A method for predicting ingredient-target interactions based on web crawlers and multimodal features
The data of multiple Chinese medicine databases were obtained through network crawlers, multimodal features were fused and reliable negative sample sets were constructed, which solved the problem of incomplete data and single characteristics in the prediction of Chinese medicine component-target interactions, and improved the accuracy and practicality of the prediction. value.
Patent Information
- Application Number
- CN202210703362.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-06-21
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2042-06-21
AI Technical Summary
The existing traditional Chinese medicine component-target interaction prediction methods have problems of incomplete data and single characteristics, resulting in high false negative risk of prediction results and failure to effectively utilize data from multiple databases.
Using prediction methods based on network crawlers and multimodal features, the reliable negative sample set is constructed by automatically obtaining data related to traditional Chinese medicine ingredients and targets, and the descriptor characteristics, sequence similarity characteristics and network topological characteristics of components and targets are integrated, and the reliable negative sample set is constructed. Use extreme gradient boost classifiers to train binary classification models.
The integration of data integrity and multimodal features is achieved, the risk of a single feature is reduced, and a reliable negative sample set is provided, which improves the accuracy and practical value of predictions.
Smart Images

Figure CN114944191B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence technology for traditional Chinese medicine, and in particular to a component-target interaction prediction method based on a web crawler and multimodal features. Background Art
[0002] Traditional Chinese medicine has rich clinical experience and significant efficacy, but the mechanism of action of Chinese medicine components and therapeutic targets is still poorly understood, which poses a great challenge to clinical precision treatment. Explaining the material basis of Chinese medicine efficacy from a microscopic mechanism has become a key research content of TCM modernization. Therefore, it is crucial to identify ingredient-target interaction (ITI) in understanding the mechanism of action of TCM. Traditional Chinese medicine has the characteristics of multiple components and multiple targets. The relationship between many potential Chinese medicines and targets has not yet been clarified. The time cost and economic cost of studying the action targets of each component of traditional Chinese medicine through biological experiments are huge and difficult to achieve. Therefore, studying the prediction method of ingredient-target interaction is an urgent issue for the modernization of traditional Chinese medicine.
[0003] With the emergence of public TCM analysis databases, component target databases such as TCMSP, SymMap, HERB, ETCM, and TCMID have emerged in recent years. These resources provide data support for the prediction of ITI. In recent years, methods based on chemical genomics combined with computer technology can quickly and cheaply identify potential ITIs. In current research methods, chemical genomics-based methods can be divided into three categories: network diffusion-based, matrix decomposition-based, and classification-based methods. The network diffusion-based method regards components and targets as nodes, and the relationship between components and targets as edges. At the same time, it adds similarity relationships between components and targets, and constructs a network. It uses network propagation methods such as random walks on the network to predict unknown component-target interactions. The matrix decomposition-based method uses matrix decomposition to decompose the component-target association matrix into two low-rank matrices corresponding to the feature space of components and targets.
[0004] In classification-based methods, the characteristics of component-target are generally constructed first, and then the machine learning method is used to predict the component-target interaction relationship. The extraction of features in this method determines the quality of the prediction results. At present, the selection of features is diverse. Some use molecular descriptors and protein descriptors as features; some calculate the similarity between component-component and target-target as features; some construct heterogeneous networks and use the network structure of nodes as features. Although there has been some progress in the study of component-target interaction, there are still the following problems: 1. The existing data collection is incomplete and multiple databases have not been comprehensively considered; 2. The classification prediction of component-target interaction is a two-classification problem, with only labeled positive samples and unlabeled samples. At present, research often randomly selects the same number of negative samples as positive samples from unlabeled samples, and the risk of false negative prediction results is high. 3. The current classification-based method features are not fully considered and selected in a single way, which limits the development of effective calculation methods. Summary of the invention
[0005] In order to solve the technical limitations of the existing component-target interaction prediction methods caused by incomplete data and single features, the present invention provides a component-target interaction prediction method based on web crawlers and multimodal features. Compared with the existing component-target prediction methods, this method obtains more complete data, greatly mines the potential features of Chinese medicine components and targets, reduces the risk of single features, and provides a reliable negative sample set, which has high practical value. In order to solve the above problems, its technical solution is as follows:
[0006] The present invention provides a component-target interaction prediction method based on a web crawler and multimodal features, the prediction method comprising the following steps:
[0007] S1, data collection module: using web crawlers to automatically obtain data related to the composition and targets of traditional Chinese medicine;
[0008] S2, data preprocessing module: integrate component and target related data, construct component table, target table, component-target relationship table;
[0009] S3, multimodal feature extraction module: cross-modally fuse the descriptor features, sequence similarity features, and network topology features of components and targets to obtain component and target features;
[0010] S4, reliable negative sample construction module: uses PU learning and known component-target interaction relationships to construct a reliable negative sample set;
[0011] S5, classification model building module: component-target features combined with positive and negative sample sets to train a binary classification model.
