henhao
Sign in
BiophysicsNature
Peer-reviewed journal article

Learning millisecond protein dynamics from what is missing in NMR spectra

Hannah K. Wayment-SteeleiD, Gina El NesriD, Ramith HettiarachchiiD, Adedolapo M. OjoawoiD, Hasindu KariyawasamiD, Sergey OvchinnikoviD, Dorothee KerniD

Abstract

Many proteins biological functions rely on interconversions between multiple conformations occurring at micro-to millisecond ({micro}s-ms) timescales. A lack of standardized, large-scale experimental data has hindered obtaining a more predictive understanding of these motions. After curating >100 Nuclear Magnetic Resonance (NMR) relaxation datasets, we realized an observable for {micro}s-ms dynamics might be hiding in plain sight. Millisecond dynamics can cause NMR signals to broaden beyond detection, leaving some residues not assigned in the chemical shift datasets of ~10,000 proteins deposited in the Biological Magnetic Resonance Data Bank (BMRB)1. We made the bold assumption that residues missing assignments are exchange-broadened due to {micro}s-ms motions and trained various deep learning models to predict missing assignments. Strikingly, these models also predict exchange measured via NMR relaxation experiments, indicative of {micro}s-ms dynamics. The best of these models, which we named Dyna-1, leverages an intermediate layer of the multimodal language model ESM-32. Notably, dynamics directly linked to biological function, including enzyme catalysis and ligand binding, are particularly well predicted by Dyna-1, which parallels our findings that residues experiencing {micro}s-ms exchange are more conserved. We anticipate the datasets and models presented here will be transformative in unlocking the common language of dynamics and function.

PDFDOICodeData
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
16
FWCI
18.72
Field percentile
top 1%
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 3 (1 Jul 2026) · CC BY 4.0

Trained on missing NMR amide assignments in the Biological Magnetic Resonance Data Bank, Dyna-1 (ESM-3 layer 22 embeddings) predicts micro-to-millisecond conformational exchange seen in relaxation data.

Predicting micro-to-millisecond (µs-ms) protein motions lags behind structure prediction because no large, standardized experimental dynamics data exist to train or evaluate models. The work curates relaxation benchmarks (RelaxDB, RelaxDB-CPMG) and reuses missing NMR assignments in 9,381 BMRB proteins as a scalable proxy label for exchange, which makes supervised learning of millisecond dynamics possible.

Key results

  1. ShownFrozen embeddings from pretrained protein models, passed to a transformer classifier, predict which residues lack 15N backbone assignments. They beat baselines built on amino acid, secondary structure, SASA, normal-mode and conservation features. ESM-3 given both sequence and structure scored highest.

    Validation AUROC (80% identity split): ESM-3 layer 22 with sequence and structure 0.77; ESM-2 last layer 0.75; AF2 pair representation (structure only, no MSA) 0.71. Retraining on the most stringent split (30% sequence identity, 0.5 TM-score; smallest set ~2700 proteins) lowered Dyna-1 validation AUROC only to 0.74. · Predicting missing NMR assignments; Baseline training; Figure legends

  2. ShownDyna-1 carries over from predicting missing assignments to predicting relaxation-detected exchange (Rex) in assigned residues of the held-out RelaxDB benchmark. It works best on conserved core residues and on motifs tied to catalysis or ligand binding.

    RelaxDB AUROC (missing plus Rex) was 0.63, rising to 0.66 after removing phosphate-buffer artefacts: 10 of the 22 proteins with AUROC ≤ 0.5 (45%) were phosphate-interacting proteins measured in phosphate buffer. BPTI AUROC was 0.92. A binned analysis (n = 13,810 residues) favored conserved core residues. · Dyna-1 predicts µs-ms exchange; Figure legends

  3. ShownIn RelaxDB, conservation tracks µs-ms exchange. Exchanging residues are more conserved than unlabeled ones, which are more conserved than ps-ns mobile residues. Residues missing assignments match Rex residues in conservation, which supports missing assignments as a proxy for exchange.

