Pith. sign in

REVIEW 4 major objections 139 references

Diffusion over DCT spectral volumes of protein motion yields fast, temperature-aware trajectories with strong flexibility accuracy.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-11 21:26 UTC pith:CG3NH5CC

load-bearing objection Solid spectral-trajectory idea with real speed and RMSF numbers, but the geometry post-processing undercuts the speed claim and the low-k recovery is still incomplete. the 4 major comments →

arxiv 2607.04134 v1 pith:CG3NH5CC submitted 2026-07-05 q-bio.BM q-bio.QM

Spectral Diffusion for Protein Dynamics

Hew Phipps , Matteo Cagiada , Santiago D. Villalba , Charlotte M. Deane This is my paper
classification q-bio.BM q-bio.QM
keywords protein dynamicsdiffusion modelsspectral volumesDCTRMSFMD emulationtemperature-conditioned generationmdCATH
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that protein dynamics are better generated as Discrete Cosine Transform spectral volumes than as raw time series of coordinates. Fourier modes separate slow collective conformational changes from fast local jitter, so a diffusion model conditioned on structure and temperature can denoise whole trajectory windows at once. The resulting method, DynaMode, trained on mdCATH monomers, produces temporally ordered Cα trajectories for proteins under 576 residues and temperatures from roughly 300–450 K. On the held-out test set it reaches an RMSF Pearson correlation of 0.844 and samples on the order of one second per 250-frame window, outperforming several published MD emulators on flexibility metrics while remaining temporally coherent. The same low-frequency coefficients also serve as a built-in measure of per-residue flexibility, because by Parseval’s relation they approximate RMSF and carry directional and timescale information. The work matters to anyone who needs rapid dynamical ensembles without running nanosecond molecular-dynamics trajectories for every new structure or temperature.

Core claim

Denoising structure- and temperature-conditioned DCT spectral volumes of Cα displacement trajectories produces temporally coherent protein ensembles that capture multiscale dynamics, achieving RMSF Pearson r of 0.844 and pairwise RMSD r of 0.854 on held-out mdCATH data while generating roughly 250 frames in about one second.

What carries the argument

DCT spectral volumes of native-frame displacements: an orthonormal frequency representation of trajectory windows in which low modes encode collective flexibility (and analytically approximate RMSF via Parseval) and high modes encode local jitter; diffusion with hierarchical log-SNR noise and a spectral-convolution trunk plus low-frequency calibration head operates directly in this basis.

Load-bearing premise

That imperfect spectral predictions can still become physically valid trajectories after a lightweight coordinate refiner, bond-length projection, and optional energy minimisation, even though raw inverse transforms show high clash rates and high-frequency truncation already destroys geometry.

What would settle it

On a larger held-out set of unseen monomers and temperatures, measure nonbonded clash rates and RMSF correlation of raw (unminimised) inverse-DCT trajectories; if clashes remain near the reported 90% of frames and RMSF r falls well below 0.8, the claim that spectral diffusion alone yields usable dynamics fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 0 minor

Summary. The paper introduces DynaMode, a diffusion model that generates protein dynamics by denoising Discrete Cosine Transform (DCT) spectral volumes of Cα displacement trajectories, conditioned on an input structure and temperature. The spectral representation is motivated as a physics-informed inductive bias that disentangles slow conformational modes from fast fluctuations; via Parseval’s identity the low non-DC frequencies are linked to RMSF. Trained on mdCATH (monomers ≤576 residues, 320–450 K), the model reports strong ensemble metrics on a held-out test set (global RMSF Pearson r = 0.844, pairwise RMSD r = 0.854) and ~1 s per 250-frame trajectory on a GH200 before post-processing, with weaker out-of-distribution results on ATLAS at 300 K. A spectral-convolution architecture with a low-frequency amplitude-calibration head is used; optional coordinate refinement, differentiable SHAKE, and Cα energy minimisation are applied to mitigate steric clashes.

Significance. If the central results hold under fair comparison and with physically usable trajectories, the work would be a meaningful contribution to generative MD emulation: it offers a universal (protein-agnostic) temporal basis, an explicit multiscale separation of dynamics, competitive flexibility metrics on the standard mdCATH split, and a lightweight architecture that can be very fast. The Parseval-motivated link between low-k spectral amplitude and per-residue flexibility is a clean inductive bias, and the public code is a clear strength. The main scientific value is the spectral-volume formulation itself rather than any single benchmark number; that value is currently limited by the severity of structural collapse in raw samples and by the dependence of the headline speed claim on omitting the geometry repair the paper itself shows is needed.