[0012] Furthermore, in the data collection module described in step S1, web crawler technology is used to automatically obtain descriptors, sequences and interaction data of components and targets from multiple Chinese medicine chemical component databases and medicinal chemistry databases.
[0013] Furthermore, in the data preprocessing module described in step S2, corresponding fusion rules are formulated for the Chinese medicine ingredients, targets, interaction relationships and corresponding descriptor information appearing in different databases to construct ingredient tables, target tables and ingredient-target tables.
[0014] Furthermore, the multimodal feature extraction module described in step S3 further includes:
[0015] S31, calculates the molecular descriptor features of the ingredients based on the pharmacological and molecular property data of the ingredients; calculates the protein descriptor features of the targets based on the dipeptide composition, pseudoamino acid composition, and autocorrelation of the targets;
[0016] S32, SMILES sequence similarity features of components were calculated using the Jaccard similarity coefficient, and FASTA sequence similarity features of targets were calculated using the Smith–Waterman score;
[0017] S33, constructing a heterogeneous network based on known component-target interaction relationships, and using the node2vec network embedding algorithm to extract the network topological features of components and targets;
[0018] S34, the descriptor features, sequence similarity features and network topology features of the components and targets were concatenated and then the principal component analysis method was used for front-end cross-modal feature fusion.
[0019] Furthermore, in step S31, the pharmacological and molecular property data of the ingredients include molecular weight, lipid-water distribution coefficient, hydrogen bond donor, hydrogen bond receptor, oral bioavailability, intestinal epithelial permeability, blood-brain barrier, drug-like properties, negative surface area, compound polarity, the number of rotatable bonds in the compound, and drug half-life. The above 12 data are normalized as molecular descriptor features.
[0020] In step S31, the protein descriptor features of the target include five feature classes, namely, the deviation of dipeptides from the expected mean (DDE), pseudo amino acid composition (PAAC), autocorrelation (Moran), composition of spaced amino acid pairs (CKSAAP), and grouped tripeptide composition (GTPC). After concatenation, the dimensions are reduced to d dimensions by principal component analysis and used as the target descriptor feature vector
[0021] Furthermore, in step S32, the calculation formula of the SMILES sequence similarity feature of the component is:
[0022]
[0023] In the formula, c i and c j is the component Morgan molecular fingerprint vector calculated based on the SMILES sequence of each component, |c i ∩c j | indicates component c i and c j The number of atoms in the largest common subgraph between |c i ∪c j | means c i and c j The sum of the quantities minus c i and c j The number of atoms in the largest common subgraph between i and c j The union of i The similarity with all components is reduced to d dimensions through principal component analysis, and is called the component similarity vector
[0024] In step S32, the calculation formula of the FASTA sequence similarity feature of the target is:
[0025]
[0026] In the formula, SW(t i , t j ) represents the Smith-Waterman score of the original two targets, and t j The similarity with all targets is reduced to d dimensions through principal component analysis, which is the target similarity vector
[0027] Furthermore, in step S33, the method for calculating the network topology characteristics of the components and targets includes: using the node2vec method to generate a d-dimensional feature vector of each component and target node and The details are as follows:
[0028] S331, given the ITI matrix Y, N c Indicates the total number of components, N t represents the total number of targets, Y(i, j) represents the interaction relationship between the i-th component and the j-th target, Y(i, j) = 1 means that there is an interaction relationship between the i-th component and the j-th target, otherwise Y(i, j) = 0;
[0029] S332, construct an undirected and unweighted graph G=(V, E) according to the matrix Y, where V represents a set of nodes, |V| represents the number of nodes, and E represents a set of edges. |E| represents the number of edges; when Y(i, j) = 1, it means there is e ij Make v i With v j connected;
