Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

Applications of Modular Co-Design for De Novo 3D Molecule Generation

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Megalodon, a transformer with a single equivariant structure layer trained on a joint continuous/discrete denoising objective, claims state-of-the-art 3D molecule generation on GEOM-Drugs and a 2-10x reduction in GFN2-xTB relaxation…

desk verdict Solid empirical paper with a useful new energy benchmark and strong unconditional generation, but the headline energy claim is overblown and the conditional SOTA claim needs qualification. read the letter →

arxiv 2505.18392 v1 pith:2EAYL5JP submitted 2025-05-23 cs.LG cs.AIq-bio.BMq-bio.QM

Danny Reidenbach , Filipp Nikitin , Olexandr Isayev , Saee Paliwal This is my paper · ORCID
classification cs.LGcs.AIq-bio.BMq-bio.QM
keywords denovo3Dmoleculegenerationdenoisingdiffusionflowmatchingequivariantgraphneuralnetworkstransformerarchitecturestructure-basedenergybenchmarksGEOM-Drugsscaling
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

This paper claims that Megalodon, a scalable transformer architecture augmented with one equivariant structure layer, sets a new high-water mark for de novo 3D molecule generation on GEOM-Drugs. Trained with a joint continuous and discrete denoising co-design objective, the same model reaches state-of-the-art results under both diffusion and flow matching, and it is the first 3DMG model to perform unconditional molecule generation and conditional structure generation without retraining. The strongest concrete claim is energetic: the diffusion variant produces valid molecules whose median GFN2-xTB relaxation energy is about 3.17 kcal/mol, roughly 2x lower than the best diffusion baseline and about 10x lower than the best flow baseline, approaching the 2.5 kcal/mol thermal window used to define the GEOM conformer ensembles. If these results hold, physically realistic 3D structure becomes a directly learnable target of the generative process rather than an afterthought.

What carries the argument

The central object is Megalodon's fused invariant transformer block: a multi-head attention trunk that operates on a fused invariant feature of atom types, bond types, and distances, followed by a single EGNN structure layer whose update includes a cross-product term. This is paired with a joint continuous/discrete denoising co-design objective in which the continuous coordinate track and the discrete atom, bond, and charge tracks have independent noise-time variables sampled from the same distribution, so the model can fill in structure from a 2D graph and vice versa. Self-conditioning is applied as an outer residual wrapper over the raw logits and coordinates. That combination, not any single component, carries the argument: the transformer handles discrete generalization, the equivariant layer handles geometry, and the co-design sampling lets one trained model be prompted with a 2D topology to generate 3D structure.

What would settle it

Recompute the xTB relaxation-energy benchmark on size-stratified subsets, for example 30-49, 50-72, and 73-125 atoms, with per-stratum error bars; if Megalodon's 2-10x energy advantage disappears or reverses on the larger strata, the headline energy claim would not survive.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a transformer trunk that models discrete atom and bond data, fused with a simple distance-based EGNN update for the continuous coordinates, can jointly denoise all molecular modalities in a way that prior equivariant message-passing architectures cannot. The model is trained with independent time variables for the continuous and discrete tracks plus a co-design step that sometimes noises only the structure, which lets it learn the mapping from 2D graph to 3D geometry. As a result, Megalodon achieves state-of-the-art connected validity (0.927 for diffusion, 0.948 for flow), best 3D distributional metrics, and competitive conformer recall and precision on conditional structure generation; its diffusion variant reaches a median relaxation energy of 3.17 kcal/mol versus 6.36 for EQGAT-diff and 32.96 for SemlaFlow. The paper also reports that going from 19M to 40M parameters produces up to 49x more valid large molecules and 2-10x lower relaxation energies, and it introduces new structure-quality benchmarks based on bond length, bond angle, and dihedral deviation from GFN2-xTB-relaxed geometries.

Load-bearing premise

The energy comparison assumes that the relaxation energy computed with GFN2-xTB on only the valid generated molecules is a fair and unbiased measure of structural quality across models, and that validity filtering does not preferentially keep easier or smaller molecules for any one model.

