Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

EquiFlow: Equivariant Conditional Flow Matching with Optimal Transport for 3D Molecular Conformation Prediction

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read EquiFlow predicts 3D molecular conformations more accurately than prior models by pairing conditional flow matching with optimal transport.

desk verdict Plausible new SOTA on GEOM-QM9 from OT-CFM, but missing equivariance proof and missing code keep the result conditional. read the letter →

arxiv 2412.11082 v1 pith:GNONEKUT submitted 2024-12-15 cs.LG physics.chem-phq-bio.BM

Qingwen Tian , Yuxin Xu , Yixuan Yang , Zhen Wang , Ziqi Liu , Pengju Yan , Xiaolin Li This is my paper · ORCID
classification cs.LGphysics.chem-phq-bio.BM
keywords molecularconformationpredictionconditionalflowmatchingoptimaltransportequivariantgraphneuralnetworkSE(3)equivarianceEquiformerQM9GEOM-QM9
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes EquiFlow, a model that predicts 3D molecular conformations by learning a conditional flow from Gaussian noise on atomic coordinates to real conformations, with the flow's target vector field chosen by optimal transport. It claims this is the first application of conditional flow matching to molecular conformation prediction, and that the simulation-free training plus an ODE solver makes it faster to train and sample than diffusion models. On the GEOM-QM9 benchmark, EquiFlow reports higher coverage (COV-R 95.9 vs 92.8) and lower matching error (MAT-R 0.130 Å vs 0.178 Å) than the previous best deep-learning method, and 0.17 Å RMSD on single-conformation QM9 prediction. A careful reader would care because molecular geometry governs binding and reactivity, and a fast, accurate predictor would be directly useful in drug and materials design.

What carries the argument

The load-bearing object is the equivariant conditional flow matching objective with optimal transport. A Gaussian point cloud $x_0$ is paired with a real conformation $x_1$ by solving a minibatch optimal transport problem with RMSD cost $C(x_{0i}, x_{1j}) = \sqrt{\frac{1}{K}\sum_k \|x_{0ik}-x_{1jk}\|^2}$ after zeroing center of mass and rotationally aligning the structures; the conditional target is the straight-line vector field $u_t = x_1 - x_0$ along $x_t = x_0 + t u_t$. A modified Equiformer, a Transformer-style equivariant graph neural network built on SO(3) irreducible representations, regresses this vector field from atom types, bond types, relative positions, and time, using a degree-1 (vector) prediction head. The paper's argument is that this pairing keeps the training target compatible with SE(3) equivariance while giving straighter, easier-to-learn probability paths than diffusion.

What would settle it

Train EquiFlow on GEOM-QM9 twice: once with conformations in their native orientations and once with every molecule randomly rotated before both training and evaluation. If the OT pairing is not rotation-covariant, the rotated run should show materially worse COV-R and MAT-R; additionally, with a fixed-orientation test, a random global rotation of a test molecule should not change the RMSD of the predicted conformation to the reference after re-alignment, and any such change would indicate broken equivariance.

Watch

Extended reading notes

Core claim

The central discovery claimed is that conditional flow matching with an optimal-transport pairing is a better training objective for 3D conformation prediction than the diffusion and torsion-space objectives used previously. EquiFlow pairs each sampled Gaussian coordinate set $x_0$ with a true conformation $x_1$ by solving a transport problem whose cost is the rotation-aligned RMSD between conformations, then regresses the straight-line velocity $u_t = x_1 - x_0$ with a modified Equiformer network that represents atomic positions as degree-1 irreducible vectors and takes atom types, bond types, and time as inputs. In the reported experiments this reaches COV-R 95.9 and MAT-R 0.130 Å on GEOM-QM9, with COV-P 91.8 and MAT-P 0.164 Å, and 0.17 Å RMSD on QM9 single-conformation prediction. The paper presents these numbers as evidence that combining optimal transport with equivariant flow matching improves both diversity and accuracy over prior state-of-the-art models.

Load-bearing premise

The method assumes that the optimal-transport pairing computed with rotation-aligned RMSD cost remains rotation- and translation-covariant, so the straight-line target field $u_t = x_1 - x_0$ can be learned by an SE(3)-equivariant network; the paper asserts this but does not prove it.

