Nuclear magnetic resonance spectrum metabonomics analysis method based on convolutional neural network

Abstract

The invention discloses a nuclear magnetic resonance spectrum metabonomics analysis method based on a convolutional neural network, and relates to the technical field of metabonomics, and the technical scheme is characterized in that nuclear magnetic resonance hydrogen spectrum metabolic data of a biological sample is acquired and preprocessed; performing feature construction on the preprocessed data; a deep learning model based on a convolutional neural network is constructed and trained; and explaining the trained deep learning model to obtain differential characteristics, and obtaining differential metabolite information having significant influence on diseases in the biological sample. According to the method, original metabonomics data can be directly processed through the convolutional neural network model without depending on compound identification based on prior experience, so that information omission and identification deviation in the identification process are avoided; meanwhile, the problem of signal alignment deviation caused by heterogeneity of experimental samples is avoided, and high-accuracy sample classification and high-sensitivity differential metabolite extraction can be achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of metabolomics, and more specifically, to a nuclear magnetic resonance spectroscopy metabolomics analysis method based on convolutional neural network. Background Art

[0002] Nuclear magnetic resonance (NMR) spectroscopy is an important analytical method in the field of metabolomics, with high repeatability and excellent quantitative ability. Extracting metabolomics information from NMR data and screening out important biomarkers are crucial steps.

[0003] In existing studies, NMR metabolomics data are mainly analyzed through global chemometric analysis and quantitative analysis. Global chemometric analysis is a multivariate analysis after preprocessing the data. The problem is that the spectrum cannot achieve accurate alignment of all signals, and the spectral baseline and noise will introduce a lot of uncertainty to the analysis, which limits the sensitivity of this method to extract differences. Especially for biological samples such as urine, under sample conditions such as pH, ionic strength, temperature and macromolecule concentration, the chemical shift of molecules is highly sensitive to the environment, making peak alignment more difficult.

[0004] Quantitative analysis first identifies and quantifies metabolites, extracts some known metabolite information from the spectrum based on prior experience through metabolomics platform analysis, and then conducts further analysis; its identification and quantification process relies on a lot of prior experience, and the results include a lot of human factors, and the sensitivity and resolution are limited, so that the current analysis scheme can only realize the utilization of a small part of the information, and at the same time causes some key differential metabolites to be ignored.

[0005] Therefore, the present invention aims to provide a convolutional neural network-based nuclear magnetic resonance spectroscopy metabolomics analysis method to solve the above problems. Summary of the invention

[0006] The purpose of the present invention is to provide a convolutional neural network-based nuclear magnetic resonance spectroscopy metabolomics analysis method. The present invention directly processes the metabolomics raw data through a convolutional neural network model, does not rely on prior experience for compound identification, and avoids information omission and identification bias in the identification process; at the same time, a deep learning method based on a convolutional neural network is used to avoid the signal alignment bias problem caused by the heterogeneity of experimental samples, and can achieve high-accuracy sample classification and high-sensitivity extraction of differential metabolites.

[0007] The above technical objectives of the present invention are achieved through the following technical solutions: A nuclear magnetic resonance spectroscopy metabolomics analysis method based on convolutional neural network, comprising the following steps:

[0008] S1. Obtain the nuclear magnetic resonance hydrogen spectrum metabolic data of biological samples and perform preprocessing;

[0009] S2, constructing features of the preprocessed nuclear magnetic resonance hydrogen spectrum metabolic data to obtain a nuclear magnetic resonance hydrogen spectrum metabolic data set;

[0010] S3. Build a deep learning model based on convolutional neural network and train the deep learning model using the nuclear magnetic resonance hydrogen spectrum metabolism dataset;

[0011] S4. Perform interpretation and analysis on the trained deep learning model to generate differential features of the deep learning model, and then sort the differential features to obtain differential metabolite information in biological samples that has a significant impact on the disease.

[0012] The present invention is further configured as follows: obtaining the nuclear magnetic resonance hydrogen spectrum metabolic data of the biological sample in step S1 specifically obtains the 1H NMR original spectrum of the biological sample, and the biological sample includes a disease group biological sample and a control group biological sample.