Editorial extensions

If this is right

  • One trained Megalodon can both generate molecules from noise and, when prompted with a 2D graph, produce 3D conformers, so conditional structure generation no longer requires a separately trained conformer model.
  • The diffusion variant's median relaxation energy of 3.17 kcal/mol puts generated structures near the 2.5 kcal/mol thermal window, meaning downstream ranking and screening could skip repeated minimization for many molecules.
  • Doubling the parameter count to 40M improves validity on large molecules by up to 49x, suggesting further scaling could extend de novo generation into the over-100-atom regime relevant to degrader design.
  • The flow-matching variant matches or beats diffusion on 2D stability while using 25x fewer inference steps, at the cost of worse bond-length and energy accuracy.
  • Bond-length precision, not just topology or atom stability, appears to be the decisive factor for low relaxation energy, pointing to structure-level metrics as the right benchmark for future 3DMG models.

Reading between the lines

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

  • The co-design training trick, separate noise-time variables for continuous and discrete tracks with occasional structure-only noising, should transfer directly to other multimodal generative problems such as protein-ligand co-design, where one modality is expected to inform another.
  • Because the energy table reports no error bars or size stratification, the claimed 2-10x energy advantage is not fully established; a size-stratified rerun would either confirm it or reveal that the gain is concentrated in small molecules.
  • If the energetic proximity to the thermal window is real, an obvious next step is to add an explicit energy or relaxation penalty during training, which could push generated structures below the 2.5 kcal/mol threshold.
  • The transformer's ability to generalize to molecule sizes with negligible training support suggests that architecture choice, not just dataset coverage, may determine extrapolation in molecular generation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper presents Megalodon, a scalable transformer-based architecture for de novo 3D molecule generation, augmented with equivariant structure layers and trained with a joint continuous/discrete denoising co-design objective. The authors evaluate diffusion and flow-matching variants on GEOM-Drugs, reporting strong unconditional 2D/3D generation metrics (Table 1), size-scaling improvements (Fig. 3), conditional structure generation results (Table 4), and new relaxation-energy benchmarks computed with GFN2-xTB (Table 5). The paper claims state-of-the-art results in unconditional generation, conditional structure generation, and structure-based energy benchmarks, and further claims that Megalodon is the first model that can do both unconditional and conditional generation without retraining.

Significance. If the results hold, the paper is a useful empirical contribution: Megalodon is a relatively simple and scalable architecture; the baselines are retrained under matched training conditions; the size-sweep analysis draws attention to extrapolation beyond the training size distribution; and the new energy-based benchmark is a constructive step toward physics-aware evaluation. However, two load-bearing claims need further support. The headline '2-10x lower energy' is not established by Table 5 as currently analyzed, because relaxation energies are computed on validity-filtered sets with different size distributions and without uncertainty quantification. In addition, the conditional structure generation SOTA claim is contradicted by the recall metrics in Table 4, where Torsional Diffusion is better on recall coverage and recall AMR.

