From sequence to protein structure and conformational dynamics with AI/ML

Alexander M. Ille; Emily Anas; Michael B. Mathews; Stephen K. Burley

REVIEW

From sequence to protein structure and conformational dynamics with AI/ML

Alexander M. Ille, Emily Anas, Michael B. Mathews, Stephen K. Burley

Biomolecules · Submitted 18 Apr 2025 · 9 references

Pith's verdict: A transparent perspective proposing sequence-based prediction of conformational ensembles; the idea is plausible but the central claim is overstated without qualification.

T0 · AI review

How this review was made: An AI referee (deepseek-v4-flash) read the paper on August 16, 2026 and answered a fixed list of questions about it. T0 is the starting level every paper gets. The levels describe how deeply Pith has checked the paper, never the authors.

T0 AI review An AI read the paper and answered a fixed list of questions.
T1 Audit Automatic checks that the references and files the paper relies on exist and match.
T2 Compute The paper's code or calculations re-run with the same results.
T3 Formal statement The claims written out in a formal math language a computer can read.
T4 Proof Those claims proved, and the proof checked by computer.

How Pith checks papers →

· 2 major objections · 5 minor · Ask about this paper · The compiled record · Read the full review →

Challenge this review Comment (0) Read on arXiv ↗

Cite ▾

arXiv ID2504.14059 (version 1)

Pith IDOYA2BVOM, Pith's permanent record of this paper and its review

CitationAlexander M. Ille, Emily Anas, Michael B. Mathews, Stephen K. Burley. "From sequence to protein structure and conformational dynamics with AI/ML." arXiv:2504.14059 (2025). https://pith.science/paper/OYA2BVOM

Cite this reviewBibTeX for Pith's review of this paper

Claim paper

Why Pith says this

Highlight any text to comment on it.

Editor's note

read the letter

Dear colleague,

Quick take: this is a perspective, not a research result. The authors openly propose that 1D sequence determines not just static structure but also conformational dynamics, and they sketch an architecture for an AI/ML model that would train on sequence plus NMR ensembles. No implementation, no new data, no derivation. That is fine if read as a research agenda.

What is genuinely useful: the paper pulls together the recent MSA-subsampling AlphaFold2 variants (Del Alamo, Monteiro da Silva, Wayment-Steele) and the Mansoor VAE work, and points out that nobody has yet combined sequence input with NMR-derived ensembles for end-to-end ensemble prediction. The observation that there are over 10,000 single-chain NMR ensembles in the PDB is concrete and useful. Figure 1b's comparison of NMR versus AF2-subsampled ensembles for IL-13 is illustrative, though not a systematic benchmark. The writing is clear and the authors are honest about the hypothesis status.

The soft spot is the load-bearing claim. The main text says "we posit that 1D amino acid sequence determines both 3D structure and protein conformational dynamics" without the qualifier the abstract includes ("at least in part"). The stress-test note is right: the myoglobin example actually shows the same sequence adopting different ensembles depending on ligand binding, so sequence alone does not determine the ensemble; sequence plus environment does. The more defensible claim is Anfinsen's thermodynamic hypothesis with the environment folded in. The leap from AlphaFold2's success at static structure prediction to ensemble-population prediction is analogical, not empirical. Also, the argument "if sequence did not determine structure, AlphaFold2 would not have been so successful" is too simple—AF2 also uses MSA co-evolution and training on structures. None of this kills the proposal, but the strong form of the claim is unsupported.

My overall read: as a perspective, it is fine and worth engaging. As a research result, there is nothing to verify. The proposed architecture is a suggestion, not a method.

Recommendation: send it to peer review as a perspective/hypothesis paper, but require the authors to qualify the central claim and to discuss the conditional nature of conformational ensembles. It deserves serious refereeing, not desk rejection, because the agenda is timely and the literature synthesis is useful.

