Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

Protein Language Model Embeddings Improve Generalization of Implicit Transfer Operators

T0 review · 5 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read A coarse-grained transition-operator model conditioned on protein language model embeddings can reproduce equilibrium distributions of unseen fast-folding proteins with roughly tenfold less data and compute than a leading Boltzmann emulator

desk verdict PLaTITO is a solid incremental advance—pLM conditioning clearly helps the controlled ablation—but the SOTA-over-BioEmu claim is undermined by a temperature mismatch, a 5x rollout change, and missing significance tests. read the letter →

arxiv 2602.11216 v2 pith:YNCNYT77 submitted 2026-02-11 cs.LG physics.bio-ph

Panagiotis Antoniadis , Beatrice Pavesi , Simon Olsson , Ole Winther This is my paper
classification cs.LGphysics.bio-ph
keywords implicittransferoperatorflowmatchingmoleculardynamicssurrogateproteinlanguagemodelembeddingsequilibriumsamplingfast-foldingproteinscoarse-grainedtransferablegenerative
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that implicit transfer operators—models that learn the transition density between molecular states a nanosecond apart—can be made broadly transferable and data-efficient when conditioned on pretrained protein language model embeddings. Training a coarse-grained C-alpha backbone flow model on about 56 milliseconds of off-equilibrium molecular dynamics trajectories produces equilibrium samples of out-of-distribution fast-folding proteins that match or exceed a leading Boltzmann emulator baseline, at nearly a tenth of the compute. The same model recovers non-Arrhenius temperature-dependent folding and unfolding rates, suggesting it has learned physically meaningful kinetics rather than merely matching a static density. The contribution is an architecture and conditioning recipe that shifts the cost of learning dynamics from MD data to cheaply precomputed sequence embeddings.

What carries the argument

The central object is the implicit transfer operator: a conditional generative model of the transition density p(x_{t+Δt}|x_t, Δt, S, T) learned by rectified conditional flow matching, where a velocity field interpolates between Gaussian noise and the future backbone configuration. Iterating this learned one-step operator over rollout steps generates trajectories that approach the equilibrium distribution. The conditioning network computes a per-residue representation from the current backbone, time step, temperature, and sequence, optionally augmented by pretrained protein language model embeddings; the velocity network then predicts the flow velocity from this conditioning and the noise st

What would settle it

Run a Chapman-Kolmogorov consistency test on one of the fast-folding proteins: generate two-step transition samples by composing the learned 1 ns operator twice and compare the resulting distribution to the model's direct prediction with a 2 ns step. A mismatch indicates the learned operator is not a consistent Markov semi-group, undermining the claim that its iterated stationary distribution equals the MD equilibrium. Alternatively, simulate a protein whose side-chain mobility is known to gate backbone transitions; if the model's equilibrium distribution deviates from the MD reference, the C-

Watch

Extended reading notes

Core claim

PLaTITO learns a conditional flow-matching model of the 1 nanosecond transition density in coarse-grained C-alpha coordinates, conditioned on residue-level embeddings from a pretrained protein language model. On equilibrium sampling benchmarks, the largest variant attains the best reported free-energy accuracy and coverage on unseen fast-folding proteins, outperforming a leading Boltzmann emulator baseline across all metrics while trained with nearly 10 times less MD data (56 ms versus 216 ms) and roughly 8 times less GPU time (1,100 versus 9,216 hours). The paper also shows the learned temperature-conditioned operator reproduces non-Arrhenius folding and unfolding rates, and samples apo-lik

Load-bearing premise

The load-bearing assumption is that the C-alpha backbone coordinates alone form a Markov state for the 1 nanosecond transition density, so that iterating the learned flow model converges to the true Boltzmann equilibrium; the paper concedes it offers no formal guarantees of detailed balance, semi-group consistency, or stability.

Editorial extensions

If this is right

  • With protein language model conditioning, transferable molecular dynamics surrogates reach state-of-the-art equilibrium sampling on unseen proteins with nearly an order of magnitude less training data and compute than Boltzmann emulators.
  • The temperature-conditioned operator yields non-Arrhenius folding and unfolding rates, indicating it approximates the underlying transition propagator rather than a simple exponential rate law.
  • Because sequence embeddings are precomputed offline, adding this conditioning costs almost nothing at training or test time over the base model, and scaling both model capacity and embedding size further improves results.
  • Kinetic observables such as folding and unfolding mean first-passage times systematically underestimate MD reference timescales, which the paper attributes to a variational principle and interprets as a scaling limitation rather than a fundamental ceiling.
  • The model can explore apo-like conformations from holo initial states in some cryptic binding pocket systems where the baseline fails, though it does not yet form metastable basins in the most difficult cases.

