Disease marker metabolite screening method, device, equipment and storage medium

Abstract

The present application discloses a method, device, equipment and storage medium for screening metabolites of disease markers, the method comprising: obtaining the breath of healthy subjects and disease subjects, and constructing respective compound spectra, standardizing the respective compound spectra, and preliminarily screening candidate differential metabolites based on the false positive discovery rate FDR value and the variable projection importance index VIP value, and further screening the target differential metabolites for judging the disease from each candidate differential metabolite through the breath metabolite judgment model. It can be seen that the standardized treatment of the compound spectra can screen out most of the invalid compounds, greatly narrow the determination range of the disease markers, and re-screening in the candidate differential metabolites can obtain more accurate disease markers and more effectively assist the judgment of disease detection.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and more specifically, to a method, device, equipment and storage medium for screening metabolites of disease markers. Background Art

[0002] Metabolomics is a discipline that studies the composition and changes of metabolites in organisms. By analyzing the metabolites in organisms, we can understand the types, contents and changing patterns of metabolites in organisms, and thus explore issues such as the physiology, pathology and environmental adaptation of organisms.

[0003] Exhaled breath is a type of metabolite, and exhaled breath diagnostics is a method of diagnosing disease by analyzing the chemical composition of exhaled breath. Exhaled breath is not only easy to obtain and non-invasive, but also contains volatile organic compounds (VOCs) in exhaled breath metabolites, which are a new type of biomarker. As a disease marker, VOCs can be used to identify differences in metabolites in different tissues, different states, or different environments. By analyzing metabolites in different samples, disease diagnostic markers related to specific states can be screened out, thus providing a reference for disease diagnosis, drug development, environmental monitoring, and other aspects.

[0004] Exhaled breath metabolites may contain a variety of disease-related metabolites. How to more accurately screen out exhaled breath metabolite markers of diseases to assist in disease detection is an issue that requires attention. Summary of the invention

[0005] In view of the above problems, the present application provides a method, device, equipment and storage medium for screening metabolites of disease markers to more accurately screen out exhaled metabolite markers of diseases.

[0006] In order to achieve the above objectives, the specific plan is proposed as follows:

[0007] A method for screening metabolites of disease markers, comprising:

[0008] obtaining exhaled breath from a healthy subject and exhaled breath from a subject suffering from a target disease;

[0009] Detecting the exhaled breath of the healthy subject and the exhaled breath of the subject suffering from the target disease by using a gas chromatograph coupled to a mass spectrometer to obtain a first exhaled breath compound spectrum of the healthy subject and a second exhaled breath compound spectrum of the subject suffering from the target disease;

[0010] For each breath compound spectrum in the first breath compound spectrum and the second breath compound spectrum, performing normalization processing, peak removal processing, peak filling processing and Z-score processing on the breath compound spectrum to obtain a standardized breath compound spectrum;

[0011] Calculating the false positive discovery rate (FDR) value and the variable projection importance index (VIP) value of each compound peak in the standardized breath compound spectrum, and determining the compounds corresponding to the compound peaks in the standardized breath compound spectrum whose FDR values ​​are less than a preset FDR threshold and whose VIP values ​​are greater than a preset VIP threshold as candidate differential metabolites;

[0012] The information of each candidate differential metabolite, the disease information of the target disease and the clinical research information of the target disease are input into a pre-trained breath metabolite judgment model, and the target differential metabolite of the target disease is output as a disease marker of the target disease.

[0013] Optionally, the training process of the breath metabolite determination model includes:

[0014] Acquire a training sample, wherein the training sample includes a plurality of training data, each training data includes disease information of a single disease, clinical research information of the disease, and a metabolite set, and the metabolite set of each training data includes information of each of the candidate differential metabolites;

[0015] Using the training samples, constructing a node interaction network including metabolite information, disease information and clinical research information;

[0016] Based on the node interaction network, construct a pre-trained network;

[0017] The pre-trained network is trained using the multiple training data to obtain a breath metabolite determination model.

[0018] Optionally, the pre-trained network includes a three-layer graph convolutional network and a single-layer fully connected network;

[0019] The pre-trained network is trained using the plurality of training data to obtain a breath metabolite determination model, including:

[0020] Using each piece of training data to train the three-layer graph convolutional network, so that the three-layer graph convolutional network can infer the node prediction results of the prediction nodes in the node interaction network for the training data;

[0021] Predicting the node prediction results of each piece of training data through the single-layer fully connected network to obtain the training prediction results of the training data;

[0022] Based on the training prediction results of each training data, the cross entropy loss formula is used to calculate the loss value of the pre-trained network when training the training data, so as to optimize the model parameters of the pre-trained network according to the loss value, until the pre-trained network completes the training of all training data, thereby obtaining a trained exhaled metabolite judgment model.

[0023] Optionally, the node interaction network includes metabolite nodes, disease nodes, clinical research nodes, chemical reaction location nodes, metabolic pathway correlation nodes, biochemical reaction correlation nodes, enzyme correlation nodes, clinical research support nodes, interaction nodes, marker possibility nodes, clinical research enhanced connection nodes and prediction nodes;

[0024] The node feature embedding of the metabolite node is:

[0025]

[0026] Wherein, M is the total metabolite set consisting of the SMILES structural formulas of all metabolites in the metabolite set of all training data in the training sample, m i is the SMILES structure of the ith metabolite, f m (·) is the structure vector encoder, k is the preset dimension;

[0027] The node feature embedding of the disease node is:

[0028] h d =GKS(d),h d ∈R k ,D={d1,d2,...,d i}

[0029] Where D is the disease set consisting of all diseases in all training data in the training sample, d i is the ith disease in the disease set, GKS(d) is the Gaussian kernel similarity between disease d and other diseases;

[0030] The node feature embedding of the clinical research node is:

[0031]

[0032] Where S is the clinical research set consisting of all clinical research information of all training data in the training sample, and f cs (·) is the clinical text coder;

[0033] For each of the chemical reaction location node, the metabolic pathway correlation node, the biochemical reaction correlation node, the enzyme correlation node, and the clinical research support node, the node feature embedding of the node is:

[0034] h * =f * (h input ),h * ∈R k

[0035]

[0036]

[0037] *={SOM,p,r,e,ds}

[0038]

[0039] Among them, f * (·) is a three-layer fully connected network with activation and residual connections, and FC(·) is used to train f * (·) is the fully connected layer operation for output prediction, and Sigmoid(·) is used to train f * (·) is the output prediction sigmoid activation function operation, SOM is the subscript code of the chemical reaction location node, p is the subscript code of the metabolic pathway correlation node, r is the subscript code of the biochemical reaction correlation node, e is the subscript code of the enzyme correlation node, ds is the subscript code of the clinical research support node, y * The prediction results of the nodes coded as *;

[0040] The node feature embedding of the interacting nodes is:

[0041] h I =f I (h d ,h cs ),h I ∈R k

[0042] Among them, f I (·) is a three-layer fully connected network with activation and residual connections;

[0043] The node feature embedding of the marker possibility node is:

[0044] h bp =f bp (h SOM ,h p ,h r ,h e ),hbp ∈R k

[0045] Among them, f bp (·) is a three-layer fully connected network with activation and residual connections;

[0046] The node feature embedding of the clinical research enhanced contact node is:

[0047] h a =f a (h ds ,h I ),h a ∈R k

[0048] Among them, f a (·) is a three-layer fully connected network with activation and residual connections;

[0049] The node feature embedding of the prediction node is:

[0050] h c =f c (h bp ,h a ),h c ∈R k

[0051] Among them, f c (·) is a three-layer fully connected network with activation and residual connections.