Referee report

Significance

If the central hypothesis holds, the paper identifies a timely research direction with broad potential impact: sequence-based prediction of conformational ensembles could extend the AlphaFold2 paradigm from static snapshots to functionally important dynamics. The authors are appropriately cautious in labeling their claim a hypothesis, and they provide a concrete, if conceptual, architecture and a plausible training-data resource (PDB NMR ensembles). The paper is clearly written and well referenced, and the discussion of MSAs, AF2 subsampling, and VAE latent-space sampling is accurate. Its main scholarly contribution is to crystallize an emerging research program and to propose a specific model design, which is a legitimate contribution for a perspective article.

Major objections (2)

[Biological sequence information and its relationship with protein structure and dynamics (p.3)] The sentence 'we posit that 1D amino acid sequence determines both 3D structure and protein conformational dynamics' is stated without the environmental qualifier that the abstract itself provides ('at least in part') and that the myoglobin example immediately requires: ligand binding, pH, post-translational modifications, and temperature all shape conformational ensembles. This ambiguity is load-bearing because the proposed model is to be trained on sequence alone, yet if the environment co-determines the ensemble, a sequence-only model is under-specified unless 'determines' is explicitly defined as conditional on a fixed environment. The main text should either adopt the qualified formulation throughout or explain why sequence alone suffices despite the myoglobin counterexample.
[Prediction of protein structure and dynamics with AI/ML (p.6)] The inference that AlphaFold2's success 'bolsters support for this hypothesis' and the extension of that reasoning to conformational dynamics is a non-sequitur. AlphaFold2 predicts a single low-energy structure; MSA-subsampling variants produce multiple structures, but their sampling frequencies have not been shown to match experimental thermodynamic populations—they plausibly reflect evolutionary covariance rather than equilibrium weights. The paper should explicitly acknowledge this gap and formulate the sequence-determines-dynamics claim as a falsifiable prediction (e.g., that ensembles predicted from sequence reproduce NMR populations under fixed conditions), rather than as a corollary of static-structure prediction success.

Minor objections (5)

[Figure 1b (p.5)] The figure shows an RMSF comparison between the NMR and AlphaFold2-predicted ensembles of interleukin 13, but the text does not describe how the ensembles were aligned or whether the RMSF difference is statistically meaningful; adding at least a qualitative description of the comparison method would help the reader interpret the figure.
[Page 3, paragraph 1] There is a typo: 'T ogether' should read 'Together.'
[Figure 1d caption (p.5)] The caption contains 'PBD ID 2LSO'; this should be 'PDB ID 2LSO.'
[Page 3, myoglobin example] The sentence 'structural re-arrangement is required for molecular oxygen binding' is an oversimplification; the cited literature (Kendrew et al., 1960; Shulman et al., 1970) documents ligand-induced conformational changes, but the phrasing suggests a static requirement rather than a dynamic population shift. Consider rewording to 'structural rearrangements accompany oxygen binding.'
[Page 6, paragraph on AlphaFold2 training data] The statement that AlphaFold2 was 'trained on structures determined by X-ray crystallography and cryogenic electron microscopy' is accurate for the original model's filtered training set, but the sentence could be clarified to note that some NMR structures were also present in the broader PDB dataset used; this nuance matters because the paper later proposes NMR data as an underused resource.

How Pith checks papers → · Full audit of this paper →

The paper

The paper claims

This paper posits that a protein's amino acid sequence determines not only its folded 3D structure but also its full repertoire of conformational motions, and argues that AI/ML models can be built to predict ensembles of protein…

The reading

The paper extends the classic sequence-to-structure hypothesis to conformational dynamics: it posits that the one-dimensional amino acid sequence determines both the three-dimensional fold and the distribution of conformations a protein samples. If the hypothesis holds, the motions that govern protein function would be readable from sequence data and predictable by machine learning, rather than requiring expensive experiments or physics-based simulations for each new protein. The authors point to the success of AlphaFold2, RoseTTAFold, and ESMFold as evidence that the sequence-to-structure half is sound, and to recent demonstrations that stochastic MSA subsampling makes AlphaFold2 produce multiple conformations resembling experimental states as early evidence for the dynamics half. They then propose a conceptual model architecture, combining attention-based components with a variational autoencoder, that would be trained on sequence data plus NMR-derived conformational ensembles, of which they note over 10,000 single-chain examples are available in the Protein Data Bank.

