Pith. sign in

REVIEW 4 major objections 6 minor 41 references

DyneTrion: A Spatio-temporally Coherent Generative Emulator for Protein Dynamics Across Timescales

T0 review · 4 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read DyneTrion claims that a single diffusion emulator can replace molecular dynamics for protein motion on timescales from nanoseconds to microseconds, preserving both short-timescale statistics and long-timescale free-energy landscapes.

desk verdict A competent short-timescale protein dynamics emulator with a valuable new dataset, but the microsecond 'extrapolation' claim is unsupported by temporal metrics and may be in-distribution. read the letter →

arxiv 2607.15309 v1 pith:EU3N56K5 submitted 2026-07-15 q-bio.QM cs.LG

Kaihui Cheng , Zhiqiang Cai , Peng Tu , Yisong Yao , Limei Han , Libo Wu , Siyu Zhu , Tzuhsiung Yang
show 1 more author
Yuan Qi
This is my paper · ORCID
classification q-bio.QMcs.LG
keywords proteindynamicsmolecularemulationgenerativemodeldiffusiontrajectorygenerationensemblesfree-energylandscapetemporalcoherence
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

DyneTrion is a generative emulator intended to stand in for molecular dynamics (MD) simulations of protein motion. The paper's central claim is that one architecture can jointly enforce geometric symmetry, structural consistency, and temporal coherence—three properties earlier trajectory generators achieved only partially or separately. Across 100-ns benchmarks the model reproduces MD-derived flexibility, ensemble distributions, and interaction observables while staying stereochemically valid; on microsecond trajectories it is reported to preserve free-energy landscapes and metastable-state populations, and to propagate apo structures toward holo conformations. If these claims hold, DyneTrion would give structural biology a fast, general route from static structures to time-resolved, ensemble-faithful models, and the paper's new dynamicPDB dataset provides a common benchmark for testing such long-horizon claims.

What carries the argument

The load-bearing mechanism is a tri-attention score-based diffusion model over time-indexed backbone frames. Invariant point attention (IPA) updates residue representations in a way that respects rotational and translational symmetry, preventing rigid-body drift; spatial attention, anchored to a fixed high-quality reference structure, preserves global packing and stereochemistry over long horizons; temporal attention, fed by motion information from consecutive frames, models correlated evolution across time. Rotations are diffused with a variance-exploding SDE on SO(3) and translations with a variance-preserving SDE on R^3, and the model is trained with denoising score matching plus torsion-

What would settle it

Take a held-out multi-domain protein with a documented slow allosteric transition that is not in dynamicPDB, generate a 1-µs-equivalent trajectory, and compare the free-energy landscape and the population of the known intermediate state against explicit-solvent MD. If the generated ensemble never visits the experimentally observed intermediate, or if the metastable-state populations differ from MD by more than the reported error (≈0.51 kcal/mol MAE, JSD ≈0.5), the long-timescale generalization claim fails.

Watch

Extended reading notes

Core claim

Working from a sequence and two consecutive frames as motion information, DyneTrion runs a denoising diffusion process over time-indexed SE(3) rigid-body frames, with an invariant point attention module for SE(3)-robust geometric updates, a spatial transformer that anchors generation to a fixed reference conformation, and a temporal transformer that keeps frames smoothly correlated. The authors report that this combination removes the characteristic failure modes of earlier emulators—rigid-body drift, accumulating packing errors, and temporal decorrelation. On short timescales DyneTrion achieves the highest correlations with MD among compared models on flexibility and distributional metrics

Load-bearing premise

The load-bearing premise is that a model trained only to denoise coordinates—with no force-field or energy supervision—can learn the true thermodynamic weights of protein motion, so its long-timescale free energies and metastable populations are correct even for proteins and transitions not in its training set.

