A Deep Learning-Based Virtual Decoupling Method for One-Dimensional Proton NMR Spectra

Abstract

This invention provides a deep learning-based virtual decoupling method for one-dimensional proton NMR spectra. The method mainly involves using a deep learning model to virtually decouple one-dimensional proton NMR spectra. First, training samples are generated for the deep learning model. Second, the structure of the deep learning model is designed. Then, the deep learning model is trained using the training samples to obtain a trained model. Finally, the trained deep learning model is used to virtually decouple the one-dimensional proton NMR spectrum. The main advantage of this invention is that it eliminates the need to design pulse sequences; one-dimensional proton spectrum decoupling can be achieved solely through a deep learning model.

Description

Technical Field

[0001] This invention pertains to nuclear magnetic resonance (NMR) spectroscopy signal processing methods, and is a deep learning-based virtual decoupling method for one-dimensional proton NMR spectra. Background Technology

[0002] Nuclear magnetic resonance spectroscopy is a powerful analytical technique often used for analyzing molecular structure and dynamic environments. Specific atomic nuclei (such as...) 1 H and 13 C) Under the influence of an external magnetic field, nuclear spins align, absorbing and emitting electromagnetic radiation of specific frequencies, thus forming a nuclear magnetic resonance (NMR) spectrum, providing information about the molecular structure. Coupling splits are a common phenomenon in NMR spectroscopy, caused by interactions between adjacent spins. They lead to crowded peaks and decreased resolution, causing significant difficulties in spectral interpretation. To remove the negative effects of coupling splits and restore pure chemical shift spectra, traditional methods primarily use pulse sequences, including BIRD, ZS, and PSYCHE pulse sequences, for decoupling. However, traditional decoupling methods are time-consuming, experimentally challenging, and the resulting pure chemical shift spectra suffer from spurious peaks and decreased resolution, which greatly limits the application of pure chemical shift spectroscopy. Summary of the Invention

[0003] The main technical problem to be solved by this invention is to provide a virtual decoupling method for one-dimensional proton NMR spectra based on deep learning, which can decouple one-dimensional proton NMR spectra to obtain a clean, purely chemical shift spectrum.

[0004] To address the technical challenges of long experimental time and high difficulty in decoupling pulse sequences, this invention adopts the following technical solution:

[0005] A deep learning-based virtual decoupling method for one-dimensional proton NMR spectra includes the following steps:

[0006] S1: The training and test sets required to generate a deep learning model;

[0007] S2: Design the deep learning model architecture and build the model;

[0008] S3: The deep learning model is trained using the training set and the test set to obtain a converged and trained deep learning model.

[0009] S4: The trained deep learning model is tested using the test set and the one-dimensional hydrogen spectrum obtained from the experiment.

[0010] In a preferred embodiment, step S1 specifically includes:

[0011] S11: Based on the characteristic that the free induction decay signal is composed of the superposition of complex exponential signals, a mathematical model is used to generate a simulated one-dimensional nuclear magnetic resonance hydrogen spectrum with Voigt peak shape and noise and distortion.

[0012] The parameters of the simulated one-dimensional nuclear magnetic resonance hydrogen spectrum include: number of coupling systems, number of spins in the coupling system, spectral width, linewidth, peak position, coupling constant, noise intensity, and distortion degree;

[0013] S12: Generate a pure chemical shift spectrum with no coupling split peaks, no noise and no distortion, corresponding to the one-dimensional nuclear magnetic resonance hydrogen spectrum described in S11.

[0014] S13: Based on S12, generate the peak positions and intensities in the pure chemical shift spectrum, and generate two-channel tag data. Each peak occupies three data points. The data size of the first channel is 0 or 1, indicating whether the point is a peak. The data of the second channel is the peak intensity.

[0015] S14: Logarithmically scale the peak intensity of the tag spectrum to keep them on the same order of magnitude;

[0016] S15: Repeat S11-S14 to generate a set amount of training and test sets.

