Beyond traditional magnetic resonance processing with artificial intelligence | Communications Chemistry

Communications Chemistry volume 7, Article number: 244 (2024) Cite this article

Metrics details

Smart signal processing approaches using Artificial Intelligence are gaining momentum in NMR applications. In this study, we demonstrate that AI offers new opportunities beyond tasks addressed by traditional techniques. We developed and trained artificial neural networks to solve three problems that until now were deemed “impossible”: quadrature detection using only Echo (or Anti-Echo) modulation from the traditional Echo/Anti-Echo scheme; accessing uncertainty of signal intensity at each point in a spectrum processed by any given method; and defining a reference-free score for quantitative access of NMR spectrum quality. Our findings highlight the potential of AI techniques to revolutionize NMR processing and analysis.

NMR spectroscopy is a powerful analytical technique widely used to acquire atomic-level information about molecular structure, dynamics, and interactions1,2. To derive meaningful insights from the acquired spectra, NMR data processing plays a vital role. Artificial Intelligence (AI), and specifically Deep Learning (DL), presents a compelling alternative to traditional methods in NMR processing3,4. Although early demonstrations of machine learning in NMR date back to the 1970s5, practical applications have evolved significantly with recent advancements in algorithms and computer hardware. In most cases, DL in NMR data processing focuses on surpassing the existing algorithmic techniques for fast and high-quality solving of traditional tasks such as spectra reconstruction from Non-Uniformly Sampled (NUS) time domain signals6,7,8,9, virtual homo-decoupling9,10,11,12, obtaining pure shift spectra13,14,15, spectra denoising16,17, and automating peak picking18,19. In this study, we address an intriguing question of whether DL can go beyond the traditional problems and offer new ways of spectra processing and analysis4, and possibly give us insights for designing new signal processing algorithms9.

We demonstrate a Magnetic Resonance processing with Artificial intelligence (MR-Ai) solution for the seemingly impossible task of recovering a high-quality spectrum from an incomplete phase-modulated quadrature detection experiment, where only one of the two P- or N-type parts of phase-modulated quadrature detection experiments is available. Furthermore, we show that MR-Ai is able to perform valuable statistical analyses of spectra reconstructed by other methods and thus provides a new reference-free metric of the spectrum’s quality.

Traditionally, in multidimensional (nD) NMR experiments, frequency discrimination and obtaining pure, absorptive phase signals rely on quadrature detection. This involves acquiring two data points per time increment and per spectral dimension. For a 2D experiment, where the signal evolves in two-time dimensions t1 and t2, the amplitude-modulated quadrature detection20,21 is implemented by acquiring two separate complex 2D data sets in the form of cosine and sine modulation:

Data set1 (cos modulated): cos(Ω1t1)exp(iΩ2t2)

Data set2 (sin modulated): sin(Ω1t1)exp(iΩ2t2)

where Ωn is the signal frequency in the nth dimension.

In contrast, the phase-modulated data usually obtained from gradient coherence order selection experiments are encoded as either Echo (P-type data) or Anti-Echo (N-type data) coherence:

Data set1 (P-type): exp( + iΩ1t1)exp(iΩ2t2)

Data set2 (N-type): exp( − iΩ1t1)exp(iΩ2t2)

Individually, the Fourier transform of each of these complex 2D datasets produces frequency-discriminated spectra albeit with a phase-twisted lineshape of the peaks (Fig. 1a), which is not amenable for normal analysis. Until now, it has been understood that the only way to obtain pure absorptive phase signals in Echo-Anti-Echo experiments is by using both P- and N-type data22,23,24. In this work, we demonstrate that MR-Ai can effectively recognize the twisted lineshapes and convert them into the pure absorption form (Fig. 1b). To the best of our knowledge, none of the traditional methods demonstrated this capability so far.

(a) Echo spectrum with a phase-twist lineshape, (b) normal spectrum with a pure absorptive peak in the frequency domain, and (c and d) the complex time domain Virtual Echo presentations obtained by the inverse Fourier transform of the real parts of corresponding spectra. P and N represent the P-type and N-type time domain signal, while \(\widetilde{P}\) and \(\widetilde{N}\) indicate the time reverse and conjugation of P-type and N-type data sets respectively.

In our recent publication, we introduced a Deep Neural Network (DNN) architecture called WNN, specifically designed to grasp 1D patterns over the entire NMR spectrum in the frequency domain, such as specific patterns of NUS aliasing artifacts and peak multiples in homo-decoupling9. In this work, we present an updated version of our WNN architecture (see Methods for more details) capable of capturing 2D patterns, including the phase twisted peaks associated with the P- (or N-) type data, as a pattern recognition problem in the frequency domain. Figure 2a demonstrates the excellent performance of Echo and Anti-Echo reconstructions of 1H-15N correlation spectra by MR-Ai on MALT1 (45 kDa) protein25 – Similar results for Ubiquitin (7 kDa), Azurin (14 kDa)26, and Tau (disordered, 441 amino acids)27 are found in Supplementary Figs. 1 to 4.

