metrics.

Thank you again for your thoughtful feedback. We have addressed each of the concerns you raised in our point-by-point response, and we look forward to hearing your thoughts. With less than 24 hours left in the discussion period, we would be grateful if you could let us know soon whether any concerns still affect your assessment of the paper, so that we have time to address them.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.