[0030] S333, perform a random walk on the graph G. Taking node c0=u as an example, perform random walk sampling on the neighborhood of node c0. Assuming that the current walk is from node t to node v, the probability calculation formula for walking to the next node x is as follows:
[0031]
[0032] Where Z is the normalization constant, π vx =α pq (t, x) is the unnormalized transition probability between node v and node x, α pq The calculation formula of (t, x) is as follows:
[0033]
[0034] Among them, d tx represents the shortest distance from node t to node x, so the preference of random walk is controlled by parameters p and q to generate a random walk sequence of node u
[0035] S334, after obtaining the sequence of each node by random walk, the learning goal is to use a low-dimensional vector to represent the node u and maximize the sequence N s The probability of neighbor nodes appearing in (u), the optimization function is as follows:
[0036]
[0037] Where f(u) is the final node embedding representation. This method is used to train each node to obtain a d-dimensional feature vector and
[0038] Furthermore, in the reliable negative sample construction module described in step S4, part of the data is randomly selected as spy samples using the known component-target interaction relationship, and the predicted probabilities of all samples are generated and sorted using random forest, and those with lower probabilities than the spy samples are taken as a reliable negative sample set.
[0039] Furthermore, in the classification model construction module described in step S5, the component and target feature vectors obtained in S3 are concatenated according to the combination of positive and negative sample sets to obtain component-target feature vectors, and an extreme gradient boosting classifier is used to train a binary classification model. The trained model can be used for component-target interaction prediction.
[0040] The component-target interaction prediction method based on web crawlers and multimodal features provided by the present invention has the following beneficial effects:
[0041] 1. The prediction method provided by the present invention collects data from multiple Chinese medicine chemical component databases and medicinal chemistry databases through a web crawler, and integrates the component table, target table and their relationship table to ensure the integrity of the data.
[0042] Second, a single feature may have the problem of missing features or insufficient representation capabilities. The multimodal feature fusion method based on descriptor features, sequence similarity features, and network topology features of the present invention can effectively solve this problem.
[0043] 3. In the component-target interaction prediction problem, there are only positive samples and unknown samples with interactions, and there are no negative samples that have been experimentally verified. Existing methods usually randomly select unknown samples equal to the positive samples as negative samples, and the risk of false positives is high. The technology based on PU learning in the present invention can screen out reliable negative samples from unknown samples. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0045] Figure 1 It is a schematic flow chart of the component-target interaction prediction method based on web crawler and multimodal features of the present invention;
[0046] Figure 2 This is a flowchart of obtaining reliable negative samples in the reliable negative sample module of the component-target interaction prediction method based on web crawlers and multimodal features of the present invention, taking the spy technology of PU learning as an example. Wherein U represents the unknown sample set, P represents the positive sample set, S represents the spy sample set, RF is the random forest classification model, and N represents the reliable negative sample set. DETAILED DESCRIPTION
[0047] In order to enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention and to make the above-mentioned objects, features and advantages of the present invention more obvious and understandable, the specific implementation modes of the present invention are further described below in conjunction with the accompanying drawings.
[0048] It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention. In addition, the technical features involved in each embodiment of the present invention described below can be combined with each other as long as there is no conflict between them.
