Drug relocation prediction method based on multi-similarity network and knowledge graph

Abstract

The invention relates to the technical field of bioinformatics, in particular to a drug relocation prediction method based on a multi-similarity network and a knowledge graph, and the method comprises the steps: obtaining a multi-similarity network of drugs and diseases; fusing the multi-similarity network by using a graph regularization integration method, and extracting public attribute feature representations of the medicines and the diseases; learning topological feature representations of drugs and diseases in a heterogeneous network containing multiple entity associations by using a biomedical knowledge graph; carrying out fusion processing on the attribute feature representation and the topological feature representation to obtain a fusion feature vector; and according to the fusion feature vectors of the drugs and the diseases, constructing and training a drug and disease heterogeneous network to obtain a trained drug and disease heterogeneous network, and then reconstructing association of the drugs and the diseases to obtain a prediction result of drug and disease interaction. By fusing the multi-similarity network and introducing the mapping knowledge domain embedding method, the accuracy of drug and disease interaction prediction is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of bioinformatics, and in particular to a drug repositioning prediction method based on a multi-similarity network and a knowledge graph. Background Art

[0002] Traditional drug development is extremely time-consuming, expensive and risky. Bringing a new drug to market is a very complex, expensive and lengthy process, and the reason for such high costs is due to the high failure rate of preclinical trials. Regarding preclinical trials, about 90% of experimental drugs fail to pass Phase I clinical trials because drugs with new structures are prone to induce unpredictable adverse reactions. Therefore, it is of great practical significance to capture the relevant indications of drugs more accurately and effectively during the drug development process.

[0003] Drug repositioning, or repurposing old drugs, is a drug development method that uses existing drugs to discover new therapeutic potentials. It has the advantages of reducing drug risks, shortening clinical evaluation cycles, reducing costs and increasing efficiency. In recent years, drug repositioning has been widely used in diseases and related treatment areas, such as the discovery of anticancer drugs, the discovery of new therapies for rare diseases, overcoming drug resistance, and the development of personalized medicine, indicating that drug repositioning is increasingly becoming a promising research field in drug development.

[0004] Existing drug repositioning methods mainly use drug similarity and disease similarity information to predict potential drug-disease associations. However, most existing drug repositioning methods use single similarity information for calculation, which may have data quality problems such as missing and biased data, seriously affecting model performance and resulting in low accuracy of drug-disease interactions predicted by the model, which in turn leads to the inability to effectively capture the relevant indications of the drug. Summary of the invention

[0005] In order to solve the technical problem of low accuracy of the above-mentioned existing drug repositioning results, the purpose of the present invention is to provide a drug repositioning prediction method based on a multi-similarity network and a knowledge graph. The technical solution adopted is as follows:

[0006] An embodiment of the present invention provides a drug repositioning prediction method based on a multi-similarity network and a knowledge graph, the method comprising the following steps:

[0007] Obtaining a multi-similarity network of drugs and diseases; fusing the multi-similarity network of drugs and diseases using a graph regularization integration method, and extracting a common attribute feature representation of drugs and diseases based on the fused result;

[0008] Using a biomedical knowledge graph, learning the topological feature representation of drugs and diseases in a heterogeneous network containing multiple entity associations, including the construction of a knowledge graph and the learning of an embedded representation; fusing the attribute feature representation of drugs and diseases with the topological feature representation to obtain a fused feature vector of drugs and diseases;

[0009] According to the fusion feature vectors of drugs and diseases, a drug-disease heterogeneous network is constructed and trained to obtain a trained drug-disease heterogeneous network;

[0010] Based on the trained drug-disease heterogeneous network, the association between drugs and diseases is reconstructed to obtain the prediction results of drug-disease interaction.

[0011] Furthermore, the method of obtaining a multi-similarity network of drugs and diseases includes:

[0012] Obtain drug-disease interaction datasets, drug-related similarity datasets, and disease similarity datasets;

[0013] Based on different drug-related information in the drug-related similarity dataset, calculating drug similarities between two types of drugs in each drug pair under different dimensional features, and constructing a first drug similarity network of corresponding dimensional features based on the drug similarities;

[0014] Based on different disease-related information in the disease similarity dataset, calculating the disease similarity of two types of diseases in each disease pair under different dimensional features, and constructing a first disease similarity network of corresponding dimensional features based on the disease similarity;

[0015] Based on the interaction information between drugs and diseases in the drug-disease interaction dataset, the first spectrum similarity between two types of drugs in each drug pair is calculated, and a second drug similarity network is constructed based on the first spectrum similarity; and the second spectrum similarity between two types of diseases in each disease pair is calculated, and a second disease similarity network is constructed based on the second spectrum similarity;

[0016] The second drug similarity network and several of the first drug similarity networks are used as a multi-similarity network of drugs; the second disease similarity network and several of the first disease similarity networks are used as a multi-similarity network of diseases.

[0017] Furthermore, the method of using a graph regularization integration method to fuse the multiple similarity networks of the drugs and diseases, and extracting common attribute feature representations of the drugs and diseases based on the fused results, includes:

[0018] Based on the multi-similarity network of drugs or diseases, several candidate similarity matrices are obtained, and all the candidate similarity matrices are filtered using a heuristic similarity selection process to obtain multiple first similarity matrices after screening;

[0019] Using a diffusion denoising method, further performing denoising processing on the plurality of first similarity matrices to obtain a plurality of second similarity matrices;

[0020] By using a joint decomposition method with graph regularization, a plurality of the second similarity matrices corresponding to drugs and diseases are uniformly analyzed to learn the common attribute feature representation of drugs and diseases.

