A pseudo-labeling approach for automated assignment of unlabeled proteins exceeding 100 amino acid residues

Abstract

In the era of AI-driven protein structure prediction, there is an urgent demand for rapid, accessible experimental approaches to validate computationally predicted structural models or to improve the accuracy of AI-based structure determination. As a key experimental technique for determining protein structures in solution, conventional biomolecular NMR spectroscopy relies heavily on proteins labeled with ¹³C and ¹⁵N isotopes, a requirement that entails high experimental costs, lengthy sample preparation times, and steep technical barriers, thereby limiting its broader application. Building on our previous work, we have developed a pseudo-labeling strategy for the automated resonance assignment of unlabeled proteins. Using high-field NMR spectrometers equipped with cryoprobes, we acquired natural-abundance ¹H–¹³C and ¹H–¹⁵N heteronuclear correlation spectra directly from unlabeled samples. By integrating predicted chemical shifts of H_N, H_α, ¹³C_α, and ¹⁵N from three complementary AI tools into an iterative simulated annealing framework with an alternating scoring scheme, our method effectively resolves severe cross-peak overlap during the automated assignment process by introducing heteronuclear dimensional constraints. This approach enables reliable automated backbone resonance assignment for unlabeled proteins exceeding 100 amino acid residues. We anticipate that this labeling-free strategy will lower the technical barrier to biomolecular NMR, enabling researchers without specialized NMR expertise to employ NMR spectroscopy in fundamental life science research and drug development.

Graphical abstract

1. Introduction

The three-dimensional structure of proteins is fundamental to deciphering their biological functions and molecular mechanisms [1,2]. Nuclear magnetic resonance (NMR) spectroscopy, X-ray crystallography, and cryogenic electron microscopy (cryo-EM) remain the three major experimental techniques for determining protein structures [[3], [4], [5], [6]]. In recent years, artificial intelligence (AI) approaches have revolutionized protein structure prediction: models such as AlphaFold can accurately predict the folding of most proteins from sequence alone and even enable de novo design of proteins for specific functional needs [7,8].

Despite these computational advances, experimental validation remains indispensable, and experimental methods such as NMR spectroscopy play a critical role in evaluating and improving AI-derived structural models [9,10]. Previous work has demonstrated that even for small, well-folded globular proteins, AlphaFold may produce high-confidence yet topologically incorrect structures, which can only be identified and corrected by experimentally determined NMR structures [11]. Beyond static structure validation, NMR also captures protein conformational dynamics [[12], [13], [14]]. Researchers have further shown that NMR experimental restraints can be incorporated into AI generative frameworks to produce structural ensembles consistent with experimental measurements, overcoming a key limitation of AI models, which typically output only a single dominant conformation [15].

Among the three major techniques, solution NMR offers two prominent advantages. First, it requires only straightforward sample preparation without crystallization or vitrification [16,17], and measurements can be conducted under near-physiological buffer conditions. Second, it is well suited for investigating proteins with molecular weights below 30 kDa, a size range that remains challenging for cryo-EM to resolve [18,19]. These combined strengths make it ideal for the rapid validation of AI-predicted structures. However, for proteins with more than 50 residues, conventional solution NMR structure determination relies heavily on ¹⁵N and ¹³C isotope labeling [20,21]. Switching protein expression to isotope-labeled minimal media often results in low or undetectable protein yields, necessitating extensive optimization of induction conditions and medium composition [22]. The subsequent spectral assignment and structure calculation also present technical hurdles for researchers without specialized NMR expertise. These obstacles limit the widespread application of solution NMR in the life sciences and drug development. Therefore, developing fast, low-cost, and accessible NMR workflows to bridge AI prediction and experimental verification is a pressing need in structural biology research.

In the 1980s and 1990s, researchers relied on the main chain direct (MCD) method to assign two-dimensional NMR spectra of unlabeled proteins, which was extremely time-consuming and labor-intensive [23,24]. Although several computer-assisted methods were developed during this period, they did not become routine tools, mainly because of the limited computational power and algorithmic sophistication available at the time [24,25]. Today, the development of high-field NMR spectrometers with improved resolution and sensitivity, together with rapid advances in AI, has created new opportunities to revisit this long-standing problem.