Reading between the lines

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

  • Because the protein language model embeddings are computed from sequence alone, the same architecture could be applied to proteins with no MD data at all—training on related sequences could enable zero-shot dynamics prediction for metagenomic or uncharacterized proteins.
  • The paper's own limitation statement implies a testable extension: checking Chapman-Kolmogorov self-consistency of the learned operator would reveal how far the C-alpha Markov assumption can be pushed; if two-step and direct transitions disagree, the equilibrium benchmarks would be measuring the surrogate's stationary distribution rather than the true MD one.
  • The reported failure mode in cryptic pocket systems is described as energetic separation rather than mode coverage, suggesting that adding an energy-based correction or reweighting step could sharpen metastable basins without requiring additional molecular dynamics data.
  • The architecture is non-equivariant and backbone-only, so extending to all-atom representation or adding side-chain degrees of freedom could improve accuracy for side-chain-coupled allosteric or ligand-binding questions, at the cost of the data efficiency demonstrated here.
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

5 major / 5 minor

Summary. The paper proposes PLaTITO, a coarse-grained (Cα-only) implicit transfer operator (ITO) trained with flow matching to model transition densities p(x_{t+Δt} | x_t, S, T) for proteins. The model is trained on the mdCATH dataset and conditioned on ESM protein language model embeddings, optionally also on Proteina structure embeddings and LLM-derived annotations. The central claims are that pLM embeddings improve out-of-distribution generalization and data efficiency, that the scaled PLaTITO-Big outperforms the BioEmu Boltzmann emulator on fast-folding protein equilibrium sampling benchmarks while using substantially less MD data and compute, and that the temperature-conditioned model captures non-Arrhenius folding/unfolding kinetics.

Significance. If the central claims hold, the paper would demonstrate that a Cα-coarse-grained, transferable transition-operator model can approach MD equilibrium distributions for unseen proteins with far less trajectory data than a direct Boltzmann emulator. That would be a practically important result for generative MD. The paper has real strengths: it evaluates on an external fast-folding benchmark, uses a strict sequence-similarity train/test split, compares several conditioning ablations, reports inference throughput, and states its limitations openly. However, the central head-to-head comparison with BioEmu is currently weakened by documented evaluation asymmetries and the absence of statistical testing, so the claimed state-of-the-art status is not yet established.

major comments (5)
  1. [§5.2, Table 4] Trp-cage is sampled by all TITO models at 310 K even though the MD reference free-energy landscape is at 290 K, because 290 K is outside the 320–450 K training range. Since Table 1 aggregates free-energy errors over all 12 fast-folders, a perfect model at 310 K would still show a systematic error against the 290 K reference. This is a load-bearing evaluation asymmetry for the claim that PLaTITO-Big outperforms BioEmu 'across all metrics.' Please report per-system metrics and the aggregate excluding Trp-cage, or evaluate at 290 K via a procedure that does not rely on out-of-range conditioning.
  2. [§5.2, Table 3 and §B.2] NTL9 is generated with 5,000 rollout steps instead of the default 1,000 used in the main protocol, while BioEmu is evaluated with its default sampling hyperparameters. This is a system-specific post-hoc adjustment that can only improve convergence of the TITO models for NTL9. The footnote justification based on the MD dataset is not sufficient for an apples-to-apples comparison. Please report NTL9 at 1,000 steps, and if 5,000 steps are retained, apply an equivalent protocol to BioEmu or provide a quantitative argument that the extra rollout length does not affect the comparison.
  3. [Table 1, §5.2] No paired significance test is reported for the PLaTITO-Big versus BioEmu differences in MAE, RMSE, or Coverage across the 12 fast-folding systems. The reported ± values are aggregate dispersions over systems, not confidence intervals for the paired difference. The central claim that PLaTITO-Big 'outperforms BioEmu across all equilibrium sampling evaluation metrics' requires a paired test (or a per-system table showing the direction of the difference on a majority of systems). Please add this.
  4. [§4.1, Algorithms 2–3, Limitations] The equilibrium-sampling evaluation rests on the assumption that the Cα backbone coordinates form a sufficient Markov state for 1 ns transition densities, and that iterative rollout converges to the MD stationary distribution. The Limitations section correctly states that there are no formal guarantees of unbiased dynamics, detailed balance, or Chapman-Kolmogorov consistency. Given that this is the primary evidence for the SOTA claim, the paper needs an empirical validation of this assumption: e.g., a Chapman-Kolmogorov test on held-out systems, comparison of iterative rollouts with direct conditional samples, and per-system convergence diagnostics. Without this, the equilibrium metrics may partly reflect projection error or accumulated stochastic error rather than the learned dynamics.
  5. [Table 1 footnote, §5.2, Conclusion] The paper repeatedly states that PLaTITO-Big is trained with 'substantially less data' and a 'ten-fold reduction in training data' compared to BioEmu. The table footnote shows that BioEmu additionally used 131k AFDB structures and 502k experimental ΔG measurements. If the comparison is specifically about MD trajectory data, this should be stated explicitly; otherwise the data-efficiency claim conflates different data types. Please report the total training-data budget for both models in a consistent way.
minor comments (5)
  1. [§5.4, Figure 5] The MD reference rates shown as red squares are not described in the text: which temperatures do they correspond to, and are they from Lindorff-Larsen et al. or new simulations? Please add a source and error bars.
  2. [§5.5, Conclusion] The conclusion states that PLaTITO-Big surpasses BioEmu on 'cryptic binding pocket sampling,' but §5.5 reports recovery of the apo state in only one of the four systems and failure to form clearly metastable basins in all four. Please temper the conclusion to match the reported results.
  3. [Table 3] The row for TEMPERATURE reads '1 MD MD MD MD MD 350' with a footnote marker. This is confusing; please spell out the per-figure temperatures in the table or move the Trp-cage exception to the main text.
  4. [§4.3] Minor typo: 'can be proved useful' should be 'can prove useful.'
  5. [§A.2] For Villin, the ESM embedding is computed from the PDB sequence with NLE replaced by LEU. Please clarify whether the same substitution is used for all fast-folders with non-canonical residues and whether this could affect the comparison.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the main SOTA generalization claim is checked against external fast-folding MD benchmarks and pretrained ESM embeddings, so it does not reduce to the model's training objective; the only self-citation is the inherited ITO/TITO framework and is not load-bearing.