[0021] Furthermore, the expression for obtaining the second similarity matrix is:

[0022] In the formula, represents the second similarity matrix after the denoising of the u-th first similarity matrix, u represents the sequence number of the first similarity matrix, α represents the hyperparameter used to increase the self-similarity strength, Q (u) represents the transfer matrix corresponding to the u-th first similarity matrix, and the elements of the transfer matrix are calculated by normalizing and symmetrizing the elements in the first similarity matrix, S (u) represents the u-th first similarity matrix.

[0023] Furthermore, the joint decomposition method with graph regularization is used to uniformly analyze multiple second similarity matrices corresponding to drugs and diseases, and learn common attribute feature representations of drugs and diseases, including:

[0024] The joint decomposition method with graph regularization includes two parts: joint decomposition and graph regularization processing;

[0025] First, a joint decomposition method is used to uniformly analyze multiple second similarity matrices, separate inconsistencies and explore the correlations between the second similarity matrices, and obtain a result after joint decomposition;

[0026] Then, a graph regularization process is used to perform consistency constraints on the plurality of second similarity matrices to obtain a graph regularized result;

[0027] Combining the results after joint decomposition and graph regularization, we obtain the common attribute feature representation of drugs and diseases.

[0028] Furthermore, the expression of the result after the joint decomposition is:

[0029]

[0030] In the formula, represents the result after joint decomposition, V represents the number of second similarity matrices, v represents the sequence number of the second similarity matrix, KL(·) represents the KL divergence function, represents the second similarity matrix, Represents the reconstruction matrix, which is obtained by converting the common feature matrix X and the specific feature matrix W of the second similarity matrix v The product of is normalized, N represents the number of nodes in the similarity matrix, i represents the row number of the second similarity matrix, j represents the column number of the second similarity matrix, represents the i-th row and j-th column element of the v-th second similarity matrix, log represents the logarithmic function, x i represents the column vector in the common feature matrix X, represents the column vector in a specific feature matrix, j′ and With j and Similarly, j′ is responsible for enumerating all columns of the similarity matrix through the inner product and exponential operations to approximate the reconstructed matrix x i T represents the transpose of the column vector in the common feature matrix X, and exp represents the exponential function with a natural constant as the base;

[0031] The expression of the result after graph regularization is:

[0032]

[0033] In the formula, represents the result after graph regularization, x i and x j represents the column vector in the common feature matrix X, Tr represents the trace of the matrix, Represents the second similarity matrix The diagonal matrix of Represents the Laplacian matrix of a graph.

[0034] Furthermore, the fusing of the attribute feature representation of the drug and the disease with the topological feature representation includes:

[0035] The attribute feature representation of the drug or disease and the topological feature representation are fused by using an attention feature fusion method. The calculation formula of the attention feature fusion is:

[0036] Where Z represents the fusion feature vector of drug or disease, represents the disjoint union operation. represents the element-by-element product operation, M represents the attention weight generated using the multi-scale channel attention module, G represents the attribute feature representation of the drug or disease, and Y represents the topological feature representation of the drug or disease;

[0037]

[0038] Furthermore, the training of the drug-disease heterogeneous network includes:

[0039] The training is performed in the graph convolutional neural network. The training process includes intra-domain feature extraction and inter-domain feature extraction. The extracted intra-domain features and inter-domain features are merged to obtain the output embedding of each layer node. The expression of the output embedding is:

[0040]

[0041] In the formula, H l+1 represents the output embedding of the l-th layer node, represents the in-domain output features of the lth layer, represents the inter-domain output feature of the lth layer, H l represents the input features of the lth layer, represents the in-domain output features of the l-th layer of drugs, represents the inter-domain output features of the l-th layer of drugs, represents the input features of the l-th layer of drugs, represents the in-domain output features of the disease at layer l, represents the inter-domain output features of the disease at layer l, Represents the input features of the disease at layer l;

[0042] Then, different attentions are used to integrate the output embeddings to obtain the embedding information of drugs and diseases. The expression of the embedding information is:

[0043] In the formula, H R ∈R N×k represents the final embedding of the drug, R N×k represents the final embedding size of N drugs with k nearest neighbor similar nodes, H D ∈R M×k represents the final embedding of the disease, R M×k represents the final embedding size of M drugs with k nearest neighbor similar nodes, L represents the hyperparameter, β l ∈R is automatically initialized by the neural network learning It represents the contribution of the embedded data of different convolutional layers to the final embedded data.

[0044] Furthermore, the reconstruction of the relationship between the drug and the disease includes:

[0045] The definition of the decoder function used in reconstructing the association between drugs and diseases is as follows:

[0046]

[0047] In the formula, represents the predicted probability score matrix, H R represents the final embedding of the drug, H D represents the final embedding of the disease, sigmoid represents the activation function, and H D T represents the transpose of the final embedding of disease, drug r i With disease j The correlation prediction score between The corresponding (i,j) term of is given;

[0048] At the same time, the parameters are learned by minimizing the weighted binary cross entropy loss. The loss function is defined as follows:

[0049] In the formula, loss represents the loss function, N represents the number of drug nodes, and M represents the number of disease nodes. Indicates drug r i and diseases j The association prediction score of , log represents the logarithmic function, (i, j) represents the drug-disease pair, S + represents the set of all known drug-disease association information, S - represents the set of all unknown or unobserved drug-disease association information, and the balance factor Used to reduce the impact of data imbalance, |S - | and |S + |It is S + With S - The number of pairs in ;

[0050] Finally, the Adam optimizer is used to optimize the drug-disease heterogeneous network and initialize the weights of each parameter in the drug-disease heterogeneous network.