[0017] In a preferred embodiment, step S2 specifically includes:

[0018] S21: Design and build the encoder module: Each layer of the encoder module consists of a multi-scale feature extraction module and three CBR modules; the CBR module consists of convolutional layers with a kernel size of 3 and a stride of 1; each convolutional layer is followed by a batch normalization layer and a non-linear ReLU layer; each layer is followed by a max pooling layer with a size of 2 and a stride of 2 for downsampling.

[0019] S22: Design and build the decoder module: Each layer of the decoder module consists of an upsampling convolution module, an attention gate module, a multi-scale feature extraction module, and three CBR modules;

[0020] The upsampling convolution module consists of a deconvolution layer with 4 kernels and a stride of 2 and a CBR module. The output of the decoder module of the previous layer is expanded in dimension through the upsampling convolution module and input into the attention gate module together with the output of the encoder of the same layer. The output of the attention gate module and the output of the upsampling convolution module are concatenated in the channel dimension. Then, there is a multi-scale feature extraction module and three CBR modules.

[0021] S23: Design and build a multi-scale feature extraction module: The multi-scale feature extraction module consists of four convolutional channels, namely: (1) a convolutional layer with a kernel of 1 and a stride of 1; (2) a convolutional layer with a kernel of 1 and a stride of 1 superimposed with a convolutional layer with a kernel of 8 and a stride of 1; (3) a convolutional layer with a kernel of 1 and a stride of 1 superimposed with a convolutional layer with a kernel of 32 and a stride of 1; (4) a convolutional layer with a kernel of 1 and a stride of 1 superimposed with a convolutional layer with a kernel of 64 and a stride of 1; each convolutional layer is followed by a batch normalization layer and a first nonlinear layer ReLU;

[0022] S24: The encoder and decoder modules respectively perform feature extraction and feature recovery;

[0023] S25: Design and build the attention gate module: The attention gate module includes two independent convolutional layers with a kernel size of 1 and a stride of 1. The outputs of the two layers are added together and then passed through a second nonlinear layer ReLU, a convolutional layer with a kernel size of 1 and a stride of 1, and a first nonlinear layer Sigmoid. The encoder output and the output of the upsampled convolutional module in the decoder are passed through the attention gate module, and the output of the first nonlinear layer Sigmoid is multiplied by the encoder output to obtain the output of the attention gate module.

[0024] S26: Design and build the output module: The output module consists of two convolutional layers with 3 kernels and a stride of 1; the first convolutional layer is followed by the third nonlinear layer ReLU and the batch normalization layer; the output of the second convolutional layer has the first channel followed by the second nonlinear layer Sigmoid, and the second channel followed by the fourth nonlinear layer ReLU.

[0025] In a preferred embodiment, step S3 specifically includes:

[0026] S31: Design the loss function: The loss function consists of two parts. The first part is the loss function for spectral peak position information, which is composed of the binary cross-entropy loss function. The second part is the loss function for spectral peak intensity information, which is composed of the mean square error loss function. The total loss function is composed of the weighted sum of the two.

[0027] S32: Single network training: Input the one-dimensional hydrogen spectrum from the training set into the model, use the model's output and labels to calculate the loss function described in S31, and use the Adam optimizer to update the parameters in the model based on this loss function; at the same time, use the test set to calculate the loss function after each training session and select the best model;

[0028] S33: Repeat the training steps described in S32 multiple times until the number of training iterations reaches the preset maximum, or the loss function decreases to convergence, and then stop training to obtain a trained deep learning model.

[0029] In a preferred embodiment, step S4 specifically includes:

[0030] S41: Design evaluation metrics: The accuracy of peak position information is evaluated by calculating true positives (TP), false positives (FP), true negatives (TN), false negatives (FN), precision, recall, and F1 score; the accuracy of peak intensity information is evaluated by calculating the normalized mean square error of the spectrum; the meanings of the evaluation metrics are as follows:

[0031] TP: Determine that this point is a spectral peak, and the position corresponding to the label is also a spectral peak;

[0032] FP: This point is identified as a spectral peak, but the corresponding label position is the baseline;

[0033] TN: This point is determined to be the baseline, and the location corresponding to the label is also the baseline;

[0034] FN: This point is determined to be the baseline, but the corresponding position of the label is a spectral peak;