full rationale

Derivation chain inspected: Eq. (3)-(4) train v_θ to regress x_{t+Δt} − ε under a rectified-flow interpolant; conditioning includes x_t, S, T, Δt and optionally ESM/Proteina embeddings; Algorithm 3 iterates the learned transition model; Table 1 compares the resulting free-energy surfaces to reference MD for 12 fast-folding proteins. The benchmark systems are external and held out: §4.4 removes proteins with ≥40% sequence similarity to test proteins from the mdCATH training set, and the fast-folder data (Lindorff-Larsen et al., 2011, CHARMM22) are not used in training or in fitting any parameter. ESM-Cambrian embeddings are pretrained on protein sequence data, so the observed OOD improvement is an empirical claim about an external representation, not a quantity defined by the model's own loss. The ITO/TITO framework is taken from prior work with overlapping authorship (Schreiner et al., 2023; Diez et al., 2025a), and this is the only notable self-citation; it is not load-bearing because the framework's usefulness is here validated against external MD references, and no 'uniqueness theorem' or fitted ansatz is imported from those citations. §6 explicitly concedes 'no formal guarantees of unbiased dynamics, detailed balance, semi-group self-consistency (Chapman-Kolmogorov), or stability'; that is an assumption/limitation, not a circular reduction. The Trp-cage 310 K vs 290 K reference (Table 4) and NTL9 5,000-step rollout (Table 3) are protocol asymmetries that affect the credibility of the 'across all metrics' comparison, but they do not make the predicted equilibrium equal to the training target by construction. No step reduces Eq. X to Eq. Y by definition or renames a fitted parameter as a prediction. Score 2 reflects the minor self-citation, not a circular derivation.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The model introduces no new physical entities. The central claim rests on standard flow-matching theory plus domain assumptions about the Markovianity of coarse-grained dynamics and the usefulness of pLM embeddings. No model free parameters are fitted to the fast-folder benchmark itself.

free parameters (3)
  • Trp-cage sampling temperature = 310 K
    Reference MD for Trp-cage is 290 K; the model samples at 310 K because 290 K lies outside the 320–450 K training range (Appendix B.1, Table 4). This changes the target distribution while metrics are computed against the 290 K reference.
  • NTL9 rollout length = 5,000 rollout steps
    Default is 1,000; NTL9 is given 5,000 'to reach equilibrium' (Table 3 footnote). This is a per-system post-hoc adjustment that changes effective sampling effort.
  • Equilibrium benchmark projection/histogram choices = TICA lag 10 ns; 4 TICA components; final 10 ns of each 1 µs rollout retained; free-energy binning
    These hand-chosen analysis settings determine MAE/RMSE/Coverage values and are not sensitivity-tested in the paper.
assumptions (5)
  • domain assumption Langevin dynamics with potential U has the Boltzmann distribution as invariant measure.
    Invoked in §2.1 to identify equilibrium with the Boltzmann distribution.
  • domain assumption Conditional flow matching on linear interpolants provides a valid generative model of the transition density p(x_{t+Δt}|x_t).
    The CFM objective in Eqs. (2)–(3) is assumed sufficient; no convergence guarantee is given for iterative rollout.
  • domain assumption C-alpha-coordinate state is Markovian enough at 1 ns time steps that p(x_{t+Δt}|x_t, S, T) captures the dynamics needed for equilibrium sampling.
    Adopted in §4.1; the Limitations section concedes no Chapman-Kolmogorov consistency or detailed-balance guarantee.
  • domain assumption Pretrained ESM embeddings contain thermodynamic/evolutionary information useful for conformational dynamics.
    Operationalized in §4.3.I; supported only by the ablation in this paper and prior stability-prediction literature.
  • domain assumption mdCATH off-equilibrium trajectories at 320–450 K are a sufficient training distribution for OOD generalization to fast-folding proteins at 290–370 K.
    This transferability premise underlies the entire evaluation and is not independently justified beyond the reported results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Protein Language Model Embeddings Improve Generalization of Implicit Transfer Operators." pith.science (2026). https://pith.science/paper/YNCNYT77

@misc{pith2026260211216,
  author       = {Pith},
  title        = {Pith review of: Protein Language Model Embeddings Improve Generalization of Implicit Transfer Operators},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YNCNYT77}},
  note         = {Machine review of arXiv:2602.11216}
}
read the original abstract