What carries the argument

The central object is the hypothesized mapping from amino acid sequence to a distribution over three-dimensional conformations, rather than to a single structure. The proposed machinery is a conceptual AI/ML architecture that combines attention-based sequence and structure processing, as in AlphaFold2, with a variational autoencoder that samples a latent space of conformational variation, as in the Mansoor et al. approach. The argument's hinge is that NMR spectroscopy and molecular dynamics simulations provide methodologically independent experimental links between sequence and dynamics, and that NMR ensemble data already in the Protein Data Bank can serve as training signal for such a model.

Extended reading notes

Core claim

The paper's central claim, stated in its own words, is that '1D amino acid sequence determines both 3D structure and protein conformational dynamics.' This is a deliberate extension of the Sequence Hypothesis and the Thermodynamic Hypothesis from a single native structure to the entire ensemble of states a protein occupies. The paper's evidence is indirect: AlphaFold2's accuracy shows that sequence carries enough information to determine a folded structure, and recent adaptations of AlphaFold2 that subsample multiple sequence alignments, along with a variational autoencoder trained on crystallographic and MD-derived conformations, show that multiple conformations can also be generated from learned sequence- or structure-based representations. The paper does not report a trained model; it conceptualizes an architecture and identifies NMR-derived ensembles as the key untapped training resource for building one.

Load-bearing premise

The paper's claim stands or falls on whether a protein's amino acid sequence alone carries enough information to fix not only the folded structure but also the full set of conformational states the protein can visit and their relative likelihoods.

Editorial extensions

If this is right

A sequence-to-ensemble model would let researchers predict the conformational states of a protein directly from its amino acid sequence, without an experimentally determined starting structure.
NMR-derived ensemble data, currently underused relative to crystallographic and cryoEM structures, would become a central training resource for structure prediction, expanding the value of those deposits.
The approach would make prediction of functionally important states, such as cryptic pockets or allosteric conformations, a routine part of sequence-based analysis, provided those states are represented in the training ensemble data.
The existing demonstrations that MSA subsampling with AlphaFold2 yields experimental-like conformations already hint that conformational diversity is latent in sequence information, supporting the feasibility of the proposed direction.

Reading between the lines

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

If the hypothesis is correct but sequence alone does not fix relative populations—for example when ligands or post-translational modifications shift the ensemble—an extension of the paper's proposal would condition the model on those factors while still using sequence as the primary input; the paper leaves this conditional variant implicit.
A near-term test of the proposal would be to train the described architecture on a subset of the PDB's NMR ensembles and benchmark predicted ensembles against held-out NMR entries; the paper's Figure 1c suggests enough data exist for such a benchmark.
The same reasoning could extend naturally to intrinsically disordered proteins, where the sequence determines a broad ensemble rather than a folded state; such proteins would be a stringent and informative test of the hypothesis.

Original abstract

The 2024 Nobel Prize in Chemistry was awarded in part for protein structure prediction using AlphaFold2, an artificial intelligence/machine learning (AI/ML) model trained on vast amounts of sequence and 3D structure data. AlphaFold2 and related models, including RoseTTAFold and ESMFold, employ specialized neural network architectures driven by attention mechanisms to infer relationships between sequence and structure. At a fundamental level, these AI/ML models operate on the long-standing hypothesis that the structure of a protein is determined by its amino acid sequence. More recently, AlphaFold2 has been adapted for the prediction of multiple protein conformations by subsampling multiple sequence alignments (MSAs). The deterministic relationship between sequence and structure was hypothesized over half a century ago with profound implications for the biological sciences ever since. Based on this relationship, we hypothesize that protein conformational dynamics are also determined, at least in part, by amino acid sequence and that this relationship may be leveraged for construction of AI/ML models dedicated to predicting ensembles of protein structures (i.e., distinct conformations). Accordingly, we conceptualized an AI/ML model architecture which may be trained on sequence data in combination with conformationally-sensitive structure data, coming primarily from nuclear magnetic resonance (NMR) spectroscopy. Sequence-informed prediction of protein structural dynamics has the potential to emerge as a transformative capability across the biological sciences, and its implementation could very well be on the horizon.