Previously, we developed a method combining the MCD method and chemical shifts of H_N and H_α predicted by AI tools to assign two-dimensional spectra of unlabeled human ubiquitin [23,26]. However, this method was only applicable to proteins less than 100 amino acid residues. Herein, we present a pseudo-labeling approach that enables spectral assignment of unlabeled proteins with over 100 amino acid residues by incorporating ¹⁵N and ¹³Cα chemical shifts generated by multiple AI prediction software packages.

In our previous studies on automated assignment of two-dimensional NMR spectra of unlabeled human ubiquitin, cross-peaks in the fingerprint region (H_N-H_α) of the ¹H–¹H correlated spectroscopy (COSY) were identified and grouped into discrete spin systems, each corresponding to one amino acid residue of target protein. For each spin system, all associated proton resonances were assigned based on correlation patterns in the ¹H–¹H total correlation spectroscopy (TOCSY), and corresponding nuclear Overhauser effect (NOE) cross-peaks with consistent H_N chemical shifts were extracted from the ¹H–¹H nuclear Overhauser effect spectroscopy (NOESY) to provide sequential connectivity constraints.

Simulated annealing (SA) [27] coupled with Monte Carlo (MC) [28] sampling is a stochastic global optimization framework tailored for combinatorial optimization problems with exponentially large solution spaces, such as mapping NMR spin systems to amino acid residues along a protein sequence. In this hybrid approach, MC sampling acts as the core perturbation engine: it generates new candidate configurations by introducing random, targeted rearrangements to the current assignment state—typically by swapping the sequence positions of two unfixed spin systems to produce a novel permutation of spin system-to-residue mapping. Each newly generated permutation is then evaluated against a predefined scoring function that integrates multiple constraints, including amino acid type compatibility, NOE connectivity consistency, and chemical shift similarity.

The overall scoring function is composed of three weighted terms, each derived from independent experimental characterization or computational prediction. The first term is the amino acid type score, generated by a combination of random forest (RF) [29] and support vector machine (SVM) [30] models trained on protein chemical shift data from the Biological Magnetic Resonance Data Bank (BMRB) database [31]. For each spin system, the combined model outputs a confidence score for multiple standard amino acid types, reflecting how well the proton chemical shift profile of the spin system matches the established statistical distribution of the corresponding residue. The second term assesses NOE connectivity consistency, extracted from the ¹H–¹H NOESY spectrum. This term rewards assignment configurations where sequentially adjacent spin systems exhibit characteristic sequential NOE cross-peaks, as these inter-residue proton correlations verify the spatial proximity of neighboring residues along the polypeptide backbone. The third term measures chemical shift similarity, based on H_N and H_α chemical shifts predicted by the UCBShift2.0 software using solution NMR structure of human ubiquitin from Protein Data Bank [[32], [33], [34]]. Through principal component analysis (PCA) [35], this term computes the distance between the experimental H_N-H_α chemical shift vector of each spin system and that predicted for each candidate residue in the reduced principal component space. Distances below the threshold of 0.5 are converted into a matching score, with smaller distances yielding higher scores to quantify the structural agreement between the spin system and its assigned residue position.

The SA framework is employed to search for the optimal assignment that maximizes the overall scoring function over all possible mappings from spin systems to residue positions. To avoid being trapped in local maxima, the algorithm does not perform a simple greedy ascent; instead, it probabilistically accepts solutions that decrease the total score (i.e., worsen the objective) according to the Metropolis criterion [36]: a transition that lowers the score by ΔS < 0 is accepted with probability exp(ΔS/T), where ΔS denotes the change in the score decrease and T is the current temperature parameter. Better assignments that increase the total score are always accepted unconditionally. At the initial high temperature, the algorithm tolerates moderately inferior assignments, enabling broad and unbiased exploration of the entire solution space. As the temperature is gradually reduced following a predefined cooling schedule, the acceptance probability for score-decreasing moves steadily declines, progressively steering the search toward the global maximum. In each round, the algorithm performs 100 independent replicate runs. Any spin system whose assignment appears in the same position in more than 90% of these independent runs is considered highly confident and is fixed in subsequent rounds, which effectively reduces the dimensionality of the search space and accelerates convergence in later optimization cycles.