Editorial extensions

If this is right

  • Static structure predictors gain a time dimension: given one structure, DyneTrion can generate trajectories whose ensembles match microsecond MD, at a reported 2-3 orders of magnitude lower GPU cost.
  • Ensemble generation and trajectory generation become the same tool, matching specialized samplers on precision/recall while retaining dynamical signatures like residue-resolved flexibility.
  • Apo-to-holo propagation means conformational changes on ligand binding can be explored without simulating the binding process itself.
  • The dynamicPDB benchmark (over 10,000 proteins, up to 1-µs trajectories) makes long-horizon extrapolation claims testable and comparable across future models.
  • If the free-energy results hold, DyneTrion could serve as a fast filter to identify candidate proteins and states for more expensive explicit-solvent MD.

Reading between the lines

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

  • The absence of force/energy supervision implies the central bet is that denoising-score matching on coordinates alone recovers Boltzmann weights; a direct test would be comparing transition rates (kinetics), not just populations, against MD.
  • The reference-anchored design suggests the model's long-horizon stability may depend on starting from an experimentally resolved structure; for predicted structures with larger errors, an iterative-reference variant may be needed.
  • The current training covers single-chain soluble proteins under fixed conditions, so the same framework would need extension to multi-chain complexes and variable temperature/solvent before the claims generalize to those regimes.
  • An immediate stress test is a multi-domain protein with a known slow allosteric transition, absent from the current test set, to see whether the emulator can cross the relevant free-energy barrier at all.
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

4 major / 6 minor

Summary. The paper introduces DyneTrion, a score-based generative model for emulating protein MD trajectories. The architecture combines invariant point attention, spatial attention anchored to a reference structure, and temporal attention, trained with denoising score matching on rigid-body frame representations. The authors report short-timescale benchmarks on 100-ns MD trajectories across 100 test proteins, where DyneTrion outperforms MDGEN, UniSim, and ProTDyn on flexibility, distributional, ensemble-observable, and stereochemical metrics. They also introduce dynamicPDB, a claimed large-scale MD dataset, and evaluate long-timescale performance on 20 microsecond trajectories, 72 apo-to-holo systems, and 12 fast-folding proteins. The central claim is that DyneTrion is a general-purpose surrogate for atomistic MD that can generate spatio-temporally coherent trajectories from nanoseconds to microseconds at large speedups.

Significance. If the central claims were fully supported, this would be a significant contribution: it would be one of the first single generative models to jointly target geometric equivariance, structural consistency, and temporal coherence in protein dynamics, with a publicly available codebase and a new large-scale MD dataset. The short-timescale benchmarking is comprehensive in metric coverage and shows consistent point-estimate improvements over strong baselines. However, the long-timescale claims — which are central to the title and abstract — rest on stationary ensemble metrics and lack temporal validation, and the dataset description is internally inconsistent. The Discussion explicitly acknowledges that the model learns dynamics implicitly without force or energy supervision; this makes the long-timescale thermodynamic claims particularly fragile and in need of direct kinetic validation. The contribution is promising but the manuscript as written does not establish the central 'across timescales' claim.

