ARTINA showed that a deep-learning pipeline can take raw multidimensional protein NMR spectra and deliver resonance assignments and structures with no human intervention, reaching near-expert quality on 100 monomeric solution-state proteins. Its authors posed explicit enhancement paths: use of reduced spectral sets, integration of additional experimental restraints, end-to-end training of the model stack, and generalization to protein–ligand/protein–protein complexes, RNA, solid-state, and in-cell NMR. Later work by the same group resolved the reduced-spectra and AlphaFold-prior component for chemical-shift assignment. The surviving open core is to extend validated human-free raw-spectra automation to richer experimental input types, new target classes, and a single end-to-end trained system.

Background

Nuclear magnetic resonance (NMR) spectroscopy determines the structures of proteins in solution by measuring the resonances of atomic nuclei and converting them into distance and angle information. In a typical study, a series of two- to four-dimensional spectra (for example triple-resonance experiments that correlate backbone nuclei such as 1^1H, 13^{13}C, and 15^{15}N, and NOESY spectra whose cross-peaks, called nuclear Overhauser effects or NOEs, encode inter-proton distances) must be analyzed: signals must be identified ('peak picking'), mapped to specific atoms in the protein sequence ('resonance assignment'), and converted into structural restraints from which a three-dimensional structure is calculated with programs such as CYANA. Traditionally every step is done manually by trained spectroscopists over weeks or months. The deep-learning method ARTINA (Klukowski, Riek and Güntert, Nature Communications, 2022) automated this entire workflow: a family of deep residual convolutional neural networks (ResNets) detects and deconvolutes peaks in raw spectra, a graph neural network and a gradient-boosted-tree model perform assignment and structure ranking, and the pipeline runs without any human intervention or protein-specific parameters, producing assignments and a structure bundle within hours. Its quality was benchmarked on 100 proteins against human-expert results deposited in the Biological Magnetic Resonance Data Bank (BMRB) and the Protein Data Bank (PDB) structure repository. The paper's Discussion states concrete enhancement paths. First, the benchmark used all spectra from earlier manual studies although the sets are highly redundant, and the authors announced a future study on how much NMR measurement time could be saved with smaller spectral sets — a question with precedent in pre-deep-learning work (Scott, López-Méndez and Güntert, 2006, who determined a domain structure from varying spectrum subsets). Second, the collection of independently trained machine-learning models could be integrated into a single coherent system trained as a whole. Third, the method could consume additional experimental input data — known partial or stereospecific assignments (distinguishing experimentally which of two chemically similar groups a signal belongs to), vicinal 3J^3J couplings (dihedral-angle-dependent spin–spin interactions), residual dipolar couplings (RDCs, orientation-dependent interactions measured in weakly aligned samples), paramagnetic data (distance effects of attached spin labels, e.g. paramagnetic relaxation enhancement), and hydrogen bonds — plus structural priors from AlphaFold (Jumper et al., 2021), the highly accurate protein-structure prediction system. Fourth, the range of applications could be generalized to small-molecule–protein and protein–protein complexes, RNA, solid-state NMR — including magic-angle spinning (MAS), in which anisotropic interactions are averaged by fast sample rotation, enabling insoluble assemblies such as amyloid fibrils to be studied — and in-cell NMR (spectroscopy of proteins inside living cells). Which of these directions have since been demonstrated, and which remain open, defines the problem's current status.

Problem Statement

The problem, as posed by the ARTINA outlook (Nature Communications 2022) with the reduced-spectra/measurement-time and AlphaFold-prior component since demonstrated for chemical-shift assignment, is to remove the remaining scope limitations of fully automated, human-intervention-free NMR analysis of raw multidimensional spectra — currently valid only for monomeric solution-state proteins and a modular pipeline. Concretely: (1) make additional experimental information usable as input to the automated workflow — known partial assignments, stereospecific assignments, vicinal 3J^3J couplings, residual dipolar couplings, paramagnetic data, and hydrogen bonds; (2) generalize the workflow to new target classes — small-molecule–protein and protein–protein complexes, RNA, solid-state NMR, and in-cell NMR — with automated results validated against human-expert references in each class; and (3) replace the separately trained collection of machine-learning models (and the conventional algorithms it still relies on) with a single coherent system trained as a whole. These directions jointly serve one scientific objective: extending validated no-human-intervention NMR assignment and structure determination from standalone solution-state proteins toward the general biomolecular NMR analysis problem. Automation retains the source benchmark's standard throughout: no manual interventions and no protein-specific parameters to be adapted, with assignment and structure quality assessed against human-expert reference results. Progress on any one restraint type or target class is partial progress toward that shared objective, not a resolution of the whole: the restraint types listed in (1) and the target classes listed in (2) are all required rather than illustrative, and the task is resolved only when all three directions — every listed restraint type, every listed target class, and the end-to-end system — have been demonstrated by an automated method meeting this standard.

