A method for predicting active ingredient groups in traditional Chinese medicine prescriptions based on heterogeneous network embedding
The biomolecular information network of traditional Chinese medicine prescriptions was constructed through heterogeneous network embedding method to predict active ingredient groups, solving the problems of low screening efficiency of traditional Chinese medicine prescriptions and insufficient information capture, and achieving efficient and accurate screening of traditional Chinese medicine active ingredients.
Patent Information
- Application Number
- CN202111346064.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-11-12
- Publication Date
- 2024-11-19
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The screening efficiency of active ingredient groups of traditional Chinese medicine prescriptions is low, the existing analytical methods cannot effectively capture multi-level information of heterogeneous networks, and the disease treatment mechanism is unclear and the target information is incomplete.
Using a method based on heterogeneous network embedding, a drug molecule information network is constructed, and the embedded representation of Chinese medicine prescriptions is generated through heterogeneous network embedding learning. The clustering algorithm is used to predict the active ingredient group, and the correlation between Western medicine molecules and Chinese medicine prescriptions is generated to generate labels and Update network parameters and implement node embedding.
It improves the efficiency and accuracy of screening active ingredients of traditional Chinese medicine, can effectively utilize multi-level information in the network, reduce the experimental failure rate, and helps the development of traditional Chinese medicine prescriptions.
Smart Images

Figure CN114038499B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for predicting active ingredient groups in a traditional Chinese medicine pharmacy, and in particular to a method for predicting active ingredient groups in a traditional Chinese medicine prescription based on heterogeneous network embedding. Background Art
[0002] Traditional Chinese medicine has complex ingredients, and the screening method of active ingredient groups of traditional Chinese medicine based only on wet experiments has low screening efficiency. With the rapid rise of systems biology, scholars at home and abroad began to explore the systematic research methods of diseases and drugs from the perspective of complex networks, and the new discipline of network pharmacology began and developed rapidly. Network pharmacology uses computers to analyze molecular information networks containing nodes such as drug molecules, targets, pathways, and go terms to predict the relationship between drugs and diseases, and the relationship between drugs and targets. It is also used to predict the active ingredient groups of traditional Chinese medicine compound prescriptions. At present, most studies on the molecular information network of traditional Chinese medicine prescriptions are still only on the centrality analysis of the network nodes of molecules and disease common targets, the construction of the interaction network of molecules and disease common targets, the screening of key targets through node centrality analysis, and the prediction results of the active ingredient groups of traditional Chinese medicine compound prescriptions related to key targets. However, this method has its limitations: first, the treatment mechanism of many complex diseases is still unclear, and the target information related to the disease is incomplete in the database; second, when drug molecules treat diseases, they may not directly act on disease-related targets, but act on the disease indirectly through the interaction between drug targets.
[0003] Although the therapeutic mechanisms of many complex diseases have not been fully elucidated, with the continuous deepening of research, many Western medicine chemical molecules based on specific therapeutic mechanisms have been discovered. When drugs treat diseases, they work through targets. These Western medicine molecules have a certain therapeutic effect on diseases. Therefore, their mechanism for treating complex diseases may be contained in the molecular information network composed of these drug molecules, human targets, related pathways, and GO terms. The Chinese medicine biomolecular information network is essentially a heterogeneous information network that integrates multi-level information such as Chinese medicine ingredients, targets, pathways, and diseases. The heterogeneous network embedding method can conveniently use the current machine learning methods to complete various prediction tasks by integrating the semantic information under different meta-paths, characterizing the relationship between nodes from multiple dimensions, and mapping the nodes in the network to a low-dimensional vector space. Summary of the invention
[0004] Purpose of the invention: The purpose of the present invention is to address the problem that the screening efficiency of active ingredient groups in traditional Chinese medicine prescriptions is low and the existing methods for analyzing molecular information networks of traditional Chinese medicine prescriptions cannot capture the multi-level information of heterogeneous networks. A method for predicting active ingredient groups in traditional Chinese medicine prescriptions based on heterogeneous network embedding (Heterogeneous Network Embedding for Prediction of Active Ingredients Group in Chinese Formula, HEPA) is proposed;
[0005] Technical solution: To achieve the above-mentioned purpose, the present invention provides a method for predicting active ingredient groups of traditional Chinese medicine prescriptions based on heterogeneous network embedding, which comprises the following steps executed in sequence:
[0006] Step 1. Construction of drug molecule information network: Collect the targets, pathways, and GO term information of molecules in traditional Chinese medicine prescriptions and Western medicine molecules from public databases. At the same time, collect the association between diseases and Western medicine molecules from public data and, based on the above data, construct a drug molecule information network; based on the constructed drug molecule information network, generate labels for Western medicine molecules according to the Western medicine molecule-disease relationship and the correlation between Western medicine molecules and diseases that can be treated by traditional Chinese medicine prescriptions, and use one-hot encoding to represent them.