[0052] Optionally, the initialization node feature embedding of the node interaction network is:

[0053] H 0 =[h m ,h d ,h cs ,h SOM ,h p ,h r ,h e ,h ds ,h I ,h bp ,h a ,h c ],H 0 ∈R 12×k .

[0054] Optionally, the three-layer graph convolutional network includes a two-layer fully connected network;

[0055] The intermediate layer formula of the three-layer graph convolutional network is:

[0056] H n =ReLU(B n+(A⊙N)(H n-1 W n ))

[0057] Among them, H n is the intermediate variable of the nth layer of the three-layer graph convolutional network, ReLU(·) is the ReLU activation function used by the three-layer graph convolutional network, W n is the learnable parameter corresponding to the nth layer of the three-layer graph convolutional network, W n ∈R k×k , B n is the deviation corresponding to the nth layer of the three-layer graph convolutional network, B n ∈R k×k , A is B n adjacency matrix, N is the node attention matrix, N∈R k×k ;

[0058] The formula of the two-layer fully connected network is:

[0059]

[0060] Among them, N i,j is the element in the i-th row and j-th column of the node attention matrix, and is the feature encoding of any two nodes when the three-layer graph convolutional network is initialized, f att (·) is the function of the fully connected network of the two layers, and concat(·) is the connection function.

[0061] Optionally, predicting the node prediction result of each piece of training data through the single-layer fully connected network to obtain the training prediction result of the training data includes:

[0062] Use the following formula to predict the node prediction results of each training data to obtain the training prediction results of the training data:

[0063]

[0064] in, is the node prediction result of the three-layer graph convolutional network, y pred is the training prediction result, Sigmoid(·) is the Sigmoid activation function used by the single-layer fully connected network, and FC(·) is the fully connected layer operation of the single-layer fully connected network.

[0065] Optionally, the cross entropy loss formula is:

[0066] ζ=-ylog(y pred )-(1-y)log(1-y pred)

[0067] Among them, y is the preset prediction result target value, y pred is the training prediction result of each training data, and ζ is the loss value of the pre-trained network when training the training data.

[0068] A metabolite screening device for disease markers, comprising:

[0069] an exhaled breath acquisition unit, for acquiring the exhaled breath of a healthy subject and the exhaled breath of a subject suffering from a target disease;

[0070] a spectrum acquisition unit, configured to detect the exhaled breath of the healthy subject and the exhaled breath of the subject suffering from the target disease by means of a gas chromatograph coupled to a mass spectrometer, and obtain a first exhaled breath compound spectrum of the healthy subject and a second exhaled breath compound spectrum of the subject suffering from the target disease;

[0071] a spectrum processing unit, configured to perform normalization processing, peak removal processing, peak filling processing, and Z-score processing on each of the first breath compound spectrum and the second breath compound spectrum to obtain a standardized breath compound spectrum;

[0072] A statistical analysis value calculation unit, used to calculate the false positive discovery rate FDR value and the variable projection importance index VIP value of each compound peak in the standardized breath compound spectrum, and determine the compound corresponding to the compound peak whose FDR value in the standardized breath compound spectrum is less than a preset FDR threshold and whose VIP value is greater than a preset VIP threshold as a candidate differential metabolite;

[0073] The model output unit is used to input the information of each candidate differential metabolite, the disease information of the target disease and the clinical research information of the target disease into a pre-trained breath metabolite judgment model, and output the target differential metabolite of the target disease as a disease marker of the target disease.

[0074] Optionally, the device further comprises:

[0075] A training sample acquisition unit, used to acquire training samples, wherein the training samples include multiple training data, each of which includes disease information of a single disease, clinical research information of the disease, and a metabolite set, and the metabolite set of each training data includes information of each of the candidate differential metabolites;

[0076] A node interaction network construction unit, used to construct a node interaction network including metabolite information, disease information and clinical research information using the training samples;

[0077] A graph convolutional network construction unit, used to construct a pre-trained network based on the node interaction network;

[0078] The network training unit is used to train the pre-trained network through the multiple training data to obtain a breath metabolite judgment model.

[0079] Optionally, the pre-trained network includes a three-layer graph convolutional network and a single-layer fully connected network;

[0080] The network training unit comprises:

[0081] A training unit, used to train the three-layer graph convolutional network using each piece of training data, so that the three-layer graph convolutional network infers the node prediction result of the prediction node in the node interaction network for the training data;

[0082] A node prediction unit, used to predict the node prediction result of each piece of training data through the single-layer fully connected network to obtain the training prediction result of the training data;

[0083] The entropy loss feedback unit is used to calculate the loss value of the pre-trained network when training the training data based on the training prediction results of each training data using the cross-entropy loss formula, so as to optimize the model parameters of the pre-trained network according to the loss value, until the pre-trained network completes the training of all training data, thereby obtaining a trained exhaled metabolite judgment model.

[0084] Optionally, the node interaction network includes metabolite nodes, disease nodes, clinical research nodes, chemical reaction location nodes, metabolic pathway correlation nodes, biochemical reaction correlation nodes, enzyme correlation nodes, clinical research support nodes, interaction nodes, marker possibility nodes, clinical research enhanced connection nodes and prediction nodes;

[0085] The node feature embedding of the metabolite node is:

[0086]

[0087] Wherein, M is the total metabolite set consisting of the SMILES structural formulas of all metabolites in the metabolite set of all training data in the training sample, m i is the SMILES structure of the ith metabolite, f m (·) is the structure vector encoder, k is the preset dimension;

[0088] The node feature embedding of the disease node is:

[0089] h d =GKS(d),h d ∈Rk ,D={d1,d2,...,d i}

[0090] Where D is the disease set consisting of all diseases in all training data in the training sample, d i is the ith disease in the disease set, GKS(d) is the Gaussian kernel similarity between disease d and other diseases;

[0091] The node feature embedding of the clinical research node is:

[0092]

[0093] Where S is the clinical research set consisting of all clinical research information of all training data in the training sample, and f cs (·) is the clinical text coder;

[0094] For each of the chemical reaction location node, the metabolic pathway correlation node, the biochemical reaction correlation node, the enzyme correlation node, and the clinical research support node, the node feature embedding of the node is:

[0095] h * =f * (h input ),h * ∈R k

[0096]

[0097]

[0098] *={SOM,p,r,e,ds}

[0099]

[0100] Among them, f * (·) is a three-layer fully connected network with activation and residual connections, and FC(·) is used to train f * (·) is the fully connected layer operation for output prediction, and Sigmoid(·) is used to train f * (·) is the output prediction sigmoid activation function operation, SOM is the subscript code of the chemical reaction location node, p is the subscript code of the metabolic pathway correlation node, r is the subscript code of the biochemical reaction correlation node, e is the subscript code of the enzyme correlation node, ds is the subscript code of the clinical research support node, y * The prediction results of the nodes coded as *;

[0101] The node feature embedding of the interacting nodes is:

[0102] h I =f I (h d ,h cs ),h I ∈R k

[0103] Among them, f I (·) is a three-layer fully connected network with activation and residual connections;

[0104] The node feature embedding of the marker possibility node is:

[0105] h bp =f bp (h SOM ,h p ,h r ,h e ),h bp ∈R k

[0106] Among them, f bp (·) is a three-layer fully connected network with activation and residual connections;

[0107] The node feature embedding of the clinical research enhanced contact node is:

[0108] h a =f a (h ds ,h I ),h a ∈R k

[0109] Among them, f a (·) is a three-layer fully connected network with activation and residual connections;

[0110] The node feature embedding of the prediction node is:

[0111] h c =f c (h bp ,h a ),h c ∈R k

[0112] Among them, f c (·) is a three-layer fully connected network with activation and residual connections.

[0113] Optionally, the initialization node feature embedding of the node interaction network is:

[0114] H 0 =[hm ,h d ,h cs ,h SOM ,h p ,h r ,h e ,h ds ,h I ,h bp ,h a ,h c ],H 0 ∈R 12×k .