While the method described above enables successful automated assignment of two-dimensional NMR spectra for unlabeled human ubiquitin (76 aa), its performance deteriorates markedly when applied to proteins with more than 100 amino acid residues, where the algorithm suffers from severe convergence failure. Specifically, after a few rounds of iterative refinement, no spin system is assigned to the same sequence position in over 90% of independent replicate runs, meaning that the workflow cannot advance to subsequent rounds.

Two primary factors account for this convergence breakdown. First, for larger proteins, AI chemical shift prediction software produces considerably larger deviations for a subset of amino acid residues. These inaccuracies weaken the discriminatory power of the chemical shift similarity term in the scoring function, thereby increasing ambiguity throughout the assignment search process. Second, for proteins exceeding 100 residues, peak overlap in two-dimensional homonuclear NMR spectra becomes markedly severe. Such severe spectral overlap compromises the accuracy of spin system grouping and the reliability of extracted NOE cross-peak constraints, further amplifying the combinatorial complexity of the assignment problem and preventing the algorithm from reaching stable convergence.

2. Methods

To overcome the convergence failure and achieve fully automated two-dimensional NMR resonance assignment for unlabeled proteins exceeding 100 amino acid residues, we introduce three targeted enhancements to the previous assignment framework. First, chemical shift predictions from multiple independent AI tools are integrated to mitigate systematic prediction deviations and improve the robustness of the chemical shift similarity term in the scoring function. Second, a pseudo-labeling strategy is developed, wherein AI-predicted ¹³C_α and ¹⁵N chemical shifts of the target protein are incorporated into the scoring function as auxiliary dimensional constraints, thereby helping to solve assignment ambiguities stemming from severe peak overlap in homonuclear two-dimensional spectra. Third, the overall computational workflow and iterative optimization protocol are refined to improve the search efficiency and convergence stability of the simulated annealing algorithm when handling the drastically expanded combinatorial space of larger proteins. The detailed implementation of the proposed methodology is described in the following sections, using hen egg white lysozyme (HEWL) [37] as the model system to illustrate the complete workflow.

2.1. NMR spectroscopy

HEWL NMR samples were prepared at a final concentration of approximately 0.5 mM in 50 mM sodium phosphate buffer (pH 5.0) containing 10% (v/v) D₂O. All NMR experiments, including ¹H–¹H NOESY, ¹H–¹H COSY, ¹H–¹H TOCSY, ¹H–¹³C heteronuclear single quantum coherence (HSQC), and selective optimized flip-angle short transient (SOFAST) ¹H–¹⁵N heteronuclear multiple quantum correlation (HMQC), were acquired on a Bruker AVANCE III HD 950 MHz spectrometer equipped with a TCI cryoprobe at 308 K [[38], [39], [40]]. Water suppression for the ¹H–¹H COSY spectrum was achieved using a 3-9-19 pulse sequence, whereas the ¹H–¹H TOCSY and ¹H–¹H NOESY experiments employed selective excitation combined with gradient pulses for water suppression [41,42]. Given that the protein was not isotopically labeled, the ¹H–¹³C HSQC spectrum was acquired with 128 scans, and the SOFAST ¹H–¹⁵N HMQC spectrum was acquired with 2048 scans. Both heteronuclear experiments required acquisition times of over 14 h to achieve good signal-to-noise ratios.

2.2. Pseudo-labelling approach

As the initial step of the assignment workflow, spin systems are extracted from the acquired NMR spectra to generate the corresponding computational input files. Specifically, H_N-H_α spin systems are identified based on ¹H–¹H COSY and ¹H–¹H TOCSY correlations, H_α-C_α spin systems are identified based on ¹H–¹³C HSQC spectra, and H_N–N spin systems are confirmed using SOFAST ¹H–¹⁵N HMQC spectra. Theoretically, each H_N-H_α, H_α-C_α, and H_N–N spin system maps to one corresponding amino acid residue of the target protein. Nevertheless, the exact residue-wise mapping remains undetermined at this stage, so all spin systems are initially assigned sequential random identifiers (1, 2, 3, 4, …) (Fig. 1).

Fig. 1.

2D ¹H–¹H TOCSY(A), ¹H–¹⁵N SOFAST HMQC(B) and ¹H–¹³C HSQC(C) spectra of HEWL. Peaks in the three spectra were assigned to H_N-H_α, H_N–N, and H_α-C_α based spin systems respectively, and labeled with random numerical identifiers.