Editorial extensions

If this is right

  • Single-conformation prediction at 0.17 Å RMSD on QM9 implies the model is accurate enough for geometry-sensitive downstream tasks on small molecules, if the number reproduces under independent evaluation.
  • GEOM-QM9 COV-R of 95.9 and MAT-R of 0.130 Å would mean the generated ensembles cover essentially all reference conformations with lower error than torsion-based diffusion (92.8, 0.178 Å).
  • Because training is simulation-free and sampling uses an ODE, the same accuracy would come with faster training and inference than score-based diffusion; the paper states this speed advantage explicitly.
  • Using bond types and higher-degree equivariant features (up to degree 6) would show that chemistry-informed edge features and high-degree tensors help conformation prediction, not just property prediction.
  • The method is end-to-end for conformations given a 2D molecular graph, avoiding the precomputed rigid substructures required by torsion-based approaches.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the equivariance of the OT pairing holds, the same recipe should transfer to larger molecules and to conformer ensembles of protein fragments, since the architecture does not depend on molecule size; the paper only demonstrates up to QM9-sized heavy-atom counts.
  • The reported COV-P (91.8) is slightly below Torsional Diffusion (92.7), so the improvement is not uniform across all precision metrics; a practitioner interested in precision at the 0.5 Å threshold should look at per-molecule distributions before switching methods.
  • Because the model needs bond types, its applicability to conformer generation of molecules with unknown bond orders or reaction intermediates is limited unless bond assignment is provided.
  • A direct stress test would be to train EquiFlow on randomly rotated versus fixed-orientation conformations; if the Kabsch-based OT pairing is not truly rotation-covariant, the rotated training run should underperform.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. EquiFlow applies conditional flow matching with optimal transport (OT-CFM) to 3D molecular conformation prediction. The training procedure pairs Gaussian noise conformations with true conformations using an OT cost based on Kabsch-aligned RMSD, then regresses a modified EquiformerV2 network to the straight-line conditional vector field u_t = x1_j - x0_i. The paper reports state-of-the-art or competitive multi-conformation results on GEOM-QM9 (COV-R 95.9%, MAT-R 0.130 Å, MAT-P 0.164 Å) and a single-conformation QM9 RMSD of 0.17 Å, and argues that the ODE-based sampler is faster than SDE-based diffusion.

Significance. If the reported results are reproducible and the equivariance issue below is resolved, EquiFlow would be a solid contribution: it combines the efficient training of CFM with a high-degree equivariant backbone, and the use of a Kabsch-aligned RMSD cost in the OT coupling is a sensible adaptation to molecular symmetry. The empirical numbers on GEOM-QM9 are competitive with strong published baselines. However, the manuscript currently lacks a proof that the OT-CFM training objective is SE(3)-equivariant, reports no variance estimates or code, provides no runtime measurements despite speed being a central motivation, and presents the single-conformation claim without baselines. These gaps prevent a full verification of the central accuracy and efficiency claims.