[0115] Optionally, the three-layer graph convolutional network includes a two-layer fully connected network;

[0116] The intermediate layer formula of the three-layer graph convolutional network is:

[0117] H n =ReLU(B n +(A⊙N)(H n-1 W n ))

[0118] Among them, H n is the intermediate variable of the nth layer of the three-layer graph convolutional network, ReLU(·) is the ReLU activation function used by the three-layer graph convolutional network, W n is the learnable parameter corresponding to the nth layer of the three-layer graph convolutional network, W n ∈R k×k , B n is the deviation corresponding to the nth layer of the three-layer graph convolutional network, B n ∈R k×k , A is B n adjacency matrix, N is the node attention matrix, N∈R k×k ;

[0119] The formula of the two-layer fully connected network is:

[0120]

[0121] Among them, N i,j is the element in the i-th row and j-th column of the node attention matrix, and is the feature encoding of any two nodes when the three-layer graph convolutional network is initialized, f att (·) is the function of the fully connected network of the two layers, and concat(·) is the connection function.

[0122] Optionally, the node prediction unit includes:

[0123] The node prediction subunit is used to predict the node prediction result of each training data using the following formula to obtain the training prediction result of the training data:

[0124]

[0125] in, is the node prediction result of the three-layer graph convolutional network, y pred is the training prediction result, Sigmoid(·) is the Sigmoid activation function used by the single-layer fully connected network, and FC(·) is the fully connected layer operation of the single-layer fully connected network.

[0126] Optionally, the cross entropy loss formula is:

[0127] ζ=-ylog(y pred )-(1-y)log(1-y pred )

[0128] Among them, y is the preset prediction result target value, y pred is the training prediction result of each training data, and ζ is the loss value of the pre-trained network when training the training data.

[0129] A metabolite screening device for disease markers, comprising a memory and a processor;

[0130] The memory is used to store programs;

[0131] The processor is used to execute the program to implement the various steps of the metabolite screening method for disease markers as described above.

[0132] A storage medium stores a computer program, which, when executed by a processor, implements the various steps of the metabolite screening method for disease markers as described above.

[0133] By means of the above technical solution, the present application obtains the breath of healthy subjects and disease subjects, constructs their respective compound spectra, and after standardizing their respective compound spectra, initially screens candidate differential metabolites based on the false positive discovery rate FDR value and the variable projection importance index VIP value, and further screens the target differential metabolites for judging the disease from each candidate differential metabolite through the breath metabolite judgment model. It can be seen that the standardized treatment of the compound spectra can screen out most of the invalid compounds, greatly narrow the scope of determination of disease markers, and re-screening in the candidate differential metabolites can obtain more accurate disease markers and more effectively assist in the judgment of disease detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0134] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present application. Also, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0135] Figure 1 A schematic diagram of a process for implementing metabolite screening of disease markers provided in an embodiment of the present application;

[0136] Figure 2 A schematic diagram of a flow chart for implementing the training of a breath metabolite determination model provided in an embodiment of the present application;

[0137] Figure 3 A schematic diagram of a node interaction network provided in an embodiment of the present application;

[0138] Figure 4 A schematic diagram of the structure of a device for implementing metabolite screening of disease markers provided in an embodiment of the present application;

[0139] Figure 5 A schematic diagram of the structure of a device for implementing metabolite screening of disease markers provided in an embodiment of the present application. DETAILED DESCRIPTION

[0140] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0141] The present application scheme can be implemented based on a terminal with data processing capabilities, which can be a disease marker screening system, and the disease marker screening system can be equipped with a breath collection device.

[0142] Next, combine Figure 1 As shown, the metabolite screening method of the disease marker of the present application may include the following steps:

[0143] Step S110: Acquire the exhaled breath of a healthy subject and the exhaled breath of a subject suffering from a target disease.

[0144] Specifically, a healthy subject may refer to a subject that does not suffer from any disease. The breath collection device may be used to collect breath from a subject that is predetermined to be in a healthy state, and to collect breath from a subject that is predetermined to suffer from a target disease.

[0145] Step S120, detecting the exhaled breath of the healthy subject and the exhaled breath of the subject suffering from the target disease by using a gas chromatograph coupled to a mass spectrometer to obtain a first exhaled breath compound spectrum of the healthy subject and a second exhaled breath compound spectrum of the subject suffering from the target disease.

[0146] Specifically, the exhaled breath of healthy subjects and subjects suffering from the target disease can be first detected by gas chromatograph coupled to mass spectrometry to obtain a first compound spectrum signal of the healthy subjects and a second compound spectrum signal of the subjects suffering from the target disease, and then the noise of the first compound spectrum signal and the second compound spectrum signal can be removed to obtain the first compound spectrum signal after noise removal and the second compound spectrum signal after noise removal.

[0147] It is understandable that in order to reduce the impact of instrument detection fluctuations, environmental interference and human errors on the spectrum and improve the stability and effectiveness of data analysis results, the spectrum data can be preprocessed before starting data analysis. The preprocessing process can include noise removal and baseline correction.

[0148] Specifically, for each of the first compound spectrum signal and the second compound spectrum signal, if there is a target signal point in the compound spectrum signal, the signal value of the target signal point is replaced by the signal mean of two adjacent signal points of the target signal point to obtain the compound spectrum signal after noise removal, wherein the signal value of the target signal point is greater than the preset signal ratio of the signal value of any adjacent signal point of the target signal point.

[0149] Furthermore, baseline calibration may be performed on the first compound spectrum signal after noise removal and the second compound spectrum signal after noise removal to obtain the baseline-calibrated first compound spectrum signal and the baseline-calibrated second compound spectrum signal.

[0150] It is understandable that, affected by the instrument itself, the baseline value in the spectrum data will usually gradually deviate from 0, affecting the peak detection and quantification process. Then the baseline drift can be fitted and the fitted curve can be used for calibration.

[0151] Furthermore, a first breath compound spectrum of a healthy subject may be plotted based on the baseline-calibrated first compound spectrum signal, and a second breath compound spectrum of a subject suffering from a target disease may be plotted based on the baseline-calibrated second compound spectrum signal.

[0152] Step S130 , for each of the first breath compound spectrum and the second breath compound spectrum, perform normalization processing, peak removal processing, peak filling processing and Z-score processing on the breath compound spectrum to obtain a standardized breath compound spectrum.

[0153] It is understandable that each compound peak in the breath compound spectrum may be represented by different features, so the normalization process can uniformly convert each compound peak into a relative content value.

[0154] Due to differences in subjects and sampling environments, the metabolite peaks detected in each sample are also different. Some compounds / metabolites cannot be identified by the algorithm because their concentration is too low or they are seriously interfered by background noise, which ultimately leads to a peak area of ​​zero. This type of peak only has area values ​​in a small number of samples, and most sample values ​​are missing. When the value of the peak is missing in most samples, it will cause errors or failures in the data analysis method, so peak removal can remove such peaks.

[0155] Although peaks with many missing values ​​are removed, the remaining peaks still have some missing values, so the peak filling process can fill the missing values. Specifically, the missing values ​​can be filled with 1 / 2 of the minimum non-zero value in the group to ensure the difference between groups.

[0156] Z-score processing can map the data of the breath compound spectrum to a standard normal distribution. The purpose is to make the weight of each peak consistent in the subsequent modeling process and eliminate the deviation caused by the numerical magnitude.

[0157] Step S140, calculating the false positive discovery rate FDR value and the variable projection importance index VIP value of each compound peak in the standardized breath compound spectrum, and determining the compounds corresponding to the compound peaks in the standardized breath compound spectrum whose FDR values ​​are less than a preset FDR threshold and whose VIP values ​​are greater than a preset VIP threshold as candidate differential metabolites.