[0051] The present invention has the following beneficial effects:

[0052] The embodiment of the present invention proposes a drug repositioning prediction method based on a multi-similarity network and a knowledge graph to accurately discover candidate drugs for diseases. First, in order to fully consider the different representation differences of entities in a multivariate environment, graph regularization is applied to integrate and fuse data information from multiple data sources; secondly, most of the current research on drug-disease interaction prediction issues only considers drug and disease entities, ignoring the association between drugs or diseases and other entities during the drug's effectiveness. By introducing a knowledge graph embedding module, the association between drugs and diseases and multiple medical entities is constructed to obtain the global topological feature representation of drug and disease entities; finally, the attribute feature information of drugs and diseases obtained by fusing multi-similarity data is combined with the global topological information obtained by knowledge graph embedding, and passed to the graph convolutional network prediction model to identify potential therapeutic indications for drugs. The present invention integrates multi-source similarity data information to construct a comprehensive attribute feature representation of drugs and diseases, which helps the model to more comprehensively understand the association structure and complementary information between different similarity views. At the same time, the knowledge graph embedding method is introduced to enable the model to effectively obtain the global topological characteristics of drug and disease entities. By deeply mining the complex relationship network in the knowledge graph, the mutual correlation and influence between drugs and diseases and multiple entities in the entire biomedical knowledge graph can be more comprehensively considered, effectively improving the model's predictive performance for drug-disease interactions. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. 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 paying creative work.

[0054] Figure 1 A flowchart of the steps of a drug repositioning prediction method based on a multi-similarity network and a knowledge graph according to an embodiment of the present invention;

[0055] Figure 2 A flowchart of the steps of fusing multiple similarity networks of drugs and diseases using a graph regularization integration method in an embodiment of the present invention;

[0056] Figure 3 Schematic diagram of an iterative attention feature fusion method in an embodiment of the present invention;

[0057] Figure 4 Schematic diagram of a graph convolutional network prediction model in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects of the technical solutions proposed by the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0059] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0060] In order to improve the accuracy of drug repositioning prediction results, based on drug similarity, disease similarity, and the association information between drugs and diseases, valuable information can be provided for drug-disease interaction prediction. First, a suitable similarity measurement method is used to calculate the similarity between drug pairs and disease pairs; then, a graph regularization integration method is used to integrate the multi-source similarity information of drugs and diseases, and the integrated attribute feature representation is extracted; then, based on the biomedical knowledge graph, the topological feature representation of drugs and diseases is learned in a heterogeneous network containing multiple entity associations; finally, the acquired features are fused using the attention feature fusion method, and the drug-disease association is predicted through a graph convolutional network. By comparing with other methods and testing on the data set, it is shown that this method has good performance in drug-disease interaction prediction.

[0061] Specifically, this embodiment provides a drug repositioning prediction method based on a multi-similarity network and a knowledge graph, such as Figure 1 As shown, the following steps are included:

[0062] S1, obtain the multi-similarity network of drugs and diseases; use the graph regularization integration method to fuse the multi-similarity network of drugs and diseases, and extract the common attribute feature representation of drugs and diseases based on the fusion results.

[0063] First, a multi-similarity network of drugs and diseases is obtained.

[0064] The first step is to obtain the drug-disease interaction dataset, drug-related similarity dataset and disease similarity dataset.

[0065] Before drug repositioning is realized, it is necessary to obtain benchmark data for drug repositioning so that a drug-disease heterogeneous network can be constructed through subsequent processing and analysis of the benchmark data to obtain more accurate drug repositioning results.

[0066] In this embodiment, the SND (Similarity Network Dataset) benchmark dataset is used. The data in the benchmark dataset is collected from various biological and biomedical data sources and consists of three different types of information, namely drug-disease interaction data, drug-related similarity data, and disease similarity data; the information of the same type in the benchmark dataset is divided together to form a drug-disease interaction dataset, a drug-related similarity dataset, and a disease similarity dataset. Among them, the drug-disease interaction data (i.e., gold standard data) comes from two widely used data sources, namely DrugBank and repoDB. The gold standard data includes 867 approved drugs, 803 diseases, and 8684 clinically reported or experimentally verified drug-disease interactions, with a sparsity of up to 98.75%.

[0067] In the second step, based on the different drug-related information in the drug-related similarity dataset, the drug similarity of the two types of drugs in each drug pair under different dimensional features is calculated, and the first drug similarity network of the corresponding dimensional features is constructed based on the drug similarity.

[0068] In this embodiment, the drug-related similarity data has a network with 10 different dimensional features, that is, a first drug similarity network with 10 dimensional features. The information based on which the drug similarity is calculated includes: (1) target protein interaction, (2) side effects, (3) chemical structure, (4) GO molecular function, (5) GO biological process, (6) GO cellular component, (7) metabolic enzyme, (8) protein sequence, (9) anatomical therapeutic chemical classification code, and (10) drug pairing interaction. There are many ways to calculate the similarity of two types of drugs, depending on the properties of the drugs and the available data. For example, when analyzing the similarity of drug chemical structures, the cosine similarity calculation method can be used. The similarity calculation process is a prior art and is not within the scope of protection of the present invention, so it will not be elaborated here.