Backbone chemical shifts of HEWL, including H_N, H_α, C_α, and N resonances of each residue, are predicted using AI-based chemical shift prediction tools. Here, UCBShift2.0 is used as a representative example to illustrate the workflow, while the full pipeline integrates predictions from UCBShift2.0, EFG-CS [43], and PLMs [44], as described later. From these predicted chemical shifts, the expected positions of H_N-H_α, H_α-C_α, and H_N–N cross-peaks are generated for every residue of the target protein. PCA is then applied to calculate the distance between experimentally detected cross-peaks and their predicted counterparts in the PCA-transformed space (Fig. 2).

Fig. 2.

Heat map illustrating the distribution of pairwise distances in principal component space between experimental resonances of H_N-H_α(A), H_N–N(B) and H_α-C_α(C) spin systems (vertical axis) and predicted residue-specific chemical shifts of HEWL (horizontal axis). The color bar transitions from yellow to red for distances 0 to 0.5, with all distances exceeding 0.5 rendered in blue.

To normalize chemical shift dynamic ranges and ensure dimensional comparability between each heteronuclear nucleus and its coupled proton resonance, scaling factors are applied prior to distance calculation. Specifically, ¹⁵N chemical shifts coupled with H_N protons are scaled by a factor of 0.14 [45], whereas ¹³C chemical shifts coupled with H_α protons are scaled by a factor of 0.1.

The computed PCA distances are further converted into matching scores according to a predefined tiered scoring scheme. For instance, an experimental H_N–H_α cross-peak with a PCA distance below 0.1 from its predicted residue counterpart yields a score of 50; a distance between 0.4 and 0.5 yields a score of 10, and other intermediate distance intervals are scored accordingly; distances greater than 0.5 are regarded as non-matching and assigned a score of 0.

To further improve the accuracy of resonance assignment, two additional AI-based tools, EFG-CS and PLMs, were employed to independently predict the positions of H_N–H_α, H_α–C_α, and H_N–N cross-peaks for HEWL. For each tool, the distances between experimentally observed cross-peaks and their predicted counterparts were calculated in the same PCA-transformed space, and the resulting PCA distances were converted into matching scores following the same tiered scoring scheme. The matching scores from UCBShift2.0, EFG-CS, and PLMs were then summed to generate a final consensus score. In this ensemble approach, if a given spin system is consistently matched to the same residue across multiple predictors, the corresponding scores accumulate, which effectively enhances the discriminative power of the scoring function for correct residue assignments.

To address the spectral overlap and chemical shift degeneracy inherent to larger proteins, the assignment workflow was initiated using H_N–H_α spin systems as core anchors. First, all H_N–N spin systems with matching H_N chemical shifts were linked to each H_N–H_α spin system via their shared H_N resonances; correspondingly, all H_α–C_α spin systems with matching H_α chemical shifts were associated with the same H_N–H_α spin system via their common H_α resonances. Because of chemical shift degeneracy, multiple H_N–N or H_α–C_α spin systems may be linked to a single H_N–H_α anchor; all such linked spin systems are retained without prior discrimination. When the predicted ¹³C_α and ¹⁵N chemical shifts were required in the SA calculations, candidate residue matches were collected from every linked spin system, and the intersection of these candidate sets was extracted, with the matching scores for the shared candidate residues summed. If no residue was common to all three spin-system types, pairwise intersections among the candidate sets were adopted instead, with the corresponding matching scores accumulated accordingly (Fig. 3). The disambiguation is achieved entirely through this computational intersection operation: a candidate residue accumulates a high score only when it is consistently supported by predicted shifts across all three linked spin-system types (the triple-intersection region highlighted in red in Fig. 3), whereas pairwise intersections (yellow regions) serve as a fallback that still integrates evidence from two independent spin-system types. This strategy allows AI-predicted ¹³C_α and ¹⁵N chemical shift information to facilitate NMR resonance assignment in the computational pipeline together with predicted H_N and H_α chemical shifts, without the need for isotopic ¹³C or ¹⁵N labeling of the target protein.

Fig. 3.