major comments (4)
  1. [§2.3, Eqs. (34)–(35), Fig. 4] The microsecond evaluation contains no temporal validation. The free-energy MAE, metastable-population histograms, RMSD/RMSF, and contact/SASA overlaps are all invariant under time reordering; the apo-to-holo maximin TM-score (Fig. 4B) is a set-coverage statistic over generated structures and does not demonstrate a time-ordered path from apo to holo. None of these quantities require the generated sequence to be dynamically plausible. Since the title and abstract claim a 'spatio-temporally coherent emulator across timescales', the long-timescale claim is not established by the reported evidence. Additionally, in Eq. (34) N is stated to be the number of generated trajectories, but the summand is a per-bin free-energy difference; this definition must be corrected. Please add time-correlation functions, transition rates, or autocorrelations of collective variables at 1 µs, and a time-ordered
  2. [§4.5, Fig. 6, vs. Abstract] The dataset description is internally inconsistent. The abstract and introduction claim dynamicPDB contains 'over 10,000 proteins' with '10-ps resolution', while Methods §4.5 states 'approximately 4.0k protein dynamic trajectories' with coordinates recorded 'every 1 ps' (1.0M frames). Figure 6 shows training n=3720, 100-ns test n=100, and 1-µs test n=20. Moreover, the sentence 'generating 16 structures with 400ps intervals per inference' is ambiguous about the training/inference horizon. This ambiguity directly affects the meaning of 'extrapolation' to 1 µs: if training trajectories are full 1 µs simulations, then the 1 µs test is cross-protein generalization, not temporal extrapolation. Please reconcile all numbers and specify the exact temporal window used for training examples and the autoregressive generation procedure.
  3. [Table 1] No error bars, confidence intervals, or repeated-seed statistics are reported. The claim that DyneTrion 'achieves the highest overall agreement with MD' across flexibility, distributional accuracy, and ensemble observables rests entirely on point estimates. Given the table has no uncertainty quantification, it is impossible to assess whether the differences relative to MDGEN are significant. Please report means ± s.d. or confidence intervals over multiple model seeds or bootstrap resamples, and ideally paired statistical tests, for the central short-timescale metrics.
  4. [Fig. 4 and §2.3] The long-timescale results lack any baseline comparison. The microsecond free-energy MAE, apo-to-holo maximin TM-score, and fast-folding free-energy surfaces are reported without comparison to MDGEN, AlphaFlow, BioEmu, Str2Str, or other generative/ensemble methods. Since Fig. 5 shows that AlphaFlow achieves comparable ensemble precision and recall on the 20 dynamicPDB proteins, the reader cannot determine whether the long-timescale behavior is unique to DyneTrion or common to strong ensemble samplers. Please add at least one baseline for the µs free-energy task and the apo-to-holo task.
minor comments (6)
  1. [§4.3, Eq. (7)] The loss weights w1=0.25 and w2=1 are set empirically, but no sensitivity analysis or ablation is provided for these values.
  2. [§4.5] There is an unresolved placeholder citation 'following [?]' for the SDE hyperparameter choices (β_min, β_max, σ_min, σ_max).
  3. [General] Typos: 'AlphaFlod2' (§4.2), 'Embeder' (§4.2), 'axiliary' (§4.3), 'All-aton' (Introduction), 'the the average lDDT' (§4.6.2).
  4. [Fig. 6] The label 'T est Data 1 s' should read '1 µs' to match the text.
  5. [Availability] The code repository is provided, but no URL or accession mechanism is given for the dynamicPDB dataset itself. Please provide a dataset release link or availability statement.
  6. [Fig. 2D] The y-axis of the GPU-hour plot is not explicitly defined; please clarify whether this is wall-clock time, GPU-hours, or normalized units, and describe the MD baseline used.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DyneTrion is evaluated on cluster-held-out MD targets, and the long-timescale ensemble metrics, while stationary, are computed from generated samples rather than from fitted parameters.

full rationale

DyneTrion's claims are empirical benchmarks, not first-principles derivations. The model is trained by denoising score matching (Eq. 5) on MD frames and tested on proteins from clusters held out at 40% sequence identity ('the train/test split was performed at the cluster level to prevent information leakage'). No fitted parameter is relabeled as a prediction: the free-energy MAE (Eqs. 34–35), TICA-population overlaps, apo–holo maximin TM-scores, and RMSF correlations are all computed from generated ensembles and compared with reference MD values. Any training/evaluation overlap is controlled by the cluster split, so the nanosecond and microsecond results are genuine out-of-sample predictions. The architecture reuses AlphaFold2's Evoformer/IPA, but this is standard engineering and is not a self-citation chain; there is no load-bearing uniqueness theorem or self-citation. The Discussion's limitation ('DyneTrion currently learns dynamics implicitly without force or energy supervision') and the stationary character of the 1-µs metrics (histogram free energies, metastable populations, maximin TM-score) do mean the spatio-temporal coherence claim is under-supported at microsecond scale—a validation gap, not a circular reduction. I find no step where an output is equivalent by construction to an input.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

No new physical entities are postulated. The model architecture and dynamicPDB dataset are artifacts, not entities with independent falsifiable handles. The central assumptions are the representativeness of the MD training data and the sufficiency of coordinate-only denoising to learn thermodynamic distributions.

