A multi-source data efficient fusion drug target interaction prediction method

Abstract

The application provides a drug target interaction prediction method based on efficient fusion of multi-source data, first, multi-source data is fused or spliced by classifying drug and target multi-source data. For the original data, according to the biological characteristics of the drug or the target, several categories are divided respectively. When there are multiple networks in the category, the node feature is fused or spliced according to the characteristics of the network. Secondly, the network embedding method obtains high-quality and rich node low-dimensional feature representation. Finally, the forward unlabeled learning is used to predict the potential DTIs. According to the performance of the data, the data is classified, and the classified data is processed with different weights in the fusion process, so that the drugs and targets of different categories can be adaptively fused. Before the pharmaceutical biological test, the candidate drug target is learned by computer technology, so that the time and money cost of the biological control test are greatly reduced, and the pharmaceutical cycle is shortened.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of bioinformatics, and relates to a fusion, feature extraction and prediction method of drug targets. BACKGROUND

[0002] Drugs act on the target of disease to achieve pharmacological effect. Therefore, accurate identification of drug target interaction (DTI) is of great significance to understand the mechanism of drug treatment and discover new drugs for treating diseases. At present, the computational method combining drug and target multi-source data can effectively reduce the failure rate of pre-control test of biopharmaceutical process. This kind of method mainly uses machine learning framework to learn the characteristics of drugs and predict targets. In the aspect of multi-source data processing, it is usually easy to splice the features of different source data. Therefore, how to fully utilize the contribution of different source data to the efficient fusion of DTI prediction is the key to improve the prediction accuracy of DTI.

[0003] In the past few decades, a large number of computational methods for predicting DTI have been developed. In the computational strategy of DTI prediction, multiple sources of drug and target data are often considered. There are mainly two forms of traditional single-source data prediction of DTI, including ligand-based method and target-based method. Ligand-based target prediction refers to the analysis and comparison of the structural features of unknown target and known target drugs, and the prediction of potential target according to the similarity principle. The choice of similarity usually depends on the chemical structure or physical and chemical properties of ligand. However, the number of known target protein ligands in a single data source is insufficient, and the prediction result of ligand-based method may be unreliable. Similarly, target-based method needs to know the three-dimensional structure of the target, but it is very time-consuming to obtain the three-dimensional structure of protein at present, and some targets cannot obtain three-dimensional structure, which makes it difficult for target-based method to be used in the whole genome.

[0004] Currently, there are also studies that use multi-source data to establish a deepDTnet model to predict DTI, which uses a simple model to splice multiple network data equally. Specifically, DeepDTnet includes two key steps: (1) applying a deep neural network algorithm for network embedding to embed each vertex in the network into a low-dimensional vector space; (2) due to the lack of publicly available negative samples, a PU matrix completion algorithm is used, which is a vector space projection scheme, to predict new drug-target interactions. DeepDTnet embeds 15 types of chemical networks, genomic networks, phenotype networks, and cellular networks, and applies a deep neural network algorithm to learn a low-dimensional vector representation of the features of each node. After learning the feature matrix of drugs and targets, DeepDTnet applies PU matrix completion to find the best projection from the drug space to the target (protein) space, so that the projected feature vectors of the drugs are geometrically close to the feature vectors of their known interacting targets. Finally, DeepDTnet infers new targets for drugs based on the geometric closeness of the projected feature vectors of the drugs in the projection space.

[0005] Prediction of drug-target protein interactions is of great significance to improve the speed and accuracy of new drug discovery. Traditional experimental methods are very expensive and time-consuming in identifying new DTIs. High-efficiency computational methods provide stronger guidance for improving the efficiency and accuracy of DTI prediction. It provides accurate candidate drugs for drug discovery, further reducing the cost and time of drug discovery. Currently, a large number of studies have focused on DTI prediction and have made significant achievements.

[0006] With the increasing availability of public datasets, a large amount of multi-source data is used to study the properties of drugs and targets to predict DTIs. Multi-source data of drugs and targets not only contains inherent features of drugs and targets, but also contains network topology information based on other attributes such as drug side effects. Therefore, considering the topology and feature information of multi-source data, exploring the class network weight of integrated networks, and how to effectively fuse multi-source data are challenges for accurate identification of DTIs. SUMMARY

[0007] The present application provides a method for predicting drug-target interactions based on efficient fusion of multi-source data, comprising the following steps:

[0008] Step one: classify the topological structure data and feature attribute data of drugs and targets according to biological characteristics, respectively, combine the classified networks of drugs and targets, and input them into the subsequent positive unlabeled prediction model to obtain different combination prediction result performance indicators;