(a) 2D 1H-15N TROSY spectra of MALT125 Echo reconstruction using MR-Ai with predicted uncertainty in pink color - The insets show zoomed part and corresponding reference with the actual absolute error between reference and reconstruction in red color. Bar graphs (b) depicts RMSD as a traditional reference-based evaluation metric and (c) Boxplots show normalized uncertainty as the intelligent reference-free metric for evaluating the spectrum reconstruction quality. The orange bar and the green triangle indicate the median and mean of the uncertainty distribution, respectively.

Figure 1 illustrates that the task to rectify the phase twist problem encountered in Echo (or Anti-Echo) spectra can be viewed as a specific case of NUS reconstruction. Indeed, Fig. 1c, d show the time domain equivalents of the twisted (Fig. 1a) and absorptive (Fig. 1b) line shapes. The time domain presentation, obtained by the two-dimensional inverse Fourier Transform of the real part of the spectrum shown in Fig. 1a, b and in the following called the Virtual Echo (VE)28, clearly shows the roles of the P-type and N-type data. There are four regions defined by the signs of t1 and t2. P-type data corresponds to the upper-right region where both t1 and t2 are positive (P), and the time-reversed conjugated P-type data corresponds to the lower-left region where both t1 and t2 are negative (\(\widetilde{P}\)). Similarly, N-type data corresponds to the region where t1 is negative and t2 is positive (N), while the time-reversed and conjugated N-type data is in the region where t1 is positive and t2 is negative (\(\widetilde{N}\)). Therefore, rectification of the twisted line shape in the frequency domain is equivalent to completing the missing half of the signal in the VE time domain. The problem is akin to spectral reconstruction from NUS data and, hence, can be performed using the Compressed Sensing Iterative Soft Thresholding algorithm (CS-IST)29,30, a representative traditional NUS reconstruction technique. The CS algorithm maximizes the sparsity of the spectrum, thus filling in the missing data to produce the most compact absorptive form of the signal while suppressing the wider and less sparse dispersive features of the twisted signal. Previously, we demonstrated that for performing data completion in the NUS spectra reconstruction, the MR-Ai utilized pattern recognition9. In this work, we point out that the seemingly pattern reconstruction problem of rectifying the Echo (or Anti-Echo) twisted line shapes is akin to the data completion in the time domain and can be addressed by traditional algorithms such as CS.

Both reconstruction methods, MR-Ai and CS-IST, reproduce the spectrum with high quality using either P- or N-type data. Figure 2b (and in Supplementary Fig. 9) shows a spectra quality metric point-to-point RMSD (and \({R}_{2}^{s}\)) between the reconstruction and reference spectra for MALT1 (Tau, Azurin, and Ubiquitin). It was demonstrated that the results from the RMSD (and \({R}_{2}^{s}\)) metric correspond well to the results obtained using a more extended and advanced set of the NUScon metrics9,31. With its lower RMSD (and higher \({R}_{2}^{s}\)), the MR-Ai displays visibly better results compared to the CS for both Echo and Anti-Echo reconstructions. The even better quality score obtained for the reconstruction from time equivalent 50% NUS experiment (Fig. 2b) indicates that well-randomized NUS is a better time-saving strategy than acquiring only N- or P-type data.

Figure 2b shows that the quality scores for the Anti-Echo reconstruction are higher than for the Echo regardless of the reconstruction method MR-Ai or CS, which is also reproduced for three other proteins shown in Supplementary Fig. 9. At first glance, this is a surprising result since from the theory we expect the quality of the reconstructions to be the same for Echo and Anti-Echo. However, in practice, we should note that these are separate experiments with somewhat different pulse sequences32, which may lead to imbalances between the two spectra. We reproduced the result in simulations (Fig. 6b) where the Echo signal had a somewhat lower amplitude than the Anti-Echo. Then, the better result for the Anti-Echo can be explained by the residual unbalanced contribution of the Anti-Echo part in the traditionally processed reference spectrum. This result underlines the value of the reconstruction approach from the individual Echo or Anti-Echo parts in cases of imbalance between the two or if only one can be practically obtained.

For any physical measurement, such as the intensity at a point in a reconstructed spectrum, estimating the error is equally important for quantitative analysis and distinguishing true signals from noise and artifacts. This task is particularly challenging when using nonlinear processing methods such as CS and DNN, since the RMSD of the baseline noise can no longer be used as a reliable error estimate. Traditionally, this problem is solved by the brute force approach of repeated measures or post-experiment resampling of the data31,33. DNN offers a much more efficient alternative34. It is possible to train a network to predict the quality of the results generated by any method35 by employing the negative log-likelihood (NLL) as the loss function during the training stage36.

where for point i in the spectrum, yi, represents the ground truth value; μi is the value produced by the applied method, and σi is the uncertainty represented by the standard deviation of a normal distribution. During the training, σi is learned, while yi and μi are known.

