Method for obtaining a high-quality, purely chemical NMR shift spectrum

Abstract

Method for obtaining a high-quality, purely chemical NMR shift spectrum, characterized in that it comprises the following steps: 1) Using the Matlab program to simulate a free induction attenuation signal and thus generate a purely chemical shift spectrum dataset for random NMR waves, and performing a Fourier transform of the FID signal to obtain a purely chemical NMR shift spectrum dataset and a corresponding labeling dataset; 2) Simulating the real-time ZS spectra as input data for a neural network by generating different chemical shift values, different signal-to-noise ratios, different spectrum peak intensities, different lateral relaxations, and different coupling strengths, and simultaneously simulating an ideal pure shift spectrum whose chemical shift, spectrum peak intensity, and lateral relaxation correspond to the simulated real-time ZS spectra, without coupling intensity information, without noise, and without data splice pseudopeaks as label data; creating a dataset with the input data and the label dataset, and dividing the dataset into a training set and a test set, which contains test data; 3) Perform preprocessing of data normalization for all datasets, including training set and test set; 4) Designing a structure of a neural network model of high-quality, purely chemical shift spectra that can achieve effects such as noise reduction, pseudopeak removal, linewidth reduction, and other effects; 5) Using the test data to test the neural network model, inputting the test data after preprocessing into the neural network model to obtain high-quality pure chemical shift spectra in order to output a high-quality pure chemical NMR shift spectrum that corresponds to the real-time ZS spectrum.

Description

TECHNICAL AREA

[0001] The present invention relates to nuclear magnetic resonance, in particular a method based on the algorithm of the neural network for obtaining a high-quality purely chemical NMR (nuclear magnetic resonance) shift spectrum. STATE OF THE ART

[0002] Nuclear magnetic resonance (NMR) is widely used in biology, chemistry, physics, and other fields due to its outstanding performance in determining material structure and obtaining spatial information. One-dimensional hydrogen NMR spectroscopy (1D 1H-NMR) is one of the most important research methods in NMR spectroscopy technology. Homonuclear broadband decoupling NMR hydrogen spectroscopy technology only addresses the signal overlap defect caused by the J-coupling of the one-dimensional NMR hydrogen spectrum peaks. Pure-chemical hydrogen shift spectroscopy (pure-shift 1H NMR spectroscopy) eliminates the J-coupling effect and merges the multiple split signals into a single peak, which not only improves the resolution of the spectrum but also solves the problem of 1H NMR spectrum signal overlap.

[0003] Among the experimental methods of purely chemical shift analysis, ZS (Zangger-Sterk) technology attracts considerable attention, but conventional real-time ZS and pseudo-two-dimensional ZS technologies exhibit obvious shortcomings. Relatively speaking, real-time ZS spectrograms produce larger linewidths, while pseudo-two-dimensional ZS experiments take considerably longer than real-time ZS. The purpose of the present invention is to compensate for the shortcomings of both methods. The real-time ZS data are processed using a neural network algorithm to avoid the problem of the long experimentation time of pseudo-two-dimensional ZS and to reduce the experimental time by one to two orders of magnitude. Based on this, the various indicators of the spectrum are optimized and improved.

[0004] Deep learning works very well for reconstructing nuclear magnetic resonance spectroscopy. The network, built using neural network learning, can detect and remove unwanted information such as false peaks and noise in the spectrum. For the problem of spectrogram reconstruction, there is no single network type that can be introduced. From fully connected neural feedforward networks to convolutional neural networks and related variant structures, they are all used, and relatively satisfactory results are achieved.

[0005] Regarding the state of the art, reference is made to CN 110879980 A, from which a method for noise suppression in NMR spectra based on a neural network algorithm is known. CONTENT OF THE PRESENT INVENTION

[0006] The purpose of the present invention is to provide a method for obtaining a high-quality, purely chemical NMR shift spectrum based on the neural network algorithm, provided that a shorter experimental time is required.