major comments (4)
  1. [§3.2, Algorithm 1 lines 6-9] The paper asserts that Kabsch alignment ensures rotational invariance, but it never states the group action or proves that the OT pairing M and the target vector field u_t = x1_j - x0_i transform covariantly under SE(3). This is load-bearing: if the pairing or target is orientation-dependent, the equivariant network cannot fit a consistent objective and the reported accuracy numbers are not explained. The natural proof (simultaneously rotating both point clouds leaves the Kabsch cost invariant, so M is unchanged, and u_t rotates as R u_t) is not written down, and Algorithm 1 is ambiguous about whether the Kabsch-rotated x1 or the original x1 is used in Eq. (11). Please add a formal equivariance statement for the OT-CFM coupling and specify exactly which coordinates enter the target. Also reconcile the order of centering: §3.2 says Zero CoM is applied before computing the Kabsch cost, while Algorithm 1 centers only after OT.
  2. [Experiments, Table 1 and single-conformation paragraph] The GEOM-QM9 baseline numbers are directly taken from Jing et al. (2022), yet the manuscript modifies the dataset by grouping by SMILES and removing invalid conformations. Unless this filtered test set is provably identical to the one used by Torsional Diffusion, the comparison is not valid; the paper should state the overlap or rerun the baselines on the same filtered set. In addition, the headline single-conformation RMSD of 0.17 Å on QM9 is reported with no comparison table, no description of the exact protocol beyond the split, and no baseline numbers, so this central claim is currently unsupported.
  3. [Experimental Setups / Experimental Results] No error bars, seeds, or confidence intervals are provided for any EquiFlow result, and no code is released. Since EquiFlow's COV-P (91.8%) is below Torsional Diffusion's (92.7%) and several metric differences are small, it is impossible to assess whether the reported improvements are statistically significant. Please report results over multiple seeds with mean and standard deviation, and release code or at least a detailed configuration to make the numbers reproducible.
  4. [Abstract and Introduction] The stated motivation is avoiding slow training and enabling faster inference than diffusion models with SDEs, but no runtime measurements are reported for training or sampling, and no ODE-vs-SDE inference-time comparison is given. If speed is one of the two central claimed advantages, it must be measured; otherwise the speed claims in the abstract and introduction should be removed or substantially softened.
minor comments (5)
  1. [Figure 1 caption] The caption refers to 'ConfromFlow' instead of 'EquiFlow'.
  2. [Algorithm 1 / Training Procedure] The text says 'MES loss function' where it should say 'MSE loss function'.
  3. [Equations (12)-(15)] The COV-R and COV-P definitions use xp without an explicit existential quantifier or set-builder binding; as written, xp appears as a free variable. Please clarify that the inner condition is 'there exists xp in Sp' (or equivalently 'min over xp').
  4. [§3.2, Equation (7)] The notation C(x0_i, x1_j) does not indicate that x1 is Kabsch-rotated relative to x0 before the RMSD is computed; this is a source of the ambiguity noted in the major comment and should be made explicit in the equation or surrounding text.
  5. [Appendix A reference] The main text says the straight-line OT approach 'may pose issues in 3D molecular conformation prediction due to spatial symmetry (see details in Appendix A)', but Appendix A is a review of SE(3) equivariance and tensor products; it does not directly discuss why straight-line OT fails under molecular symmetry. Please connect the arguments more explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the central claim is an empirical benchmark result obtained from a training objective that is not constructed from the reported accuracy.

full rationale

The paper's central claim is that EquiFlow beats prior methods on QM9 and GEOM-QM9 conformation benchmarks (Table 1 and the 0.17 angstrom QM9 RMSD). The training objective is standard conditional flow matching: x0 is Gaussian noise, x1 is a training-set conformation, the pairing is chosen by an OT map with a Kabsch-aligned RMSD cost, and the network regresses u_t = x1_j - x0_i (Eq. 11, Algorithm 1). No fitted parameter is later relabeled as a prediction, and the evaluation uses held-out test conformations that are not used to construct the OT pairings. The cited flow-matching and architecture results (Lipman et al. 2023; Tong et al. 2024; Liao et al. 2024) are external prior-work results rather than self-citations, and the paper's modifications are architectural rather than definitional. The one substantive gap, the assertion that Kabsch alignment ensures rotational invariance of the OT coupling and target field (Section 'Equivariant Conditional Flow Matching with Optimal Transport', Eq. 7-8), is an omitted proof rather than a circular reduction: even if the target were not equivariant, the method would be miscalibrated, but its derivation would not reduce to its inputs. Therefore the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The free parameters are training hyperparameters and evaluation settings. The main unproven assumptions are the equivariance compatibility of the Kabsch-based OT target and the standard CFM sampling properties.

free parameters (4)
  • Maximum degree of irreps = 6
    Chosen to use higher-degree features; no ablation shows that degree 6 is responsible for the improvement.
  • CFM sigma = 0.0
    The variance of the conditional probability path is set to zero; this is a modeling choice that affects training and sampling.
  • Maximum conformations per molecule = 20
    The training procedure caps the number of conformations used per molecule at 20, with no sensitivity analysis.
  • Sample coefficient = 2
    Evaluation generates twice as many predicted conformations as true conformations, which affects COV and MAT scores.
assumptions (4)
  • standard math The OT-CFM loss has the same gradients as the flow matching loss and yields the correct marginal vector field.
    The paper relies on theorems from Lipman et al. and Tong et al. without reproving them.
  • domain assumption The modified Equiformer is SE(3)-equivariant and the Kabsch-aligned OT-CFM training target is compatible with that equivariance.
    The paper asserts this compatibility but does not prove it; it is load-bearing for the validity of the training objective.
  • domain assumption Sampling x0 from a zero-center-of-mass Gaussian and integrating the learned vector field with an ODE solver produces valid conformer samples.
    This is standard for flow matching, but the paper provides no diagnostics to verify that the learned distribution matches the target distribution.
  • domain assumption Baseline results taken from Jing et al. are directly comparable because the same dataset split and evaluation protocol are used.
    The paper states that it follows the Torsional Diffusion split, but does not provide per-method implementation checks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EquiFlow: Equivariant Conditional Flow Matching with Optimal Transport for 3D Molecular Conformation Prediction." pith.science (2026). https://pith.science/paper/GNONEKUT