[0049] Please refer to Figure 1 and Figure 2 , a component-target interaction prediction method based on a web crawler and multimodal features of this embodiment, the prediction method comprises the following steps:
[0050] S1, data collection module: using web crawlers to automatically obtain data related to the composition and targets of traditional Chinese medicine;
[0051] S2, data preprocessing module: integrate component and target related data, construct component table, target table, component-target relationship table;
[0052] S3, multimodal feature extraction module: cross-modally fuse the descriptor features, sequence similarity features, and network topology features of components and targets to obtain component and target features;
[0053] S4, reliable negative sample construction module: uses PU learning and known component-target interaction relationships to construct a reliable negative sample set;
[0054] S5, classification model building module: component-target features combined with positive and negative sample sets to train a binary classification model.
[0055] As a preferred embodiment, in the data collection module described in step S1, web crawler technology is used to automatically obtain descriptors, sequences and interaction data of components and targets from multiple Chinese medicine chemical component databases and medicinal chemistry databases.
[0056] Among them, web crawlers are characterized by fast speed and high efficiency, and can efficiently complete the structured data extraction of complex web page structures. Since the website structures of various databases are different, before data acquisition, it is necessary to conduct an in-depth analysis of the structure of the web page to find the corresponding data files.
[0057] The relevant databases crawled and their quantity statistics are shown in Table 1. The URL of the data in the selected table is queried, the data ID code is spliced and added to the URL or the data transmission interface is found through packet capture analysis, and the request library file is used to initiate a request to the server to obtain the corresponding HTML or JSON data file. The data file is parsed using xpath, selector, regular expressions, etc. to extract the specific data that is desired to be obtained from the web page, and finally all the data is obtained through multiple requests and URL iterations. For example, the fields that need to be extracted from the ingredient table of the TCMSP database include 17 fields such as ingredient name, molecular weight, lipid-water distribution coefficient, hydrogen bond donor, and acceptor. The URL rules of the ingredients in the database are analyzed to obtain the URL of each ingredient, and the corresponding web pages and fields are obtained and parsed through the web crawler, and then the new URL is passed in for the next round of data crawling until all ingredient data are crawled.
[0058] Table 1 Related databases
[0059]
[0060] Preferably, in the data preprocessing module described in step S2, corresponding fusion rules are formulated for the Chinese medicine ingredients, targets, interaction relationships and corresponding descriptor information appearing in different databases, and ingredient tables, target tables and ingredient-target tables are constructed.
[0061] Among them, after obtaining the data of the database through the web crawler, the component and target related data are fused to construct the component table, target table, and component-target relationship table. The crawled data is stored separately according to different databases and database tables. Since the fields and relationship data contained in different databases are different, rules need to be formulated to merge the data of each database to ensure the integrity of the data. Taking the fusion of TCMSP and HERB component tables as an example, after analyzing the component table fields of the two databases, when the component data in the two databases meet one of the four conditions of the same component name (not case-sensitive), the same PubChem_id, the same TCMSP_id, and the same CAS_id, it can be determined as the same component. At the same time, some field names are different but represent the same data, which requires manual identification processing. After screening out some ID fields that are not used in the experiment, all fields of the database are taken as the union in the fusion process. Table 2 shows the component table, target table, and component-target relationship fields after fusion.
[0062] Table 2 Field Settings
[0063]
[0064] Preferably, the multimodal feature extraction module described in step S3 further includes:
[0065] S31, calculates the molecular descriptor features of the ingredients based on the pharmacological and molecular property data of the ingredients; calculates the protein descriptor features of the targets based on the dipeptide composition, pseudoamino acid composition, and autocorrelation of the targets;
[0066] S32, SMILES sequence similarity features of components were calculated using the Jaccard similarity coefficient, and FASTA sequence similarity features of targets were calculated using the Smith–Waterman score;
[0067] S33, constructing a heterogeneous network based on known component-target interaction relationships, and using the node2vec network embedding algorithm to extract the network topological features of components and targets;
[0068] S34, the descriptor features, sequence similarity features and network topology features of the components and targets were concatenated and then the principal component analysis method was used for front-end cross-modal feature fusion.