Schematic diagram depicting the pseudo-labeling strategy. With H_N-H_α spin systems as the anchor, H_N–N spin systems are correlated through matched H_N chemical shifts, and H_α-C_α spin systems are subsequently connected via matched H_α chemical shifts. When computing the chemical shift similarity encompassing H_N, H_α, ¹³C_α, and ¹⁵N nuclei, the scores assigned to the same predicted residue from the three associated spin systems are aggregated(red). If no residue is common to all three spin systems, scores derived from pairwise intersections of the three spin systems are employed (yellow).

2.3. Assignment workflow

The overall computational workflow is illustrated in Fig. 4. The calculation started with the loading of all initial input files containing pre-extracted spin system data of three types and their combined chemical shift similarity files generated from three AI tools, followed by SA optimization that explored candidate assignment permutations between spin systems and amino acid residues of the target protein. The scoring function integrated three core components to evaluate each permutation: amino acid type compatibility of individual spin systems, the distribution pattern of NOE cross-peaks, and chemical shift similarity between experimental spin systems and the corresponding predicted residues of the target protein. The SA optimization proceeded iteratively until the assignment permutation yielding the highest global score was identified for each independent run.

Fig. 4.

Flowchart of the iterative simulated annealing workflow for NMR spectral assignment of unlabeled proteins.

For each iteration round, 100 parallel independent SA runs were performed. Upon completion of each round, spin systems converging to the same residue position in more than 90 out of the 100 runs were defined as confidently assigned and fixed in all subsequent iterations. The remaining ambiguous spin systems were carried forward to the next round of optimization. This iterative process continued until over 95% of all H_N–H_α spin systems were successfully assigned.

A key modification relative to our previous work is the alternating use of two complementary scoring strategies across successive iteration rounds. In odd-numbered rounds (1st, 3rd, 5th, …), the chemical shifts of H_α, H_N, C_α, and N nuclei derived from the pseudo-labeling strategy were adopted, while in even-numbered rounds (2nd, 4th, 6th, …), only the H_N and H_α chemical shift similarity score was incorporated into the scoring function.

This alternating design was motivated by the heterogeneous nature of peak overlap across different spectral dimensions. Severely overlapped resonances in the homonuclear H_N–H_α 2D spectrum may be well resolved along the ¹³C or ¹⁵N heteronuclear dimensions, and vice versa. This complementary scoring scheme effectively alleviated assignment ambiguity caused by spectral crowding: when one set of predictive features lacked sufficient discriminative power to resolve ambiguous spin system assignments, the complementary dimensional information from the alternate round provided additional resolving power to facilitate reliable residue assignment.

3. Results and discussions

After six iterative rounds of optimization, complete assignment was achieved for all 103 H_N–H_α cross-peaks of HEWL. Among these assignments, 87 were consistent with the reference assignment deposited in BMRB, yielding an overall assignment accuracy of 84%.

Building on these H_N and H_α assignments and the AI-predicted ¹³Cα and ¹⁵N chemical shifts, the corresponding H_α–C_α cross-peaks in the ¹H–¹³C HSQC spectrum and H_N–N cross-peaks in the SOFAST ¹H–¹⁵N HMQC spectrum could be further assigned, without requiring additional isotopic labeling experiments (Fig. 5).

Fig. 5.

Final assignment of the H_αH_N region in¹H–¹H TOCSY (A), H_NN region in ¹H–¹⁵N SOFAST HMQC (B), and H_αC_α region in ¹H–¹³C HSQC (C) spectra of HEWL, with peaks labeled by one-letter amino acid code followed by residue number.

The generalizability of the proposed method was further evaluated on two additional unlabeled proteins: thioredoxin (TRX, 107aa) [46] and arsenate reductase (ArsC, 139aa) [47]. Assignment accuracies of 85% and 71% were achieved for these two proteins respectively, demonstrating the broad applicability of the strategy to unlabeled proteins exceeding 100 amino acid residues (Table 1).

Table 1.

Number of assigned residues for TRX, HEWL, and ArsC in each round of calculation, with the corresponding assignment accuracy indicated in parentheses. Full resonance assignment was achieved after 3 calculation rounds for TRX, 6 rounds for HEWL, and 15 rounds for ArsC. The final row lists the total number of assigned spin systems and the associated assignment accuracy for each of the three proteins.