@misc{pith2026241211082,
  author       = {Pith},
  title        = {Pith review of: EquiFlow: Equivariant Conditional Flow Matching with Optimal Transport for 3D Molecular Conformation Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GNONEKUT}},
  note         = {Machine review of arXiv:2412.11082}
}
read the original abstract

Molecular 3D conformations play a key role in determining how molecules interact with other molecules or protein surfaces. Recent deep learning advancements have improved conformation prediction, but slow training speeds and difficulties in utilizing high-degree features limit performance. We propose EquiFlow, an equivariant conditional flow matching model with optimal transport. EquiFlow uniquely applies conditional flow matching in molecular 3D conformation prediction, leveraging simulation-free training to address slow training speeds. It uses a modified Equiformer model to encode Cartesian molecular conformations along with their atomic and bond properties into higher-degree embeddings. Additionally, EquiFlow employs an ODE solver, providing faster inference speeds compared to diffusion models with SDEs. Experiments on the QM9 dataset show that EquiFlow predicts small molecule conformations more accurately than current state-of-the-art models.

Figures

Figures reproduced from arXiv: 2412.11082 by the authors.

Figure 1
Figure 1. The architecture of ConfromFlow. The architecture involves using fixed atom and bond types as conditions, followed [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The process of Equivariant OT-CFM. We perform [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Atom, Bond, and Time Embedding Blocks in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Samples from the single-conformation prediction on QM9 dataset. We selected 10 molecules, with SMILES [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Samples from the multi-conformation prediction on GEOM-QM9 dataset. We selected 5 conformations from 2 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Feynman-Kac-Flow: Inference Steering of Conditional Flow Matching to an Energy-Tilted Posterior

    cs.LG 2025-09 conditional novelty 6.0 of 10

    Feynman-Kac particle steering, previously diffusion-only, is derived for conditional flow matching and used to generate chirality-correct chemical transition states.

Reference graph

Works this paper leans on

43 extracted references · 23 canonical work pages · cited by 1 Pith paper

  1. [1]

    P.; Kornbluth, M.; Molinari, N.; Smidt, T

    Batzner, S.; Musaelian, A.; Sun, L.; Geiger, M.; Mailoa, J. P.; Kornbluth, M.; Molinari, N.; Smidt, T. E.; and Kozinsky, B. 2022. E (3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nature communications, 13(1): 2453

    work page 2022

  2. [2]

    T.; Rubanova, Y.; Bettencourt, J.; and Duvenaud, D

    Chen, R. T.; Rubanova, Y.; Bettencourt, J.; and Duvenaud, D. K. 2018. Neural ordinary differential equations. Advances in neural information processing systems, 31

    2018

  3. [3]

    Ganea, O.; Pattanaik, L.; Coley, C.; Barzilay, R.; Jensen, K.; Green, W.; and Jaakkola, T. 2021. Geomol: Torsional geometric generation of molecular 3d conformer ensembles. Advances in Neural Information Processing Systems, 34: 13757--13769

    work page 2021

  4. [4]

    u ller, K.-R.; and Sch \

    Gebauer, N. W.; Gastegger, M.; Hessmann, S. S.; M \"u ller, K.-R.; and Sch \"u tt, K. T. 2022. Inverse design of 3d molecular structures with conditional generative neural networks. Nature communications, 13(1): 973

    work page 2022

  5. [5]

    S.; Riley, P

    Gilmer, J.; Schoenholz, S. S.; Riley, P. F.; Vinyals, O.; and Dahl, G. E. 2017. Neural message passing for quantum chemistry. In International conference on machine learning, 1263--1272. PMLR

    2017

  6. [6]

    Grathwohl, W.; Chen, R. T. Q.; Bettencourt, J.; and Duvenaud, D. 2019. Ffjord: Free-form continuous dynamics for scalable reversible generative models. In International Conference on Learning Representations

    work page 2019

  7. [7]