    Conservation over n=12,584 residues from 133 RelaxDB proteins, using two-tailed independent-samples t-tests without multiple-comparison adjustment. Missing-assignment and Rex residues were statistically indistinguishable. In RelaxDB, 5% of residues show elevated R2/R1 and 4% have missing assignments. · Residues with exchange are most conserved; Curating proxy labels for exchange; Figure legends

  4. SuggestedA HYDRONMR-based reprocessing of 15N CPMG relaxation dispersion finds unsuppressed Rex from fast microsecond exchange that conventional analysis misses. Dyna-1 predictions agree with these residues (e.g., Cyclophilin A Arg148).

    Applied to 6 proteins with complete dispersion data (CypA, beta-lactamase, Aquifex aeolicus adenylate kinase, biliverdin reductase B, K-Ras, dual-specificity phosphatase 3). Adding unsuppressed Rex labels raised Dyna-1 AUROC for 5/6 per the Fig. 5 legend, while the main text says all improved. CypA was measured at 25°C and 600 MHz, B1 33 to 1000 Hz. · Dyna-1 reveals overlooked exchange; Complete dispersion data available; Identifying unsuppressed Rex in CPMG experiments; Figure legends

  5. SuggestedIn a prospective test run after the model's release, CPMG on two mBMRB-test proteins chosen for high versus low predicted exchange agreed with Dyna-1's ranking. The chitinase showed exchange in its active site and chitin-binding loops; yjbJ showed almost none.

    15N CPMG at 900 MHz: the Bryum coronatum chitinase (BMRB 11441) had extensive exchange, while yjbJ (BMRB 5105) showed dispersion only at residues 2 and 3. Per-residue AUROC was 0.62 for the chitinase and 0.72 for yjbJ. · Prospective experimental test of Dyna-1; NMR spectroscopy; Figure legends

Methods, data and reproducibility

Frozen embeddings from ESM-2, ESM-3 layers, or the AlphaFold2 (AF2) pair representation feed a 1D transformer that outputs a per-residue logit for a missing 15N backbone assignment. Probabilities adjusted for class imbalance are scored by AUROC against Rex labels from HYDRONMR rigid-tumbling fits of R1/R2/hetNOE data and from CPMG dispersion.

