acceptodds
Withdrawn from ICLR 2027

Molecular Conformer Sampling Beyond Generations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Sampling the low-energy conformational states available to a molecule is a fundamental problem in computational chemistry and drug discovery, since molecular properties and downstream design decisions depend on the relevant three-dimensional geometries rather than on a single structure. Recent diffusion and flow-based generative models can efficiently sample conformer candidates from molecular graphs, but under a fixed output budget their samples do not cover the full set of relevant conformational states of a molecule, while practical benchmarks and downstream applications require a small ensemble that is both diverse and physically plausible. We present a post-generation selection framework that turns candidate pools from frozen, pre-trained generators into compact conformer ensembles without any retraining or fine-tuning of the underlying generator. The framework offers two selectors that share the same candidate pool and the same learned reference-likelihood classifier: CovSelect, a coverage-maximizing selector that greedily reaches as many distinct reference states as the pool allows, and RefSelect, a tunable selector that trades some of that coverage for higher precision and, at its balanced setting, matches or exceeds strong generative baselines on both measures at once. On the full GEOM-DRUGS test set, CovSelect reaches 91.33% mean recall (98.36% median), outperforming every raw generator we compare against, and RefSelect at its balanced setting exceeds the strongest raw baseline, DMT-L, on both recall and precision. Finally, across several out of distribution benchmarks (QMugs, pepconf, and Platinum), CovSelect substantially outperforms all raw generative baselines in recall and precision.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular conformer generation, generative models, diffusion models, molecular machine learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_1SpHOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a post-generation selection framework for molecular conformer sampling that aims to build compact conformer ensembles with high coverage and precision under a fixed output budget. Instead of improving the generative models themselves, the authors focus on selecting representative conformers from an oversampled candidate pool produced by frozen, pretrained generators.

The proposed framework consists of three main components: (1) a reference-likelihood classifier that estimates the probability of each candidate matching a reference conformer, (2) a latent reference-field inference method to estimate the distribution of unseen reference conformers, and (3) a redundancy-aware greedy selection strategy that balances coverage and precision. Based on this framework, the authors introduce CovSelect, which prioritizes coverage, and RefSelect, which provides a tunable trade-off between recall and precision.

Experiments on GEOM-DRUGS demonstrate improved conformer coverage compared with existing generative baselines, while RefSelect achieves a better recall–precision balance. Additional evaluations on out-of-distribution datasets, independently generated reference conformers, and different output budgets further demonstrate the effectiveness and generalizability of the proposed selection framework.

Critical strengths
  1. The paper is well-structured and clearly written. The proposed framework is presented in a logical and coherent manner.

  2. The observation that existing generative models may already produce diverse conformers but struggle to retain representative structures under a fixed output budget provides a compelling motivation for the proposed selection framework. This perspective highlights the importance of post-generation selection as a complementary direction to improving generative models.

  3. The paper provides a relatively comprehensive experimental evaluation across multiple datasets, including GEOM-DRUGS and several out-of-distribution benchmarks.

Critical weaknesses
  1. Although the proposed method demonstrates improvements under a fixed output budget, the experimental comparisons do not control for the actual number of generated conformers. Specifically, CovSelect and RefSelect select 2n conformers from an oversampled pool of approximately 16n candidates, whereas the raw generative baselines directly produce only 2n conformers. This discrepancy raises concerns about the fairness of the comparisons, as the observed improvements may be partially attributed to the substantially larger candidate pool rather than the effectiveness of the proposed selection strategy. Consequently, the current experiments do not fully establish the method's advantages under comparable computational budgets.

  2. The paper lacks a comprehensive analysis of the computational overhead introduced by the proposed framework. Beyond generating a larger candidate pool, the method requires feature extraction using pretrained interatomic potentials, pairwise RMSD computation, reference-field inference, and greedy selection. However, the paper does not report the corresponding runtime, computational costs, or scalability. Without these measurements, it is difficult to assess the practical benefits of introducing an additional selection stage, particularly when compared with directly generating more conformers using existing models.

  3. The methodological novelty appears somewhat limited, as the proposed framework primarily combines existing techniques, including gradient-boosted classification, Richardson–Lucy deconvolution, and facility-location-based greedy selection. More importantly, several key modeling choices are not sufficiently justified. For instance, the reference-field inference relies on a Poisson approximation, but the paper provides limited theoretical or empirical evidence supporting this assumption. Similarly, the choice of the coverage kernel and its influence on the inferred reference distribution are not thoroughly investigated. Given the availability of alternative probabilistic models and kernel functions, a more systematic analysis is needed to establish the necessity and robustness of these design choices.

Questions for authors
  1. Could the authors clarify how much of the reported improvement comes from the proposed selection strategy rather than the substantially larger candidate pool? Would the proposed method still outperform the baselines under matched generation budgets?

  2. What are the computational and runtime costs of the proposed framework, including candidate generation, feature extraction, reference-field inference, and greedy selection? Does the improvement in conformer coverage justify the additional computational overhead compared with directly generating more conformers?

  3. What motivates the use of a Poisson approximation for modeling the reference distribution? How sensitive is the method to this assumption and the choice of the coverage kernel? Would alternative probabilistic formulations or kernel functions yield similar results?

I will increase my scores if the author provides enough evidence.

Reviewer_eWssOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to select a size-limited conformer ensemble that balances coverage and precision from a large pool produced by existing generators. The method combines a reference-match classifier, a coverage kernel, latent reference-field estimation, and redundancy correction, and constructs the final ensemble through greedy selection without retraining the generators. CovSelect prioritizes coverage, while RefSelect provides a tunable coverage–precision tradeoff. Experiments cover GEOM-DRUGS, three out-of-distribution datasets, four molecules with independently generated CREST references, and different return budgets.

Critical strengths
  1. Separating candidate generation from fixed-budget ensemble selection addresses a meaningful problem. The method can use outputs from frozen generators without retraining them.

  2. The evaluation is fairly comprehensive, including random and geometric selection baselines on the same candidate pool, out-of-distribution tests, and return-budget experiments. These comparisons help characterize the effects and tradeoffs of selection.

Critical weaknesses
  1. The independent contribution of the core components remains unclear. The existing controls show that the full method improves over simple ranking or geometric selection, but do not sufficiently establish the additional benefit of reference-field deconvolution over directly using classifier scores in a coverage-based selector.

  2. The out-of-distribution configuration is insufficiently specified. Section 6.3 states that CovSelect uses no GEOM-trained prior, unlike the main configuration, but does not clearly explain how reference-field estimation and selection operate after removing this information. It is therefore difficult to determine which components’ generalization is demonstrated by these results.

Questions for authors
  1. Is there a control that directly uses classifier scores to construct the coverage objective without reference-field deconvolution, to isolate the contribution of the deconvolution step?

  2. Which components are retained in the out-of-distribution version of CovSelect? How are the coverage kernel, reference field, and redundancy correction defined without the GEOM-trained prior?

  3. How were the RefSelect weight and other parameters chosen? Section 8 mentions that some operating points were explored on the reported cohort; could you distinguish settings chosen on an independent validation set from those explored on the test cohort?

Reviewer_tMKfOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposed two molecular conformer selection algorithms, RefSelect and CovSelect, to select conformers from a larger pool, with the goal to improve conformer landscape coverage. The paper applies the proposed selection algorithm to a pool of generated conformers and show improvement on metrics such as recall and precision.

Critical strengths
  1. Interesting idea of training a classifier based on MLIP and pairwise RMSD to predict which conformers match a reference.
Critical weaknesses
  1. Not a fair comparison. The focus of the paper is the conformer selection algorithm, so it should be compared with other selection algorithms using the exact same pool. In most tables, the authors instead compare their selection algorithm with an much larger pool against conformer generative models with a much smaller generation budget. In the only table that's fair comparison, Table 2, the k-medoids is already not too far away from RefSelect.

  2. Unclear description of the experiment procedure, especially how the candidate pool is generated and how large it is in Table 3 and Table 4. See questions for detail.

  3. The computational cost is not reported. The framework needs 8 times more generated conformers than the baselines. It also needs MLIP-based model inference and pairwise RMSD over the whole pool, but no wall-clock time profiling is given for any stage.

Questions for authors
  1. My most important question is: can the authors add the selection algorithm baselines from Table 2 (uniform random, Farthest-point, k-medoids) with same-sized pool to Table 3, Table 4 and Table 5? Otherwise, it is hard to tell whether the improved recall/precision comes from the selection algorithm change or simply from the larger pool.

  2. How is the candidate pool generated? Section 4.1 mentioned about 8x ET-Flow plus 8x PG-TD. Please report the inference configuration of the two models. Please also include the pool size for Table 3 and Table 4, which I cannot find in the paper.

  3. A follow-up of the question above: why the authors choose to use ET-Flow and PG-TD to construct the pool, while DMT-L is another strong model that's publicly available and benchmarked? and why the author specifically choose half and half ratio between the two models selected?

  4. Can the authors run baseline selection algorithms like uniform random, Farthest-point, k-medoids, and RefSelect/CovSelect on a 16n pool generated by DMT-L alone? Will RefSelect/CovSelect still win with a different and maybe higher quality pool?

  5. Can the authors report the wall-clock time for generating the pool, the MLIP inference for scoring, the pairwise RMSD calculation and the selection, compared with the baseline select algorithms?

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

Related papers

Discussion (0)

Sign in to comment.