[0007] The present invention comprises the following steps: 1) Using the Matlab program to simulate a free induction attenuation signal (FID signal) and thus generate a purely chemical shift spectrum dataset for random NMR waves, and performing a Fourier transform of the FID signal to obtain a purely chemical NMR shift spectrum dataset and a corresponding labeling dataset; 2) Simulating the real-time ZS spectra as input data for a neural network by generating different chemical shift values, different signal-to-noise ratios, different spectrum peak intensities, different lateral relaxations, and different coupling strengths, and simultaneously simulating an ideal pure shift spectrum whose chemical shift, spectrum peak intensity, and lateral relaxation correspond to the simulated real-time ZS spectra, without coupling intensity information, without noise, and without data splice pseudopeaks as label data; creating a dataset with the input data and the label dataset, and dividing the dataset into a training set and a test set, which contains test data; 3) Perform preprocessing of data normalization for all datasets, including training set and test set; 4) Designing a structure of a neural network model of high-quality, purely chemical shift spectra that can achieve effects such as noise reduction, pseudopeak removal, linewidth reduction, and other effects; 5) Using the test data to test the neural network model, inputting the test data after preprocessing into the neural network model to obtain high-quality pure chemical shift spectra in order to output a high-quality pure chemical NMR shift spectrum that corresponds to the real-time ZS spectrum.

[0008] In step 4), the specific procedure for designing the structure of the neural network model is: using the Keras neural network creation tool developed by Google to create the Resnet model structure of the neural network and adding the channel overlay operation during the debugging process.

[0009] In step 5), experimentally acquired data are further used to test the neural network model, the test data being a set of test data randomly generated using the above-mentioned method for generating the input data set of the simulated real-time ZS; the experimentally acquired data require preprocessing by Fourier transform, phase correction, and normalization.

[0010] In the present invention, Matlab program code is used to simulate the free induction attenuation (FID) signal of the real-time ZS spectrum and the ideal pure displacement spectrum. A Fourier transform is performed to obtain the corresponding spectrum, and preprocessing such as normalization is carried out. A simulation dataset is created and divided into a training set and a test set. A neural network structure is designed, and the network parameters are set to train the network model. The simulated real-time ZS data and the experimentally acquired real-time ZS data are used as test data to test the model. The present invention avoids the problem of the long experimental time of pseudo-two-dimensional ZS and reduces the experimental time by one to two orders of magnitude.The trained and designed neural network model is used to reconstruct the real-time ZS spectrum in the experiment, thus obtaining a spectrum with high resolution, small linewidth, high signal-to-noise ratio and no pseudopeaks, thereby achieving the goal of quickly obtaining a high-quality, purely chemical hydrogen shift spectrum. BRIEF DESCRIPTION OF THE DRAWING Fig. Figure 1 shows a schematic diagram of the data splicing of the experimental real-time ZS pulse sequence. Fig. Figure 2 shows a neural network structure Resnet, which is used to obtain a high-quality, purely chemical NMR shift spectrum, and the introduced channel superposition operation. Fig. Figure 3 shows a result of the spectrum obtained by processing the one-dimensional simulation data test set without noise using the method of the present invention. The unit of LB (linewidth) is Hz. Fig. Figure 4 shows a result of the spectrum obtained by processing the one-dimensional simulation data test set with added noise in the time domain using the method of the present invention. The unit of LB (linewidth) is Hz. Fig. Figure 5 shows the correction result and a partially enlarged view of the one-dimensional, purely chemical NMR shift spectrum (4 samples) of the quinine-DMSO-d6 solution obtained by the real-time ZS experiment. The unit of LB (linewidth) is Hz. DETAILED DESCRIPTION

[0011] The present invention will be explained in more detail below in connection with figures and exemplary embodiments.