[0158] Specifically, for each compound peak in the standardized exhaled breath compound spectrum, the p-value and VIP value of the compound peak are determined by comparing the peak value of the compound peak in the first standardized exhaled breath compound spectrum with the peak value of the compound peak in the second standardized exhaled breath compound spectrum, and the p-value is converted to obtain the FDR value of the compound peak.

[0159] Step S150: input information of each candidate differential metabolite, disease information of the target disease, and clinical research information of the target disease into a pre-trained breath metabolite judgment model, and output a target differential metabolite of the target disease.

[0160] Among them, the exhaled metabolite judgment model associates a large amount of biomedical knowledge related to metabolomics. The model contains a graph neural network, and the graph structure of the graph neural network represents the relationship between a large amount of metabolomics and biomedical knowledge. Therefore, the exhaled metabolite judgment model can integrate and analyze the chemical properties, structure and other information of the metabolites themselves, as well as information such as metabolites and metabolites, diseases, and clinical research, and abstract them into computable representation vectors, and finally use the representation vectors to output the target differential metabolites.

[0161] It can be understood that the target differential metabolite is obtained based on multiple candidate differential metabolites and is finally screened by the breath metabolite judgment model, so the target differential metabolite can be used as a disease marker for the target disease.

[0162] The metabolite screening method for disease markers provided in this embodiment obtains the breath of healthy subjects and disease subjects, constructs their respective compound spectra, and after standardizing their respective compound spectra, initially screens candidate differential metabolites based on the false positive discovery rate FDR value and the variable projection importance index VIP value, and further screens the target differential metabolites for judging the disease from each candidate differential metabolite through the breath metabolite judgment model. It can be seen that the standardization of the compound spectra can screen out most of the invalid compounds, greatly narrow the determination range of the disease markers, and re-screening in the candidate differential metabolites can obtain more accurate disease markers and more effectively assist the judgment of disease detection.

[0163] In some embodiments of the present application, the exhaled metabolite judgment model mentioned in the above embodiments is introduced. The exhaled metabolite judgment model can be constructed using an end-to-end training method. The model includes a trained three-layer graph convolutional network and a trained single-layer fully connected network. The three-layer graph convolutional network can include a two-layer fully connected network, and each layer of the two-layer fully connected network uses ReLU and Sigmoid as activation functions, respectively.

[0164] On this basis, the training process of the breath metabolite judgment model is introduced, such as Figure 2 As shown, the training process may include:

[0165] Step S210: Obtain training samples.

[0166] The training sample may include multiple training data. Each training data may include disease information of a single disease, clinical research information of the disease, and a metabolite set. The metabolite set of each training data may include information of each candidate differential metabolite.

[0167] Step S220: Using the training samples, construct a node interaction network including metabolite information, disease information and clinical research information.

[0168] Specifically, the node interaction network may include metabolite nodes, disease nodes, clinical research nodes, chemical reaction location nodes, metabolic pathway correlation nodes, biochemical reaction correlation nodes, enzyme correlation nodes, clinical research support nodes, interaction nodes, marker possibility nodes, clinical research enhanced connection nodes, and prediction nodes.

[0169] Among them, the node interaction network can include four types of nodes. The first type of node is the input node, including metabolite nodes, disease nodes and clinical research nodes. The second type of node is the external knowledge node, including chemical reaction location nodes, metabolic pathway correlation nodes, biochemical reaction correlation nodes, enzyme correlation nodes, clinical research support nodes and interaction nodes. The third type of node is the information summary node, including marker possibility nodes and clinical research enhanced connection nodes. The fourth type of node is a node associated with two nodes of the third type of node, which is a prediction node.

[0170] The node interaction network and the interaction relationship between each node are as follows: Figure 3 shown.

[0171] The node feature embeddings of the input nodes are pre-generated by the corresponding feature encoder and input from the outside.

[0172] Next, the node feature embedding of each input node is introduced.

[0173] The node feature embedding of the metabolite node is:

[0174]

[0175] Wherein, M is the total metabolite set consisting of the SMILES structural formulas of all metabolites in the metabolite set of all training data in the training sample, m i is the SMILES structure of the ith metabolite, f m (·) is a structural vector encoder, which can be specifically a SMILES-BERT structural vector encoder. SMILES-BERT is a BERT model focusing on molecular chemical structure. Its input is a SMILES structural expression, and its output is a vector embedding rich in molecular structural properties. k is a preset dimension.

[0176] The node feature embedding of the disease node is:

[0177] h d =GKS(d),h d ∈R k ,D={d1,d2,...,di}

[0178] Where D is the disease set consisting of all diseases in all training data in the training sample, d i is the ith disease in the disease set, and GKS(d) is the Gaussian kernel similarity between disease d and other diseases.

[0179] Specifically, the expression of GKS(d) is:

[0180] GKS(d)=[g(d,d0),...,g(d,d j )],d j ∈D

[0181]

[0182] Among them, Adj D (d) is the adjacency vector of disease d and other diseases, and α is a parameter calculated based on the adjacency vector. α can be calculated by the following iterative convergence formula:

[0183]

[0184] Among them, the relationship between different diseases can be pre-acquired through the MeSH (Medical Subject Headlings) database.

[0185] The node feature embedding of the clinical research node is:

[0186]

[0187] Where S is the clinical research set consisting of all clinical research information of all training data in the training sample, and f cs (·) is a clinical text encoder, which may be the ClinicalBERT clinical text encoder. The clinical research data may be pre-acquired through the ClinicalTrials.gov PRS (ClinicalTrials.gov Protocol Registration and Results System).

[0188] External knowledge nodes are used to introduce knowledge such as chemical reaction location (SOM), metabolic pathway (Pathway), biochemical reaction (Reaction), enzyme (Enzyme), disease research support (Disease Support) and interaction (Interaction). The above factors will affect the judgment of whether the metabolite is a disease diagnostic marker.

[0189] Chemical reaction sites indicate the response sites of metabolite structures when participating in biochemical reactions. Metabolic pathways connect diseases and metabolites. Biochemical reactions constitute metabolic pathways. Enzymes affect the rate and extent of biochemical reactions. Clinical research provides support for a deeper understanding of diseases. This information connects metabolites, diseases, and clinical research.

[0190] Next, the node feature embedding of each external knowledge node is introduced.

[0191] For each of the chemical reaction location node, the metabolic pathway correlation node, the biochemical reaction correlation node, the enzyme correlation node, and the clinical research support node, the node feature embedding of the node is:

[0192] h * =f * (h input ),h * ∈R k

[0193]

[0194]

[0195] *={SOM,p,r,e,ds}

[0196]

[0197] Among them, f * (·) is a three-layer fully connected network with activation and residual connections, and FC(·) is used to train f * (·) is the fully connected layer operation for output prediction, and Sigmoid(·) is used to train f * (·) is the output prediction sigmoid activation function operation, SOM is the subscript code of the chemical reaction location node, p is the subscript code of the metabolic pathway correlation node, r is the subscript code of the biochemical reaction correlation node, e is the subscript code of the enzyme correlation node, ds is the subscript code of the clinical research support node, y * The prediction results of the nodes coded as *.

[0198] It is understandable that since the influence sources of external knowledge nodes are not uniform, input Generally refers to the source, the specific input relationship is as follows Figure 3 shown.

[0199] The training data of external knowledge nodes can come from the KEGG (Kyoto Encyclopedia of Genes and Genomes) database and the HMDB (Human Metabolome Database) database. These databases provide the association between metabolites, metabolite reaction sites, chemical reactions, enzymes, metabolic pathways, and diseases. Through the relationship between external knowledge nodes and input nodes, a classification task data set (metabolite-reaction site, metabolite-disease-metabolic pathway, metabolite-disease-chemical reaction, metabolite-disease-clinical research-enzyme, disease-clinical research) was established, and the classification model was trained.