[0009] Step two: analyze the network combination test index, respectively, the basic fusion or splicing of drug and target multi-source data, including network screening, network fusion, entropy weighted fusion and reverse softmax weighted fusion; when there are multiple networks in the category, the splicing and fusion of the class network containing multiple networks are compared and analyzed, and the node features are selected or spliced according to the characteristics of the network;

[0010] Step three: input the fused or spliced drug and target network into the network embedding model respectively, and calculate the low-dimensional feature representation of drug and target, that is, the feature vector extracted by the network;

[0011] Step four: input the low-dimensional feature representation of drug and target into the positive unlabeled prediction model to obtain the prediction performance index and the predicted drug target action score.

[0012] Further, the basic fusion method in step two is similarity network fusion, and the drug network fusion steps are as follows:

[0013] (1) Calculate the similarity matrix of the heterogeneous association matrix based on Jaccard coefficient;

[0014] (2) Calculate the edge weight matrix, normalized matrix and local affinity matrix of each drug network;

[0015] (3) Update each network t times;

[0016] (4) After t iterations, calculate the fusion matrix.

[0017] Further, the step two realizes four improved methods on the basis of the basic fusion method-similarity network fusion, as follows:

[0018] (1) Selection fusion, according to the network combination test performance of step one, the class network of drug and target is screened, and then the screened network is fused based on the basic fusion method, so as to input the positive unlabeled prediction model;

[0019] (2) Entropy-based weighted fusion, before fusing the drug and target network respectively, calculate the entropy value of each class network to be fused as the network weight of the input fusion method;

[0020] (3) Weighted fusion based on prediction performance, based on one of the combination test performance indexes of step one, AUROC, the softmax value of the class network of drug and target is calculated, and each class network to be fused is weighted by sofamax and input into the similarity network fusion model;

[0021] (4) Selection fusion, combining the fusion strategies of the above (2) and (3) from two angles, so that the performance of the prediction model is better.

[0022] Further, the matrix calculation in step (1) is as follows:

[0023] Matrix element W l (d i , d j ) is defined as follows:

[0024]

[0025] Where L represents the number of drug similarity networks, M l represents the adjacency matrix of drug similarity network l (l = 1, 2…L), p 2 (d i , d j ) represents the Euclidean distance between drugs d i and d j , d i is the similarity vector between the ith drug and all other drugs, and μ is a hyperparameter, and ε ij is described as follows:

[0026]

[0027] Where N i = {d j | M l (d i , d j ) > 0} represents the neighbors of drug d i in network 1, and mean (p (d i , N i )) is the average distance between d i and its neighbors.

[0028] All data types are normalized by calculating the matrix, and the normalized matrix E l (d i , d j ) is described as follows:

[0029]

[0030] The local affinity S l (d i , d j ) is measured using K nearest neighbors (KNN) as follows:

[0031]

[0032] Where represents the K neighbors of drug d i .

[0033] Further, the average value of matrix E l after t iterations in step (3) is obtained, and the state matrix E(G) As follows:

[0034]

[0035] Further, the entropy-weighted fusion in step two, first calculates the entropy of each network, for any matrix M l , the entropy of node d i is defined as follows:

[0036]

[0037] Where m ij represents an element of matrix M l . Then the average entropy of all rows is obtained as:

[0038]

[0039] Finally, update M l with entropy as weight as follows:

[0040] EW l = M l x E l .

[0041] Further, the reverse softmax-weighted fusion in step two, for using the combined result data to calculate the softmax function.

[0042] C d is the number of drug classes and C t is the number of target classes, indicates one of the index values AUROC of the drug class c i and the target class c j of the combined input prediction model, and the softmax value of each type of drug data is calculated as follows:

[0043]

[0044] Where c i = 1, 2, … C d , c j = 1, 2, … C t and

[0045] Then take the average of all class drugs SoftMax value AD(i), and then use the SoftMax function to normalize the drug value, the specific case is shown as follows:

[0046]

[0047] Further, the network embedding model used in step three for dimensionality reduction of the highly sparse matrix is as follows:

[0048] (1) Random surfing to obtain information of each node;

[0049] (2) Calculate the positive point-wise mutual information matrix;

[0050] (3) Stack the denoising autoencoder model, and each stack is an autoencoder to extract network feature representation.