free parameters (3)
  • Diffusion SDE endpoints (β_min, β_max, σ_min, σ_max) = 0.1, 20, 0.1, 1.5
    Chosen in Section 4.5 ('following [?]') to define the forward noise process for rotations and translations.
  • Loss weights w1, w2 = 0.25, 1
    Set in Section 4.3; auxiliary losses LΩ and L2D are active only for t < 1/4.
  • Training/inference schedule numbers = 100 steps, 4 blocks, 256 hidden, batch 4, 450 epochs, 16 frames at 400 ps
    Architectural and training choices in Section 4.5; the inference length is not reconciled with the 100 ns and 1 µs trajectory claims.
assumptions (6)
  • standard math SE(3) diffusion with independent rotation and translation Brownian motions
    Used in Section 4.1 to construct forward/backward SDEs; follows prior frame diffusion work [29].
  • standard math Denoising score matching estimates the true score
    Minimizing Eq. 5 under standard regularity conditions yields the score of the data distribution.
  • domain assumption MD with Amber-ff14SB/TIP3P/OpenMM approximates real protein dynamics
    All ground-truth labels derive from these simulations (Section 4.5); any force-field bias propagates to every benchmark claim.
  • domain assumption Training distribution (single-chain, <500 residues, no membrane proteins) transfers to fast folders and apo/holo
    Apo-holo and fast-folding evaluations (Section 2.3) use proteins that may violate the training-filtering criteria in Section 4.4.
  • domain assumption Fixed reference structure remains a valid spatial anchor
    Fig. 2A shows clash rates degrade when the reference is updated; the production protocol keeps an initial PDB structure as the anchor.
  • domain assumption Dynamics can be learned implicitly from coordinates without energy/force supervision
    Discussion acknowledges DyneTrion is trained without force or energy supervision; the model must recover the effective Boltzmann distribution from denoising alone.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DyneTrion: A Spatio-temporally Coherent Generative Emulator for Protein Dynamics Across Timescales." pith.science (2026). https://pith.science/paper/EU3N56K5

@misc{pith2026260715309,
  author       = {Pith},
  title        = {Pith review of: DyneTrion: A Spatio-temporally Coherent Generative Emulator for Protein Dynamics Across Timescales},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EU3N56K5}},
  note         = {Machine review of arXiv:2607.15309}
}
abstract

Proteins function through coordinated motion across multiple spatial and temporal scales, underpinning processes such as ligand binding, allostery, and catalysis. However, accessing long-timescale conformational change through molecular dynamics (MD) simulations remains prohibitively expensive for systematic exploration across diverse systems. Here, we present DyneTrion, a generative protein dynamics emulator that jointly enforces geometric symmetry, structural consistency and temporal coherence within a single framework. DyneTrion uses a tri-attention architecture that integrates invariant point attention (IPA) for SE(3)-robust geometric updates, spatial attention anchored to a reference conformation to preserve structural integrity, and temporal attention to model correlated evolution across time frames. Across 100-ns MD trajectory simulation benchmarks, DyneTrion reproduces MD-derived flexibility, ensemble distributions and interaction observables while maintaining stereochemical validity during extrapolation. To evaluate long time-scale generalization, we introduce dynamicPDB, a dataset of over 10,000 proteins with up to 1-$\mu$s all-atom trajectories at 10-ps resolution and accompanying physical annotations. On microsecond trajectories, DyneTrion preserves free-energy landscapes and metastable-state populations, and it supports large conformational propagation in apo-to-holo transitions and fast folders. Together, DyneTrion provides a scalable path from static structure prediction toward time-resolved, ensemble-faithful protein modeling. The code is publicly available at https://github.com/fudan-generative-vision/DyneTrion

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 1 linked inside Pith

  1. [1]