The modified MR-Ai architecture employing the NLL as the loss function is illustrated in Fig. 3. In this work, we trained three MR-Ai to predict intensity uncertainties in the Echo (or Anti-Echo) spectra reconstructed by MR-Ai and CS, as well as NUS spectra reconstructed by CS. In Fig. 2a (and in Supplementary Figs. 5 to 8), the pink color represents the estimated uncertainty with a 95% confidence interval (CI) overlaid on the reconstructed spectrum. The true error, which is the difference between the reference and reconstruction for the shown spectrum, is superimposed in red for visual comparison.

The ground truth synthetic spectra (y) are used to produce incomplete or corrupted spectra, e.g. Echo or NUS as inputs for the WNN training. The training is performed by minimizing the negative log-likelihood (NLL) loss function, which is a measure of mismatch between the predicted uncertainties (σ) and statistics on the observed difference between the ground truth (y) and the spectra reconstructions (μ). At the production stage, limited to the part of the scheme within the dashed line, the trained WNN predicts uncertainty for the spectrum reconstructed by the trained nonlinear method.

The predicted uncertainty can be used as a reference-free score of the spectrum quality, akin to the predicted protein structure accuracy pTM-score calculated by AlphaFold37. In Fig. 2c (and Supplementary Fig. 10), box plots display the estimated normalized sigma for all points with the mean (green triangles) and median (orange bars). The box plots correlate well with the reference-based scores shown in Fig. 2b. Moreover, the box plot sores for the Echo and Anti-Echo reconstructions are nearly the same, as they should be, whereas the reference-based scores depicted in Fig. 2b wrongly prefer the Anti-Echo because of the biased reference as described above.

In this work, we introduce the MR-Ai toolbox, which offers intelligent NMR processing capabilities beyond traditional techniques. The reconstruction of spectra using the incomplete Echo/Anti-Echo quadrature detection pair can be advantageous for saving experimental time and in cases where obtaining a complete and intensity-balanced signal for traditional quadrature is problematic. The predicted uncertainties of spectral intensities and quantitative reference-free spectrum quality metric will aid in the development of new spectrum processing algorithms and may become a crucial component in methods utilizing targeted acquisitions approaches38,39. Our results demonstrate the potential of AI to expand the scope of traditional NMR signal processing and analysis.

In our previous work, we demonstrated the effectiveness of the WaveNet-based NMR Network (WNN) in capturing regular 1D patterns in 2D NMR spectra, including point spread function (PSF) patterns in non-uniformly sampled (NUS) spectra and peak multiplets resulting from scalar coupling9. For the Echo (or Anti-Echo) reconstruction, where the phase-twist lineshape manifests as a 2D pattern, we designed a 2D WNN architecture to accommodate this feature. Within the MR-Ai framework for Echo (or Anti-Echo) reconstruction, the overall network architecture comprises two primary components: WNNs and correction steps.

The architecture of the MR-Ai for Echo (or Anti-Echo) reconstruction consists of five distinct WNNs (Fig. 4a). Although a single network produces a relatively good spectrum, better quantitative reconstructions often require refinement by several iterations6,9. The WNNs are individually trained in sequential order, with each utilizing the output from the pre-trained upstream network as input. The inputs and outputs for each WNN are normalized using the Euclidean norm value from the input before training. The initial spectrum \({{{{\bf{S}}}}}_{{{{\rm{Echo}}}}}\in {{\mathbb{R}}}^{2k\times 2l}\), containing strong phase-twist lineshape is given as input to the first WNN. Each consecutive WNN diminishes inherent artifacts and generates the spectrum \({\widetilde{{{{\bf{S}}}}}}^{i}\in {{\mathbb{R}}}^{2k\times 2l}\), which progressively aligns more closely with the correct spectrum. As input to each subsequent WNNi (i = 1, 2, 3, 4), we computed the corrected spectrum, \({{{{\bf{S}}}}}_{\,{\mbox{Cor}}\,}^{i}\in {{\mathbb{R}}}^{2k\times 2l}\):