[0069] The third step is to calculate the disease similarity of the two types of diseases in each disease pair under different dimensional features based on the different disease-related information in the disease similarity dataset, and construct the first disease similarity network of the corresponding dimensional features based on the disease similarity.

[0070] In this embodiment, the disease similarity data has a network of 14 different dimensional features, that is, a first disease similarity network of 14 dimensional features, and the information based on the disease similarity includes: (1) control genes, (2) HPO genes, (3) literature-based genes, (4) curated variants, (5) literature-based variants, (6) microRNAs, (7) non-long coding RNAs, (8) HPO phenotypes, (9) IS-A classification, (10) information theory similarity, (11) GO terms, (12) implicit semantic similarity, (13) semantic and gene function similarity, and (14) curated association type ontology. There are many ways to calculate the similarity of two types of diseases, which depend on the nature of the disease and the available data, and are not specifically limited.

[0071] The fourth step is to calculate the first spectral similarity between the two types of drugs in each drug pair based on the drug-disease interaction data set, and construct a second drug similarity network based on the first spectral similarity; and calculate the second spectral similarity between the two types of diseases in each disease pair, and construct a second disease similarity network based on the second spectral similarity.

[0072] In this embodiment, based on the drug-disease interaction data set, the GIP Gaussian kernel similarity between drug pairs and the GIP (Gaussian Kernel Similarity) Gaussian kernel similarity between disease pairs are calculated respectively, and then the corresponding networks are constructed, that is, the second drug similarity network and the second disease similarity network are determined.

[0073] The calculation formula for the first spectrum similarity between two types of drugs in each drug pair can be:

[0074] In the formula, rs (GIP) (r1, r2) represents the first spectrum similarity between drug r1 and drug r2 in the drug pair, exp represents an exponential function with a natural constant as the base, γ represents a parameter for controlling the kernel bandwidth, which can be set by cross-validation. In this embodiment, γ can be set to 1, n r represents the number of drugs contained in the drug-disease interaction dataset, d(r1) represents the number of diseases that interact with drug r1, d(r2) represents the number of diseases that interact with drug r2, ||d(r1)-d(r2)|| 2 represents the square of the Euclidean distance between d(r1) and d(r2), d(r i ) indicates drug r i Number of interacting diseases.

[0075] The calculation formula for the second spectrum similarity between two types of diseases in each disease pair can be:

[0076] Where, ds (GIP) (d1, d2) represents the second spectrum similarity between disease d1 and disease d2 within the disease pair, n d represents the number of diseases contained in the drug-disease interaction dataset, l(d1) represents the number of drugs that interact with disease d1, l(d2) represents the number of drugs that interact with disease d2, ||l(d1)-l(d2)|| 2 represents the square of the Euclidean distance between l(d1) and l(d2), l(d j ) indicates disease j Number of interacting drugs.

[0077] The fifth step is to use the second drug similarity network and several first drug similarity networks as a multi-similarity network of drugs; and to use the second disease similarity network and several first disease similarity networks as a multi-similarity network of diseases.

[0078] Secondly, the graph regularization integration method is used to fuse the multiple similarity networks of drugs and diseases, and the common attribute feature representations of drugs and diseases are extracted based on the fused results.

[0079] First of all, it should be explained that similarity network fusion is a method for fusing multiple similarity networks. It integrates different types of data information into a new fusion network by considering the similarity information between different networks, so as to fully consider the complementarity of different types of data features, that is, to fully consider the different representation differences of entities in a multivariate environment, thereby providing a more comprehensive and comprehensive similarity measurement result.

[0080] like Figure 2 As shown, the fusion of multiple similarity networks of drugs and diseases using the graph regularization integration method can include the following steps:

[0081] S201, obtaining a number of candidate similarity matrices based on a multi-similarity network of drugs or diseases, and filtering all the candidate similarity matrices using a heuristic similarity selection process to obtain a plurality of screened first similarity matrices.

[0082] In this embodiment, based on the similarity information in the multi-similarity network, the multi-similarity network is converted into a corresponding similarity matrix, and several candidate similarity matrices can be obtained, and the elements in the similarity matrix are the similarity information of drug pairs or disease pairs; in order to screen out a valuable first similarity matrix from all the candidate similarity matrices, that is, to filter out the candidate similarity matrices with very little similarity information and serious redundancy, a heuristic similarity selection process is used to achieve it.

[0083] The heuristic similarity selection process includes calculating the average entropy of candidate similarity matrices, removing candidate similarity matrices with less similarity information, calculating the similarity between the selected similarity matrices, and removing redundant candidate similarity matrices based on the similarity. The detailed implementation process of the heuristic similarity selection process is prior art and is not within the scope of protection of the present invention, and will not be elaborated here.

[0084] S202, using a diffusion denoising method to further perform denoising processing on the multiple first similarity matrices to obtain multiple second similarity matrices.

[0085] In this embodiment, through the similarity selection step of step S201, the candidate similarity matrix with very little similarity information and serious redundancy can be filtered out, but the noise information in the first similarity matrix still exists. Since the transitional similarity selection may lead to the loss of important features, this embodiment uses the diffusion denoising method to further perform denoising on the first similarity matrix after screening, and uses the first similarity matrix after denoising as the second similarity matrix. Among them, the implementation process of diffusion denoising is a prior art, which is not within the protection scope of the present invention and will not be elaborated in detail here.