[0051] Further, after the data fusion processing in step three, the feature vector of the fused network, i.e., the low-dimensional feature representation of the drug and the target, is extracted, and the steps are as follows:

[0052] First, calculate the information of each node in the network:

[0053] p k =a·p k-1 Tr+(1-α)p0,

[0054] Where p is a vector, and Tr is the transition probability.

[0055] Second, calculate the positive point-wise mutual information of all nodes as follows:

[0056]

[0057] Where Pco is the probability co-occurrence matrix obtained in the first step;

[0058] Third, stack the denoising autoencoder to extract the feature target function:

[0059]

[0060] Where WE y is the weight matrix.

[0061] Further, the optimization problem of the positive unlabeled prediction model in step four is parameterized as:

[0062]

[0063] Where r ij =1 represents the known drug-target interaction; Ω = Ω + + Ω - represents all training samples, Ω + is the positive sample, Ω - is the unobserved sample; Z = U × V T is a low-rank matrix.

[0064] Compared with the prior art, the present application has the following advantages:

[0065] The present application proposes a framework consisting of the following three main components. First, by classifying the multi-source data of drugs and targets, the multi-source data can be fused (including selective fusion and weighted fusion) or spliced. For the original data, it includes topological graphs (such as drug-drug interaction, drug side effect and drug disease correlation network, target-target interaction and target disease correlation network) and feature graphs (such as drug similarity network and target similarity network), and the drug or target related network is divided into several categories according to the biological characteristics of the drug or target. When there are multiple networks in the category, the present application selects to fuse the node features or splice the node features according to the characteristics of the network. Second, the network embedding method obtains high-quality and rich node low-dimensional feature representation. Finally, the present application uses forward unlabeled learning to predict potential DTIs (Drug Target interactions). The purpose is to explore the optimal solution for predicting DTIs by classifying drugs and targets according to their topological structure and feature networks from multiple data sources.

[0066] The present application classifies the data according to its performance, and good results are obtained, and different weights are processed in the fusion process after the classification of the data, so that different categories of drugs and targets can be adaptively fused. Before the pharmaceutical biological test, the candidate drug target is learned by computer technology, so as to greatly reduce the time and money cost of biological control test, and shorten the pharmaceutical cycle. BRIEF DESCRIPTION OF DRAWINGS

[0067] The present application will be further described below in conjunction with the drawings.

[0068] Figure 1 It is the framework of the drug-target interaction prediction method and system based on efficient fusion of multi-source data of the present application;

[0069] Figure 2 It is a classification diagram of target related data;

[0070] Figure 3 It is a classification diagram of drug related data;

[0071] Figure 4 It is a comparison diagram of drug class network combination test; wherein *T4S represents that the fourth class network of the target contains multiple networks, and there is a selection problem of fusion and splicing, and the splicing is selected after comparison test;

[0072] Figure 5 It is a comparison diagram of target class network combination test; *D2F and D6F respectively represent that the second class and the sixth class of the drug network contain multiple networks, and the fusion effect is better when the fusion is selected in the selection of fusion and splicing;

[0073] Figure 6is a similarity network fusion algorithm taking drugs as an example;

[0074] Figure 7 is a test comparison chart of selective fusion; in precision, recall, F1-score, AUROC and AUPR, for the convenience of description, the abbreviations in the model are represented as: D, T are drugs and targets respectively; F and S are fusion and splicing; DE describes deletion; D or T followed by a number indicates the type of data being deleted.

[0075] Figure 8 is a test comparison chart of weighted fusion; weighted fusion in precision, recall, F1-score, AUROC and AUPR. The abbreviations WE, WEC and WAC represent non-classification network entropy, classification network entropy and weighted test based on prediction performance respectively;

[0076] Figure 9 is a test comparison chart of selective weighted fusion; selective weighted fusion in precision, recall, F1-score, AUROC and AUPR.

[0077] Figure 10 is a biological analysis of the predicted drug target interaction results of the model with the best effect in Figure 9 . DETAILED DESCRIPTION

[0078] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Therefore, the detailed description of the embodiments of the present application provided in the following drawings is not intended to limit the scope of the claimed application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0079] Case 1:

[0080] DTI prediction needs more accurate comprehensive analysis of multi-source data of drugs and targets, therefore, multi-source data network can improve more comprehensive information than single data network. At the same time, multiple data sources may also bring some noise, so data processing of multi-source data is particularly important. Therefore, the related content of data fusion method is discussed in the present application. The final result of comprehensive analysis shows that the selection of weighted fusion has high fusion performance. And through the concept of class network, we also find a new angle of fusion method. In the present application, we use the popular fusion strategy and entropy-based weighted fusion method to improve the prediction accuracy. According to the current research, the data sources of drugs and targets are not limited to this, such as drug-induced gene expression profile, drug pathway profile, etc. Therefore, more drug and target data sources are studied, and the robustness of the strategy is verified by supplementing the richness of drugs and targets with multiple networks.