here \({\widetilde{{{{\bf{S}}}}}}_{\,{\mbox{Anti-Echo}}\,}^{i}\) represents the Anti-Echo part of the predicted spectrum (\({\widetilde{{{{\bf{S}}}}}}^{i}\)) produced by zeroing (ZE) of the Echo part in its time domain VE presentation28 (i.e. zeroing P and \(\widetilde{P}\) in Fig. 1d). Note that the VE is the 2D complex (\({{\mathbb{C}}}^{2k\times 2l}\)) time-domain signal obtained by the inverse 2D Fourier transform of a real part (\({{\mathbb{R}}}^{2k\times 2l}\)) of the spectrum. At step i, C(i) = 1 − 0.05 × 21−i is a real scalar step-dependant factor. In other words, the correction uses clear separation of the Echo and Anti-Echo parts in the VE time domain presentation of the spectrum and slightly attenuates the predicted Anti-Echo part \({\widetilde{{{{\bf{S}}}}}}_{\,{\mbox{Anti-Echo}}\,}^{i}\) while restoring the Echo part to the input values SEcho. The correction is not applied after the final step. We found that the correction significantly increases quality of the reconstructed spectrum.

a MR-Ai network architecture with five 2D WNNs and correction steps. b Scheme of the WNN module used in MR-Ai (c) Training- and cross-validation losses for the WNNs during training.

The WNN architecture employed in MR-Ai was inspired by the WaveNet architecture, originally designed for analyzing long audio signals in the time domain40. Similar to WaveNet, WNN utilizes dilated convolutional layers that skip a specified number of data points and thus can be seen as convolutional layers with gaps. By assigning various dilation sizes to different convolutional layers, it is possible to build a block that behaves like a convolutional layer with an extensive kernel size. In this architecture, except for the first layer where a 2 × 4 kernel size with 50 filters is applied without any dilation rate, subsequent layers, such as the mth layer, employ a (2m−1, 2m) dilation rate along with a 2 × 2 kernel size with 50 filters. The WNN consists of 5 layers utilizing Rectified Linear Unit (ReLU) activation functions41 with no padding between the layers (Fig. 4b). In this configuration, the output layer’s dimensions are 32 × 64, which, due to the specific layers chosen and the absence of padding, corresponds to the input layer size of 63 × 127. We generated the network graphs using TensorFlow python package42 with the Keras frontend. The training was performed within TensorFlow using the stochastic ADAM optimizer43 with the default parameters and 0.0001 learning rate, Mean Square Error loss function, mini-batch size equal to 64, and the number of epochs equal to 1000 or less when a monitored metric has stopped improving on the validation data set (see Supplementary Methods 1.1 for Python Code).

We trained WNNs on the NMRbox server44 (128 cores 2 TB memory), equipped with 4 NVIDIA A100 TENSOR CORE GPU graphics cards. The final training and cross-validation losses for all five WNNs used for Echo reconstruction are shown in Fig. 4c.

In this study, we utilized MR-Ai to predict the uncertainty of intensities at every point of the spectrum produced by a given reconstruction method. Several methods for estimating uncertainty by DNN are available34. Due to the lightweight nature of the WNN, the small number of training parameters, and the fast training process, it was possible to adopt a Gaussian mixture model-based approach for the uncertainty estimation36. We achieved this by including a Gaussian distribution layer within TensorFlow (TFP Probabilistic Layers) into the WNN and utilizing the negative log-likelihood (NLL) as the loss function (Eq. (1)). The WNN utilizes the reconstructed spectrum intensities as the fixed means of Gaussian distributions (μ) and determines σ as the corresponding uncertainty based on input data (Fig. 3). To ensure that the estimated uncertainty (σ) is consistently positive and non-zero, we applied a ReLU activation function prior to the Gaussian distribution layer with standard regularization factors in TensorFlow (see Supplementary Methods 1.2 for Python Code). Here, we trained three MR-Ai models for uncertainty estimation, specifically for MR-Ai Echo (or Anti-Echo), CS Echo (or Anti-Echo), and CS NUS. The final training and cross-validation losses for these networks are shown in Fig. 5.

Training- and cross-validation losses for the WNNs during training MR-Ai models for uncertainty estimation.

For training the DNNs we used synthetic data. The 2D NMR hyper-complex time domain signal \({{{{\bf{X}}}}}_{{{{\rm{FID}}}}}\in {{\mathbb{H}}}^{k\times l}\), referred to as the Free Induction Decay (FID), is represented as a combination of P − and N − complex signals \({{{{\bf{X}}}}}_{{{{\rm{P}}}}}\in {{\mathbb{C}}}^{k\times l}\) and \({{{{\bf{X}}}}}_{{{{\rm{N}}}}}\in {{\mathbb{C}}}^{k\times l}\), respectively:

where sum is over N exponentials, and the nth exponential has the amplitudes An, phases \({\phi }_{{1}_{n}}\) and \({\phi }_{{2}_{n}}\), relaxation times \({\tau }_{{1}_{n}}\) and \({\tau }_{{2}_{n}}\), and frequencies \({\omega }_{{1}_{n}}\) and \({\omega }_{{2}_{n}}\) in the indirect and direct dimensions, respectively. The evolution times t1 and t2 are given by the series indexed 0, 1, ..., T-1, where T = k, l is the number of complex points in each dimension. XP and XN parts are defined by + and − signs before the imaginary unit in the first exponent, respectively. The desired number of different FIDs for the training set is simulated by randomly varying the above parameters in the ranges summarized in Table 1. These parameters are representative for 2D 1H-15N correlation spectra. However, the DNN would require re-training for a spectrum, where the signal features are outside of the ranges in the table, for example, having larger phase distortions or wider spread of peak intensities as may be observed in NOESY spectra or metabolomics data. The very high dynamic range of the signal intensities may even require changes in the DNN architecture.