Molecular dynamics (MD) is a central computational tool in physics, chemistry, and biology, enabling quantitative prediction of experimental observables as expectations over high-dimensional molecular distributions such as Boltzmann distributions and transition densities. However, conventional MD is fundamentally limited by the high computational cost required to generate independent samples. Generative molecular dynamics (GenMD) has recently emerged as an alternative, learning surrogates of molecular distributions either from data or through interaction with energy models. While these methods enable efficient sampling, their transferability across molecular systems is often limited. In this work, we show that incorporating auxiliary sources of information can improve the data efficiency and generalization of transferable implicit transfer operators (TITO) for molecular dynamics. We find that coarse-grained TITO models are substantially more data-efficient than Boltzmann Emulators, and that incorporating protein language model (pLM) embeddings further improves out-of-distribution generalization. Our approach, PLaTITO, achieves state-of-the-art performance on equilibrium sampling benchmarks for out-of-distribution protein systems, including fast-folding proteins. We further study the impact of additional conditioning signals such as structural embeddings, temperature, and large-language-model-derived embeddings on model performance.

Figures

Figures reproduced from arXiv: 2602.11216 by the authors.

Figure 1
Figure 1. PLaTITO generalizes to unseen protein systems while improving data efficiency. Given the molecular state of a protein system at physical time t, defined by backbone coordinates xt, amino-acid sequence S and temperature T, our proposed TITO models approximate the long-time transition density p(xt+∆t | xt, S, T, ∆t) for a given time step ∆t. To improve data efficiency, auxiliary representations are incorporated during… view at source ↗
Figure 3
Figure 3. Scalability of PLaTITO-Big with training compute. Equilibrium sampling metrics improve as training compute in￾creases, indicating effective scaling behavior. The red dashed line corresponds to the performance of BioEmu. Notably, PLaTITO￾Big converges within approximately 1,100 GPU hours, that is substantially less than the training cost required by BioEmu (9,216 GPU hours) highlighting the computational efficiency o… view at source ↗
Figure 4
Figure 4. Top: Free-energy surfaces projected into the two slow￾est TICA components of A3D estimated by PLaTITO-Big from 1,000 independent trajectories initialized either from unfolded (top row) or folded (middle) conformations. Distributions are shown at increasing rollout times (left to right) and compared to the MD reference distribution (rightmost column). Below: Time-trace of a long 120 µs trajectory by PLaTITO-Big proje… view at source ↗
Figures from the paper (7 more)
Figure 6
Figure 6. Figure 6: Free-energy surfaces of the local RMSD to the apo (y-axis) and holo (x-axis) reference states estimated from 100 independent trajectories sampled by PLaTITO-Big. Trajectories are initialized from the apo state (middle) and the holo state (right). PLaTITO-Big samples ap…
Figure 7
Figure 7. Figure 7: Free energy surfaces projected into the two slowest TICA components of BBA, Villin, Trp-cage, BBL, A3D, WW domain, NTL9 and Protein G sampled by our proposed TITO models compared to the MD reference distributions (left). 18 [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: Free energy surfaces projected into the two slowest TICA components of Protein B, Homeodomain and λ-repressor sampled by our proposed TITO models compared to the MD reference distributions (left) 19 [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: Free energy surfaces projected into the two slowest TICA components of BBA, Villin, Trp-cage, BBL, A3D, WW domain, NTL9 and Protein G sampled by PLaTITO-Big (middle) and compared to the MD reference distributions (left) and BioEmu (right). Squares (□) and triangles (△)…
Figure 10
Figure 10. Figure 10: Free energy surfaces projected into the two slowest TICA components of Protein B, Homeodomain and λ-repressor sampled by PLaTITO-Big (middle) and compared to the MD reference distributions (left) and BioEmu (right). Squares (□) and triangles (△) denote folded and unfo…
Figure 11
Figure 11. Figure 11: Free-energy surfaces projected into the two slowest TICA components for multiple fast-folding proteins, estimated by PLaTITO-Big from 1,000 independent trajectories initialized from unfolded conformations. Distributions are shown at increasing rollout times (10 ns, 10…
Figure 12
Figure 12. Figure 12: Free-energy surfaces of the local RMSD to the apo (y-axis) and holo (x-axis) reference states estimated from 100 independent trajectories sampled by PLaTITO-Big. Trajectories are initialized from the apo state (middle) and the holo state (right). Dashed lines denote t…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Design-CP: Context Parallelism for Design of Protein Nanoparticles

    cs.LG 2026-07 conditional novelty 6.0 of 10

    Context-parallel inference for RFdiffusion 3 enables end-to-end all-atom design of large symmetric protein nanoparticles on multi-GPU hardware without retraining.

Reference graph

Works this paper leans on

64 extracted references · 7 canonical work pages · cited by 1 Pith paper

  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]

    D., Doucet, A., Bronstein, M

    Akhound-Sadegh, T., Lee, J., Bose, J., Bortoli, V. D., Doucet, A., Bronstein, M. M., Beaini, D., Ravanbakhsh, S., Neklyudov, K., and Tong, A. Progressive inference-time annealing of diffusion models for sampling from boltzmann densities. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. URL https://openreview.net/forum?...