[0081] As shown in Figure 1 , it is a framework diagram of the present application. The original data of drugs and targets in the dashed box is classified and defined according to biological performance, and fusion and splicing are selected. The data processed by fusion operation is input into the embedding model, and the feature vector, i.e. low-dimensional feature representation, is output. The low-dimensional feature representation of drugs and targets is input into the predictor to predict the drug-target interaction score matrix S.

[0082] Case 2:

[0083] As shown in Figure 2 and Figure 3 , first, the data about drugs and targets contains two properties, topological structure data and feature attribute data. Figure 2 and Figure 3 List the networks contained in the classification network of drugs and targets, where both drugs and targets have Go similar class networks and all contain three contents, i.e. molecular function, cell component and biological process. The second class network of drugs, the drug association network, contains disease and side effect association networks to describe drugs. Through biological performance classification, a total of 9 networks of drugs are divided into 6 class networks, and a total of 6 networks of targets are divided into 4 class networks.

[0084] As shown in Figure 4 and Figure 5 , comparative analysis diagrams of drugs and targets are respectively made. First, the comparative analysis of the class network of drugs and targets containing only one network, such as Figure 4 A and 5A. Second, for the class network of drugs and targets containing multiple networks, such as the second and sixth class networks of drugs, and the fourth class network of targets. Therefore, when combining the class network of drugs and targets respectively, we make comparative analysis of splicing and fusion of the class network containing multiple networks, such as Figure 4 B and Figure 5 B. Through the comparison of Figure 4 andFigure 5 Based on comprehensive analysis, firstly, for network-like drugs containing multiple networks, we determined that fusion is more effective than splicing for the second and sixth types of networks, while for the fourth type of target networks, splicing is more effective than fusion. Therefore, for network-like drugs containing multiple networks, the second and sixth types of networks choose to fuse themselves, while for the fourth type of target networks, they choose to splice themselves. Secondly, considering all networks, we can conclude that the first and second types of networks for both drugs and targets consistently produce poor prediction results. These types of networks may be due to noise interference in the prediction effect, or they may have a low contribution to the prediction effect.

[0085] Implementation Case 3:

[0086] In Implementation Case 2, the splicing involves horizontally combining multiple matrices into a single matrix, thus merging multiple networks. However, for the fusion of multiple networks, it is as follows: Figure 6 A description of a similarity network fusion algorithm, using drug fusion as an example:

[0087] The first step is to transform heterogeneous networks into homogeneous and similar networks based on the Jaccard coefficient;

[0088] The second step is to assume there are L drug networks, and let M... l Let W represent the adjacency matrix of the drug similarity network l (l = 1, 2, ..., L), and let W be the matrix element. l (d i d j The definition is as follows:

[0089]

[0090] Where ρ 2 (d i d j ) indicates drug d i and d j The Euclidean distance between them, d i Let μ be the similarity vector between the i-th drug and all other drugs, μ be a hyperparameter, and ε be a similarity vector between the i-th drug and all other drugs. ij The description is as follows:

[0091]

[0092] Where N i ={d j |M l (d i d j )>0} represents the neighbors of the drug in network 1, mean(ρ(d)>0} i N i )) is d i The average distance between it and its neighbors;

[0093] All data types are normalized by computing a matrix, the normalized matrix E l (d i , d j ) is described as follows:

[0094]

[0095] The local affinity S l (d i , d j ) is:

[0096]

[0097] Where K represents the K neighbors of drug d i ;

[0098] The third step is to iteratively update the similarity matrix E l for each data type as follows:

[0099]

[0100] The fourth step is to obtain the state matrix E (G) by computing the average of the matrix El after t iterations as follows:

[0101]

[0102] Implementation Case 4:

[0103] For the similarity network fusion of implementation case 3, there are four methods in the present application to improve it:

[0104] First, selection fusion, according to implementation case 2, we analyzed that in the drug and target class network, some class networks have low contribution to the prediction effect, therefore, we screened the drug and target class network, removed the class network with large noise influence, and compared the test analysis, such as Figure 7 , we can see that the prediction performance index after deleting the first and second class networks of the drug is relatively optimal.