[0035] Accuracy:

[0036]

[0037] Recall rate:

[0038]

[0039] F1 score:

[0040]

[0041] Normalized mean square error:

[0042]

[0043] In equation (4), x is the one-dimensional data obtained from the model output, y is the corresponding reference pure chemical shift spectrum data, and N is the length of the spectrum data;

[0044] The accuracy of the spectral peak position information can be quantitatively judged by equations (1), (2), and (3). The closer the precision and recall are to 100%, and the closer the F1 score is to 1, the more accurate the spectral peak position is. Equation (4) can quantitatively judge the regression error and accuracy of the spectral peak intensity information. The closer the normalized mean square error is to 0, the closer the model output spectrum is to the reference spectrum.

[0045] S42: Input the one-dimensional proton spectrum samples obtained from the test set or experiments into the model. The model outputs data containing peak position and intensity information. Based on the obtained data and the peak model, construct peaks at specific positions in the spectrum to obtain a pure chemical shift spectrum. Quantitatively evaluate the obtained spectrum and data.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] The method proposed in this invention can process one-dimensional proton NMR spectra based on deep learning models and the characteristics of NMR coupling split peaks, thereby obtaining pure chemical shift spectra, which greatly improves the interpretability of one-dimensional proton NMR spectra. Furthermore, compared with the pure chemical shift spectra obtained by traditional pulse sequence decoupling, the spectra obtained by this method show improvements in signal-to-noise ratio, distortion reduction, and linewidth. In addition, this method has broad applicability to different spectral widths and coupling systems, and also exhibits good robustness. Attached Figure Description

[0048] Figure 1 This is the structure for deep learning models.

[0049] Figure 2 This is a comparison image of the simulated spectrum before and after decoupling.

[0050] Figure 3 This is a comparison image before and after decoupling the experimental spectrum. Detailed Implementation

[0051] The specific steps in the implementation of this invention are as follows:

[0052] Step 1: Generate the training and test sets required for the deep learning model.

[0053] Based on the characteristic that the free-induction decay signal is composed of the superposition of complex exponential signals, a mathematical model is used to generate a simulated one-dimensional hydrogen NMR spectrum with Voigt peak shape and noise and distortion. The parameters of the simulated one-dimensional hydrogen NMR spectrum include: the number of coupling systems is [2,12], the number of spins of the coupling system is [1,5], the spectral width is [2048,4096] Hz, the linewidth is [5,12] Hz, the peak position, the coupling constant is [5,20] Hz, the signal-to-noise ratio (SNR) is [20.97,35.74] dB, and the degree of distortion includes phase distortion and baseline distortion. The formula for calculating the SNR is as follows: All the above parameters were randomly generated; the final generated data was 2×8192, which were the real and imaginary parts of the one-dimensional proton spectrum.

[0054] Generate pure chemical shift spectra with no coupling splits, no noise, and no distortion corresponding to one-dimensional proton NMR spectra; generate two-channel tag data based on the peak positions and intensities in the pure chemical shift spectra, with each peak occupying three data points; the data size of the first channel is 0 or 1, representing whether the point is a peak; the data of the second channel is the peak intensity; logarithmically scale the peak intensities in the tags to keep them on the same order of magnitude.

[0055] Repeat the above steps to generate a training set of 140,000 and a test set of 20,000.

[0056] Step 2: Design the deep learning model architecture and build the model.

[0057] Deep learning models such as Figure 1 As shown, the model consists of four parts: encoder, decoder, attention gate module, and output module.

[0058] Each layer of the encoder module consists of a multi-scale feature extraction module and three CBR (Conv, BatchNorm, ReLU) modules; the CBR module consists of convolutional layers with a kernel size of 3 and a stride of 1; each convolutional layer is followed by a batch normalization layer and a non-linear ReLU layer; each layer is followed by a max pooling layer with a size of 2 and a stride of 2 for downsampling.