    Nucleic Acids Res

    Burley, S.K., Berman, H.M., Bhikadiya, C., Bi, C., Chen, L., Di Costanzo, L., Christie, C., Dalenberg, K., Duarte, J.M., Dutta, S.,et al.: Rcsb protein data bank: biological macromolecular structures enabling research and education in fundamental biology, biomedicine, biotechnology and energy. Nucleic Acids Res. 47(D1), 464–474 (2019)

    2019

  2. [2]

    Nature596(7873), 583–589 (2021)

    Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R.,ˇZ´ ıdek, A., Potapenko, A.,et al.: Highly accurate protein structure prediction with alphafold. Nature596(7873), 583–589 (2021)

    2021

  3. [3]

    Nature, 1–3 (2024)

    Abramson, J., Adler, J., Dunger, J., Evans, R., Green, T., Pritzel, A., Ron- neberger, O., Willmore, L., Ballard, A.J., Bambrick, J., et al.: Accurate structure prediction of biomolecular interactions with alphafold 3. Nature, 1–3 (2024)

    2024

  4. [4]

    Rives, A., Meier, J., Sercu, T., Goyal, S., Lin, Z., Liu, J., Guo, D., Ott, M., Zitnick, C.L., Ma, J.,et al.: Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences. Proc. Natl. Acad. Sci. U.S.A.118(15), 2016239118 (2021)

    2021

  5. [5]

    Science379(6637), 1123–1130 (2023)

    Lin, Z., Akin, H., Rao, R., Hie, B., Zhu, Z., Lu, W., Smetanin, N., Verkuil, R., Kabeli, O., Shmueli, Y., Santos Costa, A., Fazel-Zarandi, M., Sercu, T., Candido, S., Rives, A.: Evolutionary-scale prediction of atomic-level protein structure with a language model. Science379(6637), 1123–1130 (2023)

    2023

  6. [6]

    Science387(6736), 850–858 (2025)

    Hayes, T., Rao, R., Akin, H., Sofroniew, N.J., Oktay, D., Lin, Z., Verkuil, R., Tran, V.Q., Deaton, J., Wiggert, M.,et al.: Simulating 500 million years of evolution with a language model. Science387(6736), 850–858 (2025)

    2025

  7. [7]

    Science373(6557), 871–876 (2021)

    Baek, M., DiMaio, F., Anishchenko, I., Dauparas, J., Ovchinnikov, S., Lee, G.R., Wang, J., Cong, Q., Kinch, L.N., Schaeffer, R.D., Mill´ an, C., Park, H., Adams, C., Glassman, C.R., DeGiovanni, A., Pereira, J.H., Rodrigues, A.V., Dijk, A.A., Ebrecht, A.C., Opperman, D.J., Sagmeister, T., Buhlheller, C., Pavkov-Keller, T., Rathinaswamy, M.K., Dalwadi, U., ...

    2021

  8. [8]

    Nature 347(6294), 631–639 (1990)

    Karplus, M., Petsko, G.A.: Molecular dynamics simulations in biology. Nature 347(6294), 631–639 (1990)

    1990

Show all 41 references
  1. [9]

    Teilum, K., Olsen, J.G., Kragelund, B.B.: Functional aspects of protein flexibility. Cell. Mol. Life Sci.66, 2231–2247 (2009)

    2009

  2. [10]

    Klepeis, J.L., Lindorff-Larsen, K., Dror, R.O., Shaw, D.E.: Long-timescale molec- ular dynamics simulations of protein structure and function. Curr. Opin. Struct. Biol.19(2), 120–127 (2009)

    2009

  3. [11]

    Bahar, I., Lezon, T.R., Yang, L.-W., Eyal, E.: Global dynamics of proteins: bridging between structure and function. Annu. Rev. Biophys.39, 23–42 (2010)

    2010

  4. [12]

    Hospital, A., Go˜ ni, J.R., Orozco, M., Gelp´ ı, J.L.: Molecular dynamics simulations: advances and applications. Adv. Appl. Bioinform. Chem., 37–47 (2015)

    2015

  5. [13]

    Neuron 99(6), 1129–1143 (2018)

    Hollingsworth, S.A., Dror, R.O.: Molecular dynamics simulation for all. Neuron 99(6), 1129–1143 (2018)