    2025

  3. [3]

    Albergo, M. S. and Vanden-Eijnden, E. Building normalizing flows with stochastic interpolants. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=li7qeBbCR1t

    2023

  4. [4]

    Kinetic analysis of folding and unfolding the 56 amino acid igg-binding domain of streptococcal protein g

    Alexander, P., Orban, J., and Bryan, P. Kinetic analysis of folding and unfolding the 56 amino acid igg-binding domain of streptococcal protein g. Biochemistry, 31 0 (32): 0 7243--7248, 1992

    1992

  5. [5]

    Pytorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation

    Ansel, J., Yang, E., He, H., Gimelshein, N., Jain, A., Voznesensky, M., Bao, B., Bell, P., Berard, D., Burovski, E., et al. Pytorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation. In Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems...

    2024

  6. [6]

    U ber die reaktionsgeschwindigkeit bei der inversion von rohrzucker durch s \

    Arrhenius, S. \"U ber die reaktionsgeschwindigkeit bei der inversion von rohrzucker durch s \"a uren. Zeitschrift f \"u r physikalische Chemie , 4 0 (1): 0 226--248, 1889

  7. [7]

    Two for one: Diffusion models and force fields for coarse-grained molecular dynamics

    Arts, M., Garcia Satorras, V., Huang, C.-W., Z\" u gner, D., Federici, M., Clementi, C., Noé, F., Pinsler, R., and van den Berg, R. Two for one: Diffusion models and force fields for coarse-grained molecular dynamics. Journal of Chemical Theory and Computation, 19 0 (18): 0 6151–6159, September 2023. ISSN 1549-9626. doi:10.1021/acs.jctc.3c00702. URL http:...

    doi:10.1021/acs.jctc.3c00702 2023

  8. [8]

    Bonneau, K., Pasos‐Trejo, A. S., Plainer, M., Sagresti, L., Venturin, J., Zaporozhets, I., Caruso, A., Rolando, E., Guljas, A., Klein, L., Schebek, M., Albani, F., López‐Ríos de Castro, R., El Machachi, Z., Giambagli, L., and Clementi, C. Breaking the barriers of molecular dynamics with deep‐learning: Opportunities, pitfalls, and how to navigate them. WIR...

    work page doi:10.1002/wcms.70064 2026

Show all 64 references
  1. [9]

    Costa, A. D. S., Mitnikov, I., Pellegrini, F., Daigavane, A., Geiger, M., Cao, Z., Kreis, K., Smidt, T., Kucukbenli, E., and JACOBSON, J. Equijump: Protein dynamics simulation via SO (3)-equivariant stochastic interpolants. In ICLR 2025 Workshop on Generative and Experimental ...

    2025

  2. [10]

    Costa, A. d. S., Ponnapati, M., Rubin, D., Smidt, T., and Jacobson, J. Accelerating protein molecular dynamics simulation with deepjump. arXiv preprint arXiv:2509.13294, 2025 b

    2025

  3. [11]

    P., Davidson, D., Saremi, S., Rackers, J., and Kleinhenz, J

    Daigavane, A., Vani, B. P., Davidson, D., Saremi, S., Rackers, J., and Kleinhenz, J. Jamun: Bridging smoothed molecular dynamics and score-based learning for conformational ensembles, 2024

    2024

  4. [12]

    V., Atance, S

    Diez, J. V., Atance, S. R., Engkvist, O., and Olsson, S. Generation of conformational ensembles of small molecules via surrogate model-assisted molecular dynamics. Machine Learning: Science and Technology, 5 0 (2): 0 025010, 2024

    2024

  5. [13]

    V., Schreiner, M., and Olsson, S

    Diez, J. V., Schreiner, M., and Olsson, S. Transferable generative models bridge femtosecond to nanosecond time-step molecular dynamics. arXiv preprint arXiv:2510.07589, 2025 a

    2025

  6. [14]

    V., Schreiner, M

    Diez, J. V., Schreiner, M. J., Engkvist, O., and Olsson, S. Boltzmann priors for implicit transfer operators. In The Thirteenth International Conference on Learning Representations, 2025 b . URL https://openreview.net/forum?id=pRCOZllZdT

    2025

  7. [15]

    Esm cambrian: Revealing the mysteries of proteins with unsupervised learning, 2024

    ESM Team . Esm cambrian: Revealing the mysteries of proteins with unsupervised learning, 2024. URL https://evolutionaryscale.ai/blog/esm-cambrian

    2024

  8. [16]

    K., Brock, K., Gal, Y., and Marks, D

    Frazer, J., Notin, P., Dias, M., Gomez, A., Min, J. K., Brock, K., Gal, Y., and Marks, D. S. Disease variant prediction with deep generative models of evolutionary data. Nature, 599 0 (7883): 0 91–95, October 2021. ISSN 1476-4687. doi:10.1038/s41586-021-04043-8. URL http://dx....