[0007] Step 2. Embedding of drug molecule information network: Based on the drug molecule information network obtained in step 1, the drug molecule information network is a heterogeneous network. Through the heterogeneous network embedding learning method, the embedding representation of the molecules of the traditional Chinese medicine prescription and the related western medicine molecules of the diseases treated by the traditional Chinese medicine prescription is obtained;
[0008] Step 3. Cluster analysis to predict the active ingredient group of the Chinese medicine prescription: Based on the embedded representation obtained in step 2, a clustering algorithm is used to analyze the embedded representation of Chinese medicine molecules and Western medicine molecules, thereby obtaining multiple clusters, and the basic information of each cluster is counted. The basic information of the cluster includes the number of Chinese medicine molecules and the number of Western medicine molecules, and the Chinese medicine molecules that can be clustered with Western medicine molecules are taken as the active ingredient group;
[0009] Furthermore, the step 1 specifically includes the following steps:
[0010] Step 1.1: Data collection of Chinese medicine prescriptions: First, for each Chinese medicine prescription, the chemical molecule information contained in the Chinese medicine prescription is collected through the database, including the name of the molecule, CAS number, and related targets. The drug molecules in the prescription are preliminarily screened according to the oral bioavailability (OB)>=30% and the drug-likeness (DL)>=0.18 to obtain the component molecule set of the Chinese medicine prescription to be studied; then, the target information of each component is retrieved through the molecular target database;
[0011] Step 1.2: Western medicine data collection: Collect the names of diseases that can be treated by Western medicine molecules and the target information related to the molecules through public databases;
[0012] Step 1.3: After obtaining the target information of traditional Chinese medicine prescriptions and molecular targets of Western medicine, obtain the target-related KEGG pathway and GO term information from the public database through web crawler technology;
[0013] Step 1.4: construct a drug molecule information network based on the collected data, wherein the nodes of the drug molecule information network are diseases, prescriptions, Chinese medicine molecules, Western medicine molecules, targets, pathways, and GO terms, and the edges of the drug molecule information network are prescription-Chinese medicine molecules, molecules-targets, targets-pathways, targets-GO terms, prescriptions-diseases, and Western medicine molecules-diseases;
[0014] Step 1.5: Based on the constructed drug molecule information network, according to the relationship between western medicine molecules and diseases, and the correlation between the diseases that can be treated by western medicine molecules and traditional Chinese medicine prescriptions, generate labels for western medicine molecules.
[0015] Furthermore, the step 2 specifically includes the following steps:
[0016] Step 2.1: for the drug molecule information network constructed in step 1, generate a one-hot encoding for all nodes in the network as the initial features of the nodes in the network;
[0017] Step 2.2 Heterogeneous mutual attention calculation:
[0018] For the drug molecule information network constructed in step 1, for a node t, its neighbor nodes s∈N(t) may belong to different node types, and their mutual attention needs to be calculated according to the type of their edges. For each edge e=(s,t) in the network, calculate the h-head attention. The specific calculation process is as follows:
[0019]
[0020]
[0021]
[0022]
[0023] in, h is the number of attention heads, is the vector dimension of each attention head, τ(*) represents mapping node * to its category set, φ(e) maps edge e to the category set of the edge, H l-1 [*] indicates the representation of node * at layer l-1;
[0024] Step 2.3 Heterogeneous Information Propagation
[0025] For the drug molecule information network constructed in step 1, message propagation is performed from the source node to the target node, and the initial message uses the one-hot encoding generated for all nodes in step 2.1. For the edge e = (s, t), the multi-head message can be calculated as:
[0026]
[0027]
[0028] in h is the number of attention heads, is the vector dimension of each attention head.
