Rapid protein assignments and structures from raw NMR spectra with the deep learning technique ARTINA

Method

• ARTINA uses neural networks to identify peak positions and overlapping signals, alongside other machine-learning models for chemical-shift estimation and structure-proposal selection.Existing algorithms handle resonance assignment, torsion-angle restraint generation, NOESY assignment, and structure calculation.Source: completing the NMR measurements. Our… · p. 2
• The workflow combines machine-learning components with established algorithms for resonance assignment, restraint generation, NOESY assignment, and structure calculation.These include FLYA, TALOS-N, and CYANA.Source: completing the NMR measurements. Our… · p. 2

Training data

• ARTINA’s development required collecting and standardizing complete multidimensional NMR spectra for 100 proteins because such spectra are generally not archived in public repositories.Source: completing the NMR measurements. Our… · p. 2

Results

• ARTINA’s automated structure-determination results were reported as on par with those achieved through weeks or months of human expert labor.Source: completing the NMR measurements. Our… · p. 3

Results

• The standardized benchmark comprises 1329 spectra from 100 proteins, addressing the scarcity of large, standardized protein-NMR spectral datasets.The benchmark covers proteins of 35–175 residues.Source: Results · p. 3, manually identified chemical shifts… · p. 3

manually identified chemical shifts (usually depositions at the Biological Magnetic Resonance Data

• 1.44 Å median backbone RMSD to the PDB reference was achieved across the 100-protein benchmark, with additional validation indicating equivalent overall quality to reference structures.The average deviation between actual and predicted RMSDs was 0.35 Å, with a linear correlation coefficient of 0.77.Source: manually identified chemical shifts… · pp. 3–4

the manually prepared assignments, including both “strong” (high-reliability) and “weak” (tentative)

• 90.39% of chemical shifts were correctly assigned on average; backbone accuracy was 96.03%, compared with 86.50% for side chains.Lysine/arginine and aromatic side chains had lower accuracies than other assignments.Source: manually identified chemical shifts… · p. 4, the manually prepared assignments,… · p. 4

performance of deep learning-based visual spectrum analysis, presented qualitatively in Figs. 6–7,

• 1,168,739 cross-peaks were identified with high confidence (≥ 0.50), and ARTINA can return peak positions and intensities for downstream tasks.Source: performance of deep learning-based… · pp. 5, 7
• The largest structural deviations were associated mainly with displaced terminal secondary-structure elements or inaccurate loop conformations; excluding loops gave the lowest RMSD for 66 of 100 proteins.For 18 of the 19 proteins whose RMSD decreased by more than 0.5 Å after excluding a region, a terminal element or loop was the primary discrepancy source.Source: performance of deep learning-based… · pp. 5–6
• Refinement cycles reduced median backbone RMSD from 1.56 Å in cycle 1 to 1.44 Å in cycle 3, with assignment accuracy improving across cycles, particularly for side chains.Large, challenging systems benefited substantially from later cycles.Source: performance of deep learning-based… · p. 6
• Adding 4D CC-NOESY improved mean backbone RMSD by 0.15 Å, from 1.88 Å to 1.73 Å, with less than 1% impact on chemical-shift assignment accuracy.Effects varied: three proteins improved by more than 1 Å, while 2KCD’s RMSD decreased by more than 1 Å when 4D CC-NOESY was excluded; the authors suggest 2D/3D information is generally sufficient.Source: performance of deep learning-based… · pp. 6–7
• With NOESY spectra, side-chain assignment accuracy was 86.00%, compared with 73.70% without them; backbone accuracy rose from 92.40% with backbone spectra alone to 96.60% with the full set.The full-set results were from 45 benchmark proteins with the specified backbone experiments available.Source: performance of deep learning-based… · p. 7

• ARTINA delivers peak identification, resonance assignments, and structures for 100 proteins in hours rather than weeks or months of human work.The workflow has no manual interventions or protein-specific parameters to adapt.Source: Discussion · p. 7
• ARTINA can also provide peak picking and resonance assignments for NMR studies of ligand binding or dynamics that do not seek a structure.The webserver handles format conversions and supports different types of high-performance hardware.Source: Discussion · p. 8
• Input preparation remains critical: chemical-shift deviations above 0.03 ppm for 1H or 0.4 ppm for 13C/15N can be detrimental.With decreasing spectral quality, ARTINA progressively misses real signals.Source: Discussion · p. 8
• Unlike human experts, ARTINA currently lacks contextual information because it identifies signals from relatively small, local spectrum excerpts.Its limits regarding protein size and spectrum quality are similar to those of a trained spectroscopist.Source: Discussion · p. 8
• The benchmark used all spectra available from prior manual structure determination, often with redundant assignment information; smaller sets may also yield good results.The authors anticipate concomitant savings in NMR measurement time, but plan to investigate this in future work.Source: Discussion · p. 8
• ARTINA could be improved by integrating its separately trained models and conventional algorithms into a coherent system trained as a whole.The authors expect its training data to expand through the website, adding quantity and variety that could support retraining.Source: Discussion · p. 8
• ARTINA may incorporate additional experimental inputs, including known partial H-bonds.The discussion also proposes using structural information with reduced spectra and extending applications to complexes, RNA, solid-state NMR, and in-cell NMR.Source: Discussion · p. 9

• Structural information, including from AlphaFold, could be combined with reduced NMR spectra for rapid structure-based assignment.Source: assignments, stereospecific… · p. 9
• ARTINA could be generalized to small molecule-protein and protein-protein complexes, RNA, solid-state NMR, and in-cell NMR.Source: assignments, stereospecific… · p. 9
• ARTINA may accelerate multidimensional NMR studies requiring spectral analysis and chemical-shift assignments, including protein-ligand interaction studies.The authors note that structure determination is only one possible application.Source: assignments, stereospecific… · p. 9