  • mBMRB: BMRB protein entities labeled per residue for a missing 15N backbone assignment, after filtering out deuterated, partially labeled, short, and largely unassigned entries · n = 9,381 proteins (from 12102 polypeptide entities) · 10.57967/hf/9277
  • mBMRB validation and test sets, held out by sequence clustering, with sequences similar to RelaxDB or RelaxDB-CPMG removed from all splits · n = 500 validation and 100 test proteins
  • RelaxDB: 15N backbone R1/R2/hetNOE datasets of apo, monomeric, folded domains, compiled from BMRB, supplements, digitized figures and private correspondence · n = 133 proteins (163 curated; 112 after phosphate-buffer removal) · 10.57967/hf/9277
  • BMRB entries with R2 relaxation data (January 2023 query), screened for monomeric, cofactor-free, non-denaturing samples · n = 344 entries queried; 91 retained
  • RelaxDB-CPMG: 15N CPMG dispersion sets (CypA, BlaC, Aquifex aeolicus AdK, BLVRB, K-Ras, VHR) plus literature-derived labels (CheY, HtrA2 PDZ, RNase, arginine kinase) · n = 6 complete datasets plus 4 literature-only proteins · K-Ras: 10.5061/dryad.j6q573nm0; BMRB 27888, 27462, 51320
  • PDB protein chains (from NMRbox, October 2024) aligned to mBMRB sequences to compare missing assignments with unresolved X-ray/EM residues and B-factors · n = 2680/5906 proteins with only NMR structures
  • Prospective 15N CPMG on recombinant 13C/15N-labeled chitinase BcChi-A and hypothetical protein yjbJ, expressed in E. coli BL21(DE3) · Bryum coronatum (chitinase); E. coli (yjbJ, YjbJ_ecoli) · n = 2 proteins · BMRB 11441, 5105; NCBI WP_058905752.1; 10.5281/zenodo.20834076
Methods & tools
ESM-3 (layer 22 embeddings, sequence plus structure input), ESM-2 (ESM2-t30-150M), AlphaFold2 (ColabDesign for structure models; model 1 outputs for the pair representation), ESMFold v1, HYDRONMR (rigid-tumbling R2/R1 estimates), ColabFold run_mmseqs2 and MMseqs2 easy-cluster, TMalign, ProDy (ANM/GNM normal mode baselines), MDTraj (Shrake-Rupley SASA) and DSSP, pynmrstar v3.3.1, PINT and ChemEx (CPMG lineshape and dispersion fitting), 15N CPMG relaxation dispersion NMR (900 MHz Bruker, TopSpin 3.2)
Code
https://github.com/WaymentSteeleLab/Dyna-1https://huggingface.co/spaces/gelnesr/Dyna-1
Data & accessions
https://github.com/WaymentSteeleLab/Dyna-1https://huggingface.co/spaces/gelnesr/Dyna-1https://doi.org/10.57967/hf/9277https://doi.org/10.5281/zenodo.20834076https://doi.org/10.5061/dryad.j6q573nm0https://bmrb.io/software/query/
Software & environment
pynmrstar v3.3.1; ESMFold v1; ESM2-t30-150M; AF2 via ColabDesign gamma branch; Seaborn v0.11.2; statannotations v0.2.3; TopSpin 3.2; frozen ESM/AF2 weights, AdamW at learning rate 1e-6, dropout 0.1
Assay conditions & compute
Each model trained on one NVIDIA A40 GPU in 12 hours to 2 days; GPU memory limited the size of the AF2-pair transformer. CPMG data were collected on a 900 MHz Bruker spectrometer with a cryogenic triple resonance probe.
Reproducibility gaps
Datasets are licensed for noncommercial use only, and a patent application covers the methods. Some RelaxDB and CPMG inputs came from private correspondence, and full dispersion data were unavailable for 4 CPMG proteins.
Limitations (6)
  • Missing assignments are an incomplete and noisy label for exchange: many assigned residues still exchange, so negatives are ambiguous and apparent false positives cannot be treated as ground truth. Misassignment errors add further label noise. · stated by the authors
  • Single-field R1/R2/hetNOE data cannot separate ps-ns motion from µs-ms exchange, so Rex is likely underestimated where both occur. Dyna-1 often gives high p(exchange) to ps-ns mobile regions (e.g., MAP kinase p38 gamma). · stated by the authors
  • Dyna-1 was trained only on proteins under 400 residues, and its ESM-3 base lacks side-chain information. It is weakest in beta-sheets (e.g., it missed putative ring-flip exchange at NtrC Tyr101), and almost three-fold more training data gave only marginal gains. · stated by the authors
  • The benchmark labels carry artefacts and blind spots: phosphate buffer inflated R2 for phosphate-binding proteins, 15N-only CPMG labels ignore 1H-only exchange, and mBMRB outliers (pH greater than 8, length greater than 250) remain uncorrected. · stated by the authors
  • Absolute discrimination is modest (RelaxDB AUROC 0.63 to 0.66; prospective chitinase AUROC 0.62), and the prospective validation covers only two proteins, so per-residue reliability on new systems is uncertain. · noted in this summary
  • Several reported AUROC gains (MerP, SMN Tudor domain, NtrC, MptpA, CypA) come from relabeling residues after looking at model predictions. Without a relabeling protocol fixed in advance, this risks confirmation bias. · noted in this summary

Who should read:Read this if you model protein conformational dynamics, analyze NMR relaxation or BMRB chemical shift data, or need a benchmark and pretrained predictor of residue-level µs-ms exchange.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISKrise '26Nesr '26Kolloff '23Kleiman '26Audagnotto '21Kümmerer '20
71% semantic matchBiophysics
Olivia E Krise, Michael P. Latham·bioRxiv·Aug 1, 2026