major comments (3)
  1. [§4.3, Table 5, Appendix E.4] The abstract's claim of 'energy levels that are 2-10x lower than those of the best prior generative models' is not supported as stated. Table 5 computes relaxation energies 'over the valid molecules from Table 1'; since ΔE_relax is an extensive quantity and Fig. 3 shows that validity is strongly size-dependent, the valid sets of different models likely have different size distributions. Appendix C.2 notes that SemlaFlow and Megalodon-flow were trained on data prefiltered to ≤72 atoms, while Megalodon was not, so the comparison can be biased by molecule size. The observed ratios (6.36 vs. 3.17 kcal/mol for EQGAT-diff, and 32.96 vs. 3.17 for SemlaFlow) cannot be attributed to per-molecule structural quality without size-stratified relaxation energies or bootstrap confidence intervals. The paper should provide size-binned results, quantify uncertainty, and restrict the headline to the range actually supported by the comparison.
  2. [§4.2, Table 4] The claim that Megalodon achieves state-of-the-art results in conditional structure generation is contradicted by the recall metrics in Table 4. Torsional Diffusion achieves higher recall coverage (mean 75.3 vs. 71.4, median 82.3 vs. 75.0) and lower recall AMR (mean 0.569 vs. 0.573, median 0.532 vs. 0.557). Megalodon is better on precision, but the SOTA claim should be qualified to precision or to 'competitive' recall unless the paper explains why recall is not the operative metric for this task.
  3. [§4.3, Appendix E.4] The new relaxation-energy benchmark conflates structural quality with validity filtering in its current form. The GEOM-Drugs ground-truth row in Table 5 has median ΔE_relax = 0.00 because ground-truth conformers were already optimized with GFN2-xTB during dataset construction, whereas generated molecules are relaxed only after passing through a validity filter. Reporting energies on all generated molecules and on validity-matched subsets would separate the model's structural accuracy from its ability to pass the filter; this is necessary before the energy benchmark can support the paper's central quantitative claim.
minor comments (5)
  1. [Section 3] The sentence introducing the second time variable contains a typo: 'createtcontinuous' should read 'create t_continuous'.
  2. [Section 3 / Table 2] The text says Megalodon has '4x more parameters' than EQGAT-diff, but the reported sizes are 40.6M and 12.3M, which is about 3.3x; please correct this ratio or explain the discrepancy.
  3. [Appendix B] There is a typo in the related-work paragraph: 'in the context our our study' should be 'in the context of our study'.
  4. [Table 4 discussion] The statement that 'EQGAT-diff is unable to generate any remotely valid structures' is too strong, since Table 4 reports EQGAT recall coverage of 0.8 and precision coverage of 0.1; the text should be softened to match the reported numbers.
  5. [§4.1 / Fig. 3] The discussion of the size-sweep would benefit from reporting the number of generated molecules per size and the statistical uncertainty of the validity percentages, since 100 molecules per length gives coarse estimates for low-validity regimes.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the energy and conditional-generation results are external empirical benchmarks measured after training; the lone self-citation is not load-bearing.

full rationale

The paper's central claims are empirical comparisons, not derivations that reduce to their own inputs. Megalodon is trained with standard DDPM/D3PM and CFM/DFM objectives taken from prior published work (Le et al., 2024; Vignac et al., 2023; Irwin et al., 2024; Campbell et al., 2024), and no parameter is fitted to the newly introduced energy benchmark. The xTB relaxation-energy metric is defined independently as Delta E_relax = E_optimized - E_initial on generated molecules, and the GEOM ground truths are themselves GFN2-xTB minima, so the evaluation is consistent with the training distribution rather than being a re-labeling of the model output. Conditional structure generation is tested on 200 held-out molecules without retraining, so it is not a fitted-input-called-prediction setup. The scaling claims are measured against external baselines (EQGAT-diff, SemlaFlow, MiDi, EDM), and the authors explicitly trained those baselines from scratch where needed to align training details. The only self-citation is Reidenbach 2024 in the Limitations section, used to note that binding affinity can be optimized without 3D structure-based methods; this is not load-bearing for any of the paper's quantitative conclusions. The potential size-distribution mismatch in Table 5, where validity-filtered sets may differ in molecule sizes, is a benchmark-fairness and statistical-validity concern, not a circularity of the kind where a prediction is equivalent to its inputs by construction. Similarly, the choice of GFN2-xTB for evaluation is a modeling assumption rather than a circular derivation. No equation in the paper defines a reported result in terms of itself, and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper relies on standard diffusion and flow matching assumptions, plus two domain-specific evaluation assumptions (xTB energy as ground truth, and the validity of prompting unconditional models). The co-design training procedure is an ad hoc choice with no ablation. No new physical entities are introduced.