[0013] The present invention is further configured as follows: the process of preprocessing in step S1 is: Fourier transforming the original spectrum of the biological sample, then correcting the chemical shift, baseline and phase, and finally removing the solvent signal to complete the preprocessing.

[0014] The present invention is further configured as follows: the specific process of feature construction in step S2 is: converting the preprocessed original spectrum of the biological sample into chemical shift-intensity information, then determining the chemical shift distribution interval, and extracting the intensity to obtain a feature matrix of the spectrum.

[0015] The present invention is further configured as follows: the deep learning model based on the convolutional neural network constructed in step S3 is used for biological sample classification; when the deep learning model is trained using the nuclear magnetic resonance hydrogen spectrum metabolic data set, the hyperparameters are optimized according to the learning effect to obtain the best classification effect.

[0016] The present invention is further configured as follows: the process of interpreting and analyzing the trained deep learning model in step S4 is: calculating the SHAP value corresponding to each chemical shift characteristic value through the trained deep learning model to evaluate the contribution of the chemical shift to the classification model, and then calculating the average SHAP value of all samples, and sorting to obtain differential metabolite information that has a significant impact on the disease.

[0017] In summary, the present invention has the following beneficial effects:

[0018] 1. The present invention can directly process the raw data of metabolomics through the convolutional neural network model, and does not rely on prior experience for compound identification, thus avoiding information omission and identification bias in the identification process; at the same time, the deep learning method based on the convolutional neural network avoids the signal alignment bias problem caused by the heterogeneity of experimental samples, and can achieve high-accuracy sample classification and high-sensitivity extraction of differential metabolites;

[0019] 2. The present invention applied the method to whole blood samples and platelet samples in a group of sepsis and non-sepsis samples, respectively. The results showed that the method was able to achieve high-accuracy data classification (sensitivity (90%, 95%), specificity (85%, 80%), and area under the curve (0.83, 0.96)); in terms of the extraction of differential metabolite information in the application examples, compared with the currently used quantitative analysis combined with non-neural network machine learning model methods, 3 and 7 additional differential metabolites were extracted, respectively, avoiding the situation where key differential metabolites are neglected. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of a process of a convolutional neural network-based nuclear magnetic resonance spectroscopy metabolomics analysis method in Example 1 of the present invention;

[0021] Figure 2 is a schematic diagram of a process of using a deep learning model for biological sample classification in Example 1 of the present invention;

[0022] Figure 3 is a schematic diagram of a process for preprocessing original NMR data in Example 2 of the present invention;

[0023] Figure 4 is a detailed neural network structure diagram of NmCNN in Example 2 of the present invention;

[0024] Figure 5 It is a schematic diagram of the confusion matrix in Example 2 of the present invention. DETAILED DESCRIPTION

[0025] The following is combined with Figure 1-5 The present invention is described in further detail.

[0026] Example 1: A convolutional neural network-based nuclear magnetic resonance spectroscopy metabolomics analysis method comprising the following steps:

[0027] S1. Obtaining the original 1H NMR spectra of the biological samples of the disease group and the control group, then performing Fourier transform on the original spectra of the biological samples, correcting the chemical shift, baseline and phase, and finally removing the solvent signal to complete the preprocessing;

[0028] S2, converting the pre-processed original spectrum of the biological sample into chemical shift-intensity information, then determining the chemical shift distribution interval, extracting the intensity to obtain the characteristic matrix of the spectrum;

[0029] S3. Construct a deep learning model based on convolutional neural network for biological sample classification. When training the deep learning model using the nuclear magnetic resonance hydrogen spectrum metabolism dataset, adjust the hyperparameters according to the learning effect to obtain the best classification effect.

[0030] S4. The process of interpreting and analyzing the trained deep learning model is as follows: the SHAP value corresponding to each chemical shift feature value is calculated through the trained deep learning model to evaluate the contribution of the chemical shift to the classification model, and then the average SHAP value of all samples is calculated, and the differential metabolite information that has a significant impact on the disease is obtained through sorting.