[0069] Preferably, in step S31, the pharmacological and molecular property data of the ingredient include molecular weight, lipid-water distribution coefficient, hydrogen bond donor, hydrogen bond acceptor, oral bioavailability, intestinal epithelial permeability, blood-brain barrier, drug-like properties, negative surface area, compound polarity, the number of rotatable bonds in the compound and drug half-life. The above 12 data are normalized as molecular descriptor features.
[0070] In step S31, the iFeature tool is used to extract sequence protein descriptor features. The tool is commonly used for feature engineering analysis and modeling of targets, RNA and DNA sequences. It can use the properties of proteins expressed in FASTA sequences from many aspects to extract five feature classes, namely, the deviation of dipeptides from the expected average (DDE), pseudo amino acid composition (PAAC), autocorrelation (Moran), composition of spaced amino acid pairs (CKSAAP) and grouped tripeptide composition (GTPC). After splicing, the dimension is reduced to d dimensions through principal component analysis, which is the target descriptor feature vector.
[0071] Preferably, in step S32, the calculation formula of the SMILES sequence similarity feature of the component is:
[0072]
[0073] In the formula, c i and c j is the component Morgan molecular fingerprint vector calculated based on the SMILES sequence of each component, |c i ∩c j | indicates component c i and c jThe number of atoms in the largest common subgraph between |c i ∪c j | means c i and c j The sum of the quantities minus c i and c j The number of atoms in the largest common subgraph between i and c j The union of i The similarity with all components is reduced to d dimensions through principal component analysis, and is called the component similarity vector
[0074] In step S32, the calculation formula of the FASTA sequence similarity feature of the target is:
[0075]
[0076] In the formula, SW(t i , t j ) represents the Smith–Waterman score of the original two targets, and t j The similarity with all targets is reduced to d dimensions through principal component analysis, which is the target similarity vector
[0077] Preferably, in step S33, the method for calculating the network topology characteristics of the components and targets includes: using the node2vec method to generate a d-dimensional feature vector of each component and target node and The details are as follows:
[0078] S331, given the ITI matrix Y, N c Indicates the total number of components, N t represents the total number of targets, Y(i, j) represents the interaction relationship between the i-th component and the j-th target, Y(i, j) = 1 means that there is an interaction relationship between the i-th component and the j-th target, otherwise Y(i, j) = 0;
[0079] S332, construct an undirected and unweighted graph G=(V, E) according to the matrix Y, where V represents a set of nodes, |V| represents the number of nodes, and E represents a set of edges. |E| represents the number of edges; when Y(i, j) = 1, it means there is e ij Make v i With v j connected;
[0080] S333, perform a random walk on the graph G. Taking node c0=u as an example, perform random walk sampling on the neighborhood of node c0. Assuming that the current walk is from node t to node v, the probability calculation formula for walking to the next node x is as follows:
[0081]
[0082] Where Z is the normalization constant, π vx =α pq (t, x) is the unnormalized transition probability between node v and node x, α pq The calculation formula of (t, x) is as follows:
[0083]
[0084] Among them, d tx represents the shortest distance from node t to node x, so the preference of random walk is controlled by parameters p and q to generate a random walk sequence of node u
[0085] S334, after obtaining the sequence of each node by random walk, the learning goal is to use a low-dimensional vector to represent the node u and maximize the sequence N s The probability of neighbor nodes appearing in (u), the optimization function is as follows:
[0086]
[0087] Where f(u) is the final node embedding representation. This method is used to train each node to obtain a d-dimensional feature vector and
[0088] Preferably, in step S34, the descriptor features, sequence similarity features and network topology features of the components and targets are horizontally spliced and front-end cross-modal feature fusion is performed. The component-related features are combined to obtain Combining target-related features