[0086] As an example, the expression of the u-th second similarity matrix may be:

[0087] In the formula, represents the second similarity matrix after the denoising of the u-th first similarity matrix, u represents the sequence number of the first similarity matrix, α represents the hyperparameter used to increase the self-similarity strength, Q (u) represents the transfer matrix corresponding to the u-th first similarity matrix, and the elements of the transfer matrix are calculated by normalizing and symmetrizing the elements in the first similarity matrix, S (u) represents the u-th first similarity matrix.

[0088] S203, using a joint decomposition method with graph regularization, a plurality of second similarity matrices corresponding to drugs and diseases are uniformly analyzed to learn common attribute feature representations of drugs and diseases.

[0089] In this embodiment, the treatment methods of drugs and diseases are consistent. In order to facilitate understanding of the method, drugs or diseases are taken as examples to perform a unified analysis on the second similarity matrix to learn the common attribute feature representations of drugs or diseases, and there can be multiple common attribute feature representations. Among them, the joint decomposition method with graph regularization includes two parts: joint decomposition and graph regularization processing.

[0090] In the first step, a joint decomposition method is used to uniformly analyze multiple second similarity matrices, separate inconsistencies and explore the correlations between the second similarity matrices to obtain the results after joint decomposition.

[0091] As an example, the expression of the result after the joint decomposition can be:

[0092]

[0093] In the formula, represents the result after joint decomposition, V represents the number of second similarity matrices, v represents the sequence number of the second similarity matrix, KL(·) represents the KL divergence function, represents the second similarity matrix, Represents the reconstruction matrix, which is obtained by converting the common feature matrix X and the specific feature matrix W of the second similarity matrix v The product of is normalized, N represents the number of nodes in the similarity matrix, i represents the row number of the second similarity matrix, j represents the column number of the second similarity matrix, represents the i-th row and j-th column element of the v-th second similarity matrix, log represents the logarithmic function, x i represents the column vector in the common feature matrix X, represents the column vector in a specific feature matrix, j′ and With j and Similarly, j′ is responsible for enumerating all columns of the similarity matrix through the inner product and exponential operations to approximate the reconstructed matrix x i T represents the transpose of the column vector in the common feature matrix X, and exp represents the exponential function with a natural constant as the base.

[0094] In the second step, graph regularization is used to perform consistency constraints on multiple second similarity matrices to obtain the graph regularized result.

[0095] It should be noted that in most networks, nodes with higher similarity scores are closer in space, while nodes with lower similarity scores are farther apart in space. Therefore, a regularization is introduced on the common feature representation to obtain prior information in the network and provide consistency constraints.

[0096] As an example, the expression of the result after graph regularization can be:

[0097]

[0098] In the formula, represents the result after graph regularization, x i and x j represents the column vector in the common feature matrix X, Tr represents the trace of the matrix, Represents the second similarity matrix The diagonal matrix of Represents the Laplacian matrix of a graph.

[0099] In the third step, the common attribute feature representation of drugs and diseases is obtained by combining the results after joint decomposition and graph regularization.

[0100] In this embodiment, the joint decomposition and graph regularization processing are combined to obtain an objective function, and the expression of the objective function can be:

[0101]

[0102] In the formula, represents the vth specific feature matrix, represents the regularization parameter, X represents the common feature matrix, and min represents the minimum function.

[0103] It should be noted that by applying the multi-similarity network fusion method, the problem of incomplete entity information coverage easily caused by using a single data source is effectively solved, making the training data used for subsequent model training richer, which is conducive to improving the prediction accuracy of the model.

[0104] So far, this embodiment has obtained the common attribute feature representation of drugs and diseases.

[0105] S2, using biomedical knowledge graphs, learns the topological feature representation of drugs and diseases in a heterogeneous network containing multiple entity associations, including the construction of knowledge graphs and the learning of embedded representations; fuses the attribute feature representation of drugs and diseases with the topological feature representation to obtain the fused feature vectors of drugs and diseases.

[0106] Specifically, we first use DRKG (Drug Repurposing Knowledge Graph) as the basic knowledge graph, and screen the drug and disease entities to obtain the entities in the required drug and disease datasets; secondly, we use ComplEx's embedding learning technology to obtain the embedded feature representation of the relevant entities in the knowledge graph. The learning of the embedded representation includes five parts: embedding initialization, score calculation, loss calculation, model optimization, and feature extraction, to obtain the topological feature representation of drugs and diseases; then, we use the attention feature fusion method to fuse the attribute feature representation of drugs and diseases with the topological feature representation to obtain the fused feature vector of drugs and diseases. The schematic diagram of the iterative attention feature fusion method is shown in the figure. Figure 3 shown.

[0107] The steps of fusing the attribute feature representation and the topological feature representation of the drug or disease using the attention feature fusion method include:

[0108] Specifically, the initial feature fusion is first performed on the two input feature vectors, and then the obtained initial features are input into the MS-CAM (Multi-Scale Class Activation Mapping) module, and finally the broadcast addition is applied for fusion. Among them, the broadcast addition is a prior art, and the present invention will not repeat its implementation process. It is mainly used for feature visualization in visual tasks, and the calculation formula of attention feature fusion is:

[0109] Where Z represents the fusion feature vector of drugs and diseases, represents the disjoint union operation. represents the element-wise product operation, M represents the attention weight generated using the multi-scale channel attention module, G represents the attribute feature, and Y represents the topological feature representation of drugs and diseases.