    2021 doi

  9. [17]

    M., Bengtsen, T., Olsen, L., Lindorff-Larsen, K., Ferkinghoff-Borg, J., and Boomsma, W

    Frellsen, J., Kassem, M. M., Bengtsen, T., Olsen, L., Lindorff-Larsen, K., Ferkinghoff-Borg, J., and Boomsma, W. Zero-shot protein stability prediction by inverse folding models: a free energy interpretation. arXiv preprint arXiv:2506.05596, 2025

    2025

  10. [18]

    J., Olsen, B., and Jaakkola, T

    Fu, X., Xie, T., Rebello, N. J., Olsen, B., and Jaakkola, T. S. Simulate time-integrated coarse-grained molecular dynamics with multi-scale graph networks. Transactions on Machine Learning Research, 2023. ISSN 2835-8856. URL https://openreview.net/forum?id=y8RZoPjEUl

    2023

  11. [19]

    Proteina: Scaling flow-based protein structure generative models

    Geffner, T., Didi, K., Zhang, Z., Reidenbach, D., Cao, Z., Yim, J., Geiger, M., Dallago, C., Kucukbenli, E., Vahdat, A., and Kreis, K. Proteina: Scaling flow-based protein structure generative models. In International Conference on Learning Representations (ICLR), 2025

    2025

  12. [20]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  13. [21]

    J., Oktay, D., Lin, Z., Verkuil, R., Tran, V

    Hayes, T., Rao, R., Akin, H., Sofroniew, N. J., Oktay, D., Lin, Z., Verkuil, R., Tran, V. Q., Deaton, J., Wiggert, M., Badkundri, R., Shafkat, I., Gong, J., Derry, A., Molina, R. S., Thomas, N., Khan, Y. A., Mishra, C., Kim, C., Bartie, L. J., Nemeth, M., Hsu, P. D., Sercu, T....

    2025 doi

  14. [22]

    Markov field models: Scaling molecular kinetics approaches to large molecular machines

    Hempel, T., Olsson, S., and Noé, F. Markov field models: Scaling molecular kinetics approaches to large molecular machines. Current Opinion in Structural Biology, 77: 0 102458, December 2022. ISSN 0959-440X. doi:10.1016/j.sbi.2022.102458. URL http://dx.doi.org/10.1016/j.sbi.20...

    2022

  15. [23]

    and Feig, M

    Janson, G. and Feig, M. Generation of protein dynamics by machine learning. Current Opinion in Structural Biology, 93: 0 103115, August 2025. ISSN 0959-440X. doi:10.1016/j.sbi.2025.103115. URL http://dx.doi.org/10.1016/j.sbi.2025.103115

    2025

  16. [24]

    Torsional diffusion for molecular conformer generation

    Jing, B., Corso, G., Chang, J., Barzilay, R., and Jaakkola, T. Torsional diffusion for molecular conformer generation. In Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., and Oh, A. (eds.), Advances in Neural Information Processing Systems, volume 35, pp.\ 24240--2...

    2022

  17. [25]

    Generative modeling of molecular dynamics trajectories

    Jing, B., St \"a rk, H., Jaakkola, T., and Berger, B. Generative modeling of molecular dynamics trajectories. Advances in Neural Information Processing Systems, 37: 0 40534--40564, 2024

    2024

  18. [26]

    and No\' e , F

    Klein, L. and No\' e , F. Transferable boltzmann generators. In Globerson, A., Mackey, L., Belgrave, D., Fan, A., Paquet, U., Tomczak, J., and Zhang, C. (eds.), Advances in Neural Information Processing Systems, volume 37, pp.\ 45281--45314. Curran Associates, Inc., 2024. URL ...

    2024

  19. [27]

    Timewarp: Transferable acceleration of molecular dynamics by learning time-coarsened dynamics

    Klein, L., Foong, A., Fjelde, T., Mlodozeniec, B., Brockschmidt, M., Nowozin, S., No \'e , F., and Tomioka, R. Timewarp: Transferable acceleration of molecular dynamics by learning time-coarsened dynamics. Advances in Neural Information Processing Systems, 36: 0 52863--52883, 2023

    2023

  20. [28]

    B., Osmark, P., Neergaard, T

    Kragelund, B. B., Osmark, P., Neergaard, T. B., Schiødt, J., Kristiansen, K., Knudsen, J., and Poulsen, F. M. The formation of a native-like structure containing eight conserved hydrophobic residues is rate limiting in two-state protein folding of acbp. Nature Structural Biolo...

    1999 doi

  21. [29]

    Langevin, P. et al. Sur la th \'e orie du mouvement brownien. CR Acad. Sci. Paris, 146 0 (530-533): 0 530, 1908

    1908

  22. [30]

    Y., Satorras, V

    Lewis, S., Hempel, T., Jim \'e nez-Luna, J., Gastegger, M., Xie, Y., Foong, A. Y., Satorras, V. G., Abdin, O., Veeling, B. S., Zaporozhets, I., et al. Scalable emulation of protein equilibrium ensembles with generative deep learning. Science, 389 0 (6761): 0 eadv9817, 2025

    2025

  23. [31]

    Out of many, one: Designing and scaffolding proteins at the scale of the structural universe with genie 2

    Lin, Y., Lee, M., Zhang, Z., and AlQuraishi, M. Out of many, one: Designing and scaffolding proteins at the scale of the structural universe with genie 2. arXiv preprint arXiv:2405.15489, 2024

    2024 arXiv

  24. [32]

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

    Lin, Z., Akin, H., Rao, R., Hie, B., Zhu, Z., Lu, W., Smetanin, N., Verkuil, R., Kabeli, O., Shmueli, Y., et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science, 379 0 (6637): 0 1123--1130, 2023