[0029] Step 2.4 Aggregation of heterogeneous information
[0030] For the drug molecule information network constructed in step 1, after the calculation of heterogeneous mutual attention in step 2.2 and heterogeneous information in step 2.3 is completed, it is necessary to aggregate the heterogeneous messages of all neighboring nodes s∈N(t) for the target node t, so as to obtain the intrinsic connection between the nodes of the drug molecule heterogeneous information network. The calculation method of heterogeneous message aggregation is as follows:
[0031]
[0032] get In order to obtain the node representation of layer l, it is necessary to use the linear mapping A-Linear according to the category of the target node. τ(t) Mapped to the distribution of the corresponding category, the representation of the node can be calculated as:
[0033]
[0034] Repeating the process from step 2.2 to step 2.4 L times, we can obtain a rich context representation H for each node of the drug molecule information network constructed in step 1. (L) .
[0035] Step 2.5 Heterogeneous Network Embedding Learning
[0036] For the drug molecule information network constructed in step 1, step 1.5 generates labels for western medicine molecules based on the constructed drug molecule information network, according to the relationship between western medicine molecules and diseases, and the correlation between diseases that can be treated by western medicine molecules and traditional Chinese medicine prescriptions. At the same time, step 2.4 generates node representations. Since western medicine molecules and traditional Chinese medicine molecules are nodes of the same type in the network, this method designs a classification task for western medicine molecule nodes to update the parameters in the network, thereby obtaining the final embedding results of all nodes in the heterogeneous network. Specifically, for the node representation H obtained in step 2.4 (L) , the representation of the molecular nodes of western medicine is used as the input of the MLP classifier, and the output corresponds to the molecular label generated in step 1.5. The western medicine molecules are randomly divided into 80% training set and 20% validation set. The loss function uses the cross entropy loss function, Adam is used as the optimizer, and then the gradient of the prediction loss is used to update the model parameters, thereby obtaining the network embedding of all molecular nodes in the drug molecule information network.
[0037] The further step 3 specifically comprises the following steps:
[0038] Step 3.1 For each molecular node obtained in step 2, the network embedding of each component molecule of the TCM prescription to be tested and the network embedding of the Western medicine molecule for the related diseases treated by the network TCM prescription are combined to obtain the molecular network embedding set related to the TCM prescription.
[0039] Step 3.2: For the molecular network embedding set related to the Chinese medicine prescription obtained in step 3.1, the spectral clustering algorithm is used to perform cluster analysis on the molecular network embedding set related to the prescription, and then the k value of the cluster is determined by the silhouette coefficient method. The silhouette coefficient is an evaluation index used to evaluate the clustering effect, which takes into account the two factors of cohesion and separation in the classification. The best k value can be determined by looking at the silhouette coefficient of the spectral clustering algorithm at different k values. For different k values, the spectral clustering algorithm is used to cluster the relevant data of each Chinese medicine prescription, and the mean of the silhouette coefficient of all sample points is calculated. The larger the mean of the silhouette coefficient, the better the cohesion and separation of the cluster. The k value that makes the coefficient larger is selected as the parameter of the spectral clustering algorithm to obtain the final cluster cluster. After obtaining the final cluster cluster for each Chinese medicine prescription, the basic information of each cluster cluster, including the number of Chinese medicine molecules and the number of Western medicine molecules, is counted, and the Chinese medicine molecules that can be clustered with most Western medicine molecules are used as the prediction results of the active ingredient group.