free parameters (3)
  • Co-design noise split (fraction of steps with structure-only noising) = 0.5
    In Sec 3, 'Half the time all data types are independently noised as normal... the other half we only add noise to the structure.' This ratio is chosen by hand and no ablation is shown.
  • Architecture hyperparameters (layers, feature dims, heads) = 10 layers; edge dim 64/256; vector features 64/128; heads 4
    Table 6 lists configurations. These are chosen by hand without sensitivity analysis and affect all reported results.
  • Inference step counts (500 diffusion, 100 flow) = 500 / 100
    Used for all benchmark comparisons; the claim of 25x fewer steps for flow matching depends on this choice.
assumptions (4)
  • domain assumption GEOM-Drugs conformers are local minima of the GFN2-xTB potential energy surface and are the correct target distribution.
    Sec 2.2 states this; the whole energy benchmark rests on it.
  • domain assumption Delta E_relax computed with GFN2-xTB on valid molecules is an unbiased structure-quality measure across models.
    Table 5 and Appendix E.4; validity-based selection may bias comparisons.
  • domain assumption Prompting an unconditional model by fixing the 2D graph in input and output is a fair conditional-generation test.
    Sec 4.2; the same procedure is applied to EQGAT-diff, which fails, but the prompt may not match EQGAT-diff's training objective.
  • ad hoc to paper Sampling t_continuous and t_discrete from the same distribution creates a 'concrete connection' between modalities that improves learning.
    Sec 3; no ablation isolates this component from the architecture change.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Applications of Modular Co-Design for De Novo 3D Molecule Generation." pith.science (2026). https://pith.science/paper/2EAYL5JP

@misc{pith2026250518392,
  author       = {Pith},
  title        = {Pith review of: Applications of Modular Co-Design for De Novo 3D Molecule Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2EAYL5JP}},
  note         = {Machine review of arXiv:2505.18392}
}
read the original abstract

De novo 3D molecule generation is a pivotal task in drug discovery. However, many recent geometric generative models struggle to produce high-quality 3D structures, even if they maintain 2D validity and topological stability. To tackle this issue and enhance the learning of effective molecular generation dynamics, we present Megalodon-a family of scalable transformer models. These models are enhanced with basic equivariant layers and trained using a joint continuous and discrete denoising co-design objective. We assess Megalodon's performance on established molecule generation benchmarks and introduce new 3D structure benchmarks that evaluate a model's capability to generate realistic molecular structures, particularly focusing on energetics. We show that Megalodon achieves state-of-the-art results in 3D molecule generation, conditional structure generation, and structure energy benchmarks using diffusion and flow matching. Furthermore, doubling the number of parameters in Megalodon to 40M significantly enhances its performance, generating up to 49x more valid large molecules and achieving energy levels that are 2-10x lower than those of the best prior generative models.

Figures

Figures reproduced from arXiv: 2505.18392 by the authors.

Figure 1
Figure 1. Megalodon Architecture: molecules are separated into 3D structures and discrete atom [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Time and interpolation comparison between Megalodon and Megalodon-flow [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Diffusion model performance as a function of molecule size. Note the ability for Megalodon [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Megalodon molecule generation dynamics generated with Imagen 2 [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: Distribution of molecule sizes reflect low-energy conformations. Nevertheless, QM9 remains a well-established and frequently used small-scale benchmark, despite the fact that its median molecular size is unrealistically small (approximately 20 atoms). For completeness,…
Figure 6
Figure 6. Figure 6: Examples of generated molecules using Megalodon: (1) Diffusion and (2) Flow Matching. [PITH_FULL_IMAGE:figures/full_fig_p024_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. VEDA: 3D Molecular Generation via Variance-Exploding Diffusion with Annealing

    physics.chem-ph 2025-11 conditional novelty 6.0 of 10

    VEDA generates 3D molecules with VE diffusion plus LMMSE preconditioning and an arcsin scheduler, reaching near-relaxed geometries with 100 sampling steps.