Protein round	TRX	HEWL	ArsC
1	64 (94%)	46 (91%)	36 (86%)
2	16 (63%)	23 (96%)	10 (90%)
3	6 (50%)	9 (78%)	14 (64%)
4	–	12 (42%)	3 (100%)
5	–	3 (33%)	5 (100%)
6	–	10 (70%)	3 (0%)
7	–	–	6 (83%)
8	–	–	3 (33%)
9	–	–	7 (100%)
10	–	–	4 (50%)
11	–	–	5 (40%)
12	–	–	1 (0%)
13	–	–	1 (100%)
14	–	–	2 (50%)
15	–	–	7 (29%)
Total	86 (85%)	103 (84%)	107 (71%)

Table 1 summarizes the number of assigned residues and the corresponding assignment accuracy at each iterative round for all three proteins. Notably, the first round of calculation assigned the largest number of residues with consistently high accuracy (94% for TRX, 91% for HEWL, and 86% for ArsC), all of which exceed the overall assignment accuracies for the respective proteins (85%, 84%, and 71%). The accuracies of subsequent rounds, however, showed no consistent trend and varied considerably across rounds and proteins. This distinctive feature makes the first-round assignment particularly well suited for applications where a subset of high-confidence assignments is sufficient. One possible application is the validation of AI-predicted protein–ligand binding sites: researchers first use AI-based tools such as AF2BIND to predict candidate interaction residues [48], and then employ NMR chemical shift titration experiments to verify these predictions. In this workflow, confirmation is achieved when the residues undergoing significant chemical shift perturbations coincide with the AI-predicted binding site, which provides a convenient and reliable means of interaction-site cross-validation.

To further evaluate the contribution of AI-predicted chemical shifts to assignment performance, we performed independent assignment calculations using chemical shifts predicted individually by each of the three tools —UCBShift2.0, EFG-CS, and PLMs—and compared the results with those obtained using the integrated predictions from all three tools (Table 2). In general, assignment accuracies achieved with any single predictor were markedly lower than those obtained with the integrated strategy. Moreover, the performance of individual predictors was strongly protein-dependent: PLMs achieved the highest accuracy on TRX (85%) but performed poorly on HEWL (7%), whereas EFG-CS was the best predictor for HEWL (68%) but yielded only 30% on ArsC. This pronounced variability across proteins underscores the complementary nature of the three prediction tools and confirms that integrating their predictions produces a substantially more robust and accurate assignment than any single tool alone.

Table 2.

Assignment accuracy of TRX, HEWL, and ArsC from resonance assignment calculations using chemical shifts predicted individually by UCBShift2.0 (UCB), EFG-CS (EFG), and PLMs.

Tool	TRX			HEWL			ArsC
	UCB	EFG	PLMs	UCB	EFG	PLMs	UCB	EFG	PLMs
Assignment accuracy	70%	65%	85%	55%	68%	7%	33%	30%	50%

Assignment errors in the integrated three-AI-tool pipeline arise primarily from two coupled effects associated with protein molecular weight. First, as the number of residues increases, spectral crowding in 2D ¹H–¹H NOESY spectra becomes progressively more severe. The severe peak overlap introduces substantial ambiguity into NOE-derived connectivity restraints, which serve as the primary experimental basis for establishing sequential ordering of spin systems along the polypeptide chain. This ambiguity substantially increases the probability of incorrect sequential arrangement of spin systems during the calculation process. Second, larger proteins generally contain a higher fraction of conformationally heterogeneous regions, including flexible loops, interdomain interfaces, and dynamic regions. The prediction accuracy of AI-based chemical shift predictors degrades noticeably in these regions, primarily because chemical shift data for them in the BMRB database are frequently compromised by severe spectral crowding, incomplete resonance assignments, and a greater probability of incorrect peak assignments. Thus, for larger proteins, the assignment pipeline becomes more dependent on chemical shift predictions with inherently lower accuracy, while the experimental NOE restraints become less reliable due to severe spectral overlap. This synergistic effect accounts for the observed trend that assignment accuracy decreases markedly with increasing protein molecular weight.