    C.; Skillman, A

    Hawkins, P. C.; Skillman, A. G.; Warren, G. L.; Ellingson, B. A.; and Stahl, M. T. 2010. Conformer generation with OMEGA: algorithm and validation using high quality structures from the Protein Databank and Cambridge Structural Database. Journal of chemical information and modeling, 50(4): 572--584

    work page 2010

  8. [8]

    Ho, J.; Jain, A.; and Abbeel, P. 2020. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 6840--6851

    2020

Show all 43 references
  1. [9]

    G.; Vignac, C.; and Welling, M

    Hoogeboom, E.; Satorras, V. G.; Vignac, C.; and Welling, M. 2022. Equivariant diffusion for molecule generation in 3d. In International conference on machine learning, 8867--8887. PMLR

    2022

  2. [10]

    Irwin, R.; Tibo, A.; Janet, J.-P.; and Olsson, S. 2024. Efficient 3D Molecular Generation with Flow Matching and Scale Optimal Transport. arXiv preprint arXiv:2406.07266

    2024 arXiv

  3. [11]

    Jing, B.; Corso, G.; Chang, J.; Barzilay, R.; and Jaakkola, T. 2022. Torsional diffusion for molecular conformer generation. Advances in Neural Information Processing Systems, 35: 24240--24253

    2022

  4. [12]

    Jing, B.; Eismann, S.; Suriana, P.; Townshend, R. J. L.; and Dror, R. 2020. Learning from protein structure with geometric vector perceptrons. In International Conference on Learning Representations

    2020

  5. [13]

    Kabsch, W. 1976. A solution for the best rotation to relate two sets of vectors. Acta Crystallographica Section A: Crystal Physics, Diffraction, Theoretical and General Crystallography, 32(5): 922--923

    1976

  6. [14]

    Landrum, G.; et al. 2013. RDKit: A software suite for cheminformatics, computational chemistry, and predictive modeling. Greg Landrum, 8(31.10): 5281

    2013

  7. [15]

    Le, T.; No \'e , F.; and Clevert, D.-A. 2022. Equivariant graph attention networks for molecular property prediction. arXiv preprint arXiv:2202.09891

    2022 arXiv

  8. [16]

    S.; and Hashimoto, T

    Li, X.; Thickstun, J.; Gulrajani, I.; Liang, P. S.; and Hashimoto, T. B. 2022. Diffusion-lm improves controllable text generation. Advances in Neural Information Processing Systems, 35: 4328--4343

    2022

  9. [17]

    Liao, Y.-L.; and Smidt, T. 2023. Equiformer: Equivariant Graph Attention Transformer for 3D Atomistic Graphs. In The Eleventh International Conference on Learning Representations

    2023

  10. [18]

    M.; Das, A.; and Smidt, T

    Liao, Y.-L.; Wood, B. M.; Das, A.; and Smidt, T. 2024. EquiformerV2: Improved Equivariant Transformer for Scaling to Higher-Degree Representations. In The Twelfth International Conference on Learning Representations

    2024

  11. [19]

    Lipman, Y.; Chen, R. T. Q.; Ben-Hamu, H.; Nickel, M.; and Le, M. 2023. Flow Matching for Generative Modeling. In The Eleventh International Conference on Learning Representations

    2023

  12. [20]

    W.; Li, X.; and Ruthotto, L

    Onken, D.; Fung, S. W.; Li, X.; and Ruthotto, L. 2021. Ot-flow: Fast and accurate continuous normalizing flows via optimal transport. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, 9223--9232

    2021

  13. [21]

    J.; Mohamed, S.; and Lakshminarayanan, B

    Papamakarios, G.; Nalisnick, E.; Rezende, D. J.; Mohamed, S.; and Lakshminarayanan, B. 2021. Normalizing flows for probabilistic modeling and inference. Journal of Machine Learning Research, 22(57): 1--64

    2021

  14. [22]

    O.; Rupp, M.; and Von Lilienfeld, O

    Ramakrishnan, R.; Dral, P. O.; Rupp, M.; and Von Lilienfeld, O. A. 2014. Quantum chemistry structures and properties of 134 kilo molecules. Scientific data, 1(1): 1--7

    2014

  15. [23]