[0040] Beneficial effects:
[0041] The present invention provides a method for predicting the active ingredient group of traditional Chinese medicine prescriptions based on heterogeneous network embedding. Traditional Chinese medicine and its prescriptions have the characteristics of multi-component, multi-pathway, multi-target synergy, etc. The complexity of its components and effects on the human body makes it very difficult to study its pharmacological substances and mechanisms of action. This method has the following advantages for the task of predicting the active ingredient group of traditional Chinese medicine:
[0042] (1) This method constructs and analyzes the biomolecular information interaction network of traditional Chinese medicine prescriptions to predict their active ingredient groups, allowing researchers to grasp the key points from a large number of drug ingredients, reduce the failure rate of experiments, improve the efficiency of screening active ingredients of traditional Chinese medicine, and contribute to the development of traditional Chinese medicine prescriptions.
[0043] (2) The molecular information network of traditional Chinese medicine prescriptions is analyzed through heterogeneous network embedding technology, the network mapped by molecular nodes is embedded, and the active ingredient group of traditional Chinese medicine is analyzed through clustering algorithm. It can utilize the multi-level and multi-dimensional information in the network and has a higher accuracy in the prediction task of the active ingredient group of traditional Chinese medicine. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 : The overall framework diagram of the HEPA model in the present invention. DETAILED DESCRIPTION
[0045] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with specific embodiments and illustrative drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention and are not used to limit the present invention.
[0046] The present invention provides a method for predicting active ingredient groups of traditional Chinese medicine prescriptions based on heterogeneous network embedding, which comprises the following steps executed in sequence:
[0047] Step 1. Construction of drug molecule information network: Collect the targets, pathways, and GO term information of molecules in traditional Chinese medicine prescriptions and Western medicine molecules from public databases. At the same time, collect the association between diseases and Western medicine molecules from public data and, based on the above data, construct a drug molecule information network; based on the constructed drug molecule information network, generate labels for Western medicine molecules according to the Western medicine molecule-disease relationship and the correlation between Western medicine molecules and diseases that can be treated by traditional Chinese medicine prescriptions, and use one-hot encoding to represent them.
[0048] Step 2. Embedding of drug molecule information network: Based on the drug molecule information network obtained in step 1, the drug molecule information network is a heterogeneous network. Through the heterogeneous network embedding learning method, the embedding representation of the molecules of the traditional Chinese medicine prescription and the related western medicine molecules of the diseases treated by the traditional Chinese medicine prescription is obtained;
[0049] Step 3. Cluster analysis to predict the active ingredient group of the Chinese medicine prescription: Based on the embedded representation obtained in step 2, a clustering algorithm is used to analyze the embedded representation of Chinese medicine molecules and Western medicine molecules, thereby obtaining multiple clusters. The basic information of each cluster, including the number of Chinese medicine molecules and the number of Western medicine molecules, is counted, and the Chinese medicine molecules that can be clustered with Western medicine molecules are taken as the active ingredient group;
[0050] Furthermore, step 1 is specifically as follows:
[0051] Step 1.1 Collection of TCM prescription data: First, for each TCM prescription, the chemical molecule information contained in the TCM prescription is collected through the database, including the name of the molecule, CAS number, and related targets. The drug molecules in the prescription are preliminarily screened according to the oral bioavailability (OB)>=30% and the drug-likeness (DL)>=0.18 to obtain the component molecule set of the TCM prescription to be studied; then, the target information of each component is retrieved through the molecular target database;
[0052] Step 1.2 Western medicine data collection: Collect the names of diseases that can be treated by Western medicine molecules and the target information related to the molecules through public databases;
[0053] Step 1.3 After obtaining the target information of traditional Chinese medicine prescriptions and molecular targets of Western medicine, obtain the target-related KEGG pathway and GO term information from the public database through web crawler technology;
[0054] Step 1.4 constructs a drug molecule information network based on the collected data, whose nodes are diseases, prescriptions, Chinese medicine molecules, Western medicine molecules, targets, pathways, and GO terms, and whose edges are prescription-Chinese medicine molecules, molecules-targets, targets-pathways, targets-GOterms, prescriptions-diseases, and Western medicine molecules-diseases;
[0055] Step 1.5 generates labels for Western medicine molecules based on the constructed drug molecule information network, according to the Western medicine molecule-disease relationship and the correlation between the diseases that can be treated by Western medicine molecules and traditional Chinese medicine prescriptions.