We added Gaussian noise to emulate the noise present in realistic NMR spectra. Subsequently, standard 2D signal processing by the Python package nmrglue45, including apodization, zero-filling, FT, and phase correction, are employed to obtain pure absorption mode spectra S (Fig. 1b) from the hypercomplex XFID.

The Virtual Echo28 presentation (VE), which is a straightforward method for generating Echo (or Anti-Echo) spectra with a phase-twist lineshape, is obtained by inverse 2D Fourier Transform (IFT) of the real part of the spectrum S (Fig. 1d). The VE representation is essentially the combination of P- and N-type complex time domain data. The Echo spectrum SEcho (Fig. 1a) is obtained by zeroing N-type part of the data (i.e. N and \(\widetilde{N}\) areas in Fig. 1c) followed by 2D FT. Similarly, the Anti-Echo SAnti-Echo spectrum is obtained by zeroing the P-type data. In this work, we synthesized 1024 [SEcho, S] datasets for training MR-Ai Echo (or Anti-Echo) reconstruction based on Table 1.

To train the MR-Ai for predicting uncertainties in the spectra reconstructed by various methods, the training dataset includes the input data, the reconstruction result obtained from the method, and reference ground truce spectra S. In a production run, the trained MR-Ai can estimate the uncertainty of the reconstruction generated by the method using only the input data, e.g., the twisted Echo spectrum or NUS spectrum with aliasing artifacts (Table 2, Fig. 3). Here, we used three training datasets based on Table 3 to train three MR-Ai for uncertainty estimation for MR-Ai Echo (or Anti-Echo), CS Echo (or Anti-Echo), and CS NUS.

For the training, multiple synthetic NUS spectra SNUS are generated with the same 50% NUS schedule of the Poisson-gap type46 sampling from uniformly sampled spectra S.

To test CS and trained MR-Ai performances, we used previously described 2D spectra for several proteins: U — 15N-13C — labeled ubiquitin (8.6 kDa)47, U — 15N-13C — labeled Cu(I) azurin (14 kDa)26, U — 15N-13C-2H methyl ILV back-protonated MALT1 (44 kDa)48, and U — 15N-13C-2H Tau (46 kDa, IDP)27. The fully sampled two-dimensional experiments used in this study are described in Table 2. We used NMRPipe49, mddnmr50, and Python package nmrglue45 for reading, writing, and traditional processing of the NMR spectra.

We utilized the Compressed Sensing Iterative Soft Thresholding algorithm (CS-IST) combined with VE, similar to recent versions of the CS module in the mddnmr software28,29. The same algorithm was used for NUS reconstruction with 50% fixed Poisson Gap random sampling NUS spectra.

We accessed the quality of the experimental spectra obtained using three reconstruction methods: MR-Ai for Echo (Anti-Echo) data; CS for Echo (Anti-Echo) data; CS for time equivalent 50% NUS.

To evaluate the similarity between reconstructed and complete reference spectra, we utilized two traditional reference-based evaluation metrics: the root-mean-square deviation (RMSD) and the correlation coefficient (\({R}_{S}^{2}\)) between reconstruction and reference spectra. First, all spectra were normalized based on their maximal peak intensity. The RMSD and \({R}_{S}^{2}\) were calculated only for spectral points with intensities above 1% of the highest peak intensity in either of the reconstruction or reference spectra, aiming to limit the potential effects of the baseline noise on the quality metrics. This approach ensures that both metrics are sensitive to false-positive and false-negative spectral artifacts, while the points with very low near/below noise intensities are ignored. It was demonstrated that the results from the RMSD and \({R}_{2}^{s}\) metrics closely correspond to the results obtained using the NUScon metrics9,31. Supplementary Fig. 9 displays the comparison between different reconstruction methods on four experimental spectra of different proteins with varying complexities.

relies on the predicted uncertainties for each point on the spectrum. Initially, all predicted uncertainties for each point were normalized based on the maximal peak intensity in the reconstructed spectrum. By comparing the predicted uncertainties only for spectral points with intensities above 1% of the highest peak intensity in the reconstruction, we can evaluate the performance of reconstruction across different methods (Supplementary Fig. 10).