[0059] Each layer of the decoder module consists of an upsampling convolutional module, an attention gate module, a multi-scale feature extraction module, and three CBR modules. The upsampling convolutional module is composed of a deconvolutional layer with a kernel of 4 and a stride of 2 and a CBR module. The output of the decoder module of the previous layer is expanded in dimension through the upsampling convolutional module and input into the attention gate module together with the output of the encoder of the same layer. The output of the attention gate module and the output of the upsampling convolutional module are concatenated in the channel dimension. Then, the multi-scale feature extraction module and the three CBR modules are followed.

[0060] The multi-scale feature extraction module consists of four convolutional channels: (1) a convolutional layer with a kernel of 1 and a stride of 1; (2) a convolutional layer with a kernel of 1 and a stride of 1 superimposed with a convolutional layer with a kernel of 8 and a stride of 1; (3) a convolutional layer with a kernel of 1 and a stride of 1 superimposed with a convolutional layer with a kernel of 32 and a stride of 1; (4) a convolutional layer with a kernel of 1 and a stride of 1 superimposed with a convolutional layer with a kernel of 64 and a stride of 1. Each convolutional layer is followed by a batch normalization layer and a nonlinear ReLU layer.

[0061] The encoder and decoder modules perform feature extraction and feature recovery respectively. The skip connection between the two modules ensures that the details lost due to downsampling are restored, and these losses are also weighted by the attention gate module, making the restored features more representative.

[0062] The attention gate module consists of: two independent convolutional layers with a kernel size of 1 and a stride of 1. The outputs of the two layers are added together and then passed through a nonlinear ReLU layer, followed by a convolutional layer with a kernel size of 1 and a stride of 1, and a nonlinear Sigmoid layer. The encoder output and the output of the upsampled convolutional module in the decoder are passed through the attention gate module, and the output of the Sigmoid layer is multiplied by the encoder output to obtain the output of the attention gate module.

[0063] The output module consists of two convolutional layers with 3 kernels and a stride of 1. The first convolutional layer is followed by a non-linear ReLU layer and a batch normalization layer. The output of the second convolutional layer is followed by a non-linear Sigmoid layer for the first channel and a non-linear ReLU layer for the second channel.

[0064] Step 3: Train the deep learning model using the training set and test set to obtain a converged and trained deep learning model.

[0065] The loss function consists of two parts: the first part is the loss function for spectral peak position information, which is composed of the binary cross-entropy loss function; the second part is the loss function for spectral peak intensity information, which is composed of the mean square error loss function; the total loss function is composed of the weighted sum of the two.

[0066] A single network training iteration involves inputting a one-dimensional hydrogen spectrum from the training set into the model, calculating the aforementioned loss function using the model's output and labels, and updating the model's parameters using the Adam optimizer based on this loss function. Simultaneously, after each training iteration, the loss function is calculated using the test set to select the best model.

[0067] Repeat the training steps multiple times until the preset maximum number of training iterations is reached, or the loss function decreases to convergence, and then stop training to obtain a trained deep learning model.

[0068] Step 4: Test the trained deep learning model using the test set and the one-dimensional hydrogen spectrum obtained from the experiment.

[0069] The evaluation metrics are designed as follows: The accuracy of peak position information is evaluated by calculating true positives (TP), false positives (FP), true negatives (TN), false negatives (FN), precision, recall, and F1 score; the accuracy of peak intensity information is evaluated by calculating the normalized mean square error of the spectrum. The meanings of these evaluation metrics are as follows:

[0070] TP: Determine that this point is a spectral peak, and the position corresponding to the label is also a spectral peak;

[0071] FP: This point is identified as a spectral peak, but the corresponding label position is the baseline;

[0072] TN: This point is determined to be the baseline, and the location corresponding to the label is also the baseline;

[0073] FN: This point is determined to be the baseline, but the corresponding position of the label is a spectral peak;

[0074] Accuracy:

[0075]

[0076] Recall rate:

[0077]

[0078] F1 score:

[0079]

[0080] Normalized mean square error:

[0081]

[0082] In equation (4), x is the one-dimensional data obtained from the model output, y is the corresponding reference pure chemical shift spectrum data, and N is the length of the spectrum data;

[0083] The accuracy of the spectral peak position information can be quantitatively judged by equations (1), (2), and (3). The closer the precision and recall are to 100%, and the closer the F1 score is to 1, the more accurate the spectral peak position is. Equation (4) can quantitatively judge the regression error and accuracy of the spectral peak intensity information. The closer the normalized mean square error is to 0, the closer the model output spectrum is to the reference spectrum.