[0056] Furthermore, step 2 is specifically as follows:
[0057] Step 2.1: for the drug molecule information network constructed in step 1, generate a one-hot encoding for all nodes in the network as the initial features of the nodes in the network;
[0058] Step 2.2 Heterogeneous mutual attention calculation:
[0059] For the drug molecule information network constructed in step 1, for a node t, its neighbor nodes s∈N(t) may belong to different node types, and their mutual attention needs to be calculated according to the type of their edges. For each edge e=(s,t) in the network, calculate the h-head attention. The specific calculation process is as follows:
[0060]
[0061]
[0062]
[0063]
[0064] in, h is the number of attention heads, is the vector dimension of each attention head, τ(*) represents mapping node * to its category set, φ(e) maps edge e to the category set of the edge, H l-1 [*] indicates the representation of node * at layer l-1;
[0065] Step 2.3 Heterogeneous Information Propagation
[0066] For the drug molecule information network constructed in step 1, message propagation is performed from the source node to the target node, and the initial message uses the one-hot encoding generated for all nodes in step 2.1. For the edge e = (s, t), the multi-head message can be calculated as:
[0067]
[0068]
[0069] in h is the number of attention heads, is the vector dimension of each attention head.
[0070] Step 2.4 Aggregation of heterogeneous information
[0071] For the drug molecule information network constructed in step 1, after the calculation of heterogeneous mutual attention in step 2.2 and heterogeneous information in step 2.3 is completed, it is necessary to aggregate the heterogeneous messages of all neighboring nodes s∈N(t) for the target node t, so as to obtain the intrinsic connection between the nodes of the drug molecule heterogeneous information network. The calculation method of heterogeneous message aggregation is as follows:
[0072]
[0073] get In order to obtain the node representation of layer l, it is necessary to use the linear mapping A-Linear according to the category of the target node. τ(t) Mapped to the distribution of the corresponding category, the representation of the node can be calculated as:
[0074]
[0075] Repeating the process from step 2.2 to step 2.4 L times, we can obtain a rich context representation H for each node of the drug molecule information network constructed in step 1. (L) .
[0076] Step 2.5 Heterogeneous Network Embedding Learning
[0077] For the drug molecule information network constructed in step 1, step 1.5 generates labels for western medicine molecules based on the constructed drug molecule information network, according to the relationship between western medicine molecules and diseases, and the correlation between diseases that can be treated by western medicine molecules and traditional Chinese medicine prescriptions. At the same time, step 2.4 generates node representations. Since western medicine molecules and traditional Chinese medicine molecules are nodes of the same type in the network, this method designs a classification task for western medicine molecule nodes to update the parameters in the network, thereby obtaining the final embedding results of all nodes in the heterogeneous network. Specifically, for the node representation H obtained in step 2.4 (L) , the representation of the molecular nodes of western medicine is used as the input of the MLP classifier, and the output corresponds to the molecular label generated in step 1.5. The western medicine molecules are randomly divided into 80% training set and 20% validation set. The loss function uses the cross entropy loss function, Adam is used as the optimizer, and then the gradient of the prediction loss is used to update the model parameters, thereby obtaining the network embedding of all molecular nodes in the drug molecule information network.
[0078] The further step 3 is as follows:
[0079] Step 3.1 For each molecular node obtained in step 2, the network embedding of each component molecule of the TCM prescription to be tested and the network embedding of the Western medicine molecule for the related diseases treated by the network TCM prescription are combined to obtain the molecular network embedding set related to the TCM prescription.