The synthetic testing data is used to simulate two distinct scenarios Fig. 6. In the first case, P- and N-type data parts were amplitude-balanced and consequently, produced an ideal peak line shape. In the second scenario, imbalanced in the amplitude of P- and N-type data (in this case, P-type is 10% smaller than N-type) resulted in an imperfect peak in the traditionally processed reference spectrum with a visible slight residual phase twisted line shape. This line imperfection distorts the reference-based metric by favoring the reconstruction of the corresponding Echo (or Anti-Echo) spectrum regardless of the used method MR-Ai or CS (Fig. 6b). The pSQ metric (Fig. 6c) does not use the distorted reference spectrum and thus shows expected equal quality for the Echo and Anti-Echo spectra reconstructed by MR-Ai and CS.

a The effects of balanced and imbalanced P- and N-type data on outcomes in normal phase-modulated data processing. MR-Ai and CS reconstruction performance under conditions of balanced and imbalanced P- and N-type synthetic data by using (b) Traditional reference-based evaluation metrics and (c) Predicted reference-free quality metric shown as a box plot of normalized spectra uncertainties. The orange bar and the green triangle indicate the median and mean of the uncertainty distribution, respectively.

The data required and trained WNNs to reproduce our results are available at https://github.com/Amir-Jahangiri/MRAi_2D. Any additional supporting data are available from the corresponding authors upon reasonable request.

The source codes for WNN architectures are available in Supplementary Methods. Additionally, the code for Echo reconstruction and uncertainty estimation to reproduce our results are available at https://github.com/Amir-Jahangiri/MRAi_2D. Any additional supporting codes are available from the corresponding authors upon reasonable request.

Cavanagh, J., Fairbrother, W. J., Palmer III, A. G. & Skelton, N. J.Protein NMR spectroscopy: principles and practice (Academic press, 1996).

Claridge, T. D.High-resolution NMR techniques in organic chemistry, vol. 27 (Elsevier, 2016).

Chen, D., Wang, Z., Guo, D., Orekhov, V. & Qu, X. Review and prospect: deep learning in nuclear magnetic resonance spectroscopy. Chem.–A Eur. J. 26, 10391–10401 (2020).

Article CAS Google Scholar

Shukla, V. K., Heller, G. T. & Hansen, D. F. Biomolecular nmr spectroscopy in the era of artificial intelligence. Structure 31, 1360–1374 (2023).

Article CAS PubMed Google Scholar

Reilly, C. & Kowalski, B. Nuclear magnetic resonance spectral interpretation by pattern recognition. J. Phys. Chem. 75, 1402–1411 (1971).

Article CAS Google Scholar

Qu, X. et al. Accelerated nuclear magnetic resonance spectroscopy with deep learning. Angew. Chem. 132, 10383–10386 (2020).

Article Google Scholar

Hansen, D. Using deep neural networks to reconstruct non-uniformly sampled nmr spectra. J. Biomol. NMR 73, 577–585 (2019).

Article CAS PubMed PubMed Central Google Scholar

Karunanithy, G. & Hansen, D. Fid-net: A versatile deep neural network architecture for nmr spectral reconstruction and virtual decoupling. J. Biomol. NMR 75, 179–191 (2021).

Article CAS PubMed PubMed Central Google Scholar

Jahangiri, A. et al. Nmr spectrum reconstruction as a pattern recognition problem. J. Magn. Reson. 346, 107342 (2023).

Article CAS PubMed Google Scholar

Karunanithy, G., Mackenzie, H. W. & Hansen, D. F. Virtual homonuclear decoupling in direct detection nuclear magnetic resonance experiments using deep neural networks. J. Am. Chem. Soc. 143, 16935–16942 (2021).

Article CAS PubMed Google Scholar

Kazimierczuk, K. et al. Resolution enhancement in nmr spectra by deconvolution with compressed sensing reconstruction. Chem. Commun. 56, 14585–14588 (2020).

Article CAS Google Scholar

Qiu, T. et al. Resolution enhancement of nmr by decoupling with the low-rank hankel model. Chem. Commun. 59, 5475–5478 (2023).

Article CAS Google Scholar

Zheng, X. et al. Fast acquisition of high-quality nuclear magnetic resonance pure shift spectroscopy via a deep neural network. J. Phys. Chem. Lett. 13, 2101–2106 (2022).

Article CAS PubMed Google Scholar

Zhan, H. et al. Fast pure shift nmr spectroscopy using attention-assisted deep neural network. Advanced Science 2309810 (2024).

Zhan, H., Liu, J., Fang, Q., Chen, X. & Hu, L. Accelerated pure shift nmr spectroscopy with deep learning. Anal. Chem. 96, 1515–1521 (2024).