[0012] The present invention uses a Matlab program to simulate an FID signal and generate an NMR simulation signal spectrum. This is used to train a neural network to obtain the model, and then experiments are performed on an NMR instrument to obtain a real-time ZS spectrum. Finally, the experimental spectrum is fed into the network to obtain a high-quality, purely chemical shift spectrum. The specific real-time method is as follows: 1) Deriving the evolutionary formulas of relaxation, chemical shift and J-coupling of the ideal pure shift spectrum and the simulated real-time ZS spectrum according to the pulse sequence and using the evolutionary formulas to derive the corresponding algorithm formula of the corresponding FID signal in the time domain; 2) FID signal data block splice data of the simulated real-time ZS spectrum are in Fig. 1 is shown, where t0 is the sampling time of the data block, d t where is the time interval during non-sampling and n is the number of repeated selections of flip and sample; the midpoint of each data block t0 is the reunion of the J-coupling, and there will be an evolution of the chemical shift at t0 and dd, and the J-coupling will evolve only within t0 due to the selection of flip; therefore, the expression of the FID signal of each data block is as follows: FID(t0)=[∏Jcos(πJt0)]*[∑fej2πfl0]*e−t0T2 FID(ti)=[∏Jcos(πJt1)]*[∑fej2πfl2]*e−t3T2 where FID (t0) is the signal expression of the FID data block of duration t0 / 2, FID (t i) is the signal expression of the i-th FID data block of duration t0; j is the imaginary symbol, T2 is the transverse relaxation time, and f is the frequency corresponding to the chemical shift of each peak. The summation is performed on the chemical shifts of each storage nucleus, and the quadrature is performed on the coupling constants of each storage nucleus and all nuclei. The meanings of t0, t1, t2, and t3 in formulas (2.1) and (2.2) are more difficult to explain, so the MATLAB expression is quoted here. t0=0: 1: (na / 2−1) t1=(−na / 2): 1:(na / 2−1) t2=na*(1 / 2+i−1): 1:(na*(1 / 2+i)−1) t3=(na*(1 / 2+i−1)+dd*i): 1:(na*(1 / 2+i)+dd*i−1) where na is the corresponding number of sampling points within t0, and dd is the number of sampling interval points within d. t and i refers to the i-th FID data block with duration t0; during calculation, the time is replaced by the number of points; 3) The FID signal derivation of the ideal purely chemical shift spectrum should only consider the evolution of the chemical shift and the relaxation process. The derivation formula is as follows: FID(t)=[∑fej2πfl]*e−tT2 where t is the continuous time of the entire sampling; 4) After generating the FID signal of the simulated real-time ZS spectrum and the ideal pure shift spectrum, random Gaussian white noise is added to the FID signal of the simulated real-time ZS spectrum to simulate the noise generated in the real-time ZS experiment. Then, the Fourier transform is performed on the two simulated FID signals, and the real part is used for normalization. Finally, an ideal pure shift spectrum and a simulated real-time ZS spectrum with noise are obtained. 5) Design a neural network model structure of high-quality, purely chemical shift spectra capable of achieving noise reduction, pseudopeak removal, linewidth reduction, and other effects; the network structure is as shown in Fig. 2 shown; The loss function used in the training process is NMSE (normalized mean squared error); the initial learning rate is set to 0.01; the learning rate strategy is that if the loss value does not change, the learning rate is reduced to 0.1 times the original learning rate; the activation function is the leaky ReLU function. The optimization algorithm adopts the Adam algorithm; 6) Using the simulation test data mentioned above to test and verify the neural network model, and calculating the spectral signal-to-noise ratio (SRV) and spectral linewidth (LB); 7) Using the above-mentioned real-time ZS test data to test and verify the neural network model, and calculating the spectral signal-to-noise ratio (SRV) and spectral linewidth (LB);

[0013] Fig. Figure 1 shows a schematic diagram of the data splicing of the experimental real-time ZS pulse sequence referenced in the design of the simulation dataset of the present invention. In the figure, t0 is the sampling time of the data block, dt is the time interval during non-sampling, and n is the number of repeated selections of flip and sample. The midpoint of each data block t0 is the reunification of the J-coupling, and there will be an evolution of the chemical shift at t0 and dt, with the J-coupling evolving only within t0 due to the selection of flip.

[0014] Fig. Figure 2 shows a neural network structure used by the present invention, wherein the main structure is the neural Resnet network, according to the requirements of the network training process, the channel superposition operation is introduced on this basis to prevent the loss of flat features.

[0015] Fig. Figure 3 shows a result of the spectrum obtained by processing the one-dimensional simulation data test set without noise using the method of the present invention. The abscissa of the simulation data represents the number of points. Figures (a), (b), and (c) represent the simulated real-time ZS spectrum without noise, the result of the mesh correction simulation spectrum, and the simulated ideal pure displacement spectrum, respectively. The total number of points in the simulated spectrum is 2025, and the spectrum width is 2000 Hz.