  2. Predictive Feature Caching for Training-free Acceleration of Molecular Geometry Generation

    cs.LG 2025-10 conditional novelty 4.0 of 10

    Predictive feature caching, borrowed from image diffusion, speeds up molecular flow-matching generation by 2-3x at near-matched quality by forecasting hidden features instead of recomputing them.

Reference graph

Works this paper leans on

52 extracted references · 14 canonical work pages · cited by 2 Pith papers

  1. [1]

    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 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    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 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  3. [3]

    Ballard, Joshua Bambrick, Sebastian W

    Josh Abramson, Jonas Adler, Jack Dunger, Richard Evans, Tim Green, Alexander Pritzel, Olaf Ronneberger, Lindsay Willmore, Andrew J. Ballard, Joshua Bambrick, Sebastian W. Bodenstein, David A. Evans, Chia-Chun Hung, Michael O'Neill, David Reiman, Kathryn Tunyasuvunakool, Zachary Wu, Akvil \. e Z emgulyt \. e , Eirini Arvaniti, Charles Beattie, Ottavia Bert...

    2024

  4. [4]

    Albergo, Nicholas M

    Michael S. Albergo, Nicholas M. Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions, 2023. URL https://arxiv.org/abs/2303.08797

    arXiv 2023

  5. [5]

    Structured denoising diffusion models in discrete state-spaces

    Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 34: 0 17981--17993, 2021

    2021

  6. [6]

    Geom, energy-annotated molecular conformations for property prediction and molecular generation

    Simon Axelrod and Rafael G \'o mez-Bombarelli. Geom, energy-annotated molecular conformations for property prediction and molecular generation. Scientific Data, 9 0 (1): 0 185, 2022. doi:10.1038/s41597-022-01288-4. URL https://doi.org/10.1038/s41597-022-01288-4

    doi:10.1038/s41597-022-01288-4 2022

  7. [7]

    Gfn2-xtb---an accurate and broadly parametrized self-consistent tight-binding quantum chemical method with multipole electrostatics and density-dependent dispersion contributions

    Christoph Bannwarth, Sebastian Ehlert, and Stefan Grimme. Gfn2-xtb---an accurate and broadly parametrized self-consistent tight-binding quantum chemical method with multipole electrostatics and density-dependent dispersion contributions. Journal of Chemical Theory and Computation, 15 0 (3): 0 1652--1671, Mar 2019

    work page 2019

  8. [8]

    Langley, and Craig M

    Mikl \'o s B \'e k \'e s, David R. Langley, and Craig M. Crews. Protac targeted protein degraders: the past is prologue. Nature Reviews Drug Discovery, 21 0 (3): 0 181--200, 2022. doi:10.1038/s41573-021-00371-6. URL https://doi.org/10.1038/s41573-021-00371-6

    work page doi:10.1038/s41573-021-00371-6 2022

Show all 52 references
  1. [9]

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffr...

    2020

  2. [10]

    Generative flows on discrete state-spaces: Enabling multimodal flows with applications to protein co-design

    Andrew Campbell, Jason Yim, Regina Barzilay, Tom Rainforth, and Tommi Jaakkola. Generative flows on discrete state-spaces: Enabling multimodal flows with applications to protein co-design. arXiv preprint arXiv:2402.04997, 2024

    2024 arXiv

  3. [11]

    Analog bits: Generating discrete data using diffusion models with self-conditioning

    Ting Chen, Ruixiang Zhang, and Geoffrey Hinton. Analog bits: Generating discrete data using diffusion models with self-conditioning. arXiv preprint arXiv:2208.04202, 2022

    2022 arXiv

  4. [12]

    Diffdock: Diffusion steps, twists, and turns for molecular docking

    Gabriele Corso, Hannes Stärk, Bowen Jing, Regina Barzilay, and Tommi Jaakkola. Diffdock: Diffusion steps, twists, and turns for molecular docking. In International Conference on Learning Representations (ICLR), 2023 a

    2023

  5. [13]