Article CAS PubMed Google Scholar

Lee, H. & Kim, H. Intact metabolite spectrum mining by deep learning in proton magnetic resonance spectroscopy of the brain. Magn. Reson. Med. 82, 33–48 (2019).

Article CAS PubMed Google Scholar

Chen, D. et al. Magnetic resonance spectroscopy deep learning denoising using few in vivo data. IEEE Trans. Comput. Imaging 9, 448–458 (2023).

Article Google Scholar

Klukowski, P. et al. Nmrnet: A deep learning approach to automated peak picking of protein nmr spectra. Bioinforma. (Oxf., Engl.) 34, 2590–2597 (2018).

CAS Google Scholar

Li, D.-W., Hansen, A. L., Bruschweiler-Li, L., Yuan, C. & Brüschweiler, R. Fundamental and practical aspects of machine learning for the peak picking of biomolecular nmr spectra. J. Biomol. NMR 76, 49–57 (2022).

Article CAS PubMed PubMed Central Google Scholar

States, D., Haberkorn, R. & Ruben, D. A two-dimensional nuclear overhauser experiment with pure absorption phase in four quadrants. J. Magn. Reson. 48, 286–292 (1982).

CAS Google Scholar

Marion, D., Ikura, M., Tschudin, R. & Bax, A. Rapid recording of 2d nmr spectra without phase cycling. application to the study of hydrogen exchange in proteins. J. Magn. Reson 85, 393–399 (1989).

CAS Google Scholar

Davis, A. L., Keeler, J., Laue, E. D. & Moskau, D. Experiments for recording pure-absorption heteronuclear correlation spectra using pulsed field gradients. J. Magn. Reson. 98, 207–216 (1992).

CAS Google Scholar

Kontaxis, G., Stonehouse, J., Laue, E. & Keeler, J. The sensitivity of experiments which use gradient pulses for coherence-pathway selection. J. Magn. Reson. Ser. A 111, 70–76 (1994).

Article CAS Google Scholar

Bostock, M., Holland, D. & Nietlispach, D. Improving resolution in multidimensional nmr using random quadrature detection with compressed sensing reconstruction. J. Biomol. NMR 68, 67–77 (2017).

Article CAS PubMed Google Scholar

Han, X. et al. Assignment of ivl-methyl side chain of the ligand-free monomeric human malt1 paracaspase-igl3 domain in solution. Biomol. NMR Assign. 16, 363–371 (2022).

Article CAS PubMed PubMed Central Google Scholar

Korzhnev, D. M., Karlsson, B. G., Orekhov, V. Y. & Billeter, M. Nmr detection of multiple transitions to low-populated states in azurin. Protein Sci. 12, 56–65 (2003).

Article CAS PubMed PubMed Central Google Scholar

Lesovoy, D. M. et al. Unambiguous tracking of protein phosphorylation by fast high-resolution fosy nmr. Angew. Chem. Int. Ed. 60, 23540–23544 (2021).

Article CAS Google Scholar

Mayzel, M., Kazimierczuk, K. & Orekhov, V. Y. The causality principle in the reconstruction of sparse nmr spectra. Chem. Commun. 50, 8947–8950 (2014).

Article CAS Google Scholar

Kazimierczuk, K. & Orekhov, V. Y. Accelerated nmr spectroscopy by using compressed sensing. Angew. Chem. Int. Ed. 50, 5556–5559 (2011).

Article CAS Google Scholar

Holland, D. J., Bostock, M. J., Gladden, L. F. & Nietlispach, D. Fast multidimensional nmr spectroscopy using compressed sensing. Angew. Chem. Int. Ed. 50, 6548–6551 (2011).

Article CAS Google Scholar

Pustovalova, Y. et al. Nuscon: a community-driven platform for quantitative evaluation of nonuniform sampling in nmr. Magn. Reson. 2, 843–861 (2021).

Article CAS Google Scholar

Horne, T. J. & Morris, G. A. P-type gradient-enhanced cosy experiments show lower t1 noise than n-type. Magn. Reson. Chem. 35, 680–686 (1997).

3.0.CO;2-#" data-track-item_id="10.1002/(SICI)1097-458X(199711)35:103.0.CO;2-#" data-track-value="article reference" data-track-action="article reference" href="https://doi.org/10.1002%2F%28SICI%291097-458X%28199711%2935%3A10%3C680%3A%3AAID-OMR149%3E3.0.CO%3B2-%23" aria-label="Article reference 32" data-doi="10.1002/(SICI)1097-458X(199711)35:103.0.CO;2-#">Article CAS Google Scholar

Mayzel, M., Ahlner, A., Lundström, P. & Orekhov, V. Y. Measurement of protein backbone 13 co and 15 n relaxation dispersion at high resolution. J. Biomol. NMR 69, 1–12 (2017).