[0031] Example 2: Application experiment of a nuclear magnetic resonance spectroscopy metabolomics analysis method based on convolutional neural network

[0032] The raw nuclear magnetic resonance (NMR) data were obtained from the NMDR. The NMR data contained the characteristics of whole blood samples and platelet samples extracted from septic and non-septic patients. The data preprocessing process was as follows: Figure 3 As shown, the 1H-NMR spectrum was processed using MestReNova, and a 0.3 Hz exponential window function was applied to eliminate truncation artifacts from the fid data; followed by Fourier transform, phase correction, and baseline correction; chemical shift data correction was performed by correcting the chemical shift of the internal reference DSS to 0 ppm, and the chemical shift region less than 0.75 ppm containing the DSS signal was excluded from the statistical analysis; in addition, in order to eliminate the random effects caused by water peak suppression variation, the chemical shift region between 4.670 and 5.000 ppm was also excluded; finally, the 1H NMR spectrum was processed to obtain 8250 intervals of equal length (δ0.001 ppm), and the normalized relative intensity was used as input. This precise classification enables the subsequent convolution process to effectively capture more complex information.

[0033] In this embodiment, the NmCNN model is implemented using Python and PyTorch framework, which is widely recognized for building convolutional neural networks; the detailed neural network structure of NmCNN is as follows Figure 4As shown in the figure, the network consists of three convolutional layers, each followed by a max pooling layer. The convolutional layers consist of 32 kernels of size 5×1. These layers use ReLU as the activation function, which can effectively learn translation-invariant features from NMR spectra. The convolutional layer is used to detect the local connection of the features of the previous layer, and the following pooling layer is used to merge semantically similar features. Subsequently, the output results of the convolutional and pooling layers are fed into an unfolding layer and converted into a one-dimensional tensor for data preparation for four fully connected layers. The fully connected layers have 32864, 8192, 1024, and 128 hidden units, respectively, and use Sigmoid as their activation function. For classification tasks, the SoftMax activation function is applied at the output layer to generate the category probability distribution; for binary classification problems, binary cross entropy is selected as the loss function, and the Adam algorithm is selected as the optimizer because of its high computational efficiency and small memory requirements.

[0034] like Figure 5 As shown in the figure, the confusion matrix classifies the classification problem according to the two dimensions of actual situation and discrimination. For binary classification problems, it is calculated as a 2×2 matrix; the confusion matrix provides the basis for defining four performance indicators, namely true positives (TP), false positives (FP), false negatives (FN), and true negatives (TN). During the model training process, the accuracy and average accuracy of the model are calculated; the TPR (true positive rate) and FPR (false positive rate) indicators represent the proportion of correctly identified positive samples to all positive samples, and the proportion of negative samples misclassified as positive samples by the model to all negative samples; TPR is also called recall or sensitivity, which is used to quantify the ability of the classification model to predict positive results in actual positive samples. After training, the disease index (DI) was calculated to evaluate the potential probability of sepsis, which is defined as the inner product between the output tensor and the contribution tensor.

[0035] The deep learning model was used to interpret differential metabolites using SHapleyAdditive exPlanations (SHAP) and the SHAP Python package, which provides a comprehensive method to accurately quantify the contribution and impact of each feature on the final prediction result. The SHAP value is calculated for each prediction result, and the difference between the model prediction result and the actual output is analyzed. In a classifier with multiple outputs, an explanation is provided for each individual output; a high positive or negative SHAP value indicates that a feature has a higher importance, a positive value indicates an increase in model output, and a negative value indicates a decrease. Values ​​close to zero indicate low feature importance. The SHAP value reveals the positive or negative impact of each feature on the final prediction result.

[0036] SHAP-NMR spectra were generated using SHAP values, which can be visualized as pseudo-NMR spectra of potential characteristic metabolites. Feature names were assigned as chemical shifts, and the intensity indicated the influence factor of each chemical shift, with high intensity indicating greater importance in distinguishing between sepsis and non-septic samples.

[0037] In this embodiment, an interpretable NMR metabolomics convolutional neural network (NmCNN) model was developed based on the entire spectral analysis data for the evaluation of NMR data for sepsis screening. The method has a significant classification effect, which is manifested by high sensitivity (90%, 95%), specificity (85%, 80%) and area under the curve (0.83, 0.96), highlighting its wide applicability in processing NMR spectra and emphasizing the universality of this method; in addition, the application of SHAP analysis on the deep learning model provides a reasonable explanation for the chemical shift attributes, thereby verifying the interpretability of the NmCNN model. Establishing an interpretable neural network model provides a feasible method for extracting metabolome information from the overall metabolome data. The SHAP-NMR spectrum extracted from the network contains the contribution characteristics of the signal. The intensity in the spectrum reflects the influence of the chemical shift of different metabolites on the model output results. The set of characteristic markers of septic blood and platelet samples was successfully identified. These differential metabolites were not identified in the previous traditional quantitative analysis method, proving that this method has high sensitivity in feature extraction.