[0080] Step 3.2: For the molecular network embedding set related to the Chinese medicine prescription obtained in step 3.1, the spectral clustering algorithm is used to perform cluster analysis on the molecular network embedding set related to the prescription, and then the k value of the cluster is determined by the silhouette coefficient method. The silhouette coefficient is an evaluation index used to evaluate the clustering effect, which takes into account the two factors of cohesion and separation in the classification. The best k value can be determined by looking at the silhouette coefficient of the spectral clustering algorithm at different k values. For different k values, the spectral clustering algorithm is used to cluster the relevant data of each Chinese medicine prescription, and the mean of the silhouette coefficient of all sample points is calculated. The larger the mean of the silhouette coefficient, the better the cohesion and separation of the cluster. The k value that makes the coefficient larger is selected as the parameter of the spectral clustering algorithm to obtain the final cluster cluster. After obtaining the final cluster cluster for each Chinese medicine prescription, the basic information of each cluster cluster, including the number of Chinese medicine molecules and the number of Western medicine molecules, is counted, and the Chinese medicine molecules that can be clustered with most Western medicine molecules are used as the prediction results of the active ingredient group.
[0081] This method uses five traditional Chinese medicine prescriptions for the treatment of cardiovascular diseases as examples to verify the effectiveness of the proposed traditional Chinese medicine prescription active ingredient group prediction method. Data collection includes the following two aspects:
[0082] (1) Data collection of Chinese medicine prescriptions: This paper takes five classic prescriptions for the treatment of cardiovascular diseases as the research objects. The basic information of Chinese medicine prescriptions is shown in Table 1. In terms of molecular targets, a large number of databases on compounds and their potential targets have been accumulated in the biomedical field, including the CTD database. First, for each Chinese medicine compound, the chemical molecular information contained in the Chinese medicine compound is collected through the TCMSP database, including the name of the molecule, CAS number, and related targets. The drug molecules in the prescription are preliminarily screened according to the oral bioavailability (OB)>=30 and the drug-like property (DL)>=0.18, and the component set of the Chinese medicine prescription to be studied in this paper is obtained.
[0083] (2) Western medicine data collection: This method uses two sources of Western medicine data. One is the molecular library for the treatment of cardiovascular diseases, CVDSP, which contains information on more than 250 Western medicine molecules that can treat cardiovascular diseases, including diseases that Western medicine molecules can treat, molecular-related targets, and other information. The other is DeepDR, a public data set used to predict the relationship between drugs and diseases, which includes more than 1,000 types of drug molecules and disease association information, as well as their target information. Duplicate molecules and molecular information without targets in the two data sources were deleted, and the Western medicine molecular information from the two data sources was integrated.
[0084] After obtaining the target information of traditional Chinese medicine prescriptions and western medicine molecular targets, the Metascape platform was used to obtain the target-related KEGG pathway and GO term information. The complete data set information is shown in Table 2. Finally, the following Figure 1 The drug molecule information network shown.
[0085] Table 1 Prescription-related diseases and molecular information
[0086]
[0087] Table 2 Basic information of drug molecule information network
[0088]
[0089]
[0090] For the above-constructed data set, after constructing the drug molecule information network, this method is used to obtain the network embedding vectors of all molecular nodes. The network embedding of each component molecule of the Chinese medicine prescription to be tested and the network embedding of the Western medicine molecule of the related diseases treated by the network Chinese medicine prescription are combined to obtain the molecular network embedding set related to the Chinese medicine prescription.
[0091] For each prescription, the spectral clustering algorithm is used to perform cluster analysis on the molecular network embedding set related to the prescription, and then the k value of clustering is determined by the silhouette coefficient method. The silhouette coefficient is an evaluation index used to evaluate the clustering effect, which takes into account the two factors of cohesion and separation in classification. The best k value can be determined by looking at the silhouette coefficient of the spectral clustering algorithm at different k values. The silhouette coefficients of the clustering results with k ranging from 2 to 10 are calculated in turn, and it is found that the silhouette coefficients of these prescriptions are all maximized when k is equal to 2, so the k value selected for the clustering of these prescriptions is 2. The cluster information of the prescriptions is shown in Table 3.