[0110] In the initial feature fusion step In , the attention fusion mechanism is also used. In this way, iterative processing can achieve a more complete perception of the input feature map. The calculation formula of the iterative processing part is:

[0111]

[0112] It should be noted that the introduction of the knowledge graph embedding method enables the model to effectively obtain the global topological characteristics of drug and disease entities. By deeply mining the complex relationship network in the knowledge graph, it can more comprehensively consider the mutual correlation and influence between drugs and diseases and multiple entities in the entire biomedical knowledge graph, effectively improving the model's predictive performance for drug-disease interactions.

[0113] So far, this embodiment has obtained the fusion feature vector of drugs and diseases.

[0114] S3, calculates the similarity measure of the fused feature vectors of drugs and diseases, uses the similarity measure to construct a drug-disease heterogeneous network, and then trains the drug-disease heterogeneous network to obtain a trained drug-disease heterogeneous network.

[0115] In this embodiment, after obtaining the fused feature vector of the drug and the fused feature vector of the disease, they are passed to the graph convolutional network prediction model to identify the potential therapeutic indications of the drug. Training is performed in the graph convolutional neural network, and the training process includes intra-domain feature extraction and inter-domain feature extraction. The extracted intra-domain features and inter-domain features are merged to obtain the output embedding of each layer node.

[0116] The training process of graph convolutional network includes intra-domain feature extraction and inter-domain feature extraction, where the calculation formula of intra-domain feature extraction is:

[0117]

[0118] In the formula, represents the output feature in the drug domain of the lth layer, represents the output feature in the disease domain of the lth layer, N represents the number of drug nodes, M represents the number of disease nodes, k represents the number of nearest neighbor similar nodes of the drug, represents the drug input embedding at layer l, and represents the trainable matrix of the feature extraction module in the lth layer domain, GCN represents the graph convolution operation, and A r Represents the adjacency matrix of the drug similarity network, A d Represents the adjacency matrix of the disease similarity network; where, D=diag(∑ j A ij ), σ represents the activation function, A ij represents the element in the i-th row and j-th column of the drug-disease association network, D represents the degree matrix of the adjacency matrix A, and is used to normalize the adjacency matrix to prevent the connectivity differences between different nodes from having too much impact on the results; H represents the input embedding of the drug or disease, and W represents the trainable matrix of the in-domain feature extraction module.

[0119] The inter-domain feature extraction module for information transfer between drugs and diseases consists of a bilinear aggregator (BA) and a traditional GCN (Graph Convolutional Network Aggregator) aggregator (AGG). i , and its drug domain feature extraction module is defined as:

[0120] Where ⊙ is an element-wise product operation; Indicates drug r i The output feature of the drug domain in the first layer; W l ∈R k×k represents a trainable matrix; α l ∈R represents a trainable scalar used to balance the importance between BA and traditional GCN aggregators.

[0121] It is worth noting that for disease d j The disease domain feature extraction module is defined in the same way as the drug domain feature extraction module. To indicate disease j The l-th layer disease domain output features.

[0122] Finally, the intra-domain features and inter-domain features are combined into the output embedding, and the expression of the output embedding is:

[0123]

[0124] In the formula, H l+1 represents the output embedding of the l-th layer node, represents the in-domain output features of the lth layer, represents the inter-domain output feature of the lth layer, H l represents the input features of the lth layer, represents the in-domain output features of the l-th layer of drugs, represents the inter-domain output features of the l-th layer of drugs, represents the input features of the l-th layer of drugs, represents the in-domain output features of the disease at layer l, represents the inter-domain output features of the disease at layer l, Represents the input features of the l-th layer of diseases.

[0125] Then, different attention integration output embeddings are used to obtain the embedding information of drugs and diseases. The expression of the embedding information is:

[0126]

[0127] In the formula, H R ∈R N×k represents the final embedding of the drug, R N×k represents the final embedding size of N drugs with k nearest neighbor similar nodes, H D ∈R M×k represents the final embedding of the disease, R M×k represents the final embedding size of M drugs with k nearest neighbor similar nodes, L represents the hyperparameter, β l ∈R is automatically initialized by the neural network learning It represents the contribution of the embedded data of different convolutional layers to the final embedded data.

[0128] So far, this embodiment has obtained a trained drug-disease heterogeneous network, which is a graph convolutional network prediction model. The schematic diagram of the graph convolutional network prediction model is as follows: Figure 4 shown.

[0129] S4, based on the trained drug-disease heterogeneous network, reconstructs the association between drugs and diseases to obtain the prediction results of drug-disease interactions.

[0130] In this embodiment, in order to reconstruct the association between drugs and diseases, the decoder function used in reconstructing the association between drugs and diseases can be defined as follows:

[0131]

[0132] In the formula, represents the predicted probability score matrix, H R represents the final embedding of the drug, H D represents the final embedding of the disease, sigmoid represents the activation function, and H D T represents the transpose of the final embedding of disease, drug r i With disease j The association prediction segmentation between The corresponding (i,j) term of is given.

[0133] Given that known drug-disease associations have been manually verified, they are highly reliable and are also very important for improving prediction performance. However, the number of known drug-disease associations is far less than the number of unknown or unobserved drug-disease pairs. Therefore, the parameters are learned by minimizing the weighted binary cross entropy loss, and the loss function is defined as follows:

[0134] In the formula, loss represents the loss function, N represents the number of drug nodes, and M represents the number of disease nodes. Indicates drug r i and diseases j The association prediction score of , log represents the logarithmic function, (i, j) represents the drug-disease pair, S + represents the set of all known drug-disease association information, S - represents the set of all unknown or unobserved drug-disease association information, and the balance factor Used to reduce the impact of data imbalance, |S - | and |S + |It is S + With S - Finally, the Adam optimizer is used to optimize the drug-disease heterogeneous network and initialize the weights of each parameter in the drug-disease heterogeneous network.