    2018

  6. [14]

    Nucleic Acids Res.52(D1), 384–392 (2023)

    Vander Meersche, Y., Cretin, G., Gheeraert, A., Gelly, J.-C., Galochkina, T.: ATLAS: protein flexibility description from atomistic molecular dynamics simulations. Nucleic Acids Res.52(D1), 384–392 (2023)

    2023

  7. [15]

    Structure18, 1399–409 (2010)

    Meyer, T., D’Abramo, M., Hospital, A., Rueda, M., Ferrer-Costa, C., P´ erez, A., Carrillo, O., Camps, J., Fenollosa, C., Repchevsky, D., Gelp´ ı, J., Orozco, M.: Model (molecular dynamics extended library): A database of atomistic molecular dynamics trajectories. Structure18, ...

    2010

  8. [16]

    Protein Eng

    Kehl, C., Simms, A.M., Toofanny, R.D., Daggett, V.: Dynameomics: a multi- dimensional analysis-optimized database for dynamic protein data. Protein Eng. Des. Sel.21(6), 379–386 (2008)

    2008

  9. [17]

    Vander Meersche, Y., Cretin, G., Brevern, A.G., Gelly, J.-C., Galochkina, T.: Medusa: prediction of protein flexibility from sequence. J. Mol. Biol.433(11), 166882 (2021)

    2021

  10. [18]

    Marchetti, F., Moroni, E., Pandini, A., Colombo, G.: Machine learning prediction of allosteric drug activity from molecular dynamics. J. Phys. Chem. Lett.12(15), 3724–3732 (2021)

    2021

  11. [19]

    Advances in Neural Information Processing Systems37, 40534–40564 (2024) 26

    Jing, B., St¨ ark, H., Jaakkola, T., Berger, B.: Generative modeling of molecular dynamics trajectories. Advances in Neural Information Processing Systems37, 40534–40564 (2024) 26

    2024

  12. [20]

    In: International Conference on Machine Learning (2025)

    Yu, Z., Huang, W., Liu, Y.: Unisim: A unified simulator for time-coarsened dynamics of biomolecules. In: International Conference on Machine Learning (2025)

    2025

  13. [21]

    arXiv preprint arXiv:2510.00013 (2025)

    Liu, Y., Zheng, H., Mao, L., Wang, Y., Chen, M., Lin, G.: Protdyn: a founda- tion protein language model for thermodynamics and dynamics generation. arXiv preprint arXiv:2510.00013 (2025)

    2025

  14. [22]

    In: International Conference on Machine Learning (2024)

    Jing, B., Berger, B., Jaakkola, T.: Alphafold meets flow matching for generating protein ensembles. In: International Conference on Machine Learning (2024)

    2024

  15. [23]

    In: International Conference on Learning Representations (2024)

    Lu, J., Zhong, B., Zhang, Z., Tang, J.: Str2str: A score-based framework for zero- shot protein conformation sampling. In: International Conference on Learning Representations (2024)

    2024

  16. [24]

    Lewis, S., Hempel, T., Jim´ enez-Luna, J., Gastegger, M., Xie, Y., Foong, A.Y., Satorras, V.G., Abdin, O., Veeling, B.S., Zaporozhets, I., Chen, Y., Yang, S., Fos- ter, A.E., Schneuing, A., Nigam, J., Barbero, F., Vincent, S., Campbell, A., Yim, J., Lienen, M., Shi, Y., Zheng,...