[0092] Table 3 Prescription cluster information
[0093]
[0094] It can be seen from Table 3 that most of the relevant western medicine data for each compound can be clustered into a cluster. According to the previous analysis, the Chinese medicine molecules that can be clustered with these western medicine molecules can be regarded as the predicted results of the component group of a compound. The literature was consulted for the molecules that can be clustered with most of the western medicine molecules in the five compounds, and the statistical information was shown in Table 5. Taking the compound Danshen prescription as an example, Table 4 is the Chinese medicine molecule information table in its cluster 2, from which it can be seen that only a small number of components have not been found to be related to cardiovascular disease. At the same time, it can be known from the literature that the main active ingredient groups in the compound Danshen prescription are tanshinone, ginsenosides, and protocatechuic aldehyde. These ingredients are included in the prediction results of this article. At the same time, quercetin, luteolin, licorice and other cardiovascular-related molecules are also included in the prediction results.
[0095] Table 4 Cluster 2 information table of Fufang Danshen prescription
[0096]
[0097]
[0098] Table 5 Prescription cluster information table
[0099]
Claims
1. A method for predicting active ingredient groups in traditional Chinese medicine prescriptions based on heterogeneous network embedding, characterized in that: The steps are as follows: Step 1. Construction of drug molecule information network: Collect the target, pathway, and GO term information of molecules of traditional Chinese medicine prescriptions and western medicine molecules from public databases, and collect the association between diseases and western medicine molecules from public data and, based on the above data, construct a drug molecule information network; based on the constructed drug molecule information network, generate labels of western medicine molecules according to the relationship between western medicine molecules and diseases that can be treated by western medicine molecules and traditional Chinese medicine prescriptions, and use unique hot coding to represent them; the step 1 specifically includes the following steps: Step 1.1: Data collection of Chinese medicine prescriptions: First, for each Chinese medicine prescription, the chemical molecule information contained in the Chinese medicine prescription is collected through the database, including the name of the molecule, CAS number, and related targets; the drug molecules in the prescription are preliminarily screened according to the oral bioavailability OB>=30% and the drug-like property DL>=0.18 to obtain the component molecule set of the Chinese medicine prescription to be studied; then, the target information of each component is retrieved through the molecular target database; Step 1.2: Western medicine data collection: Collect the names of diseases that can be treated by Western medicine molecules and the target information related to the molecules through public databases; Step 1.3: After obtaining the target information of traditional Chinese medicine prescriptions and molecular targets of Western medicine, obtain the target-related KEGG pathway and GO term information from the public database through web crawler technology; Step 1.4: construct a drug molecule information network based on the collected data, wherein the nodes of the drug molecule information network are diseases, prescriptions, Chinese medicine molecules, Western medicine molecules, targets, pathways, and GO terms, and the edges of the drug molecule information network are prescription-Chinese medicine molecules, molecules-targets, targets-pathways, targets-GO terms, prescriptions-diseases, and Western medicine molecules-diseases; Step 1.5: Based on the constructed drug molecule information network, according to the relationship between western medicine molecules and diseases, and the correlation between the diseases that can be treated by western medicine molecules and traditional Chinese medicine prescriptions, generate labels for western medicine molecules; Step 2. Embedding of drug molecule information network: Based on the drug molecule information network obtained in step 1, the drug molecule information network is a heterogeneous network. Through the heterogeneous network embedding learning method, the embedding representation of the molecules of the traditional Chinese medicine prescription and the related western medicine molecules of the diseases treated by the traditional Chinese medicine prescription is obtained; the step 2 specifically includes the following steps: Step 2.1: For the drug molecule information network constructed in step 1, generate a one-hot encoding for all nodes in the network as the initial features of the nodes in the network; Step 2.2: Heterogeneous mutual attention calculation: For the drug molecule information network constructed in step 1, for a node t, its neighbor nodes s∈N(t) may belong to different node types, and their mutual attention needs to be calculated according to the type of their edges; for each edge e=(s,t) in the network, calculate the h-head attention, and the specific calculation process is as follows: in, h is the number of attention heads, is the vector dimension of each attention head, τ(*) represents mapping node * to its category set, φ(e) maps edge e to its category set, H l-1 [*] indicates the representation of node * at layer l-1; Step 2.3: Heterogeneous