[0135] Furthermore, in order to verify the effectiveness of the present invention, a ten-fold cross validation method was used to conduct experiments to test the predictive performance of the present invention for drug-disease interactions.

[0136] In the 10-fold cross-validation process, 10% of the known drug-disease association pairs and 10% of the random unknown drug-disease association pairs in the dataset were randomly selected as the test set; the remaining 90% of the drug-disease association pairs verified by clinical reports and unknown drug-disease association pairs were used to train the model. The area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPR) have been widely used in bioinformatics research. This paper uses these two indicators to evaluate the overall performance of model prediction.

[0137] In order to verify the effectiveness of the method of the present invention, the method of the present invention is compared with other six state-of-the-art drug-disease interaction prediction methods in terms of performance, including LAGCN (Layer Attention Graph Convolutional Network), DRHGCN (Drug Repositioning based on the Heterogeneous Information Fusion Graph Convolutional Network), DRWBNCF (Neighborhood and Neighborhood Interaction-based Neural Collaborative Filtering Approach), GCNAT (Graph Convolutional Network with Attention), SMGCL (Similarity Measures-based Graph Co-Contrastive Learning) and DRAGNN (Drug Repositioning based on Weighted Local Information Augmented Graph Neural Networks), which are described in detail as follows:

[0138] LAGCN is a layer-attention GCN that uses GCN to learn embeddings of drugs and diseases from a drug-disease heterogeneous network, and then integrates the learned embeddings through an attention mechanism to predict new associations;

[0139] DRHGCN uses GCN to extract inter-domain and intra-domain feature information of drugs and diseases, thereby finding new drug indications based on different network topology information of drugs and diseases in different domains;

[0140] DRWBNCF is a neural collaborative filtering method that proposes a novel weighted bilinear graph convolution operation to integrate known drug-disease associations, drug and disease neighborhoods, and neighborhood interaction information into a unified representation to infer new potential drug-disease associations;

[0141] GCNAT is a deep learning algorithm that combines graph convolutional networks and graph attention networks to build heterogeneous networks based on metabolite-disease associations, metabolite-metabolite similarities, and disease-disease similarities. It then predicts new metabolite-disease associations by combining the embeddings of multiple convolutional layers using graph attention layers and assigning different weights.

[0142] SMGCL is a graph contrastive learning method based on similarity metrics. It introduces a graph contrastive learning method to jointly train node representations to maximize consistency, thereby overcoming the problem of sparse supervision signals in traditional graph neural network methods and enhancing the predictive ability of drug-disease associations.

[0143] DRAGNN is a novel weighted local information enhancement method that improves the effectiveness of target node information collection by combining the attention mechanism and omitting self-node information aggregation, emphasizes valuable heterogeneous and homogeneous information, and introduces average pooling technology to enhance local information.

[0144] The comparison between the present invention and other methods is as follows:

[0145] Comparative experiments were conducted on the SND benchmark dataset, and all experiments used a ten-fold cross-validation scheme to evaluate the predictive performance of the model. Specifically, the original drug-disease association dataset was randomly divided into 10 subsets of equal size, each subset served as a test set in turn, and the other 9 subsets were used as training sets. For each training set, the drug-disease heterogeneous network (KGRDR model) of the present invention and other comparative models were trained to learn the parameters and weights of the model. Then, the trained model was used to predict the corresponding test set to obtain the prediction results of the experiment. This process was repeated 10 times to ensure that each subset served as a test set once, thereby obtaining the performance evaluation results of each model on different test sets. In order to maintain the fairness of the evaluation, the same random seed was used each time the dataset was randomly divided, so that different models could be fairly compared on the same training set and test set.

[0146] The specific experimental results are shown in Table 1, which shows the comparison results between the present invention and other methods.

[0147] Table 1

[0148]

[0149] As can be seen from Table 1, on the SND benchmark dataset, KGRDR finally obtained an AUROC of 0.9864, which is 2.09% higher than the second-ranked DRHGCN; KGRDR obtained an AUPR of 0.7915, which is 11.69% higher than the second-ranked DRHGCN. It is worth noting that KGRDR also achieved the highest AUROC and AUPR on the SCMFDD_S dataset. The benchmark comparison results of the two datasets show that KGRDR outperforms six state-of-the-art prediction models. In particular, the results of each ten-fold cross-validation are basically consistent, which shows that the drug-disease heterogeneous network trained by the present invention exhibits convincing performance and high robustness.

[0150] In order to study whether different modules play an important role in the model, ablation experiments were performed on the SND benchmark dataset to verify the impact of important components on the performance of KGRDR, and the following two comparison variants were designed: First, KGRDR (w / o GRI): KGRDR does not use the similarity network fusion algorithm to fuse multiple similarity data, but uses the data under a single similarity view to replace the fusion matrix obtained by graph regularization integration as the input of GCN; Second, KGRDR (w / o KG): KGRDR does not use the knowledge graph feature extraction module, but directly uses the attribute feature results obtained after multi-similarity fusion as the input of GCN.

[0151] KGRDR is compared with the two variants respectively. As shown in Table 2, the ablation experiment results show that deleting the similarity network fusion module will reduce the prediction performance of the present invention for drug-disease interaction. This is because after deleting the similarity network fusion module, KGRDR cannot fully utilize the association information and potential relationships between different similarity data when predicting drug-disease interaction. At the same time, it is also unable to consider the information of neighboring samples, causing the model to lose its in-depth understanding and comprehensive analysis capabilities of the data, thereby reducing the accuracy and reliability of the model prediction.