The compiled record

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. Numbers in the tables are the authors', checked by code. No sentence in this record was written by a model.

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 6,487 input tokens.

What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.17 · human participants 0.03 · controlled comparison 0.02 · randomized assignment 0.01 · states quantitative result 0.04 · compares with prior results 0.11 · mentions code or data availability 0.02.

The abstract, sentence by sentence

arXiv listing abstract v1. Results and the main claim are underlined; show every role.

The 2024 Nobel Prize in Chemistry was awarded in part for protein structure prediction using AlphaFold2, an artificial intelligence/machine learning (AI/ML) model trained on vast amounts of sequence and 3D structure data. AlphaFold2 and related models, including RoseTTAFold and ESMFold, employ specialized neural network architectures driven by attention mechanisms to infer relationships between sequence and structure. At a fundamental level, these AI/ML models operate on the long-standing hypothesis that the structure of a protein is determined by its amino acid sequence. More recently, AlphaFold2 has been adapted for the prediction of multiple protein conformations by subsampling multiple sequence alignments (MSAs). The deterministic relationship between sequence and structure was hypothesized over half a century ago with profound implications for the biological sciences ever since. Based on this relationship, we hypothesize that protein conformational dynamics are also determined, at least in part, by amino acid sequence and that this relationship may be leveraged for construction of AI/ML models dedicated to predicting ensembles of protein structures (i.e., distinct conformations). Accordingly, we conceptualized an AI/ML model architecture which may be trained on sequence data in combination with conformationally-sensitive structure data, coming primarily from nuclear magnetic resonance (NMR) spectroscopy. Sequence-informed prediction of protein structural dynamics has the potential to emerge as a transformative capability across the biological sciences, and its implementation could very well be on the horizon.

background contribution method result interpretation main claim

What the classifier reports for each sentence. The number is the probability it gave the label.
#	sentence	role	the authors' own work	states the main conclusion	strength
s0	The 2024 Nobel Prize in Chemistry was awarded in part for protein structure prediction using AlphaFold2, an artificial intelligence/machine learning (AI/ML) model trained on vast amounts of sequence and 3D structure data.	background 1.00	0.03	0.03	asserted 1.00
s1	AlphaFold2 and related models, including RoseTTAFold and ESMFold, employ specialized neural network architectures driven by attention mechanisms to infer relationships between sequence and structure.	background 0.99	0.04	0.03	asserted 1.00
s2	At a fundamental level, these AI/ML models operate on the long-standing hypothesis that the structure of a protein is determined by its amino acid sequence.	background 1.00	0.05	0.04	asserted 0.96
s3	More recently, AlphaFold2 has been adapted for the prediction of multiple protein conformations by subsampling multiple sequence alignments (MSAs).	background 0.98	0.16	0.04	asserted 0.99
s4	The deterministic relationship between sequence and structure was hypothesized over half a century ago with profound implications for the biological sciences ever since.	background 1.00	0.03	0.03	asserted 1.00
s5	Based on this relationship, we hypothesize that protein conformational dynamics are also determined, at least in part, by amino acid sequence and that this relationship may be leveraged for construction of AI/ML models dedicated to predicting ensembles of protein structures (i.e., distinct conformations).	contribution 0.55	0.27	0.43	hedged possibility 0.80
s6	Accordingly, we conceptualized an AI/ML model architecture which may be trained on sequence data in combination with conformationally-sensitive structure data, coming primarily from nuclear magnetic resonance (NMR) spectroscopy.	method 0.77	0.80	0.09	hedged possibility 0.53
s7	Sequence-informed prediction of protein structural dynamics has the potential to emerge as a transformative capability across the biological sciences, and its implementation could very well be on the horizon.	interpretation 0.64	0.05	0.27	hedged possibility 1.00

What kind of paper the classifier reports this is: position or commentary 0.64 · evidence from argument 0.86.