major comments (4)
  1. The load-bearing claim of “fast and accurate MD emulation” (Contributions §1.1; Abstract) is not isolated from post-hoc geometry repair. Appendix A.14 and Table 8 report raw predictions with ~20 nonbonded Cα–Cα clashes/frame and ~51% of mdCATH frames containing nonlocal backbone-trace contacts <1 Å; Figure 5 shows that the optional Cα energy minimisation that largely removes clashes multiplies wall-clock time by ~50× (1.1 s → 52.7 s). Section 4.1 / Table 4 already show that even modest high-frequency truncation collapses Cα–Cα geometry. The manuscript should report the full Table 1 / Table 2 suite both with and without minimisation, state which protocol produces the headline RMSF/pairwise numbers, and either (i) demonstrate that raw inverse-DCT trajectories are already usable for the claimed applications or (ii) revise the speed claim to the post-processed regime.
  2. Low-frequency recovery, which by Parseval (Eq. 14) and §4.2 is the primary carrier of RMSF and collective motion, remains systematically incomplete. Figure 8B–C and the case study in Figure 6L–N show under-predicted low-k amplitudes and corresponding free-energy / flexibility misalignment; the dedicated low-frequency calibration head (§A.6) improves DC correlation but does not close the mid/low gap. Because the paper’s inductive-bias argument rests on accurate low-k spectral volumes, the manuscript should quantify how much of the reported RMSF Pearson r is explained by residualised DC / low-k calibration versus full spectral fidelity, and report band-wise SpecMSE / amplitude recovery on the test set as primary diagnostics alongside RMSF.
  3. Benchmark comparisons in Tables 1–2 and A.13 are not under matched conditions. Competitor numbers (MDGen, AlphaFlow-MD, TEMPO, MarS-FM, BioEMU) are taken from published tables; ATLAS is OOD for DynaMode while several baselines train on ATLAS subsets; MarS-FM is an ensemble sampler without temporal trajectories, yet is used as a primary accuracy foil. For the central claim of superior dynamics emulation, the paper should either re-evaluate at least one open baseline on the same hardware/protocol or clearly separate “ensemble property” metrics from “temporally coherent trajectory” metrics and avoid ranking across method classes without that distinction.
  4. Temporal coherence is asserted as a distinguishing advantage over unordered ensemble samplers (§1, §2) but is not measured with a dedicated metric (e.g., time-lagged autocorrelation of collective coordinates, frame-to-frame velocity consistency, or MSM transition fidelity). Ensemble metrics (RMSF, pairwise RMSD, RMWD, weak/transient Jaccard) can be satisfied by correctly ranked flexibility without correct temporal ordering. A short analysis that isolates temporal structure of the inverse-DCT samples—before energy minimisation—would substantiate the claim that spectral diffusion yields coherent trajectories rather than only good marginal ensembles.

Circularity Check

0 steps flagged

No significant circularity: Parseval motivates the spectral target, but held-out RMSF/ensemble metrics are empirical model evaluations, not tautologies of the representation.

full rationale

DynaMode is an empirical generative model: it learns to denoise structure- and temperature-conditioned DCT spectral volumes of Cα displacement windows, inverse-transforms them, and is scored against held-out MD (sequence-dissimilar mdCATH split; OOD ATLAS). The load-bearing performance numbers (global RMSF Pearson r = 0.844, pairwise RMSD r = 0.854, etc.) are correlations between generated and reference trajectories, not quantities forced by the loss or normalisation. Parseval’s identity (Eq. 14) is a standard orthonormal-transform fact used only as inductive-bias motivation: low non-DC spectral power equals truncated RMSF², so predicting low-k volumes is ‘by construction’ reasoning about flexibility. That identity does not make the test-set correlations true; the model must still recover the correct amplitudes for unseen proteins. Train-set frequency scales, DC residualisation, and the low-k calibration head are ordinary conditioning/architecture choices; they do not fit the reported test metrics and then re-label them as predictions. There is no uniqueness theorem, self-citation chain, or ansatz smuggled in as a first-principles derivation. Weaknesses (high raw clash rates, post-hoc minimisation, low-k under-prediction) are correctness/validity issues, not circular reductions of the claimed results to their inputs.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 2 invented entities

The central claim rests on standard diffusion and DCT mathematics, domain assumptions about MD data and Cα sufficiency, plus a large set of hand-chosen architectural and schedule hyperparameters that are not derived from first principles. No new physical entities are postulated; the invented pieces are the model architecture and the spectral-volume representation for this task.

free parameters (6)
  • anisotropy strength γ = 0.5
    Set to 0.5 to enforce hierarchical low-to-high frequency denoising; chosen by the authors, not derived.
  • window length τ and crop length ℓ = τ=256, ℓ=576
    Fixed at 256 frames and 576 residues; determine the spectral volume size and model capacity.
  • frequency-band partitions for spectral mixing = four bands as listed
    Hard-coded blocks [0,8],[8,32],[32,128],[128,256] that control the FNO-style trunk; design choice.
  • low-frequency calibration head width M = M < 8
    Number of lowest modes recalibrated by the residual amplitude head; selected for best validation performance.
  • noise schedule and CFG dropout = cosine + 50 DDIM steps
    Log-SNR-shifted cosine schedule, 1000 training steps, 50 inference steps, 15% CFG dropout; standard but free choices.
  • energy-minimisation force constants = kb=10000, knb=250, rnb=3.5 Å etc.
    kb, kθ, kϕ, knb and clash thresholds used in the optional Cα minimiser; hand-tuned for clash reduction.
axioms (4)
  • domain assumption Orthonormal DCT-II along the time axis of finite MD windows is a suitable universal basis that separates slow collective modes from fast jitter and is more robust to non-equilibrium boundaries than DFT.
    Stated in Sections 2–3 and supported by truncation experiments in 4.1; not a theorem for protein dynamics, only an empirical inductive bias.
  • standard math Parseval’s identity implies that non-DC spectral power approximates RMSF², so accurate low-k prediction is sufficient for flexibility metrics.
    Equation (14) is standard; the modelling claim that this makes the spectral target superior is a domain interpretation.
  • domain assumption Cα-only trajectories plus post-hoc geometry cleanup adequately represent the dynamics of interest for the reported ensemble metrics.
    Implicit throughout training and evaluation; all-atom methods are acknowledged as more expensive alternatives.
  • domain assumption mdCATH 1 ns-resolution windows at 320–450 K plus the chosen train/val/test split are representative enough for generalisation claims, including limited OOD tests at 300 K.
    Data section and evaluation protocol; sequence-similarity filtering is used but temperature and timescale coverage remain limited.
invented entities (2)
  • DynaMode spectral-convolution architecture with low-frequency amplitude-calibration head no independent evidence
    purpose: To denoise full DCT spectral volumes efficiently while specially protecting the low-k amplitudes that dominate flexibility.
    Custom FNO-style trunk plus residual gain head; no independent physical existence outside the model.
  • Structure- and temperature-conditioned DCT spectral volume as diffusion target for protein trajectories no independent evidence
    purpose: To provide a multiscale, temporally coherent representation that can be inverse-transformed into ordered structures.
    Adaptation of spectral-volume ideas from image dynamics to MD; the entity is the modelling choice itself.