References

  1. Klukowski, P.; Riek, R.; Güntert, P. (2022). Rapid protein assignments and structures from raw NMR spectra with the deep learning technique ARTINA. Nature Communications 13, 6151. DOI 10.1038/s41467-022-33879-5. https://doi.org/10.1038/s41467-022-33879-5 (primary source)
  2. Klukowski, P.; Riek, R.; Güntert, P. (2023). Time-optimized protein NMR assignment with an integrative deep learning approach using AlphaFold and chemical shift prediction. Science Advances 9, eadi9323. DOI 10.1126/sciadv.adi9323. https://doi.org/10.1126/sciadv.adi9323
  3. Wetton, H.; Klukowski, P.; Riek, R.; Güntert, P. (2023). Chemical shift transfer: an effective strategy for protein NMR assignment with ARTINA. Frontiers in Molecular Biosciences 10, 1244029. DOI 10.3389/fmolb.2023.1244029. https://doi.org/10.3389/fmolb.2023.1244029
  4. Klukowski, P.; Damberger, F. F.; Allain, F. H.-T.; Iwai, H.; Kadavath, H.; Ramelot, T. A.; Montelione, G. T.; Riek, R.; Güntert, P. (2024). The 100-protein NMR spectra dataset: A resource for biomolecular NMR data analysis. Scientific Data 11, 30. DOI 10.1038/s41597-023-02879-5. https://doi.org/10.1038/s41597-023-02879-5
  5. Klukowski, P.; Riek, R.; Güntert, P. (2023). NMRtist: an online platform for automated biomolecular NMR spectra analysis. Bioinformatics 39, btad066. DOI 10.1093/bioinformatics/btad066. https://doi.org/10.1093/bioinformatics/btad066
  6. Rodina, N.; Sarkar, R.; Tsakalos, D.; Suladze, S.; Niu, Z.; Reif, B. (2024). Manual and automatic assignment of two different Aβ40 amyloid fibril polymorphs using MAS solid-state NMR spectroscopy. Biomolecular NMR Assignments 18, 201–212. DOI 10.1007/s12104-024-10189-z. https://doi.org/10.1007/s12104-024-10189-z
  7. Liu, S.; Chu, H.; Xie, Y.; Wu, F.; Mu, F.; Wei, J.; Ni, N.; Wang, C.; Zhang, J.; Chen, M.; Li, J.; Yu, F.; Fu, H.; Wang, S.; Tian, C.; Wang, Z.; Gao, Y. Q. (2025). Assisting and accelerating NMR assignment with restrained structure prediction. Communications Biology 8, 1067. DOI 10.1038/s42003-025-08466-1. https://doi.org/10.1038/s42003-025-08466-1
  8. Rasulov, U.; Wang, H. K.; Viennet, T.; Droemer, M. A.; Matosin, S.; Schindler, S.; Sun, Z.-Y.; Mureddu, L.; Vuister, G. W.; Robson, S. A.; Arthanari, H.; Kuprov, I. (2024). Protein NMR assignment by isotope pattern recognition. Science Advances 10, eado0403. DOI 10.1126/sciadv.ado0403. https://doi.org/10.1126/sciadv.ado0403
  9. Scott, A.; López-Méndez, B.; Güntert, P. (2006). Fully automated structure determinations of the Fes SH2 domain using different sets of NMR spectra. Magnetic Resonance in Chemistry 44, S83–S88. DOI 10.1002/mrc.1813. https://doi.org/10.1002/mrc.1813
  10. Jumper, J.; Evans, R.; Pritzel, A.; Green, T.; Figurnov, M.; Ronneberger, O.; Tunyasuvunakool, K.; Bates, R.; Žídek, A.; Potapenko, A.; Bridgland, A.; Meyer, C.; Kohl, S. A. A.; Ballard, A. J.; Cowie, A.; Romera-Paredes, B.; Nikolov, S.; Jain, R.; Adler, J.; Back, T.; Petersen, S.; Reiman, D.; Clancy, E.; Zielinski, M.; Steinegger, M.; Pacholska, M.; Berghammer, T.; Bodenstein, S.; Silver, D.; Vinyals, O.; Senior, A. W.; Kavukcuoglu, K.; Kohli, P.; Hassabis, D. (2021). Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589. DOI 10.1038/s41586-021-03819-2. https://doi.org/10.1038/s41586-021-03819-2
  11. Zhan, H.; Huang, Y.; Chen, Z. (2026). Recent progress in artificial intelligence enabled NMR spectroscopy: Methodologies, implementations, quality assessments, and prospects. Applied Physics Reviews 13, 011322. DOI 10.1063/5.0277355. https://doi.org/10.1063/5.0277355
  12. Müntener, T.; Abramson, D.; Stern, E.; Hertel, I.; Jankevicius, G.; Mas, G.; Folkers, G. E.; Wicky, B. I. M.; Hiller, S. (2026). Large-scale exploration of protein space by automated NMR. bioRxiv preprint. DOI 10.64898/2026.02.16.706194. https://doi.org/10.64898/2026.02.16.706194