[0200] The node feature embedding of the interacting nodes is:

[0201] h I =f I (h d ,h cs ), h I ∈R k

[0202] Among them, f I (·) is a three-layer fully connected network with activation and residual connections.

[0203] Next, the node feature embedding of each information aggregation node is introduced.

[0204] The node feature embedding of the landmark possibility node is:

[0205] h bp =f bp (h SOM ,h p ,h r ,h e ), h bp ∈R k

[0206] Among them, f bp (·) is a three-layer fully connected network with activation and residual connections.

[0207] The node feature embedding of the clinical research enhanced contact node is:

[0208] h a =f a (h ds ,h I ), h a ∈R k

[0209] Among them, f a (·) is a three-layer fully connected network with activation and residual connections.

[0210] Next, the node feature embedding of the prediction node is introduced.

[0211] The node feature embedding of the prediction node is:

[0212] h c =f c (h bp ,h a ), h c ∈R k

[0213] Among them, f c (·) is a three-layer fully connected network with activation and residual connections.

[0214] Based on the node feature embedding of each node, the initial node feature embedding of the node interaction network can be obtained as:

[0215] H 0 =[h m ,h d ,h cs ,h SOM ,h p ,h r ,h e ,h ds ,h I ,h bp ,h a ,h c ],H 0 ∈R 12×k .

[0216] Step S230: construct a pre-trained network based on the node interaction network.

[0217] Specifically, the pre-trained network may include a three-layer graph convolutional network and a single-layer fully connected network. The three-layer graph convolutional network may include a two-layer fully connected network.

[0218] Among them, the intermediate layer formula of the three-layer graph convolutional network is:

[0219] H n =ReLU(B n +(A⊙N)(H n-1 W n ))

[0220] Among them, H n is the intermediate variable of the nth layer of the three-layer graph convolutional network, ReLU(·) is the ReLU activation function used by the three-layer graph convolutional network, W n is the learnable parameter corresponding to the nth layer of the three-layer graph convolutional network, W n ∈R k×k, B n is the deviation corresponding to the nth layer of the three-layer graph convolutional network, B n ∈R k×k , A is B n adjacency matrix, N is the node attention matrix, N∈R k×k ;

[0221] The formula of the two-layer fully connected network is:

[0222]

[0223] Among them, N i,j is the element in the i-th row and j-th column of the node attention matrix, and is the feature encoding of any two nodes when the three-layer graph convolutional network is initialized, f att (·) is the function of the fully connected network of the two layers, and concat(·) is the connection function.

[0224] Step S240: training the pre-trained network using multiple training data to obtain a breath metabolite determination model.

[0225] The metabolite screening method for disease markers provided in this embodiment integrates and associates a large amount of biomedical knowledge related to metabolomics, and represents the relationship between nodes as a graph structure, so as to use a graph neural network to characterize, learn and model the metabolites, so that the node interaction network has the chemical properties, structure and other information of the metabolites themselves, as well as the association information between metabolites, diseases, clinical research, etc., which enables the pre-trained network to be trained more accurately under multiple training data.

[0226] In some embodiments of the present application, the process of training the pre-trained network through multiple training data to obtain the breath metabolite judgment model in step S240 is introduced, and the process may include:

[0227] S1. Use each piece of training data to train a three-layer graph convolutional network, so that the three-layer graph convolutional network can infer the node prediction results of the predicted nodes in the node interaction network for the training data.

[0228] Specifically, in each training of the three-layer graph convolutional network, prediction is performed through each training data to obtain the node prediction result The last layer H of the three-layer graph convolutional network can be 2 Output.

[0229] S2. Predict the node prediction results of each training data through a single-layer fully connected network to obtain the training prediction results of the training data.

[0230] Specifically, the node prediction result of each training data can be predicted using the following formula to obtain the training prediction result of the training data:

[0231]

[0232] in, is the node prediction result of the three-layer graph convolutional network, y pred is the training prediction result, y pred The dimension includes all metabolites involved in the modeling, Sigmoid (·) is the Sigmoid activation function used by the single-layer fully connected network, and FC (·) is the fully connected layer operation of the single-layer fully connected network.

[0233] S3. Based on the training prediction results of each training data, the cross entropy loss formula is used to calculate the loss value of the pre-trained network when training the training data, so as to optimize the model parameters of the pre-trained network according to the loss value, until the pre-trained network completes the training of all training data, and obtains a trained exhaled metabolite judgment model.

[0234] Specifically, the cross entropy loss formula can be:

[0235] ζ=-ylog(y pred )-(1-y)log(1-y pred )

[0236] Among them, y is the preset prediction result target value, y pred is the training prediction result for each training data, and ζ is the loss value of the pre-trained network when training the training data.

[0237] Next, this example compares the effect achieved by the metabolite screening method for disease markers mentioned in the previous example with the disease marker screening effect achieved by other methods.

[0238] The exhaled metabolomics data of three different diseases were selected for comparative testing. The information of each data set is shown in the table below:

[0239]

[0240]

[0241] The comparative test uses the following three methods to screen for potential disease diagnostic markers:

[0242] The first method is to use only statistical methods to determine disease diagnostic markers: using the FDR value (converted from the p value generated by the independent T test) and the VIP value generated by the partial least squares discriminant analysis (PLS-DA). If the FDR value of a metabolite is less than 0.05 and the VIP value is greater than 1, it is considered to be different and is a disease marker.

[0243] The second method is to use only the breath metabolite judgment model for detecting diseases to determine disease diagnostic markers: the breath metabolite judgment model mentioned in the above embodiment is used to determine whether each metabolite is a disease diagnostic marker.

[0244] The third method is to use the first method for preliminary screening, and then use the second method to determine whether the candidate differential metabolites are disease diagnostic markers.

[0245] Considering the need to eliminate the differences in model settings as much as possible, the differential metabolites screened were all selected to model the disease using the Logistic regression model. Further considering the possibility of eliminating overfitting, 5-fold cross validation was selected for model training.

[0246] The ROC-AUC value, sensitivity and specificity of the final modeling results will be used as evaluation indicators. The calculation method of each evaluation indicator is:

[0247] ROC_AUC = Area? of? ROC

[0248]

[0249]

[0250] Among them, TP, TN, FP and FN are the true positive number (True Positive), the true negative number (True Negative), the false positive number (False Positive) and the false negative number (False Negative) respectively.

[0251] The final test results are shown in the following table:

[0252]

[0253]

[0254] It can be seen that the third screening method, that is, the metabolite screening method for disease markers provided in this application, can effectively screen out differential metabolites on the given three test data sets, and ultimately significantly improve the evaluation indicators of disease modeling.

[0255] The following is a description of the device for implementing metabolite screening of disease markers provided in the embodiments of the present application. The device for implementing metabolite screening of disease markers described below and the method for implementing metabolite screening of disease markers described above can be referenced to each other.

[0256] See also Figure 4 , Figure 4 This is a schematic diagram of the structure of a device for implementing metabolite screening of disease markers disclosed in an embodiment of the present application.