    Particle guidance: non-iid diverse sampling with diffusion models

    Gabriele Corso, Yilun Xu, Valentin De Bortoli, Regina Barzilay, and Tommi Jaakkola. Particle guidance: non-iid diverse sampling with diffusion models. arXiv preprint arXiv:2310.13102, 2023 b

    2023 arXiv

  6. [14]

    Plinder: The protein-ligand interactions dataset and evaluation resource

    Janani Durairaj, Yusuf Adeshina, Zhonglin Cao, Xuejin Zhang, Vladas Oleinikovas, Thomas Duignan, Zachary McClure, Xavier Robin, Gabriel Studer, Daniel Kovtun, Emanuele Rossi, Guoqing Zhou, Srimukh Veccham, Clemens Isert, Yuxing Peng, Prabindh Sundareson, Mehmet Akdel, Gabriele...

    2024

  7. [15]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M \"u ller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image syn...

    2024

  8. [16]

    Energy windows for computed compound conformers: covering artefacts or truly large reorganization energies? Future Medicinal Chemistry, 11 0 (2): 0 97--118, 2019

    Nicolas Foloppe and I-Jen Chen. Energy windows for computed compound conformers: covering artefacts or truly large reorganization energies? Future Medicinal Chemistry, 11 0 (2): 0 97--118, 2019

    2019

  9. [17]

    Wei, David Duvenaud, José Miguel Hernández-Lobato, Benjamín Sánchez-Lengeling, Dennis Sheberla, Jorge Aguilera-Iparraguirre, Timothy D

    Rafael Gómez-Bombarelli, Jennifer N. Wei, David Duvenaud, José Miguel Hernández-Lobato, Benjamín Sánchez-Lengeling, Dennis Sheberla, Jorge Aguilera-Iparraguirre, Timothy D. Hirzel, Ryan P. Adams, and Alán Aspuru-Guzik. Automatic chemical design using a data-driven continuous r...

    2018 doi

  10. [18]

    Thomas A. Halgren. Merck molecular force field. i. basis, form, scope, parameterization, and performance of mmff94. Journal of Computational Chemistry, 17 0 (5-6): 0 490--519, 1996. doi:https://doi.org/10.1002/(SICI)1096-987X(199604)17:5/6<490::AID-JCC1>3.0.CO;2-P

    1996 doi

  11. [19]

    Sofroniew, Deniz Oktay, Zeming Lin, Robert Verkuil, Vincent Q

    Thomas Hayes, Roshan Rao, Halil Akin, Nicholas J. Sofroniew, Deniz Oktay, Zeming Lin, Robert Verkuil, Vincent Q. Tran, Jonathan Deaton, Marius Wiggert, Rohil Badkundri, Irhum Shafkat, Jun Gong, Alexander Derry, Raul S. Molina, Neil Thomas, Yousuf Khan, Chetan Mishra, Carolyn K...

    2024 doi

  12. [20]

    Query-key normalization for transformers

    Alex Henry, Prudhvi Raj Dachapally, Shubham Shantaram Pawar, and Yuxuan Chen. Query-key normalization for transformers. In Findings of the Association for Computational Linguistics: EMNLP 2020, pp.\ 4246–4253. Association for Computational Linguistics, 2020. doi:10.18653/v1/20...

    2020 doi

  13. [21]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239, 2020

    2006 arXiv

  14. [22]

    Equivariant diffusion for molecule generation in 3d

    Emiel Hoogeboom, V ctor Garcia Satorras, Cl \'e ment Vignac, and Max Welling. Equivariant diffusion for molecule generation in 3d. In International conference on machine learning, pp.\ 8867--8887. PMLR, 2022

    2022

  15. [23]

    Efficient 3d molecular generation with flow matching and scale optimal transport, 2024

    Ross Irwin, Alessandro Tibo, Jon Paul Janet, and Simon Olsson. Efficient 3d molecular generation with flow matching and scale optimal transport, 2024

    2024

  16. [24]