What was examined, and what was not

sources read

arXiv listing abstract v1

not examined

full text
figures and tables
references and citing sentences
relations to other papers

levels run

P abstract

not run

S sections
T body sentences
C citations
R cross-paper relations

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. Every number in the measurements table is the authors', checked by code. Record as JSON

References

Works cited (9)

[1]

1 From sequence to protein structure and conformational dynamics with AI/ML Alexander M. Ille1,*, Emily Anas2, Michael B. Mathews3,4, and Stephen K. Burley5,6,7,8,9 1Rutgers Cancer Institute, Rutgers, The State University of New Jersey, Newark, NJ 2College of Computing, Georgia Institute of T echnology, Atlanta, GA 3Department of Medicine, Rutgers New Jer...

View on Pith 2024
[2]

Towards sequence/structure-centric prediction of protein conformational dynamics with AI/ML. (a) Biological sequence information (red) encodes biophysical properties (blue) across a spectrum that spans from information storage to biological function. (b) NMR-determined (beige) (PDB ID 1GA3) and AlphaFold-predicted (teal) conformational ensembles of interl...

View on Pith 2022
[5]

to process the decoded structural information into 3D coordinate structures, sequence data were not used for training the underlying model (Mansoor et al., 2024). Nevertheless, the Mansoor et al. approach does provide a promising methodology for generating informative conformational ensembles from structural input. Collectively, these approaches offer enc...

View on Pith 2024
[11]

https://doi.org/10.7554/eLife.75751 Frauenfelder, H., Sligar, S. G., & Wolynes, P . G. (1991). The energy landscapes and motions of proteins. Science, 254(5038), 1598-1603. https://doi.org/10.1126/science.1749933 Henzler-Wildman, K., & Kern, D. (2007). Dynamic personalities of proteins. Nature, 450(7172), 964-972. https://doi.org/10.1038/nature06522 Hwang...

View on Pith Pith review arXiv doi:10.48550/arxiv.1312.6114 1991
[2018]

have been profound. This relationship will continue to play an important role at the nexus of AI/ML, data science, and biology. Acknowledgments Molecular graphics were prepared using UCSF ChimeraX, developed by the Resource for Biocomputing, Visualization, and Informatics at the University of California, San Francisco, with support from the National Insti...

View on Pith 2021
[2021]

and ESMFold (Lin et al., 2023), rely on biological sequence data for prediction of protein structure via attention-based (Vaswani et al.,

View on Pith 2023
[2024]

and other sources was also instrumental. AlphaFold2 and RoseTTAFold rely on multiple sequence alignments (MSAs), which carry information about co-evolution of pairs of amino acid residues, to inform 3D structure prediction. Remarkably, structure prediction accuracy arises in the early sequence-based stages of the AlphaFold2 model architecture (Jumper et a...

View on Pith 2021
[2025]

Nucleic Acids Res. https://doi.org/10.1093/nar/gkae1010 Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N.,…Polosukhin, I. (2017). Attention is All you Need Advances in Neural Information Processing Systems, https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf Watson, J. D., & Cric...