pith-pipeline@v1.1.0-grok45 · 32329 in / 3366 out tokens · 43427 ms · 2026-07-11T21:26:23.549334+00:00 · methodology

0 comments
read the original abstract

Generative models present a promising alternative to expensive molecular dynamics for computationally querying protein dynamics, yet many existing approaches treat ensembles as unordered snapshots rather than temporally coherent trajectories, or scale poorly with protein size. We present a new physics-informed representation using Fourier transforms as an inductive bias for the multiscale temporal nature of protein dynamics. Diffusion in the spectral domain allows for disentangling of dynamics into slow conformational modes and fast atomic jitter, enabling rapid and improved prediction of dynamics across a range of temperatures. This is facilitated by denoising of structure and temperature conditioned spectral volumes where the low frequencies directly encode per-residue flexibility. Trained on the mdCATH dataset, we evaluate our model, DynaMode, on a held-out test set achieving strong performance across a set of ensemble-based metrics including a Root Mean Squared Fluctuation (RMSF) pearson $r$ of $0.844$. Code is available at https://github.com/HPuntu/DynaMode.

Figures

Figures reproduced from arXiv: 2607.04134 by Charlotte M. Deane, Hew Phipps, Matteo Cagiada, Santiago D. Villalba.

Figure 1
Figure 1. Figure 1: DynaMode is a diffusion model that iteratively denoises a spectral volume representation of protein dynamics given an input structure and temperature. The predicted spectral volume is inverse DCT-II transformed into a trajectory of structures over time. 1.1. Contributions 1. Fast and Accurate MD Emulation We achieve superior performance on key metrics for protein dynamics and ensemble properties including … view at source ↗
Figure 2
Figure 2. Figure 2: The DCT transform gives τ frequency coefficients for each Cα coordinate channel (x,y,z) of each residue. A Illustrative cartoon depiction 12asA00, a 327 residue mdCATH domain coloured by the per-residue spectral power of the lowest k = 8 frequencies over a 450K trajectory. Dashed box: Close up of ASP100, the residue with the highest spectral power. B DCT spectral amplitudes for a number of different freque… view at source ↗
Figure 3
Figure 3. Figure 3: The log-SNR shifted cosine noise schedule enforces hierarchical low to high frequency denoising. Model-space SNR (dB) is the SNR after applying frequency normalisation. 3.3. Diffusion on Spectral Volumes Schedule and noise. For a clean spectral target Z0 the forward diffusion process is defined Zt = √ α¯t Z0+ √ 1 − α¯t (w⊙ϵ), ϵ ∼ N (0, I). (10) Here α¯t is given by a log-SNR-shifted cosine sched￾ule (Nicho… view at source ↗
Figure 4
Figure 4. Figure 4: Spectral transformation with DCT followed by truncation leads to significant reconstruction errors. A Cα − Cα distances break down with minimal spectral truncation and this scales with simulation temperature. B While spectral truncation destroys structural validity, DCT (top) is more robust to this than DFT (bottom) at the trajectory boundaries (first and last 5 frames). C A strong positive relationship be… view at source ↗
Figure 5
Figure 5. Figure 5: A Median inference time per generated 500-frame tra￾jectory on the same GH200 hardware, with error bars showing the interquartile range. Bottom: Structural-validity distributions across targets, reporting nonbonded Cα–Cα clashes below 3.5 A˚ per frame (B) and the percentage of frames containing a nonlocal backbone trace distance below 1.0 A ( ˚ C). low-frequency residual correction head in the model (Sec￾t… view at source ↗
Figure 6
Figure 6. Figure 6: We show three case study domains for DynaMode: 1. (top) mdCATH domain ID 1aabA00 is an 83 residue chain from the train set. 2. (middle) mdCATH domain ID 2fm7A00 is a small 62 residue unfolder from the validation set. 3. (bottom) mdCATH domain ID 4c23B01 from the test set is a 234 residue globular protein. Predicted trajectories were post-inference energy minimised. For each we show structures representing … view at source ↗
Figure 7
Figure 7. Figure 7: Conditioned DCT frequency-scale statistics used for spectral normalisation. A All-frequency ECDF heatmap over all K = 256 DCT modes, with the selected normalisation scale (red line) and stored Q0.75 amplitude (dashed) overlaid. B Low-frequency log-amplitude distribution summary shown as one vertical violin per mode, again with same scale overlaid. C Temperature-conditioned Q0.75 amplitudes for the first 32… view at source ↗
Figure 8
Figure 8. Figure 8: Residue flexibility (RMSF) and spectral volumes prediction accuracy across temperatures on the mdCATH test set. A Violin plots show the distribution of RMSF Pearson and Spearman correlations (predicted vs reference MD trajectory following the protocol in Appendix A.11 across temperatures. We also show spectral volume prediction accuracy across different regions of the spectral volume defined by grouping fr… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