[0105] Second, entropy-based weighted fusion, as shown in Figure 8 is the index analysis of the class network entropy weighted and non-class network entropy weighted model. First, calculate the entropy of each network. For any matrix M l , the entropy of node d i is defined as follows:

[0106]

[0107] Where mij represents one element of the matrix M l . Then, the average entropy of all rows is obtained as:

[0108]

[0109] Finally, update M l as follows:

[0110] EW l = M l × E l ;

[0111] Third, the weighted fusion based on the prediction performance, as shown in Figure 8 , in order to simulate the autonomous learning of back propagation, the combined result data is used to calculate the softmax function.

[0112] Suppose there are C d categories of drugs and C t categories of targets, let denote the index value of one of the combined input prediction model drug category c i and target category c j AUROC. The softmax value of each type of drug data is calculated as follows:

[0113]

[0114] where c i = 1, 2, … C d , c j = 1, 2, … C t and

[0115] Then take the average of all category drugs SoftMax value AD(i), and use the SoftMax function to normalize the drug value. The specific case is shown in the following table:

[0116]

[0117] Fourth, as shown in Figure 9 , the weighted fusion is selected, which is a comprehensive fusion method improved in the second and third, and the selection of fusion and weighted fusion is combined, which is better as shown in Figure 7 and Figure 8 .

[0118] Case 5:

[0119] After the data fusion processing, the feature vector of the fused network is extracted, that is, the low-dimensional feature representation of the drug and target, and the steps are as follows:

[0120] First, calculate the information of each node in the network:

[0121] p k = a p k-1 Tr + (1-a) p0,

[0122] where p is a vector, Tr is the transition probability;

[0123] Second, calculate the positive pointwise mutual information of all nodes as follows:

[0124]

[0125] where Pco is the probability co-occurrence matrix obtained in the first step process;

[0126] Third, stack the de-noising auto-encoder to extract the feature target function:

[0127]

[0128] where WE y is the weight matrix.

[0129] Case 6:

[0130] Finally, the feature representation of the drug target is input into the prediction model - positive unlabeled prediction, and the optimization problem of the model is parameterized as:

[0131]

[0132] where r ij = 1 indicates known drug target interactions; Ω = Ω + + Ω - represents all training samples, Ω + is the positive sample, Ω - is the unobserved sample; Z = U x V T is a low-rank matrix.

[0133] The above only describes the preferred embodiments of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting drug target interactions through efficient fusion of multi-source data, characterized in that: Includes the following steps: Step 1: Classify the topological structure data and feature attribute data of drugs and targets according to biological characteristics, and input the drug and target classification networks into the positive label-free prediction model to obtain the performance index of different combination prediction results; Step 2: Analyze the combined network test indicators and perform basic fusion or splicing of drug and target multi-source data, including network screening and network fusion. Network fusion includes entropy-weighted fusion and reverse softmax-weighted fusion. When there are multiple networks in a category, a comparative analysis of splicing and fusion is performed on network categories containing multiple networks. Based on the characteristics of the network, the fusion node features or splicing node features are selected. Step 3: Input the fused or spliced ​​drug and target networks into the network embedding model respectively, and calculate the low-dimensional feature representations of the drug and target; Step 4: Input the low-dimensional feature representations of the drug and target into the positive label-free prediction model to obtain the prediction performance index and the predicted drug-target effect score; The topology data includes drug topology data and target topology data. Drug topology data includes drug-drug interaction networks, drug-disease association networks, and drug side effect association networks. Target topology data includes target-target interaction networks and target-disease association networks. The characteristic attribute data includes drug characteristic attribute data and target characteristic attribute data; the drug characteristic attribute data includes drug similarity networks, which include drug chemical structure similarity networks, drug ATC-based similarity networks, drug target sequence similarity networks, and drug target GO similarity networks, wherein the drug target GO similarity network includes molecular function, cellular components, and biological processes; the target characteristic attribute data includes target similarity networks, which include target sequence similarity networks and GO-based similarity networks, wherein the GO-based similarity network includes molecular function, cellular components, and biological processes.

2. The method for predicting drug target interactions based on efficient fusion of multi-source data according to claim 1, characterized in that: The basic fusion method in step two is similarity network fusion, and the steps for drug network fusion are as follows: (1) Calculate the similarity matrix of heterogeneous correlation matrices based on Jaccard coefficients; (2) Calculate the edge weight matrix, normalization matrix, and local affinity matrix of various drug networks; (3) Each type of network is iterated and updated t times; (4) After t iterations, calculate the fusion matrix.