    2025

  17. [25]

    Advances in Neural Information Processing Systems32(2019)

    Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems32(2019)

    2019

  18. [26]

    arXiv preprint arXiv:2304.02198 (2023)

    Jing, B., Erives, E., Pao-Huang, P., Corso, G., Berger, B., Jaakkola, T.: Eigenfold: Generative protein structure prediction with diffusion models. arXiv preprint arXiv:2304.02198 (2023)

    2023 arXiv

  19. [27]

    Science334(6055), 517–520 (2011)

    Lindorff-Larsen, K., Piana, S., Dror, R.O., Shaw, D.E.: How fast-folding proteins fold. Science334(6055), 517–520 (2011)

    2011

  20. [28]

    Anderson, B.D.: Reverse-time diffusion equation models. Stoch. Process. Their Appl.12(3), 313–326 (1982)

    1982

  21. [29]

    In: International Conference on Machine Learning, pp

    Yim, J., Trippe, B.L., De Bortoli, V., Mathieu, E., Doucet, A., Barzilay, R., Jaakkola, T.: Se (3) diffusion model with application to protein backbone genera- tion. In: International Conference on Machine Learning, pp. 40001–40039 (2023). PMLR

    2023

  22. [30]

    Advances in Neural Information Processing Systems30(2017)

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, /suppress L., Polosukhin, I.: Attention is all you need. Advances in Neural Information Processing Systems30(2017)

    2017

  23. [31]

    Biological Crystallography58(6), 899–907 (2002) 27

    Berman, H.M., Battistuz, T., Bhat, T.N., Bluhm, W.F., Bourne, P.E., Burkhardt, K., Feng, Z., Gilliland, G.L., Iype, L., Jain, S.,et al.: The protein data bank. Biological Crystallography58(6), 899–907 (2002) 27

    2002

  24. [32]

    Lindahl, E., Sansom, M.S.: Membrane proteins: molecular dynamics simulations. Curr. Opin. Struct. Biol.18(4), 425–431 (2008)

    2008

  25. [33]

    Bioinformatics22(5), 623–625 (2006)

    Lomize, M.A., Lomize, A.L., Pogozheva, I.D., Mosberg, H.I.: Opm: orientations of proteins in membranes database. Bioinformatics22(5), 623–625 (2006)

    2006

  26. [34]

    Nucleic Acids Res.41(D1), 524–529 (2012)

    Kozma, D., Simon, I., Tusnady, G.E.: Pdbtm: Protein data bank of transmem- brane proteins after 8 years. Nucleic Acids Res.41(D1), 524–529 (2012)

    2012

  27. [35]

    Nucleic Acids Res.47(D1), 390–397 (2019)

    Newport, T.D., Sansom, M.S.P., Stansfeld, P.J.: The memprotmd database: a resource for membrane-embedded protein structures and their lipid interactions. Nucleic Acids Res.47(D1), 390–397 (2019)

    2019

  28. [36]

    Bioinformatics38(5), 1452– 1454 (2022)

    Bittrich, S., Rose, Y., Segura, J., Lowe, R., Westbrook, J.D., Duarte, J.M., Burley, S.K.: Rcsb protein data bank: improved annotation, search and visualization of membrane protein structures archived in the pdb. Bioinformatics38(5), 1452– 1454 (2022)

    2022

  29. [37]

    Biopolymers22(12), 2577–2637 (1983)

    Kabsch, W., Sander, C.: Dictionary of protein secondary structure: pattern recognition of hydrogen-bonded and geometrical features. Biopolymers22(12), 2577–2637 (1983)

    1983

  30. [38]

    Webb, B., Sali, A.: Comparative protein structure modeling using modeller. Curr. Protoc. Bioinform.54(1), 5–6 (2016)

    2016

  31. [39]

    PLoS Comput

    Eastman, P., Swails, J., Chodera, J.D., McGibbon, R.T., Zhao, Y., Beauchamp, K.A., Wang, L.-P., Simmonett, A.C., Harrigan, M.P., Stern, C.D., Wiewiora, R.P., Brooks, B.R., Pande, V.S.: Openmm 7: Rapid development of high per- formance algorithms for molecular dynamics. PLoS Co...

    2017

  32. [40]

    Steinegger, M., S¨ oding, J.: Mmseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nat. Biotechnol.35(11), 1026–1028 (2017)

    2017

  33. [41]

    P´ erez-Hern´ andez, G., Paul, F., Giorgino, T., De Fabritiis, G., No´ e, F.: Identi- fication of slow molecular order parameters for markov model construction. J. Chem. Phys.139(1) (2013) 28

    2013

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.