[0257] like Figure 4 As shown, the device may include:

[0258] An exhaled breath acquisition unit 11, used to acquire the exhaled breath of a healthy subject and the exhaled breath of a subject suffering from a target disease;

[0259] a spectrum acquisition unit 12, configured to detect the exhaled breath of the healthy subject and the exhaled breath of the subject suffering from the target disease by means of a gas chromatograph coupled to a mass spectrometer, and obtain a first exhaled breath compound spectrum of the healthy subject and a second exhaled breath compound spectrum of the subject suffering from the target disease;

[0260] a spectrum processing unit 13, configured to perform normalization processing, peak removal processing, peak filling processing and Z-score processing on each of the first breath compound spectrum and the second breath compound spectrum to obtain a standardized breath compound spectrum;

[0261] The statistical analysis value calculation unit 14 is used to calculate the false positive discovery rate FDR value and the variable projection importance index VIP value of each compound peak in the standardized breath compound spectrum, and determine the compound corresponding to the compound peak whose FDR value in the standardized breath compound spectrum is less than a preset FDR threshold and whose VIP value is greater than a preset VIP threshold as a candidate differential metabolite;

[0262] The model output unit 15 is used to input the information of each candidate differential metabolite, the disease information of the target disease and the clinical research information of the target disease into a pre-trained breath metabolite judgment model, and output the target differential metabolite of the target disease as a disease marker of the target disease.

[0263] Optionally, the device further comprises:

[0264] A training sample acquisition unit, used to acquire training samples, wherein the training samples include multiple training data, each of which includes disease information of a single disease, clinical research information of the disease, and a metabolite set, and the metabolite set of each training data includes information of each of the candidate differential metabolites;

[0265] A node interaction network construction unit, used to construct a node interaction network including metabolite information, disease information and clinical research information using the training samples;

[0266] A graph convolutional network construction unit, used to construct a pre-trained network based on the node interaction network;

[0267] The network training unit is used to train the pre-trained network through the multiple training data to obtain a breath metabolite judgment model.

[0268] Optionally, the pre-trained network includes a three-layer graph convolutional network and a single-layer fully connected network;

[0269] The network training unit comprises:

[0270] A training unit, used to train the three-layer graph convolutional network using each piece of training data, so that the three-layer graph convolutional network infers the node prediction result of the prediction node in the node interaction network for the training data;

[0271] A node prediction unit, used to predict the node prediction result of each piece of training data through the single-layer fully connected network to obtain the training prediction result of the training data;

[0272] The entropy loss feedback unit is used to calculate the loss value of the pre-trained network when training the training data based on the training prediction results of each training data using the cross-entropy loss formula, so as to optimize the model parameters of the pre-trained network according to the loss value, until the pre-trained network completes the training of all training data, thereby obtaining a trained exhaled metabolite judgment model.

[0273] Optionally, the node interaction network includes metabolite nodes, disease nodes, clinical research nodes, chemical reaction location nodes, metabolic pathway correlation nodes, biochemical reaction correlation nodes, enzyme correlation nodes, clinical research support nodes, interaction nodes, marker possibility nodes, clinical research enhanced connection nodes and prediction nodes;

[0274] The node feature embedding of the metabolite node is:

[0275]

[0276] Wherein, M is the total metabolite set consisting of the SMILES structural formulas of all metabolites in the metabolite set of all training data in the training sample, m i is the SMILES structure of the ith metabolite, f m (·) is the structure vector encoder, k is the preset dimension;

[0277] The node feature embedding of the disease node is:

[0278] h d =GKS(d),h d ∈R k ,D={d1,d2,...,d i}

[0279] Where D is the disease set consisting of all diseases in all training data in the training sample, d i is the ith disease in the disease set, GKS(d) is the Gaussian kernel similarity between disease d and other diseases;

[0280] The node feature embedding of the clinical research node is:

[0281]

[0282] Where S is the clinical research set consisting of all clinical research information of all training data in the training sample, and f cs (·) is the clinical text coder;

[0283] For each of the chemical reaction location node, the metabolic pathway correlation node, the biochemical reaction correlation node, the enzyme correlation node, and the clinical research support node, the node feature embedding of the node is:

[0284] h * =f * (h input ),h * ∈R k

[0285]

[0286]

[0287] *={SOM,p,r,e,ds}

[0288]

[0289] Among them, f * (·) is a three-layer fully connected network with activation and residual connections, and FC(·) is used to train f * (·) is the fully connected layer operation for output prediction, and Sigmoid(·) is used to train f *(·) is the output prediction sigmoid activation function operation, SOM is the subscript code of the chemical reaction location node, p is the subscript code of the metabolic pathway correlation node, r is the subscript code of the biochemical reaction correlation node, e is the subscript code of the enzyme correlation node, ds is the subscript code of the clinical research support node, y * The prediction results of the nodes coded as *;

[0290] The node feature embedding of the interacting nodes is:

[0291] h I =f I (h d ,h cs ), h I ∈R k

[0292] Among them, f I (·) is a three-layer fully connected network with activation and residual connections;

[0293] The node feature embedding of the marker possibility node is:

[0294] h bp =f bp (h SOM ,h p ,h r ,h e ), h bp ∈R k

[0295] Among them, f bp (·) is a three-layer fully connected network with activation and residual connections;

[0296] The node feature embedding of the clinical research enhanced contact node is:

[0297] h a =f a (h ds ,h I ), h a ∈R k

[0298] Among them, f a (·) is a three-layer fully connected network with activation and residual connections;

[0299] The node feature embedding of the prediction node is:

[0300] h c =f c (h bp ,h a ), h c ∈Rk

[0301] Among them, f c (·) is a three-layer fully connected network with activation and residual connections.

[0302] Optionally, the initialization node feature embedding of the node interaction network is:

[0303] H 0 =[h m ,h d ,h cs ,h SOM ,h p ,h r ,h e ,h ds ,h I ,h bp ,h a ,h c ], H 0 ∈R 12×k .

[0304] Optionally, the three-layer graph convolutional network includes a two-layer fully connected network;

[0305] The intermediate layer formula of the three-layer graph convolutional network is:

[0306] H n =ReLU(B n +(A⊙N)(H n-1 W n ))

[0307] Among them, H n is the intermediate variable of the nth layer of the three-layer graph convolutional network, ReLU(·) is the ReLU activation function used by the three-layer graph convolutional network, W n is the learnable parameter corresponding to the nth layer of the three-layer graph convolutional network, W n ∈R k×k , B n is the deviation corresponding to the nth layer of the three-layer graph convolutional network, B n ∈R k×k , A is B n adjacency matrix, N is the node attention matrix, N∈R k×k ;

[0308] The formula of the two-layer fully connected network is:

[0309]

[0310] Among them, N i,j is the element in the i-th row and j-th column of the node attention matrix, and is the feature encoding of any two nodes when the three-layer graph convolutional network is initialized, f att (·) is the function of the fully connected network of the two layers, and concat(·) is the connection function.

[0311] Optionally, the node prediction unit includes:

[0312] The node prediction subunit is used to predict the node prediction result of each training data using the following formula to obtain the training prediction result of the training data:

[0313]

[0314] in, is the node prediction result of the three-layer graph convolutional network, y pred is the training prediction result, Sigmoid(·) is the Sigmoid activation function used by the single-layer fully connected network, and FC(·) is the fully connected layer operation of the single-layer fully connected network.

[0315] Optionally, the cross entropy loss formula is:

[0316] ζ=-ylog(y pred )-(1-y)log(1-y pred )

[0317] Among them, y is the preset prediction result target value, y pred is the training prediction result of each training data, and ζ is the loss value of the pre-trained network when training the training data.