[0089] Preferably, in the reliable negative sample construction module described in step S4, part of the data is randomly selected as spy samples using the known component-target interaction relationship, and the prediction probability of all samples is generated and sorted using random forest, and the samples with a probability lower than that of the spy samples are used as a reliable negative sample set. Since there are no experimentally verified negative samples in the prediction of component-target interaction, most people randomly select a part of samples with unknown relationships as negative samples, which has a high risk of false positives. Therefore, the present invention uses the spy technology in PU learning to screen potential component-target negative samples to construct a reliable negative sample set;
[0090] When Y in the component-target adjacency matrix Y ij =1, C i With T j Considered as a positive sample of a Chinese medicine ingredient-target pair, when Y ij =0, it is regarded as an unknown sample, and a positive sample set P and an unknown sample set U are constructed. In order to select reliable negative samples from the unknown sample set U, a spy technique is used to construct a reliable negative sample set N. First, 15% of the positive samples are randomly selected from P as set S and put into set U as spy samples (spy). At this time, the positive sample set and the unknown sample set become PS and U+S; secondly, PS and U+S are classified using the random forest algorithm, a classifier is trained, and the predicted probability of all samples is given; finally, the minimum value predicted in the spy sample is used as the threshold, and all samples in U below this threshold are considered to be reliable negative samples, and their Y ij Change to -1, and finally get the reliable negative sample set N. The feature vector β of the component-target is obtained from the positive sample set P and the reliable negative sample set N ij =[C i , T j ].
[0091] Preferably, in the classification model construction module described in step S5, the component and target feature vectors obtained in S3 are spliced according to the combination of positive and negative sample sets to obtain component-target feature vectors, and the extreme gradient boosting classifier is used to train the binary classification model, and the trained model can be used for component-target interaction prediction. The present invention adopts the ensemble learning algorithm extreme gradient boosting (eXtreme Gradient Boosting, XGBoost) model, the base learner of XGBoost adopts CART decision tree, and k trees are applied in parallel to form an ensemble tree model, and the optimization objective function is composed of a loss function and a regularization term, as shown in the following formula:
[0092]
[0093] Where k is the number of trees, T is the number of leaves, and w is the value of the node. Starting from the root node of the tree, the model uses a greedy algorithm to calculate β ij The benefit is based on each sample feature vector β ij Sort the training set data and select the β with the largest benefit ij As the split feature is mapped to the corresponding leaf node, the leaf node is generated recursively to build a decision tree, and then the weight of the decision tree leaf node is calculated by the optimization objective function. Each tree continuously fits the residual predicted by the previous tree and assigns different weights to each tree. After training, the sum of the prediction scores of all trees is node C i With T j The probability of having interactions;
[0094] The model was evaluated using the ten-fold crossover method and the area under the ROC curve (AUC) value, and the best model, model_best, was saved. The component-target information of the required prediction data was input into the model to determine whether there was an interaction relationship between the component and the target.
[0095] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.
[0096] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations made to these embodiments without departing from the principles and spirit of the present invention are still within the scope of protection of the present invention.