3. The method for predicting drug target interactions through efficient fusion of multi-source data according to claim 1, characterized in that: Step two implements four improved methods based on the basic fusion method—the similarity network fusion method, as follows: (1) Select fusion: Select the network class for drug and target based on the network combination test performance in step one, and then fuse the selected network based on the basic fusion method, so as to input the positive label-free prediction model. (2) Entropy-based weighted fusion: Before fusing drug and target networks separately, the entropy value of each class of network to be fused is calculated as the network weight input to the fusion method; (3) Based on the weighted fusion of prediction performance, based on AUROC, one of the combined experimental performance indicators in step one, the softmax values ​​of the drug and target class networks are calculated respectively, and each class network to be fused is weighted by softmax and then input into the similarity network fusion model. (4) Choose weighted fusion, combining the fusion strategies from the two perspectives (2) and (3) above, to make the prediction model perform better.

4. The method for predicting drug target interactions through efficient fusion of multi-source data according to claim 2, characterized in that: The matrix calculation in step (1) is as follows: Matrix element The definition is as follows: ； Where L represents the number of drug similarity networks. Drug similarity network The adjacency matrix, Indicates drug and The Euclidean distance between them For the first i The similarity vector between a drug and all other drugs For hyperparameters, The description is as follows: ； in This represents the neighbors of a drug in network l. yes The average distance between it and its neighbors; Normalization of all data types is achieved by calculating a matrix; the normalized matrix... The description is as follows: ； Local affinity is measured using K nearest neighbor (KNN) algorithm. for: ； in Representative drugs d i K neighbors.

5. The method for predicting drug target interactions through efficient fusion of multi-source data according to claim 2, characterized in that: The matrix after t iterations in step (3) The average value is used to obtain the state matrix. as follows: 。 6. The method for predicting drug target interactions based on efficient fusion of multi-source data according to claim 1, characterized in that: In step two, entropy-weighted fusion first calculates the entropy for each network. For any matrix... ,node The entropy is defined as follows: ； in Representation matrix Find one element, and then obtain the average entropy of all rows as: ， Finally, using entropy as the weight, update M l as follows: ； Among them, E i For drugs The entropy value.

7. The method for predicting drug target interactions through efficient fusion of multi-source data according to claim 1, characterized in that: In step two, the reverse softmax weighted fusion is to use the combined result data to calculate the softmax function; C d The number of drugs in each category and C t For the number of category targets, This represents the drug category in the combined input prediction model. c i and target categories c j One of the indicator values, AUROC, is used to calculate the softmax value for various types of drug data as follows: ； in, and ; Then, the SoftMax values ​​of all drug categories are averaged. AD(i) Then, the drug values ​​are normalized using the SoftMax function, as shown below: 。 8. The method for predicting drug target interactions by efficient fusion of multi-source data according to claim 1, characterized in that: In step three, the network embedding model used for noise reduction and dimensionality reduction of the highly sparse matrix is ​​as follows: (1) Randomly surf to obtain information about each node; (2) Calculate the positive pointwise mutual information matrix; (3) Stacked denoising autoencoder model, where each stack is an autoencoder that extracts network feature representations.

9. The method for predicting drug target interactions through efficient fusion of multi-source data according to claim 1, characterized in that: After the data fusion process in step three, feature vectors are extracted from the fused network, which are low-dimensional feature representations of the drug and the target. The steps are as follows: First, calculate the information of each node in the network: ， in, p For vectors, Tr It is the transition probability. This represents the initial feature vector of the node. For the first i The probability that a node will continue the random browsing process; Second, the positive point-to-point mutual information of all nodes is calculated as follows: ， in It is the probability co-occurrence matrix that can be obtained in the first step. Represents a node i and j The co-occurrence probability; Third, the stacked denoising autoencoder extracts the feature objective function: ， in, It is a weight matrix. This represents the feature matrix reconstructed by the stacked autoencoder.

10. The method for predicting drug target interactions based on efficient fusion of multi-source data according to claim 1, characterized in that: In step four, the positive unlabeled prediction model has its optimization problem parameterized as follows: ， in, Indicates known drug target interactions; Represents all training samples, As a positive sample, These are unobserved samples; It is a low-rank matrix. Represented as the feature vector of the drug; The weighting parameters for unobserved samples. λ Represents the regularization coefficient. This represents the transpose of the target's eigenvectors. U and V Let represent two learnable matrices, from which the final image is recovered. , Let F be the sum of squares of the F-norm of a low-rank matrix.

Images