    Torsional diffusion for molecular conformer generation

    Bowen Jing, Gabriele Corso, Jeffrey Chang, Regina Barzilay, and Tommi Jaakkola. Torsional diffusion for molecular conformer generation. arXiv preprint arXiv:2206.01729, 2022

    2022 arXiv

  17. [25]

    Navigating the design space of equivariant diffusion-based generative models for de novo 3d molecule generation

    Tuan Le, Julian Cremer, Frank Noe, Djork-Arn \'e Clevert, and Kristof T Sch \"u tt. Navigating the design space of equivariant diffusion-based generative models for de novo 3d molecule generation. In The Twelfth International Conference on Learning Representations, 2024. URL h...

    2024

  18. [26]

    Evolutionary-scale prediction of atomic-level protein structure with a language model

    Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Nikita Smetanin, Robert Verkuil, Ori Kabeli, Yaniv Shmueli, et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science, 379 0 (6637): 0 1123--1130, 2023

    2023

  19. [27]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=PqvMRDCJT9t

    2023

  20. [28]

    Albergo, Nicholas M

    Nanye Ma, Mark Goldstein, Michael S. Albergo, Nicholas M. Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers, 2024. URL https://arxiv.org/abs/2401.08740

    2024 arXiv

  21. [29]

    O'Boyle, Michael Banck, Craig A

    Noel M. O'Boyle, Michael Banck, Craig A. James, Chris Morley, Tim Vandermeersch, and Geoffrey R. Hutchison. Open babel: An open chemical toolbox. Journal of Cheminformatics, 3 0 (1): 0 33, Oct 2011. ISSN 1758-2946. doi:10.1186/1758-2946-3-33. URL https://doi.org/10.1186/1758-2946-3-33

    2011 doi

  22. [30]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022

    2022 arXiv

  23. [31]

    M ol D iff: Addressing the atom-bond inconsistency problem in 3 D molecule diffusion generation

    Xingang Peng, Jiaqi Guan, Qiang Liu, and Jianzhu Ma. M ol D iff: Addressing the atom-bond inconsistency problem in 3 D molecule diffusion generation. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett (eds.), Proceedings o...

    2023

  24. [32]

    Pinheiro, Joshua Rackers, Joseph Kleinhenz, Michael Maser, Omar Mahmood, Andrew Martin Watkins, Stephen Ra, Vishnu Sresht, and Saeed Saremi

    Pedro O. Pinheiro, Joshua Rackers, Joseph Kleinhenz, Michael Maser, Omar Mahmood, Andrew Martin Watkins, Stephen Ra, Vishnu Sresht, and Saeed Saremi. 3d molecule generation by denoising voxel grids, 2024. URL https://arxiv.org/abs/2306.07473

    2024 arXiv

  25. [33]

    Crest—a program for the exploration of low-energy molecular chemical space

    Philipp Pracht, Stefan Grimme, Christoph Bannwarth, Fabian Bohle, Sebastian Ehlert, Gereon Feldmann, Johannes Gorges, Marcel M \"u ller, Tim Neudecker, Christoph Plett, et al. Crest—a program for the exploration of low-energy molecular chemical space. The Journal of Chemical P...

    2024

  26. [34]

    Evo SBDD : Latent evolution for accurate and efficient structure-based drug design

    Danny Reidenbach. Evo SBDD : Latent evolution for accurate and efficient structure-based drug design. In ICLR 2024 Workshop on Machine Learning for Genomics Explorations, 2024. URL https://openreview.net/forum?id=sLhUNz0uTz

    2024

  27. [35]

    E (n) equivariant graph neural networks

    V ctor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E (n) equivariant graph neural networks. In International conference on machine learning, pp.\ 9323--9332. PMLR, 2021

    2021

  28. [36]

    Structure-based drug design with equivariant diffusion models

    Arne Schneuing, Yuanqi Du, Charles Harris, Arian Jamasb, Ilia Igashov, Weitao Du, Tom Blundell, Pietro Li \'o , Carla Gomes, Max Welling, Michael Bronstein, and Bruno Correia. Structure-based drug design with equivariant diffusion models. arXiv preprint arXiv:2210.13695, 2022

    arXiv

  29. [37]