139 extracted references · 71 canonical work pages · 8 internal anchors

  1. [1]

    and Bambrick, Joshua and Bodenstein, Sebastian W

    Abramson, Josh and Adler, Jonas and Dunger, Jack and Evans, Richard and Green, Tim and Pritzel, Alexander and Ronneberger, Olaf and Willmore, Lindsay and Ballard, Andrew J. and Bambrick, Joshua and Bodenstein, Sebastian W. and Evans, David A. and Hung, Chia-Chun and O'Neill, Michael and Reiman, David and Tunyasuvunakool, Kathryn and Wu, Zachary and. Accur...

    doi:10.1038/s41586-024-07487-w

  2. [2]

    Accelerating Molecular Dynamics Simulations Using Fast

    Liang, Jiuyang and Lu, Libin and Barnett, Alex and Greengard, Leslie and Jiang, Shidong , year = 2026, month = may, journal =. Accelerating Molecular Dynamics Simulations Using Fast. doi:10.1038/s41467-026-73232-8 , urldate =

    doi:10.1038/s41467-026-73232-8 2026

  3. [3]

    and Natarajan, T

    Ahmed, N. and Natarajan, T. and Rao, K. R. , year = 1974, month = jan, journal =. Discrete. doi:10.1109/T-C.1974.223784 , urldate =

    doi:10.1109/t-c.1974.223784 1974

  4. [4]

    Nature Methods , volume =

    Unified Rational Protein Engineering with Sequence-Based Deep Representation Learning , author =. Nature Methods , volume =. doi:10.1038/s41592-019-0598-1 , urldate =

    doi:10.1038/s41592-019-0598-1

  5. [5]

    and Buggy, T

    Bagheri Zadeh, P. and Buggy, T. and Sheikh Akbari, A. , year = 2008, month = jun, journal =. Statistical,. doi:10.1049/iet-ipr:20070181 , urldate =

    work page doi:10.1049/iet-ipr:20070181 2008

  6. [7]

    Ageing Research Reviews , volume =

    A Comprehensive Review of Protein Misfolding Disorders, Underlying Mechanism, Clinical Diagnosis, and Therapeutic Strategies , author =. Ageing Research Reviews , volume =. doi:10.1016/j.arr.2023.102017 , urldate =

    work page doi:10.1016/j.arr.2023.102017 2023

  7. [8]

    and Pavlovi

    Bauer, Jacob A. and Pavlovi. Normal. Molecules , volume =. doi:10.3390/molecules24183293 , urldate =

    work page doi:10.3390/molecules24183293

  8. [10]

    and Freed, Karl F

    Baxa, Michael C. and Freed, Karl F. and Sosnick, Tobin R. , year = 2009, month = mar, journal =. Simulations of Protein Folding Transition States Using. doi:10.1016/j.jmb.2009.01.010 , urldate =

    work page doi:10.1016/j.jmb.2009.01.010 2009

  9. [11]

    Structural

    Bernad. Structural. Journal of the American Chemical Society , volume =. doi:10.1021/ja069124n , urldate =

    work page doi:10.1021/ja069124n

  10. [12]

    and Swenson, David W

    Bolhuis, Peter G. and Swenson, David W. H. , year = 2021, journal =. Transition. doi:10.1002/adts.202000237 , urldate =

    doi:10.1002/adts.202000237 2021

  11. [15]

    Diffusion

    Bortoli, Valentin De and Thornton, James and Heng, Jeremy and Doucet, Arnaud , year = 2023, month = apr, number =. Diffusion. doi:10.48550/arXiv.2106.01357 , urldate =. arXiv , keywords =:2106.01357 , primaryclass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2106.01357 2023

  12. [17]

    doi:10.48550/arXiv.2310.02391 , urldate =

    Bose, Avishek Joey and. doi:10.48550/arXiv.2310.02391 , urldate =. arXiv , keywords =:2310.02391 , primaryclass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.02391

  13. [18]

    Nature Genetics , volume =

    Genome-Wide Prediction of Disease Variant Effects with a Deep Protein Language Model , author =. Nature Genetics , volume =. doi:10.1038/s41588-023-01465-0 , urldate =

    doi:10.1038/s41588-023-01465-0

  14. [19]

    doi:10.1093/bioinformatics/btac020 , urldate =

    Brandes, Nadav and Ofer, Dan and Peleg, Yam and Rappoport, Nadav and Linial, Michal , year = 2022, month = apr, journal =. doi:10.1093/bioinformatics/btac020 , urldate =

    doi:10.1093/bioinformatics/btac020 2022

  15. [20]