Information Propagation For the drug molecule information network constructed in step 1, message propagation is performed from the source node to the target node. The initial message uses the one-hot encoding generated for all nodes in step 2.1; for the edge e = (s, t), the multi-head message is calculated as: in h is the number of attention heads, is the vector dimension of each attention head; Step 2.4: Aggregation of heterogeneous information For the drug molecule information network constructed in step 1, after the calculation of heterogeneous mutual attention in step 2.2 and heterogeneous information in step 2.3 is completed, it is necessary to aggregate the heterogeneous messages of all neighboring nodes s∈N(t) for the target node t, so as to obtain the intrinsic connection between the nodes of the drug molecule heterogeneous information network. The calculation method of heterogeneous message aggregation is as follows: get In order to obtain the node representation of layer l, it is necessary to use the linear mapping A-Linear according to the category of the target node. τ(t) Mapped to the distribution of the corresponding category, the representation of the node is calculated as: Repeat the process from step 2.2 to step 2.4 L times, and obtain a rich context representation H for each node of the drug molecule information network constructed in step 1. (L) ; Step 2.5: Heterogeneous Network Embedding Learning For the drug molecule information network constructed in step 1, step 1.5 generates labels for western medicine molecules based on the constructed drug molecule information network according to the relationship between western medicine molecules and diseases and the correlation between diseases that can be treated by western medicine molecules and traditional Chinese medicine prescriptions. At the same time, step 2.4 generates node representations. Since western medicine molecules and traditional Chinese medicine molecules are nodes of the same type in the network, this method designs a classification task for western medicine molecule nodes to update the parameters in the network, thereby obtaining the final embedding results of all nodes in the heterogeneous network. Specifically, for the node representation H obtained in step 2.4 (L) , the representation of the molecular nodes of western medicine is used as the input of the MLP classifier, and the output corresponds to the molecular label generated in step 1.5; the western medicine molecules are randomly divided into 80% training set and 20% validation set, the loss function uses the cross entropy loss function, Adam is used as the optimizer, and then the gradient of the prediction loss is used to update the model parameters, thereby obtaining the network embedding of all molecular nodes in the drug molecule information network; Step 3. Cluster analysis to predict the active ingredient group of Chinese medicine prescriptions: Based on the embedded representation obtained in step 2, a clustering algorithm is used to analyze the embedded representation of Chinese medicine molecules and Western medicine molecules, thereby obtaining multiple clusters, and the basic information of each cluster is counted. The basic information of the cluster includes the number of Chinese medicine molecules and the number of Western medicine molecules. The Chinese medicine molecules that can be clustered with Western medicine molecules are regarded as the active ingredient group.
2. The method for predicting active ingredient groups of traditional Chinese medicine prescriptions based on heterogeneous network embedding according to claim 1, characterized in that: The step 3 specifically comprises the following steps: Step 3.1: For each molecular node network embedding obtained in step 2, the network embedding of each component molecule of the TCM prescription to be tested and the network embedding of the Western medicine molecule of the related disease treated by the network TCM prescription are combined to obtain the molecular network embedding set related to the TCM prescription; Step 3.2: For the set of molecular network embeddings related to the Chinese medicine prescriptions obtained in step 3.1, the spectral clustering algorithm is used to perform cluster analysis on the molecular network embeddings related to the prescriptions, and then the k value of the clustering is determined by the silhouette coefficient method; the silhouette coefficient is an evaluation index used to evaluate the clustering effect, which takes into account the two factors of cohesion and separation in classification; the best k value is determined by looking at the silhouette coefficient of the spectral clustering algorithm under different k values; for different k values, the spectral clustering algorithm is used to cluster the relevant data of each Chinese medicine prescription, and the mean of the silhouette coefficients of all sample points is calculated. The larger the mean of the silhouette coefficient, the better the cohesion and separation of the clustering. The k value that makes the coefficient larger is selected as the parameter of the spectral clustering algorithm to obtain the final clustering cluster; after obtaining the final clustering cluster for each Chinese medicine prescription, the basic information of each clustering cluster, including the number of Chinese medicine molecules and the number of Western medicine molecules, is counted, and the Chinese medicine molecules that can be clustered with most of the Western medicine molecules are used as the prediction results of the active ingredient group.