[0084] One-dimensional proton spectrum samples obtained from the test set or experiments are input into the model. The model outputs data containing peak position and intensity information. Based on the obtained data and the peak model, peaks are constructed at specific positions in the spectrum to obtain a pure chemical shift spectrum. The obtained spectrum and data are then quantitatively evaluated.

[0085] The accompanying drawings are described in detail below:

[0086] Figure 1 Deep learning model structure;

[0087] Figure 2The output spectrum obtained by inputting the simulated one-dimensional proton NMR spectrum into the model is compared with the pure chemical shift spectrum as a label; even if some spectral peaks are very close to each other, the model can still restore the spectral peaks at all positions in the spectrum;

[0088] Figure 3 The output spectrum obtained by inputting the one-dimensional proton NMR spectrum of the estradiol sample into the model is compared with the pure chemical shift spectrum obtained by decoupling with the PSYCHE pulse sequence. Compared with the pure chemical shift spectrum obtained experimentally, the model can restore the peaks at all positions in the spectrum, and reduces the linewidth of the peaks, thus improving the resolution of the peaks.

[0089] The above embodiments are only used to further illustrate the one-dimensional nuclear magnetic resonance hydrogen spectrum virtual decoupling method of the present invention. However, the present invention is not limited to the embodiments. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the technical solution of the present invention.

Claims

1. A deep learning-based virtual decoupling method for one-dimensional proton NMR spectra, characterized in that, Includes the following steps: S1: The training and test sets required to generate a deep learning model; S2: Design the deep learning model architecture and build the model; S3: The deep learning model is trained using the training set and the test set to obtain a converged and trained deep learning model. S4: The trained deep learning model is tested using the test set and the one-dimensional hydrogen spectrum obtained from the experiment. Step S1 specifically includes: S11: Based on the characteristic that the free induction decay signal is composed of the superposition of complex exponential signals, a mathematical model is used to generate a simulated one-dimensional nuclear magnetic resonance hydrogen spectrum with Voigt peak shape and noise and distortion. The parameters of the simulated one-dimensional nuclear magnetic resonance hydrogen spectrum include: number of coupling systems, number of spins in the coupling system, spectral width, linewidth, peak position, coupling constant, noise intensity, and distortion degree; S12: Generate a pure chemical shift spectrum with no coupling split peaks, no noise and no distortion, corresponding to the one-dimensional nuclear magnetic resonance hydrogen spectrum described in S11. S13: Based on S12, generate the peak positions and intensities in the pure chemical shift spectrum, and generate two-channel tag data. Each peak occupies three data points. The data size of the first channel is 0 or 1, indicating whether the point is a peak. The data of the second channel is the peak intensity. S14: Logarithmically scale the peak intensity of the tag spectrum to keep them on the same order of magnitude; S15: Repeat S11-S14 to generate a set amount of training and test sets; Step S2 specifically includes: S21: Design and build the encoder module: Each layer of the encoder module consists of a multi-scale feature extraction module and three CBR modules; the CBR module consists of convolutional layers with a kernel size of 3 and a stride of 1; each convolutional layer is followed by a batch normalization layer and a non-linear ReLU layer; each layer is followed by a max pooling layer with a size of 2 and a stride of 2 for downsampling. S22: Design and build the decoder module: Each layer of the decoder module consists of an upsampling convolution module, an attention gate module, a multi-scale feature extraction module, and three CBR modules; The upsampling convolution module consists of a deconvolution layer with 4 kernels and a stride of 2 and a CBR module. The output of the decoder module of the previous layer is expanded in dimension through the upsampling convolution module and input into the attention gate module together with the output of the encoder of the same layer. The output of the attention gate module and the output of the upsampling convolution module are concatenated in the channel dimension. Then, there is a multi-scale feature extraction module and three CBR modules. S23: Design and build a multi-scale feature extraction module: The multi-scale feature extraction module consists of four convolutional channels, namely: (1) a convolutional layer with a kernel of 1 and a stride of 1; (2) a convolutional layer with a kernel of 1 and a stride of 1 superimposed with a convolutional layer with a kernel of 8 and a stride of 1; (3) a convolutional layer with a kernel of 1 and a stride of 1 superimposed with a convolutional layer with a kernel of 32 and a stride of 1; (4) a convolutional layer with a kernel of 1 and a stride of 1 superimposed with a convolutional layer with a kernel of 64 and a stride of 1; each convolutional layer is followed by a batch normalization layer and a first nonlinear layer ReLU; S24: The encoder and decoder modules respectively perform feature extraction and feature recovery; S25: Design and build the attention gate module: The attention gate module includes two independent convolutional layers with a kernel size of 1 and a stride of 1. The outputs of the two layers are added together and then passed through a second nonlinear layer ReLU, a convolutional layer with a kernel size of 1 and a stride of 1, and a first nonlinear layer Sigmoid. The encoder output and the output of the upsampled convolutional module in the decoder of the same layer pass through the attention gate module, and the output of the first nonlinear layer Sigmoid is multiplied by the encoder output to obtain the output of the attention gate module. S26: Design and build the output module: The output module consists of two convolutional layers with 3 kernels and a stride of 1; the first convolutional layer is followed by a third nonlinear layer ReLU and a batch normalization layer; the output of the second convolutional layer has a second nonlinear layer Sigmoid after the first channel and a fourth nonlinear layer ReLU after the second channel.