Methods

• A standardized benchmark of 1329 spectra was assembled from varied sources and formats, with data scarcity, inconsistent deposition standards, and file-to-sample matching among its collection challenges.No additional spectrum processing, such as baseline correction, was performed.Source: Methods · p. 9
• With a protein sequence and supported NMR spectra as input, ARTINA returns cross-peak positions, chemical-shift assignments, NOESY distance restraints, and a protein structure in 4–20 hours without human involvement.The computation time depends on protein size, number of spectra, and hardware load.Source: Methods · p. 10

ARTINA interchangeably executes refinement cycles. The chemical shift refinement cycle provides

• ARTINA iterates chemical-shift and structure refinement: GNN-predicted missing shifts guide another FLYA assignment, while ranked CYANA structure proposals inform a subsequent assignment run.The described refinement cycles are each executed three times.Source: positions are identified in… · p. 10, ARTINA interchangeably executes… · p. 11

positions are identified in frequency-domain NMR spectra using deep residual neural networks

• Deep residual networks identify cross-peak positions from multidimensional spectra, and a second network deconvolves selected signals into up to three peak components.The peak-picking model treats extrema as candidate objects and scores them as likely signals or artifacts; the deconvolution model is trained on synthetic examples.Source: ARTINA interchangeably executes… · pp. 11–12

Unfolding a k-dimensional spectrum is defined as a discrete optimization problem, solved

• KDE-based unfolding estimates true-peak distributions from BMRB shift data, but requires nonoverlapping regular and folded signals or different signal signs and an uncropped folded dimension.Because 4D CC-NOESY does not meet these conditions, its peaks are unfolded by matching signals in 3D 13C-NOESY and retained only when confirmed there.Source: ARTINA interchangeably executes… · p. 12, Unfolding a k-dimensional spectrum is… · p. 13

ARTINA interchangeably executes refinement cycles. The chemical shift refinement cycle provides

• A DeepGCN predicts missing chemical shifts from protein sequence and partial assignments to narrow FLYA's assignment search, while final shift values come from spectral signal positions.The model is trained on graphs derived from BMRB records.Source: Unfolding a k-dimensional spectrum is… · p. 14

Data availability

• Reference structures are identified by PDB accession codes in Fig. 2 and Supplementary Table 3.Source: Data availability · p. 15
• Spectra and reference assignments are available through the BMRB.Source: Supplementary Table 3) · p. 15
• The listed materials are peak lists, assignments, and structures, and source data for Figs. 2, 4, and 5 is available in Supplementary Tables 2, 4, and 5, respectively.Source: (https://bmrb.io/; entry IDs in… · p. 15

Code availability

• The pp-ResNet, deconv-ResNet, GNN, and GBT models are available for download in binary form.Source: The ARTINA algorithm is available as a… · p. 16
• The downloadable materials include architecture schemes, example inputs, model-input descriptions, and source code for reading model files and making predictions.Source: The ARTINA algorithm is available as a… · p. 16
• The component files provide technical specifications and permit independent use in Python.Source: The ARTINA algorithm is available as a… · p. 16

Additional information

• Fig. 3 overlays 100 automatically determined protein structures with their corresponding PDB depositions.Structures are aligned by RMSD to reference; structures without corresponding PDB depositions are marked with an asterisk.Source: Figure 3 · p. 22, Additional information · p. 22
• Fig. 4 summarizes backbone RMSD, distance restraints per residue, and chemical-shift assignment accuracy for benchmark proteins.It reports percentile values alongside protein-level bars.Source: Figure 4 · p. 23, Additional information · p. 23
• Fig. 5 compares actual and predicted RMSD to reference, with values above 4 Å truncated at 4 Å.Dotted lines mark deviations of ±1 Å between the two RMSD quantities.Source: Figure 5 · p. 24, Additional information · p. 24
• Fig. 6 and Fig. 7 present examples concerning spectrum-analysis challenges and peak-picking performance.The supplied captions identify these topics but do not give further results.Source: Figure 6 · p. 25, Figure 7 · p. 26

Table 1 Quality of assignments and structures in each refinement cycle.

• Table 1 reports assignment and structure quality across refinement cycles.A refinement cycle is one ARTINA iteration, comprising chemical-shift and structure-refinement cycles.Source: Table 1 · p. 27, Additional information · p. 27

Supplementary Fig. 1 Examples of generated spectrum fragments for deconvolution

• Supplementary Fig. 1 shows generated 64 × 32 × 5 spectrum fragments used to train the deconvolution model, with ground-truth signal positions marked.Source: Supplementary Fig. 1 Examples of… · p. 30

Supplementary Fig. 2 Visualization of 100 protein structures determined with ARTINA

• The structure visualization identifies PDB code, RMSD, and backbone and side-chain assignment accuracy, while distinguishing well-defined from disordered regions by color.The supplementary figure also depicts examples of spectrum-fragment challenges and high-RMSD structures.Source: Supplementary Fig. 2 Visualization of… · pp. 31, 37

Supplementary Table 2 Proteins and spectra used for automated structure

• Energy refinement changed backbone RMSD to the reference PDB with a 0.07 Å standard deviation, 0.16 Å maximum deviation, and 0.998 linear correlation.The reported comparison is before versus after energy refinement in explicit water using OPALp.Source: Supplementary Table 2 Proteins and… · p. 70

Supplementary Table 13 Accuracy of automated chemical shift assignment using all

• Automated spectrum analysis annotated every plane of a 16 kDa protein’s HCCH-TOCSY spectrum in under 5 minutes without human involvement.The accompanying tables report assignment accuracy using all input spectra and using all except NOESY-type spectra.Source: Supplementary Table 13 Accuracy of… · pp. 76–78