    Protein Science : A Publication of the Protein Society , volume =

    Protein Unfolding Rates Correlate as Strongly as Folding Rates with Native Structure , author =. Protein Science : A Publication of the Protein Society , volume =. doi:10.1002/pro.2606 , urldate =

    work page doi:10.1002/pro.2606

  16. [21]

    Butcher, Jasper Kenneth Veje and Krishna, Rohith and Mitra, Raktim and Brent, Rafael Isaac and Li, Yanjing and Corley, Nathaniel and Kim, Paul and Funk, Jonathan and Mathis, Simon Valentin and Salike, Saman and Muraishi, Aiko and Eisenach, Helen and Thompson, Tuscan Rock and Chen, Jie and Politanska, Yuliya and Sehgal, Enisha and Coventry, Brian and Zhang...

    doi:10.1101/2025.09.18.676967 2025

  17. [22]

    Methods in Molecular Biology (Clifton, N.J.) , volume =

    Solution-State Nuclear Magnetic Resonance Spectroscopy and Protein Folding , author =. Methods in Molecular Biology (Clifton, N.J.) , volume =. doi:10.1007/978-1-60327-223-0_7 , abstract =

    work page doi:10.1007/978-1-60327-223-0_7

  18. [23]

    doi:10.48550/arXiv.2306.01984 , urldate =

    Cachay, Salva R. doi:10.48550/arXiv.2306.01984 , urldate =

    Pith/arXiv arXiv doi:10.48550/arxiv.2306.01984

  19. [24]

    and Hussaini, M

    Canuto, C. and Hussaini, M. and Quarteroni, A. and Zang, T. , year = 2010, month = nov, urldate =. Spectral

    2010

  20. [25]

    AlphaFolding: 4D Diffusion for Dynamic Protein Structure Prediction with Reference and Motion Guidance

    Cheng, Kaihui and Liu, Ce and Su, Qingkun and Wang, Jun and Zhang, Liwei and Tang, Yining and Yao, Yao and Zhu, Siyu and Qi, Yuan , year = 2024, month = dec, publisher =. doi:10.48550/arXiv.2408.12419 , urldate =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2408.12419 2024

  21. [26]

    doi:10.1101/2023.05.24.542194 , urldate =

    An All-Atom Protein Generative Model , author =. doi:10.1101/2023.05.24.542194 , urldate =

    doi:10.1101/2023.05.24.542194 2023

  22. [27]

    Proceedings of the National Academy of Sciences , volume =

    An All-Atom Protein Generative Model , author =. Proceedings of the National Academy of Sciences , volume =. doi:10.1073/pnas.2311500121 , urldate =

    work page doi:10.1073/pnas.2311500121

  23. [28]

    and Thompson, Tuscan R

    Corley, Nathaniel and Mathis, Simon and Krishna, Rohith and Bauer, Magnus S. and Thompson, Tuscan R. and Ahern, Woody and Kazman, Maxwell W. and Brent, Rafael I. and Didi, Kieran and Kubaney, Andrew and McHugh, Lilian and Nagle, Arnav and Favor, Andrew and Kshirsagar, Meghana and Sturmfels, Pascal and Li, Yanjing and Butcher, Jasper and Qiang, Bo and Scha...

    work page doi:10.1101/2025.08.14.670328 2025

  24. [29]

    doi:10.48550/arXiv.2410.09667 , urldate =

    Costa, Allan dos Santos and Mitnikov, Ilan and Pellegrini, Franco and Daigavane, Ameya and Geiger, Mario and Cao, Zhonglin and Kreis, Karsten and Smidt, Tess and Kucukbenli, Emine and Jacobson, Joseph , year = 2024, month = dec, number =. doi:10.48550/arXiv.2410.09667 , urldate =. arXiv , keywords =:2410.09667 , primaryclass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2410.09667 2024

  25. [31]

    Beyond Static Structures:

    Cui, Xinyue and Ge, Lingyu and Chen, Xia and Lv, Zexin and Wang, Suhui and Zhou, Xiaogen and Zhang, Guijun , year = 2025, month = jul, journal =. Beyond Static Structures:. doi:10.1093/bib/bbaf340 , abstract =

    work page doi:10.1093/bib/bbaf340 2025

  26. [33]

    Journal of Molecular Biology , volume =

    Increasing Temperature Accelerates Protein Unfolding without Changing the Pathway of Unfolding , author =. Journal of Molecular Biology , volume =. doi:10.1016/s0022-2836(02)00672-1 , abstract =

    work page doi:10.1016/s0022-2836(02)00672-1

  27. [34]