[0016] Fig. Figure 4 shows a result of the spectrum obtained by processing the one-dimensional simulation data test set with added noise in the time domain using the method of the present invention. The abscissa of the simulation data represents the number of points. Figures (a), (b), and (c) represent the simulated real-time ZS spectrum with added noise, the result of the mesh correction simulation spectrum, and the simulated ideal pure displacement spectrum, respectively. The total number of points of the simulated spectrum is 2025, and the spectrum width is 2000 Hz.

[0017] Fig.Figure 5 shows the correction result and a partially enlarged view of the one-dimensional, purely chemical NMR shift spectrum (4 samples) of the quinine-DMSO-d6 solution obtained through the real-time ZS experiment. Figures (a), (b), (c), and (d) represent the results of the acquired pseudo-two-dimensional ZS spectrum of quinine samples, the acquired conventional one-dimensional spectrum of quinine samples, the acquired real-time ZS spectrum of quinine samples, and the simulation spectrum of the mesh correction, respectively. The experimental real-time ZS spectrum bandwidth (SW) of the quinine samples is 6250 Hz, the number of points after splicing is 2048, the number of complete data blocks is 32, the non-sampling time between adjacent data blocks is 0.042 s, and the number of clusters is 4.

[0018] The invention can be summarized as follows: A method for obtaining a high-quality, purely chemical NMR shift spectrum relates to nuclear magnetic resonance. In the present invention, Matlab program code is used to simulate the free induction attenuation signal of the real-time ZS spectrum and the ideal pure shift spectrum; a Fourier transform is performed to obtain the corresponding spectrum; preprocessing such as normalization is carried out; a simulation dataset is created and divided into a training set and a test set; preprocessing of the dataset is performed by normalizing the data; a neural network structure is designed; the network parameters are set to train the network model; the simulated real-time ZS data and the experimentally acquired real-time ZS data are used as test data to test the model.This avoids the problem of the long experimental time of pseudo-two-dimensional ZS and reduces the experimental time by 1 to 2 orders of magnitude. The trained and designed neural network model is used to reconstruct the real-time ZS spectrum in the experiment, thus obtaining a spectrum with high resolution, small linewidth, high signal-to-noise ratio, and no pseudopeaks. In this way, the goal of quickly obtaining a high-quality, purely chemical hydrogen shift spectrum is achieved.

Claims

[1] Method for obtaining a high-quality, purely chemical NMR shift spectrum, characterized by that it includes the following steps: 1) Using the Matlab program to simulate a free induction attenuation signal and thus generate a purely chemical shift spectrum dataset for random NMR waves, and performing a Fourier transform of the FID signal to obtain a purely chemical NMR shift spectrum dataset and a corresponding labeling dataset; 2) Simulating the real-time ZS spectra as input data for a neural network by generating different chemical shift values, different signal-to-noise ratios, different spectrum peak intensities, different lateral relaxations, and different coupling strengths, and simultaneously simulating an ideal pure shift spectrum whose chemical shift, spectrum peak intensity, and lateral relaxation correspond to the simulated real-time ZS spectra, without coupling intensity information, without noise, and without data splice pseudopeaks as label data; creating a dataset with the input data and the label dataset, and dividing the dataset into a training set and a test set, which contains test data; 3) Perform preprocessing of data normalization for all datasets, including training set and test set; 4) Designing a structure of a neural network model of high-quality, purely chemical shift spectra that can achieve effects such as noise reduction, pseudopeak removal, linewidth reduction, and other effects; 5) Using the test data to test the neural network model, inputting the test data after preprocessing into the neural network model to obtain high-quality pure chemical shift spectra in order to output a high-quality pure chemical NMR shift spectrum that corresponds to the real-time ZS spectrum. [2] Method for obtaining a high-quality, purely chemical NMR shift spectrum according to claim 1, characterized by, that in step 4) the specific procedure for designing the structure of the neural network model is: using the Keras neural network creation tool developed by Google to create the Resnet model structure of the neural network and adding the channel overlay operation during the debugging process. [3] Method for obtaining a high-quality, purely chemical NMR shift spectrum according to claim 1 and / or 2, characterized by , that in step 5) experimentally acquired data are further used to test the neural network model, wherein the test data are a set of test data randomly generated using the above-mentioned procedure for generating the input data set of the simulated real-time ZS; and wherein the experimentally acquired data require preprocessing by Fourier transform, phase correction and normalization.

Images