    Score-Based Generative Modeling through Stochastic Differential Equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equations . In International Conference on Learning Representations (ICLR), 2021

    2021

  30. [38]

    Equivariant flow matching with hybrid probability transport for 3d molecule generation

    Yuxuan Song, Jingjing Gong, Minkai Xu, Ziyao Cao, Yanyan Lan, Stefano Ermon, Hao Zhou, and Wei-Ying Ma. Equivariant flow matching with hybrid probability transport for 3d molecule generation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. URL http...

    2023

  31. [39]

    Unified generative modeling of 3d molecules via bayesian flow networks, 2024

    Yuxuan Song, Jingjing Gong, Yanru Qu, Hao Zhou, Mingyue Zheng, Jingjing Liu, and Wei-Ying Ma. Unified generative modeling of 3d molecules via bayesian flow networks, 2024. URL https://arxiv.org/abs/2403.15441

    2024 arXiv

  32. [40]

    Harmonic self-conditioned flow matching for multi-ligand docking and binding site design, 2024

    Hannes Stärk, Bowen Jing, Regina Barzilay, and Tommi Jaakkola. Harmonic self-conditioned flow matching for multi-ligand docking and binding site design, 2024. URL https://arxiv.org/abs/2310.05764

    2024 arXiv

  33. [41]

    Improving and generalizing flow-based generative models with minibatch optimal transport

    Alexander Tong, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector-Brooks, Kilian Fatras, Guy Wolf, and Yoshua Bengio. Improving and generalizing flow-based generative models with minibatch optimal transport. arXiv preprint arXiv:2302.00482, 2023

    2023 arXiv

  34. [42]

    Attention is all you need

    A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017

    2017

  35. [43]

    Midi: Mixed graph and 3d denoising diffusion for molecule generation

    Clement Vignac, Nagham Osman, Laura Toni, and Pascal Frossard. Midi: Mixed graph and 3d denoising diffusion for molecule generation. arXiv preprint arXiv:2302.09048, 2023

    2023 arXiv

  36. [44]

    A connection between score matching and denoising autoencoders

    Pascal Vincent. A connection between score matching and denoising autoencoders. Neural Computation, 23 0 (7): 0 1661--1674, 2011

    2011

  37. [45]

    Generative molecular design isn't as easy as people make it look, May 2024

    Pat Walters. Generative molecular design isn't as easy as people make it look, May 2024. URL https://practicalcheminformatics.blogspot.com/2024/05/generative-molecular-design-isnt-as.html

    2024

  38. [46]

    Elhag, Navdeep Jaitly, Joshua M

    Yuyang Wang, Ahmed A. Elhag, Navdeep Jaitly, Joshua M. Susskind, and Miguel Angel Bautista. Swallowing the bitter pill: Simplified scalable conformer generation, 2023

    2023

  39. [47]

    Geodiff: A geometric diffusion model for molecular conformation generation

    Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang. Geodiff: A geometric diffusion model for molecular conformation generation. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=PzcvxEMzvQC

    2022

  40. [48]

    Geometric latent diffusion models for 3d molecule generation, 2023

    Minkai Xu, Alexander Powers, Ron Dror, Stefano Ermon, and Jure Leskovec. Geometric latent diffusion models for 3d molecule generation, 2023. URL https://arxiv.org/abs/2305.01140

    2023 arXiv

  41. [49]

    Se (3) diffusion model with application to protein backbone generation

    Jason Yim, Brian L Trippe, Valentin De Bortoli, Emile Mathieu, Arnaud Doucet, Regina Barzilay, and Tommi Jaakkola. Se (3) diffusion model with application to protein backbone generation. arXiv preprint arXiv:2302.02277, 2023

    2023 arXiv

  42. [50]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  43. [51]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  44. [52]

    data prediction

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

    2018 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.