2. The method for virtual decoupling of one-dimensional proton NMR spectra based on deep learning according to claim 1, characterized in that, Step S3 specifically includes: S31: Design the loss function: The loss function consists of two parts. The first part is the loss function for spectral peak position information, which is composed of the binary cross-entropy loss function. The second part is the loss function for spectral peak intensity information, which is composed of the mean square error loss function. The total loss function is composed of the weighted sum of the two. S32: Single network training: Input the one-dimensional hydrogen spectrum from the training set into the model, use the model's output and labels to calculate the loss function described in S31, and use the Adam optimizer to update the parameters in the model based on this loss function; at the same time, use the test set to calculate the loss function after each training session and select the best model; S33: Repeat the training steps described in S32 multiple times until the number of training iterations reaches the preset maximum, or the loss function decreases to convergence, and then stop training to obtain a trained deep learning model.

3. The method for virtual decoupling of one-dimensional proton NMR spectra based on deep learning according to claim 1, characterized in that, Step S4 specifically includes: S41: Design evaluation metrics: The accuracy of peak position information is evaluated by calculating true positives (TP), false positives (FP), true negatives (TN), false negatives (FN), precision, recall, and F1 score; the accuracy of peak intensity information is evaluated by calculating the normalized mean square error of the spectrum; the meanings of the evaluation metrics are as follows: TP: Determine that this point is a spectral peak, and the position corresponding to the label is also a spectral peak; FP: This point is identified as a spectral peak, but the corresponding label position is the baseline; TN: This point is determined to be the baseline, and the location corresponding to the label is also the baseline; FN: This point is determined to be the baseline, but the corresponding position of the label is a spectral peak; Accuracy: Recall rate: F1 score: Normalized mean square error: In equation (4), x is the one-dimensional data obtained from the model output, y is the corresponding reference pure chemical shift spectrum data, and N is the length of the spectrum data; The accuracy of the spectral peak position information can be quantitatively judged by equations (1), (2), and (3). The closer the precision and recall are to 100%, and the closer the F1 score is to 1, the more accurate the spectral peak position is. Equation (4) can quantitatively judge the regression error and accuracy of the spectral peak intensity information. The closer the normalized mean square error is to 0, the closer the model output spectrum is to the reference spectrum. S42: Input the one-dimensional proton spectrum sample obtained from the test set or the experiment into the model. The model outputs data containing peak position and intensity information. Based on the obtained data and the peak model, construct the peak at a specific position in the spectrum to obtain a pure chemical shift spectrum. Quantitatively evaluate the obtained spectrum and data.

Images