View on Pith doi:10.1093/nar/gkae1010 2017
[2464]

https://doi.org/10.1038/s41467-024-46715-9 Pall, S., Zhmurov, A., Bauer, P ., Abraham, M., Lundborg, M., Gray, A.,…Lindahl, E. (2020). Heterogeneous parallelization and acceleration of molecular dynamics simulations in GROMACS. J Chem Phys, 153(13), 134110. https://doi.org/10.1063/5.0018516 Shulman, R. G., Wuthrich, K., Yamane, T ., Patel, D. J., & Blumbe...

View on Pith doi:10.1038/s41467-024-46715-9 2020

Discussion

Disagree with the review, or know something it missed? Leave a comment, highlight any passage above to comment on it, or challenge this review formally.

[1] [1]

1 From sequence to protein structure and conformational dynamics with AI/ML Alexander M. Ille1,*, Emily Anas2, Michael B. Mathews3,4, and Stephen K. Burley5,6,7,8,9 1Rutgers Cancer Institute, Rutgers, The State University of New Jersey, Newark, NJ 2College of Computing, Georgia Institute of T echnology, Atlanta, GA 3Department of Medicine, Rutgers New Jer...

View on Pith 2024

[2] [2]

Towards sequence/structure-centric prediction of protein conformational dynamics with AI/ML. (a) Biological sequence information (red) encodes biophysical properties (blue) across a spectrum that spans from information storage to biological function. (b) NMR-determined (beige) (PDB ID 1GA3) and AlphaFold-predicted (teal) conformational ensembles of interl...

View on Pith 2022

[3] [5]

to process the decoded structural information into 3D coordinate structures, sequence data were not used for training the underlying model (Mansoor et al., 2024). Nevertheless, the Mansoor et al. approach does provide a promising methodology for generating informative conformational ensembles from structural input. Collectively, these approaches offer enc...

View on Pith 2024

[4] [11]

https://doi.org/10.7554/eLife.75751 Frauenfelder, H., Sligar, S. G., & Wolynes, P . G. (1991). The energy landscapes and motions of proteins. Science, 254(5038), 1598-1603. https://doi.org/10.1126/science.1749933 Henzler-Wildman, K., & Kern, D. (2007). Dynamic personalities of proteins. Nature, 450(7172), 964-972. https://doi.org/10.1038/nature06522 Hwang...

View on Pith Pith review arXiv doi:10.48550/arxiv.1312.6114 1991

[5] [2018]

have been profound. This relationship will continue to play an important role at the nexus of AI/ML, data science, and biology. Acknowledgments Molecular graphics were prepared using UCSF ChimeraX, developed by the Resource for Biocomputing, Visualization, and Informatics at the University of California, San Francisco, with support from the National Insti...

View on Pith 2021

[6] [2021]

and ESMFold (Lin et al., 2023), rely on biological sequence data for prediction of protein structure via attention-based (Vaswani et al.,

View on Pith 2023

[7] [2024]

and other sources was also instrumental. AlphaFold2 and RoseTTAFold rely on multiple sequence alignments (MSAs), which carry information about co-evolution of pairs of amino acid residues, to inform 3D structure prediction. Remarkably, structure prediction accuracy arises in the early sequence-based stages of the AlphaFold2 model architecture (Jumper et a...

View on Pith 2021

[8] [2025]

Nucleic Acids Res. https://doi.org/10.1093/nar/gkae1010 Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N.,…Polosukhin, I. (2017). Attention is All you Need Advances in Neural Information Processing Systems, https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf Watson, J. D., & Cric...

View on Pith doi:10.1093/nar/gkae1010 2017

[9] [2464]

https://doi.org/10.1038/s41467-024-46715-9 Pall, S., Zhmurov, A., Bauer, P ., Abraham, M., Lundborg, M., Gray, A.,…Lindahl, E. (2020). Heterogeneous parallelization and acceleration of molecular dynamics simulations in GROMACS. J Chem Phys, 153(13), 134110. https://doi.org/10.1063/5.0018516 Shulman, R. G., Wuthrich, K., Yamane, T ., Patel, D. J., & Blumbe...

View on Pith doi:10.1038/s41467-024-46715-9 2020