    , year = 2013, month = oct, journal =

    Deng, Nan-jie and Dai, Wei and Levy, Ronald M. , year = 2013, month = oct, journal =. How. doi:10.1021/jp401962k , urldate =

    work page doi:10.1021/jp401962k 2013

  28. [35]

    and Ozkan, S

    Dill, Ken A. and Ozkan, S. Banu and Shell, M. Scott and Weikl, Thomas R. , year = 2008, month = jun, journal =. The. doi:10.1146/annurev.biophys.37.092707.153558 , urldate =

    doi:10.1146/annurev.biophys.37.092707.153558 2008

  29. [36]

    and Harvey, M

    Doerr, S. and Harvey, M. J. and No. Journal of Chemical Theory and Computation , volume =. doi:10.1021/acs.jctc.6b00049 , urldate =

    work page doi:10.1021/acs.jctc.6b00049

  30. [37]

    Design of an Optimal

    Fain, Boris and Xia, Yu and Levitt, Michael , year = 2002, month = aug, journal =. Design of an Optimal. doi:10.1110/ps.0200702 , urldate =

    work page doi:10.1110/ps.0200702 2002

  31. [38]

    Faltings, Felix and Stark, Hannes and Jaakkola, Tommi and Barzilay, Regina , year = 2025, month = jul, publisher =. Protein. doi:10.48550/arXiv.2505.08041 , urldate =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2505.08041 2025

  32. [40]

    Computational Biology and Chemistry , volume =

    A Critical Address to Advancements and Challenges in Computational Strategies for Structural Prediction of Protein in Recent Past , author =. Computational Biology and Chemistry , volume =. doi:10.1016/j.compbiolchem.2025.108430 , urldate =

    work page doi:10.1016/j.compbiolchem.2025.108430 2025

  33. [41]

    BioMD: All-atom Generative Model for Biomolecular Dynamics Simulation

    Feng, Bin and Zhang, Jiying and Zhang, Xinni and Liu, Zijing and Li, Yu , year = 2025, publisher =. doi:10.48550/ARXIV.2509.02642 , urldate =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2509.02642 2025

  34. [42]

    and Sato, Satoshi , year = 2004, month = may, journal =

    Fersht, Alan R. and Sato, Satoshi , year = 2004, month = may, journal =. doi:10.1073/pnas.0402684101 , urldate =

    work page doi:10.1073/pnas.0402684101 2004

  35. [43]

    and Daggett, Valerie , year = 2002, month = feb, journal =

    Fersht, Alan R. and Daggett, Valerie , year = 2002, month = feb, journal =. Protein. doi:10.1016/S0092-8674(02)00620-7 , urldate =

    work page doi:10.1016/s0092-8674(02)00620-7 2002

  36. [45]

    and Bogatyreva, Natalya S

    Finkelstein, Alexei V. and Bogatyreva, Natalya S. and Ivankov, Dmitry N. and Garbuzynskiy, Sergiy O. , year = 2022, month = oct, journal =. Protein Folding Problem:. doi:10.1007/s12551-022-01000-1 , urldate =

    work page doi:10.1007/s12551-022-01000-1 2022

  37. [46]

    Geffner, Tomas and Didi, Kieran and Cao, Zhonglin and Reidenbach, Danny and Zhang, Zuobai and Dallago, Christian and Kucukbenli, Emine and Kreis, Karsten and Vahdat, Arash , year = 2025, month = jul, publisher =. La-. doi:10.48550/arXiv.2507.09466 , urldate =

    Pith/arXiv arXiv doi:10.48550/arxiv.2507.09466 2025

  38. [47]

    Proteina:

    Geffner, Tomas and Didi, Kieran and Zhang, Zuobai and Reidenbach, Danny and Cao, Zhonglin and Yim, Jason and Geiger, Mario and Dallago, Christian and Kucukbenli, Emine and Vahdat, Arash and Kreis, Karsten , year = 2025, month = mar, publisher =. Proteina:. doi:10.48550/arXiv.2503.00710 , urldate =

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.00710 2025

  39. [48]

    and Galzitskaya, Oxana V

    Glyakina, Anna V. and Galzitskaya, Oxana V. , year = 2020, month = jan, journal =. How. doi:10.3390/biom10020197 , urldate =

    work page doi:10.3390/biom10020197 2020

  40. [49]

    Gottlieb, David and Shu, Chi-Wang , year = 1997, month = jan, journal =. On the. doi:10.1137/S0036144596301390 , urldate =

    doi:10.1137/s0036144596301390 1997

  41. [50]

    Hanni, Jack and Ray, Dhiman , year = 2025, month = jan, publisher =. Data. doi:10.26434/chemrxiv-2025-vswkj , urldate =

    work page doi:10.26434/chemrxiv-2025-vswkj 2025

  42. [51]

    and Salmoral, Daniel

    Hekkelman, Maarten L. and Salmoral, Daniel. Protein Science : A Publication of the Protein Society , volume =. doi:10.1002/pro.70208 , urldate =

    doi:10.1002/pro.70208

  43. [52]

    Living Journal of Computational Molecular Science , volume =

    Enhanced Sampling Methods for Molecular Dynamics Simulations , author =. Living Journal of Computational Molecular Science , volume =. doi:10.33011/livecoms.4.1.1583 , urldate =. arXiv , keywords =:2202.04164 , primaryclass =

    Pith/arXiv arXiv doi:10.33011/livecoms.4.1.1583

  44. [53]

    Classifier-

    Ho, Jonathan and Salimans, Tim , year = 2022, month = jul, number =. Classifier-. doi:10.48550/arXiv.2207.12598 , urldate =. arXiv , keywords =:2207.12598 , primaryclass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2207.12598 2022