    2023

  25. [33]

    O., and Shaw, D

    Lindorff-Larsen, K., Piana, S., Dror, R. O., and Shaw, D. E. How fast-folding proteins fold. Science, 334 0 (6055): 0 517--520, 2011

    2011

  26. [34]

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

    2023

  27. [35]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Liu, X., Gong, C., and qiang liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=XVjTT1nw5z

    2023

  28. [36]

    Vampnets for deep learning of molecular kinetics

    Mardt, A., Pasquali, L., Wu, H., and No \'e , F. Vampnets for deep learning of molecular kinetics. Nature communications, 9 0 (1): 0 5, 2018

    2018

  29. [37]

    Language models enable zero-shot prediction of the effects of mutations on protein function

    Meier, J., Rao, R., Verkuil, R., Liu, J., Sercu, T., and Rives, A. Language models enable zero-shot prediction of the effects of mutations on protein function. In Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., and Vaughan, J. W. (eds.), Advances in Neural Information Pr...

    2021

  30. [38]

    mdcath: A large-scale md dataset for data-driven computational biophysics

    Mirarchi, A., Giorgino, T., and De Fabritiis, G. mdcath: A large-scale md dataset for data-driven computational biophysics. Scientific Data, 11 0 (1): 0 1299, 2024

    2024

  31. [39]

    Thermodynamic interpolation: A generative approach to molecular thermodynamics and kinetics

    Moqvist, S., Chen, W., Schreiner, M., N\" u ske, F., and Olsson, S. Thermodynamic interpolation: A generative approach to molecular thermodynamics and kinetics. Journal of Chemical Theory and Computation, 21 0 (5): 0 2535--2545, 2025

    2025

  32. [40]

    H., Brotzakis, Z

    Murtada, M. H., Brotzakis, Z. F., and Vendruscolo, M. Md-llm-1: A large language model for molecular dynamics, 2025

    2025

  33. [41]

    Flow matching for accelerated simulation of atomic transport in crystalline materials

    Nam, J., Liu, S., Winter, G., Jun, K., Yang, S., and Gómez-Bombarelli, R. Flow matching for accelerated simulation of atomic transport in crystalline materials. Nature Machine Intelligence, 7 0 (10): 0 1625–1635, October 2025. ISSN 2522-5839. doi:10.1038/s42256-025-01125-4. UR...

    2025 doi

  34. [42]

    Boltzmann generators: Sampling equilibrium states of many-body systems with deep learning

    Noé, F., Olsson, S., K\" o hler, J., and Wu, H. Boltzmann generators: Sampling equilibrium states of many-body systems with deep learning. Science, 365 0 (6457), September 2019. ISSN 1095-9203. doi:10.1126/science.aaw1147. URL http://dx.doi.org/10.1126/science.aaw1147

    2019 doi

  35. [43]

    G., Pérez-Hernández, G., Mey, A

    N\" u ske, F., Keller, B. G., Pérez-Hernández, G., Mey, A. S. J. S., and Noé, F. Variational approach to molecular kinetics. Journal of Chemical Theory and Computation, 10 0 (4): 0 1739–1752, March 2014. ISSN 1549-9626. doi:10.1021/ct4009156. URL http://dx.doi.org/10.1021/ct4009156

    2014 doi

  36. [45]

    and No \'e , F

    Olsson, S. and No \'e , F. Dynamic graphical models of molecular kinetics. Proceedings of the National Academy of Sciences, 116 0 (30): 0 15001--15006, 2019

    2019

  37. [46]

    Identification of slow molecular order parameters for markov model construction

    P \'e rez-Hern \'a ndez, G., Paul, F., Giorgino, T., De Fabritiis, G., and No \'e , F. Identification of slow molecular order parameters for markov model construction. The Journal of chemical physics, 139 0 (1), 2013

    2013

  38. [47]

    Piana, S., Lindorff-Larsen, K., and Shaw, D. E. How robust are protein folding simulations with respect to force field parameterization? Biophysical journal, 100 0 (9): 0 L47--L49, 2011

    2011

  39. [48]

    Consistent sampling and simulation: Molecular dynamics with energy-based diffusion models

    Plainer, M., Wu, H., Klein, L., G \"u nnemann, S., and Noe, F. Consistent sampling and simulation: Molecular dynamics with energy-based diffusion models. In ICML 2025 Generative AI and Biology (GenBio) Workshop, 2025. URL https://openreview.net/forum?id=O0axp1D6Ss

    2025

  40. [49]

    D., Sch \"u tte, C., and No \'e , F

    Prinz, J.-H., Wu, H., Sarich, M., Keller, B., Senne, M., Held, M., Chodera, J. D., Sch \"u tte, C., and No \'e , F. Markov models of molecular kinetics: Generation and validation. The Journal of chemical physics, 134 0 (17), 2011

    2011

  41. [50]

    D., Noé, F., and Olsson, S

    Raich, L., Meier, K., G\" u nther, J., Christ, C. D., Noé, F., and Olsson, S. Discovery of a hidden transient state in all bromodomain families. Proceedings of the National Academy of Sciences, 118 0 (4), January 2021. ISSN 1091-6490. doi:10.1073/pnas.2017427118. URL http://dx...