[0318] The device for screening metabolites of disease markers provided in the embodiments of the present application can be applied to equipment for screening metabolites of disease markers, such as a disease marker screening system. Optionally, Figure 5 The hardware structure block diagram of the device for screening metabolites of disease markers is shown in FIG. Figure 5 , the hardware structure of the device for screening metabolites of disease markers may include: at least one processor 1, at least one communication interface 2, at least one memory 3 and at least one communication bus 4;

[0319] In the embodiment of the present application, the number of the processor 1, the communication interface 2, the memory 3, and the communication bus 4 is at least one, and the processor 1, the communication interface 2, and the memory 3 communicate with each other through the communication bus 4;

[0320] The processor 1 may be a central processing unit CPU, or an application-specific integrated circuit ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of the present invention, etc.;

[0321] The memory 3 may include a high-speed RAM memory, and may also include a non-volatile memory, such as at least one disk memory;

[0322] The memory stores a program, and the processor can call the program stored in the memory, wherein the program is used to:

[0323] obtaining exhaled breath from a healthy subject and exhaled breath from a subject suffering from a target disease;

[0324] Detecting the exhaled breath of the healthy subject and the exhaled breath of the subject suffering from the target disease by using a gas chromatograph coupled to a mass spectrometer to obtain a first exhaled breath compound spectrum of the healthy subject and a second exhaled breath compound spectrum of the subject suffering from the target disease;

[0325] For each breath compound spectrum in the first breath compound spectrum and the second breath compound spectrum, performing normalization processing, peak removal processing, peak filling processing and Z-score processing on the breath compound spectrum to obtain a standardized breath compound spectrum;

[0326] Calculating the false positive discovery rate (FDR) value and the variable projection importance index (VIP) value of each compound peak in the standardized breath compound spectrum, and determining the compounds corresponding to the compound peaks in the standardized breath compound spectrum whose FDR values ​​are less than a preset FDR threshold and whose VIP values ​​are greater than a preset VIP threshold as candidate differential metabolites;

[0327] The information of each candidate differential metabolite, the disease information of the target disease and the clinical research information of the target disease are input into a pre-trained breath metabolite judgment model, and the target differential metabolite of the target disease is output as a disease marker of the target disease.

[0328] Optionally, the detailed functions and extended functions of the program may refer to the above description.

[0329] The embodiment of the present application further provides a storage medium, which may store a program suitable for execution by a processor, wherein the program is used to:

[0330] obtaining exhaled breath from a healthy subject and exhaled breath from a subject suffering from a target disease;

[0331] Detecting the exhaled breath of the healthy subject and the exhaled breath of the subject suffering from the target disease by using a gas chromatograph coupled to a mass spectrometer to obtain a first exhaled breath compound spectrum of the healthy subject and a second exhaled breath compound spectrum of the subject suffering from the target disease;

[0332] For each breath compound spectrum in the first breath compound spectrum and the second breath compound spectrum, performing normalization processing, peak removal processing, peak filling processing and Z-score processing on the breath compound spectrum to obtain a standardized breath compound spectrum;

[0333] Calculating the false positive discovery rate (FDR) value and the variable projection importance index (VIP) value of each compound peak in the standardized breath compound spectrum, and determining the compounds corresponding to the compound peaks in the standardized breath compound spectrum whose FDR values ​​are less than a preset FDR threshold and whose VIP values ​​are greater than a preset VIP threshold as candidate differential metabolites;

[0334] The information of each candidate differential metabolite, the disease information of the target disease and the clinical research information of the target disease are input into a pre-trained breath metabolite judgment model, and the target differential metabolite of the target disease is output as a disease marker of the target disease.

[0335] Optionally, the detailed functions and extended functions of the program may refer to the above description.

[0336] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0337] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can refer to each other.

[0338] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for screening metabolites of disease markers, characterized in that: include: obtaining exhaled breath from a healthy subject and exhaled breath from a subject suffering from a target disease; Detecting the exhaled breath of the healthy subject and the exhaled breath of the subject suffering from the target disease by using a gas chromatograph coupled to a mass spectrometer to obtain a first exhaled breath compound spectrum of the healthy subject and a second exhaled breath compound spectrum of the subject suffering from the target disease; For each breath compound spectrum in the first breath compound spectrum and the second breath compound spectrum, performing normalization processing, peak removal processing, peak filling processing and Z-score processing on the breath compound spectrum to obtain a standardized breath compound spectrum; Calculating the false positive discovery rate (FDR) value and the variable projection importance index (VIP) value of each compound peak in the standardized breath compound spectrum, and determining the compounds corresponding to the compound peaks in the standardized breath compound spectrum whose FDR values ​​are less than a preset FDR threshold and whose VIP values ​​are greater than a preset VIP threshold as candidate differential metabolites; Inputting information of each candidate differential metabolite, disease information of the target disease, and clinical research information of the target disease into a pre-trained breath metabolite judgment model, and outputting a target differential metabolite of the target disease as a disease marker of the target disease; The training process of the breath metabolite judgment model includes: Acquire a training sample, wherein the training sample includes a plurality of training data, each training data includes disease information of a single disease, clinical research information of the disease, and a metabolite set, and the metabolite set of each training data includes information of each of the candidate differential metabolites; Using the training samples, constructing a node interaction network including metabolite information, disease information and clinical research information; Based on the node interaction network, construct a pre-trained network; The pre-trained network is trained using the plurality of training data to obtain a breath metabolite determination model; The pre-trained network includes a three-layer graph convolutional network and a single-layer fully connected network; The pre-trained network is trained using the plurality of training data to obtain a trained breath metabolite determination model, including: Using each piece of training data to train the three-layer graph convolutional network, so that the three-layer graph convolutional network can infer the node prediction results of the prediction nodes in the node interaction network for the training data; Predicting the node prediction results of each piece of training data through the single-layer fully connected network to obtain the training prediction results of the training data; Based on the training prediction results of each training data, a cross entropy loss formula is used to calculate the loss value of the pre-trained network when training the training data, so as to optimize the model parameters of the pre-trained network according to the loss value, until the pre-trained network completes the training of all training data, thereby obtaining a trained breath metabolite judgment model; The node interaction network includes metabolite nodes, disease nodes, clinical research nodes, chemical reaction location nodes, metabolic pathway correlation nodes, biochemical reaction correlation nodes, enzyme correlation nodes, clinical research support nodes, interaction nodes, marker possibility nodes, clinical research enhanced connection nodes and prediction nodes; The node feature embedding of the metabolite node is: Wherein, M is the total metabolite set consisting of the SMILES structural formulas of all metabolites in the metabolite set of all training data in the training sample, m i is the SMILES structure of the ith metabolite, f m (·) is the structure vector encoder, k is the preset dimension; The node feature embedding of the disease node is: h d =GKS(d),h d ∈R k ,D={d1,d2,...,d i } Where D is the disease set consisting of all diseases in all training data in the training sample, d i is the ith disease in the disease set, GKS(d) is the Gaussian kernel similarity between disease d and other diseases; The node feature embedding of the clinical research node is: Where S is the clinical research set consisting of all clinical research information of all training data in the training sample, and f cs (·) is the clinical text coder; For each of the chemical reaction location node, the metabolic pathway correlation node, the biochemical reaction correlation node, the enzyme correlation node, and the clinical research support node, the node feature embedding of the node is: h * =f * (h input ),h * ∈R k *={SOM,p,r,e,ds} Among them, f * (·) is a three-layer fully connected network with activation and residual connections, and FC(·) is used to train f * (·) is the fully connected layer operation for output prediction, and Sigmoid(·) is used to train f * (·) is the output prediction sigmoid activation function operation, SOM is the subscript code of the chemical reaction location node, p is the subscript code of the metabolic pathway correlation node, r is the subscript code of the biochemical reaction correlation node, e is the subscript code of the enzyme correlation node, ds is the subscript code of the clinical research support node, y * The prediction results of the nodes coded as *; The node feature embedding of the interacting nodes is: h I =f I (h d ,h cs ),h I ∈R k Among them, f I (·) is a three-layer fully connected network with activation and residual connections; The node feature embedding of the marker possibility node is: h bp =f bp (h SOM ,h p ,h r ,h e ),h bp ∈R k Among them, f bp (·) is a three-layer fully connected network with activation and residual connections; The node feature embedding of the clinical research enhanced contact node is: h a =f a (h ds ,h I ),h a ∈R k Among them, f a (·) is a three-layer fully connected network with activation and residual connections; The node feature embedding of the prediction node is: h c =f c (h bp ,h a ),h c ∈R k Among them, f c (·) is a three-layer fully connected network with activation and residual connections.