[0038] NMR is a non-targeted metabolomics method that can efficiently obtain comprehensive component information from biological samples, focusing on the overall metabolic state; NMR methods are highly reproducible and have a wide range of applications, with simple preprocessing procedures. The main challenge facing NMR metabolomics is the annotation of spectral signals, and the model of this method can directly extract information from NMR full metabolic spectra, thereby enhancing the ability to use NMR for metabolomics analysis; in addition, nuclear magnetic resonance spectroscopy also promotes the specific extraction of partial information in complex biological samples through pulse modulation technology, such as obtaining more accurate small molecule metabolite information based on NOESY correlation spectra between hydrogen atoms. Deep learning methods based on deep convolutional neural networks are expected to continue to play an increasingly important role in extracting NOESY-related spectral information in the field of big data mining.

[0039] This specific embodiment is merely an explanation of the present invention and is not a limitation of the present invention. After reading this specification, those skilled in the art may make non-creative modifications to the present embodiment as needed. However, as long as they are within the scope of the claims of the present invention, they are protected by the patent law.

Claims

1. A nuclear magnetic resonance spectroscopy metabolomics analysis method based on convolutional neural network, characterized by: The following steps are involved: S1. Obtain the nuclear magnetic resonance hydrogen spectrum metabolic data of biological samples and perform preprocessing; S2, constructing features of the preprocessed nuclear magnetic resonance hydrogen spectrum metabolic data to obtain a nuclear magnetic resonance hydrogen spectrum metabolic data set; S3. Build a deep learning model based on convolutional neural network and train the deep learning model using the nuclear magnetic resonance hydrogen spectrum metabolism dataset; S4. Perform interpretation and analysis on the trained deep learning model, generate differential features of the deep learning model, sort the differential features, and obtain differential metabolite information in biological samples that has a significant impact on the disease.

2. The method for metabolomics analysis based on convolutional neural network using nuclear magnetic resonance spectroscopy according to claim 1, characterized in that: Acquiring the nuclear magnetic resonance hydrogen spectrum metabolic data of the biological sample in step S1 specifically involves acquiring the 1HNMR original spectrum of the biological sample, wherein the biological sample includes a disease group biological sample and a control group biological sample.

3. The method for metabolomics analysis based on convolutional neural network using nuclear magnetic resonance spectroscopy according to claim 2, characterized in that: The process of preprocessing in step S1 is: Fourier transform the original spectrum of the biological sample, then correct the chemical shift, baseline and phase, and finally remove the solvent signal to complete the preprocessing.

4. The method for metabolomics analysis based on convolutional neural network using nuclear magnetic resonance spectroscopy according to claim 3, characterized in that: The specific process of feature construction in step S2 is: converting the pre-processed original spectrum of the biological sample into chemical shift-intensity information, then determining the chemical shift distribution interval, and obtaining the characteristic matrix of the spectrum by extracting the intensity.

5. The method for metabolomics analysis based on convolutional neural network using nuclear magnetic resonance spectroscopy according to claim 1, characterized in that: The deep learning model based on convolutional neural network constructed in step S3 is used for biological sample classification; when the deep learning model is trained using the nuclear magnetic resonance hydrogen spectrum metabolic data set, the hyperparameters are optimized according to the learning effect to obtain the best classification effect.

6. The method for metabolomics analysis based on convolutional neural network using nuclear magnetic resonance spectroscopy according to claim 1, characterized in that: The process of interpreting and analyzing the trained deep learning model in step S4 is as follows: the SHAP value corresponding to each chemical shift feature value is calculated by the trained deep learning model to evaluate the contribution of the chemical shift to the classification model, and then the average SHAP value of all samples is calculated, and the differential metabolite information that has a significant impact on the disease is obtained by sorting.

Images