    2021 doi

  42. [51]

    and Weber, M

    R \"o blitz, S. and Weber, M. Fuzzy spectral clustering by pcca+: application to markov state models and data classification. Advances in Data Analysis and Classification, 7 0 (2): 0 147--179, 2013

    2013

  43. [52]

    Scalley, M. L. and Baker, D. Protein folding kinetics exhibit an arrhenius temperature dependence when corrected for the temperature dependence of protein stability. Proceedings of the National Academy of Sciences, 94 0 (20): 0 10636--10640, 1997

    1997

  44. [53]

    Efficient mapping of phase diagrams with conditional boltzmann generators

    Schebek, M., Invernizzi, M., Noé, F., and Rogal, J. Efficient mapping of phase diagrams with conditional boltzmann generators. Machine Learning: Science and Technology, 5 0 (4): 0 045045, November 2024. ISSN 2632-2153. doi:10.1088/2632-2153/ad849d. URL http://dx.doi.org/10.108...

    2024 doi

  45. [54]

    and Schmid, F

    Schindler, T. and Schmid, F. X. Thermodynamic properties of an extremely rapid protein folding reaction. Biochemistry, 35 0 (51): 0 16833--16842, 1996

    1996

  46. [55]

    Implicit transfer operator learning: Multiple time-resolution models for molecular dynamics

    Schreiner, M., Winther, O., and Olsson, S. Implicit transfer operator learning: Multiple time-resolution models for molecular dynamics. Advances in Neural Information Processing Systems, 36: 0 36449--36462, 2023

    2023

  47. [56]

    B., Bose, J., Lin, C., Klein, L., Bronstein, M

    Tan, C. B., Bose, J., Lin, C., Klein, L., Bronstein, M. M., and Tong, A. Scalable equilibrium sampling with sequential boltzmann generators. In Frontiers in Probabilistic Inference: Learning meets Sampling, 2025. URL https://openreview.net/forum?id=8f4nfS1iko. *

    2025

  48. [57]

    Tan, Y.-J., Oliveberg, M., and Fersht, A. R. Titration properties and thermodynamics of the transition state for folding: comparison of two-state and multi-state folding pathways. Journal of molecular biology, 264 0 (2): 0 377--389, 1996

    1996

  49. [58]

    Alphafold protein structure database: massively expanding the structural coverage of protein-sequence space with high-accuracy models

    Varadi, M., Anyango, S., Deshpande, M., Nair, S., Natassia, C., Yordanova, G., Yuan, D., Stroe, O., Wood, G., Laydon, A., et al. Alphafold protein structure database: massively expanding the structural coverage of protein-sequence space with high-accuracy models. Nucleic acids...

    2022

  50. [59]

    N., Kaiser, ., and Polosukhin, I

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

    2017

  51. [60]

    R., Zavadlav, J., Praprotnik, M., and Koumoutsakos, P

    Vlachas, P. R., Zavadlav, J., Praprotnik, M., and Koumoutsakos, P. Accelerated simulations of molecular systems through learning of effective dynamics. Journal of Chemical Theory and Computation, 18 0 (1): 0 538–549, December 2021. ISSN 1549-9626. doi:10.1021/acs.jctc.1c00809....

    2021 doi

  52. [61]

    J., and Raleigh, D

    Wang, M., Tang, Y., Sato, S., Vugmeyster, L., McKnight, C. J., and Raleigh, D. P. Dynamic nmr line-shape analysis demonstrates that the villin headpiece subdomain folds on the microsecond time scale. Journal of the American Chemical Society, 125 0 (20): 0 6032–6033, April 2003...

    2003 doi

  53. [62]

    Energy-based diffusion generator for efficient sampling of boltzmann distributions

    Wang, Y., Guo, L., Wu, H., and Zhou, T. Energy-based diffusion generator for efficient sampling of boltzmann distributions. Neural Networks, 194: 0 108126, February 2026. ISSN 0893-6080. doi:10.1016/j.neunet.2025.108126. URL http://dx.doi.org/10.1016/j.neunet.2025.108126

    2026

  54. [63]

    and Noé, F

    Wu, H. and Noé, F. Variational approach for learning markov processes from time series data. Journal of Nonlinear Science, 30 0 (1): 0 23–66, August 2019. ISSN 1432-1467. doi:10.1007/s00332-019-09567-y. URL http://dx.doi.org/10.1007/s00332-019-09567-y

    2019 doi

  55. [64]

    Projected metastable markov processes and their estimation with observable operator models

    Wu, H., Prinz, J.-H., and Noé, F. Projected metastable markov processes and their estimation with observable operator models. The Journal of Chemical Physics, 143 0 (14), October 2015. ISSN 1089-7690. doi:10.1063/1.4932406. URL http://dx.doi.org/10.1063/1.4932406

    2015 doi

  56. [65]

    and Bowman, G

    Zhang, S. and Bowman, G. R. Decrypting cryptic pockets with physics-based simulations and artificial intelligence. Current Opinion in Structural Biology, 96: 0 103215, February 2026. ISSN 0959-440X. doi:10.1016/j.sbi.2025.103215. URL http://dx.doi.org/10.1016/j.sbi.2025.103215

    2026

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.