  45. [55]

    The Journal of Chemical Physics , volume =

    Quantitative Comparison of Adaptive Sampling Methods for Protein Dynamics , author =. The Journal of Chemical Physics , volume =. doi:10.1063/1.5053582 , abstract =

    work page doi:10.1063/1.5053582

  46. [56]

    and Baker, David , year = 2011, month = aug, journal =

    Huang, Po-Ssu and Ban, Yih-En Andrew and Richter, Florian and Andre, Ingemar and Vernon, Robert and Schief, William R. and Baker, David , year = 2011, month = aug, journal =. doi:10.1371/journal.pone.0024109 , urldate =

    work page doi:10.1371/journal.pone.0024109 2011

  47. [57]

    Sequence-

    Huguet, Guillaume and Vuckovic, James and Fatras, Kilian and. Sequence-. doi:10.48550/arXiv.2405.20313 , urldate =

    Pith/arXiv arXiv doi:10.48550/arxiv.2405.20313

  48. [58]

    Denoising Diffusion Probabilistic Models , booktitle =

    Ho, Jonathan and Jain, Ajay and Abbeel, Pieter , year = 2020, month = dec, series =. Denoising Diffusion Probabilistic Models , booktitle =

    2020

  49. [59]

    Proceedings of the National Academy of Sciences , volume =

    A Protein Dynamics--Based Deep Learning Model Enhances Predictions of Fitness and Epistasis , author =. Proceedings of the National Academy of Sciences , volume =. doi:10.1073/pnas.2502444122 , urldate =

    work page doi:10.1073/pnas.2502444122

  50. [60]

    Generative Models for Graph-Based Protein Design , author =

  51. [61]

    Nature , volume =

    Illuminating Protein Space with a Programmable Generative Model , author =. Nature , volume =. doi:10.1038/s41586-023-06728-8 , urldate =

    doi:10.1038/s41586-023-06728-8

  52. [62]

    Communications Chemistry , volume =

    Deep Generative Modeling of Temperature-Dependent Structural Ensembles of Proteins , author =. Communications Chemistry , volume =. doi:10.1038/s42004-025-01737-2 , urldate =

    work page doi:10.1038/s42004-025-01737-2

  53. [63]

    Folding of

    Jiang, Fan and Wu, Yun-Dong , year = 2014, month = jul, journal =. Folding of. doi:10.1021/ja502735c , urldate =

    work page doi:10.1021/ja502735c 2014

  54. [64]

    Jiang, Ping and Hansmann, Ulrich H. E. , year = 2012, month = jun, journal =. Modeling. doi:10.1021/ct3000469 , urldate =

    work page doi:10.1021/ct3000469 2012

  55. [66]

    doi:10.48550/arXiv.2402.04845 , urldate =

    Jing, Bowen and Berger, Bonnie and Jaakkola, Tommi , year = 2024, month = sep, number =. doi:10.48550/arXiv.2402.04845 , urldate =. arXiv , keywords =:2402.04845 , primaryclass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.04845 2024

  56. [67]

    Proceedings of the 41st

    Jing, Bowen and Berger, Bonnie and Jaakkola, Tommi , year = 2024, month = jul, series =. Proceedings of the 41st

    2024

  57. [68]

    doi:10.48550/ARXIV.2304.02198 , urldate =

    Jing, Bowen and Erives, Ezra and. doi:10.48550/ARXIV.2304.02198 , urldate =

    Pith/arXiv arXiv doi:10.48550/arxiv.2304.02198

  58. [69]

    Generative

    Jing, Bowen and St. Generative

  59. [70]

    Generative

    Jing, Bowen and St. Generative. doi:10.48550/arXiv.2409.17808 , urldate =

    Pith/arXiv arXiv doi:10.48550/arxiv.2409.17808

  60. [73]

    Jing, Yang and Li, Lei and Zhang, Jingtong , year = 2025, month = mar, number =. Solving. doi:10.48550/arXiv.2503.17829 , urldate =. arXiv , keywords =:2503.17829 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2503.17829 2025

  61. [74]

    doi:10.1021/acs.jctc.4c01620 , urldate =

    Jin, Yaowei and Huang, Qi and Song, Ziyang and Zheng, Mingyue and Teng, Dan and Shi, Qian , year = 2025, month = mar, journal =. doi:10.1021/acs.jctc.4c01620 , urldate =

    work page doi:10.1021/acs.jctc.4c01620 2025

  62. [75]

    and Fu, Xiang and Liao, Yi-Lun and Gharakhanyan, Vahe and Miller, Benjamin Kurt and Sriram, Anuroop and Ulissi, Zachary W

    Joshi, Chaitanya K. and Fu, Xiang and Liao, Yi-Lun and Gharakhanyan, Vahe and Miller, Benjamin Kurt and Sriram, Anuroop and Ulissi, Zachary W. , year = 2025, month = may, publisher =. All-Atom. doi:10.48550/arXiv.2503.03965 , urldate =

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.03965 2025

  63. [77]