Article CAS PubMed PubMed Central Google Scholar

Abdar, M. et al. A review of uncertainty quantification in deep learning: Techniques, applications and challenges. Inf. Fusion 76, 243–297 (2021).

Article Google Scholar

Scalia, G., Grambow, C. A., Pernici, B., Li, Y.-P. & Green, W. H. Evaluating scalable uncertainty estimation methods for deep learning-based molecular property prediction. J. Chem. Inf. model. 60, 2697–2717 (2020).

Article CAS PubMed Google Scholar

Nix, D. A. & Weigend, A. S. Estimating the mean and variance of the target probability distribution. In Proceedings of 1994 ieee international conference on neural networks (ICNN’94), vol. 1, 55–60 (IEEE, 1994).

Jumper, J. et al. Highly accurate protein structure prediction with alphafold. Nature 596, 583–589 (2021).

Article CAS PubMed PubMed Central Google Scholar

Jaravine, V. A. & Orekhov, V. Y. Targeted acquisition for real-time nmr spectroscopy. J. Am. Chem. Soc. 128, 13421–13426 (2006).

Article CAS PubMed Google Scholar

Isaksson, L. et al. Highly efficient nmr assignment of intrinsically disordered proteins: application to b-and t cell receptor domains. PLos one 8, e62947 (2013).

Article CAS PubMed PubMed Central Google Scholar

van den Oord, A. et al. Wavenet: A generative model for raw audio 1609.03499 (2016).

Agarap, A. F. Deep learning using rectified linear units (relu). arXiv preprint arXiv:1803.08375 (2018).

Abadi, M. et al. Tensorflow: Large-scale machine learning on heterogeneous distributed systems. arXiv preprint arXiv:1603.04467 (2016).

Kingma, D. P. & Ba, J. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014).

Maciejewski, M. W. et al. Nmrbox: a resource for biomolecular nmr computation. Biophys. J. 112, 1529–1534 (2017).

Article CAS PubMed PubMed Central Google Scholar

Helmus, J. J. & Jaroniec, C. P. Nmrglue: an open source python package for the analysis of multidimensional nmr data. J. biomol. NMR 55, 355–367 (2013).

Article CAS PubMed PubMed Central Google Scholar

Hyberts, S. G., Takeuchi, K. & Wagner, G. Poisson-gap sampling and forward maximum entropy reconstruction for enhancing the resolution and sensitivity of protein nmr data. J. Am. Chem. Soc. 132, 2145–2147 (2010).

Article CAS PubMed PubMed Central Google Scholar

Brzovic, P. S., Lissounov, A., Christensen, D. E., Hoyt, D. W. & Klevit, R. E. A ubch5/ubiquitin noncovalent complex is required for processive brca1-directed ubiquitination. Mol. cell 21, 873–880 (2006).

Article CAS PubMed Google Scholar

Unnerståle, S. et al. Backbone assignment of the malt1 paracaspase by solution nmr. Plos one 11, e0146496 (2016).

Article PubMed PubMed Central Google Scholar

Delaglio, F. et al. Nmrpipe: a multidimensional spectral processing system based on unix pipes. J. biomol. NMR 6, 277–293 (1995).

Article CAS PubMed Google Scholar

Orekhov, V. Y. & Jaravine, V. A. Analysis of non-uniformly sampled spectra with multi-dimensional decomposition. Prog. Nucl. Magn. Reson. Spectrosc. 59, 271–292 (2011).

Article CAS PubMed Google Scholar

Download references

The work was supported by the Swedish Research Council grants to 2019-03561 and 2023-03485 to V.O. This study used NMRbox: National Center for Biomolecular NMR Data Processing and Analysis, a Biomedical Technology Research Resource (BTRR), which is supported by NIH grant P41GM111135 (NIGMS).

Open access funding provided by University of Gothenburg.

Department of Chemistry and Molecular Biology, Swedish NMR Centre, University of Gothenburg, Gothenburg, 40530, Sweden

Amir Jahangiri & Vladislav Orekhov

You can also search for this author in PubMed Google Scholar

You can also search for this author in PubMed Google Scholar

V.O. conceptualized and supervised. A.J. performed analyses. All authors scripted, and wrote, reviewed, approved the manuscript.

Correspondence to Vladislav Orekhov.

The authors declare no competing interests.

Communications Chemistry thanks Haolin Zhan and the other anonymous reviewers for their contribution to the peer review of this work.

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions

Jahangiri, A., Orekhov, V. Beyond traditional magnetic resonance processing with artificial intelligence. Commun Chem 7, 244 (2024). https://doi.org/10.1038/s42004-024-01325-w

Download citation

Received: 15 June 2024

Accepted: 09 October 2024

Published: 27 October 2024

DOI: https://doi.org/10.1038/s42004-024-01325-w

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative