Review and Prospect: Deep Learning in Nuclear Magnetic Resonance Spectroscopy

1. Introduction

• NMR spectroscopy provides atomic-level molecular-structure information and is used to analyze molecular dynamics and interactions.The review notes its applications in chemistry, life sciences, and other fields.Source: §1 · p. 1
• Deep learning can learn features and complex nonlinear input-output mappings without prior knowledge or formal assumptions.The review describes practical NMR applications as emerging with newer algorithms and modern computing power.Source: §1 · p. 1
• Undersampling can reduce NMR measurement time, but lost information then requires reconstruction; deep learning is presented as an alternative to existing methods.The review also identifies denoising low-SNR spectra, chemical-shift prediction, and automated peak picking as potential applications.Source: §1 · p. 1
• Figure 1 illustrates deep learning by training on labeled animal images and testing whether the trained network recognizes animals.Backpropagation adjusts the network's internal parameters during training.Source: Figure 1 · p. 1

2. Basic Architectures of Deep Learning

• DNNs, CNNs, and RNNs are the three main architectures discussed for NMR applications; here, DNN refers mainly to a fully connected multilayer perceptron.Training uses forward propagation, loss calculation, backpropagation, and parameter updates; regularization helps avoid overfitting and improve generalization.Source: §2 · pp. 1–2

2.1. Deep Neural Networks

• A DNN passes input through fully connected layers, applying nonlinear activation functions so it can fit complex input-output mappings.Its structure comprises an input layer, multiple hidden layers, and an output layer.Source: §2.1 · p. 2

2.2. Convolutional Neural Networks

• CNNs apply convolution, nonlinear activation, and pooling; shared filters reduce the number of learned parameters.Pooling aggregates similar features and can also help avoid overfitting and improve generalization.Source: §2.2 · p. 2

2.3. Recurrent Neural Networks

• RNNs process sequences one element at a time, carrying a hidden state that captures preceding elements and contributes to the current output.Conventional RNNs can suffer vanishing gradients and difficulty storing very long time series; LSTM memory cells use gates for long-term storage.Source: §2.3 · p. 3

2.4. Deep Learning Libraries

• The review lists several mainstream deep-learning libraries but identifies no clear leader, noting that each has advantages.Its application overview summarizes software and hardware bases, network architectures, and shared resources.Source: §2.4 · p. 3

3. Reconstruction of the Spectra

• NUS accelerates acquisition, but established reconstruction methods may require prior assumptions and lengthy iterative computation, whereas DL learns mappings from undersampled FIDs to target spectra and reconstructs them quickly.The review attributes DL's speed to non-iterative, low-complexity algorithms that support GPU parallelization.Source: §3 · pp. 3–4
• Qu et al. [4a] used five dense CNN blocks with data consistency to remove aliasing artifacts, training successfully on synthetic NMR data generated with exponential functions [28i, 29].Figure 6 depicts the undersampled FID, artifact removal, data consistency update, and final reconstructed spectrum.Source: §3 · p. 4, Figure 6 · p. 4
• CNN computation took 4~8% of Low-Rank time for 2D spectra and 12~22% of compressed sensing time for 3D spectra.Figure 7 compares computational time for reconstructions using the three methods.Source: Figure 7 · p. 4, §3 · p. 4
• Hansen [4b] used a modified LSTM to reconstruct NUS time-domain FIDs before Fourier transformation, obtaining accurate peak intensities but 2D computational time similar to conventional methods.Unlike CNN reconstruction in the frequency domain, this approach processes time-domain signals.Source: §3 · p. 4, Figure 8 · p. 4

4. Denoising of the Spectra

• Low SNR, metabolite overlap, field inhomogeneity, and baseline distortion challenge brain spectra, while conventional time-domain denoising filters are limited by varied spin-system decay rates.Short-TE macromolecule signals can superimpose on the spectral baseline across the range.Source: §4 · p. 4
• Lee and Kim trained and tested a CNN on simulated brain spectra spanning SNR 6.90-20.74 and linewidth 10-20 Hz, then tested it on in vivo spectra from five volunteers.The simulations combined noise, line broadening, frequency/phase shifts, and spectral baseline effects to represent in vivo spectra.Source: §4 · pp. 4–5, Figure 9 · p. 5
• The CNN inferred high-SNR spectra from heavily interfered spectra, and its reported robustness may allow sub-minute 1H brain spectra acquisition for clinical applications.The method was tested on in vivo spectra from five healthy volunteers with substantially different SNR from the training simulations.Source: §4 · pp. 4–5

5.1. Chemical Shift Prediction

• Chemical shifts encode structural information but are difficult to calculate deterministically because hydrogen bonding, local electric fields, and ring-current effects contribute to their values.Source: §5.1 · p. 5
• SPARTA+ [21] improved 2–10% over the best methods, while PPM_One performed better than SPARTA+ for all atoms except C′ carbonyl carbons.Source: §5.1 · p. 5
• Liu et al. reported strong agreement for 13C, 15N, and 17O shifts, while 1H accuracy was highest and comparable to ab initio quantum chemistry methods.Their MR-3D-DenseNet predicts shifts from an atom-centered Gaussian density representation of the chemical environment.Source: §5.1 · p. 5
• TALOS-N predicted backbone torsion angles for at least 90% of residues with an error rate below 3.5% on an independent protein set.Its DNN maps chemical shifts to a 324-state torsion-angle distribution.Source: §5.1 · p. 5

5.2. Automated Peak Picking

• Peak picking remains difficult to automate because of peak overlap, low SNR, line distortions, and spectral artifacts.Source: §5.2 · p. 6
• NMR-Net achieved average precision of 0.9596 for backbone, 0.9058 for side-chain, and 0.8271 for NOESY peaks on 31 manually annotated spectra.The method classifies candidate peaks using a CNN applied to normalized 48×48 spectrum patches.Source: §5.2 · p. 6
• Bruker’s trained DNN identified regions corresponding to actual 1D 1H NMR signals in experimental data with accuracy consistent with manual selection.Source: §5.2 · p. 6

• The review reports that DL has helped relate noisy or distorted NMR signals to intact spectra and replace complex calculations or manual analyses such as chemical-shift prediction and peak picking.Source: §6 · p. 6
• DL remains difficult to interpret, and proposed efforts to open the black box include meta-learning causal structure [46] and a group-theoretical procedure [47].Source: §6 · p. 6
• Training-data shortages hinder DL development; the review points to simulated data [4a, 4b, 5, 20, 26], larger trustworthy databases, and combined simulated and experimental training.BioMagResBank and the Protein Data Bank contain over 11,900 entries with assigned shifts and coupling constants; NMRbox is cited as a data- and tool-sharing example.Source: §6 · p. 6
• The outlook includes accelerated high-quality reconstruction of high-dimensional biochemical NMR spectra and removing residual water signals and other artifacts during denoising.Source: §6 · pp. 6–7
• Potential future interpretation tasks include chemical-shift assignment, structure discovery, and describing new compounds’ physicochemical properties, with applications also extending to diffusion, dynamic, and large-scale spectral data.The review also proposes combining time- and frequency-domain inputs.Source: §6 · p. 7