    Highly Accurate Protein Structure Prediction with

    Jumper, John and Evans, Richard and Pritzel, Alexander and Green, Tim and Figurnov, Michael and Ronneberger, Olaf and Tunyasuvunakool, Kathryn and Bates, Russ and. Highly Accurate Protein Structure Prediction with. Nature , volume =. doi:10.1038/s41586-021-03819-2 , urldate =

    doi:10.1038/s41586-021-03819-2

  64. [78]

    and Bolhuis, P

    Juraszek, J. and Bolhuis, P. G. , year = 2006, month = oct, journal =. Sampling the Multiple Folding Mechanisms of. doi:10.1073/pnas.0606692103 , urldate =

    work page doi:10.1073/pnas.0606692103 2006

  65. [79]

    doi:10.48550/arXiv.2509.24779 , urldate =

    Kapu. doi:10.48550/arXiv.2509.24779 , urldate =. arXiv , keywords =:2509.24779 , primaryclass =

    work page doi:10.48550/arxiv.2509.24779

  66. [80]

    Elucidating the

    Karras, Tero and Aittala, Miika and Aila, Timo and Laine, Samuli , year = 2022, month = oct, number =. Elucidating the. doi:10.48550/arXiv.2206.00364 , urldate =. arXiv , keywords =:2206.00364 , primaryclass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2206.00364 2022

  67. [81]

    Polynomial propagators for classical molecular dynamics

    Polynomial Propagators for Classical Molecular Dynamics , author =. doi:10.48550/arXiv.2302.03516 , urldate =. arXiv , keywords =:2302.03516 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2302.03516

  68. [82]

    Kr. A Fast. Journal of Computational Chemistry , volume =. doi:10.1002/1096-987X(20010415)22:5<501::AID-JCC1021>3.0.CO;2-V , urldate =

    work page doi:10.1002/1096-987x(20010415)22:5

  69. [83]

    Compression in

    Kumar, Anand and Zhu, Xingquan and Tu, Yi-Cheng and Pandit, Sagar , editor =. Compression in. Intelligence. doi:10.1007/978-3-642-42057-3_4 , abstract =

    work page doi:10.1007/978-3-642-42057-3_4

  70. [85]

    Real-Time Protein

    Kumar, Amit and Balbach, Jochen , year = 2015, month = oct, journal =. Real-Time Protein. doi:10.1016/j.bbagen.2014.12.003 , abstract =

    work page doi:10.1016/j.bbagen.2014.12.003 2015

  71. [86]

    F1000Research , volume =

    Protein Unfolding Mechanisms and Their Effects on Folding Experiments , author =. F1000Research , volume =. doi:10.12688/f1000research.12070.1 , urldate =

    work page doi:10.12688/f1000research.12070.1

  72. [87]

    Lawrence, Jim and Bernal, Javier and Witzgall, Christoph , year = 2019, month = oct, journal =. A. doi:10.6028/jres.124.028 , urldate =

    work page doi:10.6028/jres.124.028 2019

  73. [88]

    DeepDriveMD: Deep-Learning Driven Adaptive Molecular Simulations for Protein Folding

    Lee, Hyungro and Ma, Heng and Turilli, Matteo and Bhowmik, Debsindhu and Jha, Shantenu and Ramanathan, Arvind , year = 2019, month = sep, publisher =. doi:10.48550/arXiv.1909.07817 , urldate =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1909.07817 2019

  74. [89]

    Lee, M. C. and Chan, Raymond K. W. and Adjeroh, Donald A. , year = 1997, month = dec, journal =. Quantization of. doi:10.1006/jvci.1997.0365 , urldate =

    work page doi:10.1006/jvci.1997.0365 1997

  75. [90]

    Levy, Ronald M and Haldane, Allan and Flynn, William F , year = 2017, month = apr, journal =. Potts. doi:10.1016/j.sbi.2016.11.004 , urldate =

    doi:10.1016/j.sbi.2016.11.004 2017

  76. [91]

    Science , volume =

    Scalable Emulation of Protein Equilibrium Ensembles with Generative Deep Learning , author =. Science , volume =. doi:10.1126/science.adv9817 , urldate =

    doi:10.1126/science.adv9817

  77. [92]

    Li, Zongyi and Kovachki, Nikola and Azizzadenesheli, Kamyar and Liu, Burigede and Bhattacharya, Kaushik and Stuart, Andrew and Anandkumar, Anima , year = 2021, month = may, number =. Fourier. doi:10.48550/arXiv.2010.08895 , urldate =. arXiv , keywords =:2010.08895 , primaryclass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2010.08895 2021

  78. [93]

    Generative

    Li, Zhengqi and Tucker, Richard and Snavely, Noah and Holynski, Aleksander , year = 2024, pages =. Generative. Proceedings of the

    2024

  79. [95]

    Science (New York, N.Y.) , volume =

    How Fast-Folding Proteins Fold , author =. Science (New York, N.Y.) , volume =. doi:10.1126/science.1208351 , abstract =

    doi:10.1126/science.1208351

  80. [96]

    Lipman, Yaron and Havasi, Marton and Holderrieth, Peter and Shaul, Neta and Le, Matt and Karrer, Brian and Chen, Ricky T. Q. and. Flow. doi:10.48550/arXiv.2412.06264 , urldate =

    Pith/arXiv arXiv doi:10.48550/arxiv.2412.06264

Showing first 80 references.