Claims
1. A component-target interaction prediction method based on web crawlers and multimodal features, characterized in that: The prediction method comprises the following steps: S1, data collection module: using web crawlers to automatically obtain data related to components and targets of traditional Chinese medicine, specifically: using web crawler technology to automatically obtain descriptors, sequences and interaction data of components and targets from multiple databases of chemical components of traditional Chinese medicine and medicinal chemistry databases; S2, data preprocessing module: fuse component and target related data, construct component table, target table, component-target relationship table, specifically: formulate corresponding fusion rules for the Chinese medicine components, targets, interaction relationships and corresponding descriptor information appearing in different databases, and construct component table, target table, component-target table; S3, multimodal feature extraction module: cross-modally fuse the descriptor features, sequence similarity features, and network topology features of the components and targets respectively to obtain component and target features, and: the multimodal feature extraction module also includes: S31, calculate the molecular descriptor features of the ingredients based on the pharmacological and molecular property data of the ingredients; calculate the protein descriptor features of the target based on the dipeptide composition, pseudoamino acid composition, and autocorrelation of the target; the pharmacological and molecular property data of the ingredients include molecular weight, lipid-water distribution coefficient, hydrogen bond donor, hydrogen bond receptor, oral bioavailability, intestinal epithelial permeability, blood-brain barrier, drug-like properties, negative surface area, compound polarity, the number of rotatable bonds in the compound, and drug half-life. The above 12 types of data are normalized as molecular descriptor features The protein descriptor features of the target include five feature classes: the deviation of dipeptides from the expected average, pseudo amino acid composition, autocorrelation, the composition of spaced amino acid pairs, and the grouped tripeptide composition. After splicing, the dimension is reduced to d dimensions through principal component analysis and used as the target descriptor feature vector ; S32, the SMILES sequence similarity characteristics of the components were calculated using the Jaccard similarity coefficient, and the FASTA sequence similarity characteristics of the targets were calculated using the Smith–Waterman score; the calculation formula for the SMILES sequence similarity characteristics of the components was: ; In the formula, and is the component Morgan molecular fingerprint vector calculated based on the SMILES sequence of each component, Indications and The number of atoms in the largest common subgraph between express and The sum of the quantities minus and The number of atoms in the largest common subgraph between and The union of The similarity with all components is reduced to d dimensions through principal component analysis, and is called the component similarity vector ; The calculation formula of the FASTA sequence similarity feature of the target is: ; In the formula, Represents the Smith–Waterman score of the original two targets, calculated The similarity with all targets is reduced to d dimensions through principal component analysis, which is the target similarity vector ; S33, constructing a heterogeneous network based on known component-target interaction relationships, and using the node2vec network embedding algorithm to extract the network topology features of the components and targets; wherein the method for calculating the network topology features of the components and targets includes: using the node2vec method to generate a d-dimensional feature vector for each component and target node and , as follows: S331, given the ITI matrix Y, , Indicates the total number of ingredients represents the total number of targets, represents the interaction relationship between the i-th component and the j-th target, indicates that there is an interaction relationship between the i-th component and the j-th target, otherwise ; S332, construct an undirected and unweighted graph based on matrix Y , Represents a collection of nodes. represents the number of nodes, E represents the set of edges, , represents the number of edges; when , indicating the existence Make and connected; S333, in Fig. Perform a random walk on the node For example, for the node The neighborhood of the node is randomly sampled. Walk to the node , then walk to the next node The probability calculation formula is as follows: ; in, is the normalization constant, For Node and nodes The unnormalized transition probability of The calculation formula is as follows: ; in, represents the shortest distance from node t to node x, so the parameter and Control the preference of random walk and generate the random walk sequence of node u ; S334, after obtaining the sequence of each node by random walk, the learning goal is to use a low-dimensional vector to represent the node u and maximize the sequence The probability of neighbor nodes appearing in the optimization function is as follows: ; in This method is used to train each node to obtain a d-dimensional feature vector for the final node embedding representation. and ; S34, the descriptor features, sequence similarity features and network topology features of the components and targets are concatenated and then the principal component analysis method is used for front-end cross-modal feature fusion; S4, reliable negative sample construction module: using PU learning and known component-target interaction relationships to construct a reliable negative sample set; wherein, in the reliable negative sample construction module, part of the data is randomly selected as spy samples using the known component-target interaction relationship, and the prediction probability of all samples is generated and sorted using random forest, and the samples with a lower probability than the spy samples are taken as the reliable negative sample set; S5, classification model building module: component-target features combined with positive and negative sample sets to train a binary classification model.
2. The component-target interaction prediction method according to claim 1, characterized in that: In the classification model building module described in step S5, the component and target feature vectors obtained in S3 are concatenated according to the combination of positive and negative sample sets to obtain component-target feature vectors, and an extreme gradient boosting classifier is used to train a binary classification model. The trained model can be used for component-target interaction prediction.