    Rezende, D.; and Mohamed, S. 2015. Variational inference with normalizing flows. In International conference on machine learning, 1530--1538. PMLR

    2015

  16. [24]

    Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; and Ommer, B. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 10684--10695

    2022

  17. [25]

    Sanchez-Gonzalez, A.; Godwin, J.; Pfaff, T.; Ying, R.; Leskovec, J.; and Battaglia, P. 2020. Learning to simulate complex physics with graph networks. In International conference on machine learning, 8459--8468. PMLR

    2020

  18. [26]

    G.; Hoogeboom, E.; and Welling, M

    Satorras, V. G.; Hoogeboom, E.; and Welling, M. 2021. E (n) equivariant graph neural networks. In International conference on machine learning, 9323--9332. PMLR

    2021

  19. [27]

    Shi, C.; Luo, S.; Xu, M.; and Tang, J. 2021. Learning gradient fields for molecular conformation generation. In International conference on machine learning, 9558--9568. PMLR

    2021

  20. [28]

    Simm, G.; and Hernandez-Lobato, J. M. 2020. A Generative Model for Molecular Distance Geometry. In International Conference on Machine Learning, 8949--8958. PMLR

    2020

  21. [29]

    Sohl-Dickstein, J.; Weiss, E.; Maheswaranathan, N.; and Ganguli, S. 2015. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, 2256--2265. PMLR

    2015

  22. [30]

    Song, Y.; and Ermon, S. 2019. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32

    2019

  23. [31]

    Song, Y.; Gong, J.; Xu, M.; Cao, Z.; Lan, Y.; Ermon, S.; Zhou, H.; and Ma, W.-Y. 2024. Equivariant flow matching with hybrid probability transport for 3d molecule generation. Advances in Neural Information Processing Systems, 36

    2024

  24. [32]

    G.; and Turner, C

    Tabak, E. G.; and Turner, C. V. 2013. A family of nonparametric density estimation algorithms. Communications on Pure and Applied Mathematics, 66(2): 145--164

    2013

  25. [33]

    G.; and Vanden-Eijnden, E

    Tabak, E. G.; and Vanden-Eijnden, E. 2010. Density estimation by dual ascent of the log-likelihood. Communications in Mathematical Sciences, 8(1): 217--233

    2010

  26. [34]

    Thomas, N.; Smidt, T.; Kearnes, S.; Yang, L.; Li, L.; Kohlhoff, K.; and Riley, P. 2018. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds. arXiv preprint arXiv:1802.08219

    2018 arXiv

  27. [35]

    Thölke, P.; and Fabritiis, G. D. 2022. Equivariant Transformers for Neural Network based Molecular Potentials. In International Conference on Learning Representations

    2022

  28. [36]

    Tong, A.; FATRAS, K.; Malkin, N.; Huguet, G.; Zhang, Y.; Rector-Brooks, J.; Wolf, G.; and Bengio, Y. 2024. Improving and generalizing flow-based generative models with minibatch optimal transport. Transactions on Machine Learning Research. Expert Certification

    2024

  29. [37]

    N.; Kaiser, .; and Polosukhin, I

    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, .; and Polosukhin, I. 2017. Attention is all you need. Advances in neural information processing systems, 30

    2017

  30. [38]

    M.; Caldwell, J

    Wang, J.; Wolf, R. M.; Caldwell, J. W.; Kollman, P. A.; and Case, D. A. 2004. Development and testing of a general amber force field. Journal of computational chemistry, 25(9): 1157--1174

    2004

  31. [39]

    Weiler, M.; Geiger, M.; Welling, M.; Boomsma, W.; and Cohen, T. S. 2018. 3d steerable cnns: Learning rotationally equivariant features in volumetric data. Advances in Neural Information Processing Systems, 31

    2018

  32. [40]

    Xu, M.; Wang, W.; Luo, S.; Shi, C.; Bengio, Y.; Gomez-Bombarelli, R.; and Tang, J. 2021. An end-to-end framework for molecular conformation generation via bilevel programming. In International conference on machine learning, 11537--11547. PMLR

    2021

  33. [41]

    Xu, M.; Yu, L.; Song, Y.; Shi, C.; Ermon, S.; and Tang, J. 2022. GeoDiff: A Geometric Diffusion Model for Molecular Conformation Generation. In International Conference on Learning Representations

    2022

  34. [42]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  35. [43]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.