The limitations in assignment accuracy described above may be addressed by incorporating three-dimensional (3D) protein structural models into the assignment workflow. Specifically, after preliminary resonance assignments are generated using the pipeline developed in this study, the assigned NOESY cross-peaks can be cross-validated against interpreting distance restraints derived from 3D structural models. An iterative refinement protocol can then be applied, in which candidate assignments are adjusted and re-evaluated over successive cycles against both chemical shift similarity scores and structure-based distance restraints, thereby maximizing assignment accuracy. This will be a direction for future improvement of our work.

4. Conclusion

This work introduces a computational strategy for protein backbone NMR resonance assignment that operates entirely on unlabeled protein samples, integrating multi-model AI chemical shift prediction with an iterative alternating-scoring simulated annealing algorithm. A core innovation of this strategy lies in its pseudo-labeling framework, which uses AI-predicted ¹³Cα and ¹⁵N chemical shifts to incorporate heteronuclear dimensional constraints into the assignment pipeline. Conventional backbone assignment protocols rely heavily on uniformly ¹⁵N- and ¹³C-labeled protein samples, which impose high experimental costs and long preparation cycles; for many protein systems, lengthy optimization of expression conditions in minimal inorganic media is required to produce isotope-labeled proteins with satisfactory yield. Our method circumvents these experimental barriers by utilizing high-field NMR spectrometers (≥800 MHz) equipped with cryoprobes to acquire natural-abundance ¹H–¹³C and ¹H–¹⁵N correlation spectra from unlabeled protein samples. By integrating these experimentally measured heteronuclear spectral data with AI-predicted ¹³Cα and ¹⁵N chemical shifts, we achieve multidimensional assignment constraints analogous to isotopic labeling at the computational level. By eliminating the need for isotope labeling and reducing both sample preparation and instrument acquisition time, our approach offers substantial savings in both time and financial cost compared with conventional NMR assignment workflows.

We anticipate that the developed pipeline for unlabeled protein resonance assignment will significantly broaden the applicability of NMR spectroscopy in fundamental life science research and drug discovery. By circumventing the need for isotopic labeling and lowering the technical barrier to spectral analysis through automated assignment, this strategy enables researchers without specialized expertise in biomolecular NMR to employ NMR for a wide range of protein studies, including structural analysis, protein–ligand binding characterization, and the exploration of structure–function relationships. We further expect that this approach will extend the reach of biomolecular NMR investigations beyond dedicated structural biology laboratories, serving as a robust and cost-effective analytical tool to support advances in biochemical research, drug target validation, and other areas of the life sciences.

CRediT authorship contribution statement

Xiaogang Niu: Writing – original draft, Supervision, Software, Methodology, Writing – review & editing. Xianhui Hou: Resources. Changwen Jin: Writing – review & editing, Supervision.

Data availability

The data that support the findings of this study are openly available in Science Data Bank at https://doi.org/10.57760/sciencedb.mrl.00005, CSTR: 31253.11.sciencedb.mrl.00005 or via the link https://cstr.cn/31253.11.sciencedb.mrl.00005.

Declaration of competing interest

The authors declare that a Chinese patent application (No. 202611043043.7), on which Xiaogang Niu, Xianhui Hou and Changwen Jin are listed as inventors, has been filed based on the work described in this manuscript. Changwen Jin is an editorial board member of Magnetic Resonance Letters but was not involved in the editorial review or the decision to publish this article. Apart from the above intellectual property filing and editorial membership, the authors declare no additional competing financial or commercial interests.

Biographies

Xiaogang Niu received his B.S. in Physics from Shandong Normal University in 2002, and his Ph.D. in Biology from University of Science and Technology of China in 2007. After that, he worked as postdoctoral researcher at Florida State University and the National High Magnetic Field Laboratory for three years. He is currently a Senior Engineer at Beijing NMR Center, Peking University. His primary research interests focus on investigating protein structure and dynamics using biological NMR, as well as the development of related NMR methodologies.

Changwen Jin is a Changjiang Professor and Chief Scientist of the Beijing NMR Center at Peking University. He received his B.S. from Jilin University and Ph.D. from Peking University. After postdoctoral research at Peking University and the University of Illinois at Urbana-Champaign, he joined Peking University in 2001. His primary research interests lie in structural biology, biomolecular dynamics using biological NMR, and NMR method development, with a focus on GPCR dynamic structures and drug discovery.

Appendix A. Supplementary data

The following is the Supplementary data to this article.