2. The method according to claim 1, characterized in that The initialization node feature embedding of the node interaction network is: H 0 =[h m ,h d ,h cs ,h SOM ,h p ,h r ,h e ,h ds ,h I ,h bp ,h a ,h c ],H 0 ∈R 12×k 。 3. The method according to claim 2, characterized in that The three-layer graph convolutional network includes a two-layer fully connected network; The intermediate layer formula of the three-layer graph convolutional network is: A n =ReLU(B n +(A⊙N)(H n-1 W n )) Among them, H n is the intermediate variable of the nth layer of the three-layer graph convolutional network, ReLU(·) is the ReLU activation function used by the three-layer graph convolutional network, W n is the learnable parameter corresponding to the nth layer of the three-layer graph convolutional network, W n ∈R k×k , B n is the deviation corresponding to the nth layer of the three-layer graph convolutional network, B n ∈R k×k , A is B n adjacency matrix, N is the node attention matrix, N∈R k×k ; The formula of the two-layer fully connected network is: Among them, N i,j is the element in the i-th row and j-th column of the node attention matrix, and is the feature encoding of any two nodes when the three-layer graph convolutional network is initialized, f att (·) is the function of the fully connected network of the two layers, and concat(·) is the connection function.

4. The method according to claim 3, characterized in that The node prediction result of each piece of training data is predicted by the single-layer fully connected network to obtain the training prediction result of the training data, including: Use the following formula to predict the node prediction results of each training data to obtain the training prediction results of the training data: in, is the node prediction result of the three-layer graph convolutional network, y pred is the training prediction result, Sigmoid(·) is the Sigmoid activation function used by the single-layer fully connected network, and FC(·) is the fully connected layer operation of the single-layer fully connected network.

5. The method according to any one of claims 1 to 4, characterized in that: The cross entropy loss formula is: ζ=-ylog(y pred )-(1-y)log(1-y pred ) Among them, y is the preset prediction result target value, y pred is the training prediction result of each training data, and ζ is the loss value of the pre-trained network when training the training data.

6. A metabolite screening device for disease markers, characterized in that: include: an exhaled breath acquisition unit, for acquiring the exhaled breath of a healthy subject and the exhaled breath of a subject suffering from a target disease; a spectrum acquisition unit, configured to detect the exhaled breath of the healthy subject and the exhaled breath of the subject suffering from the target disease by means of a gas chromatograph coupled to a mass spectrometer, and obtain a first exhaled breath compound spectrum of the healthy subject and a second exhaled breath compound spectrum of the subject suffering from the target disease; a spectrum processing unit, configured to perform normalization processing, peak removal processing, peak filling processing, and Z-score processing on each of the first breath compound spectrum and the second breath compound spectrum to obtain a standardized breath compound spectrum; A statistical analysis value calculation unit, used to calculate the false positive discovery rate FDR value and the variable projection importance index VIP value of each compound peak in the standardized breath compound spectrum, and determine the compound corresponding to the compound peak whose FDR value in the standardized breath compound spectrum is less than a preset FDR threshold and whose VIP value is greater than a preset VIP threshold as a candidate differential metabolite; A model output unit, used to input information of each candidate differential metabolite, disease information of the target disease and clinical research information of the target disease into a pre-built breath metabolite judgment model, and output a target differential metabolite of the target disease as a disease marker of the target disease; The device also includes: A training sample acquisition unit, used to acquire training samples, wherein the training samples include multiple training data, each of which includes disease information of a single disease, clinical research information of the disease, and a metabolite set, and the metabolite set of each training data includes information of each of the candidate differential metabolites; A node interaction network construction unit, used to construct a node interaction network including metabolite information, disease information and clinical research information using the training samples; A graph convolutional network construction unit, used to construct a pre-trained network based on the node interaction network; A network training unit, used to train the pre-trained network using the plurality of training data to obtain a breath metabolite determination model; The pre-trained network includes a three-layer graph convolutional network and a single-layer fully connected network; The network training unit comprises: A training unit, used to train the three-layer graph convolutional network using each piece of training data, so that the three-layer graph convolutional network infers the node prediction result of the prediction node in the node interaction network for the training data; A node prediction unit, used to predict the node prediction result of each piece of training data through the single-layer fully connected network to obtain the training prediction result of the training data; An entropy loss feedback unit is used to calculate the loss value of the pre-trained network under the training of each training data using a cross entropy loss formula based on the training prediction result of each training data, so as to optimize the model parameters of the pre-trained network according to the loss value, until the pre-trained network completes the training of all training data, thereby obtaining a trained breath metabolite judgment model; The node interaction network includes metabolite nodes, disease nodes, clinical research nodes, chemical reaction location nodes, metabolic pathway correlation nodes, biochemical reaction correlation nodes, enzyme correlation nodes, clinical research support nodes, interaction nodes, marker possibility nodes, clinical research enhanced connection nodes and prediction nodes; The node feature embedding of the metabolite node is: Wherein, M is the total metabolite set consisting of the SMILES structural formulas of all metabolites in the metabolite set of all training data in the training sample, m i is the SMILES structure of the ith metabolite, f m (·) is the structure vector encoder, k is the preset dimension; The node feature embedding of the disease node is: h d =GKS(d),h d ∈R k ,D={d1,d2,...,d i } Where D is the disease set consisting of all diseases in all training data in the training sample, d i is the ith disease in the disease set, GKS(d) is the Gaussian kernel similarity between disease d and other diseases; The node feature embedding of the clinical research node is: Where S is the clinical research set consisting of all clinical research information of all training data in the training sample, and f cs (·) is the clinical text coder; For each of the chemical reaction location node, the metabolic pathway correlation node, the biochemical reaction correlation node, the enzyme correlation node, and the clinical research support node, the node feature embedding of the node is: h * =f * (h input ),h * ∈R k *={SOM,p,r,e,ds} Among them, f * (·) is a three-layer fully connected network with activation and residual connections, and FC(·) is used to train f * (·) is the fully connected layer operation for output prediction, and Sigmoid() is used to train f * (·) is the output prediction sigmoid activation function operation, SOM is the subscript code of the chemical reaction location node, p is the subscript code of the metabolic pathway correlation node, r is the subscript code of the biochemical reaction correlation node, e is the subscript code of the enzyme correlation node, ds is the subscript code of the clinical research support node, y * The prediction results of the nodes coded as *; The node feature embedding of the interacting nodes is: h I =f I (h d ,h cs ),h I ∈R k Among them, f I (·) is a three-layer fully connected network with activation and residual connections; The node feature embedding of the marker possibility node is: h bp =f bp (h SOM ,h p ,h r ,h e ),h bp ∈R k Among them, f bp (·) is a three-layer fully connected network with activation and residual connections; The node feature embedding of the clinical research enhanced contact node is: h a =f a (h ds ,h I ),h a ∈R k Among them, f a (·) is a three-layer fully connected network with activation and residual connections; The node feature embedding of the prediction node is: h c =f c (h bp ,h a ),h c ∈R k Among them, f c (·) is a three-layer fully connected network with activation and residual connections.

7. A metabolite screening device for disease markers, characterized in that: including memory and processor; The memory is used to store programs; The processor is used to execute the program to implement each step of the metabolite screening method for disease markers according to any one of claims 1 to 5.

8. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, each step of the method for screening metabolites of disease markers according to any one of claims 1 to 5 is implemented.

Images