[0152] In addition, it can be seen from the ablation experiment result table that deleting the knowledge graph feature extraction module will reduce the prediction performance of the present invention for drug-disease interactions. This may be because after deleting the knowledge graph feature extraction module, KGRDR cannot fully utilize the influence of the mutual correlation between different entities in the heterogeneous network when predicting drug-disease interactions, causing the model to lose its understanding and analysis of the global topological relationship.

[0153] Table 2

[0154]

[0155] In order to further prove the performance of the graph regularization ensemble method with de-manipulated diffusion for the ensemble of multiple similarity networks, the following four multiple similarity network fusion methods were compared using ten-fold cross validation: SNF-H (Similarity Network Fusion with Hierarchical Information), DeFusion (Decomposition Fusion), EnMuGR (Enhanced Multi-Graph Representation) and SNF-NN (Similarity Network Fusion using Neural Networks). SNF-H, DeFusion and EnMuGR used the same network prediction model (DRHGCN) and compared the multiple similarity fusion methods with KGRDR on the SND benchmark dataset; SNF-NN compiled and used a large number of drug and disease similarity datasets SND for the first time, and compared the optimal network model it provided with KGRDR on the SND dataset.

[0156] As shown in Table 3, the comparative experimental results show that when using the same data set and the same network prediction model, KGRDR achieves the best AUROC and AUPR values ​​compared with SNF-H, DeFusion, EnMuGR and SNF-NN, which indicates that the method proposed in the present invention can better integrate multiple similarity networks, thereby improving the performance of the prediction model.

[0157] Table 3

[0158]

[0159] In summary, the embodiment of the present invention proposes a drug-disease association prediction method combining graph regularization integration, knowledge graph embedding and GCN to accurately discover candidate drugs for diseases. First, in order to fully consider the different representation differences of entities in a multivariate environment, graph regularization integration is applied to fuse the drug-drug and disease-disease similarities of multiple data sources. At the same time, considering the noise existing in the fusion process of the data source, the use of diffusion denoising operation can effectively process the noise information. Secondly, most of the current research on the problem of drug-disease interaction prediction only considers the drug and disease entities, ignoring the association between drugs or diseases and other entities during the drug's effectiveness. Therefore, the present invention introduces a knowledge graph embedding module to construct the association between drugs and diseases and multiple medical entities to obtain the global topological feature representation of drug and disease entities. Finally, the attention feature fusion method is used to combine the attribute feature information of drugs and diseases obtained by fusing multiple similarity data with the global topological information obtained by knowledge graph embedding, and pass it to the graph convolutional network prediction model for identifying potential therapeutic indications of drugs.

[0160] The present invention integrates multi-source similarity data information to construct a comprehensive attribute feature representation of drugs and diseases, which helps the model to more comprehensively understand the association structure and complementary information between different similarity views. At the same time, the knowledge graph embedding method is introduced to enable the model to effectively obtain the global topological characteristics of drug and disease entities. By deeply mining the complex relationship network in the knowledge graph, the mutual correlation and influence between drugs and diseases and multiple entities in the entire biomedical knowledge graph can be more comprehensively considered, effectively improving the model's predictive performance for drug-disease interactions.

[0161] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A drug repositioning prediction method based on multi-similarity network and knowledge graph, characterized in that: The following steps are involved: Obtaining a multi-similarity network of drugs and diseases; fusing the multi-similarity network of drugs and diseases using a graph regularization integration method, and extracting a common attribute feature representation of drugs and diseases based on the fused result; Using a biomedical knowledge graph, learning the topological feature representation of drugs and diseases in a heterogeneous network containing multiple entity associations, including the construction of a knowledge graph and the learning of an embedded representation; fusing the attribute feature representation of drugs and diseases with the topological feature representation to obtain a fused feature vector of drugs and diseases; According to the fusion feature vectors of drugs and diseases, a drug-disease heterogeneous network is constructed and trained to obtain a trained drug-disease heterogeneous network; Based on the trained drug-disease heterogeneous network, the association between drugs and diseases is reconstructed to obtain the prediction results of drug-disease interaction.

2. A drug repositioning prediction method based on multi-similarity network and knowledge graph according to claim 1, characterized in that: The method of obtaining a multi-similarity network of drugs and diseases includes: Obtain drug-disease interaction datasets, drug-related similarity datasets, and disease similarity datasets; Based on different drug-related information in the drug-related similarity dataset, calculating drug similarities between two types of drugs in each drug pair under different dimensional features, and constructing a first drug similarity network of corresponding dimensional features based on the drug similarities; Based on different disease-related information in the disease similarity dataset, calculating the disease similarity of two types of diseases in each disease pair under different dimensional features, and constructing a first disease similarity network of corresponding dimensional features based on the disease similarity; Based on the interaction information between drugs and diseases in the drug-disease interaction dataset, the first spectrum similarity between two types of drugs in each drug pair is calculated, and a second drug similarity network is constructed based on the first spectrum similarity; and the second spectrum similarity between two types of diseases in each disease pair is calculated, and a second disease similarity network is constructed based on the second spectrum similarity; The second drug similarity network and several of the first drug similarity networks are used as a multi-similarity network of drugs; the second disease similarity network and several of the first disease similarity networks are used as a multi-similarity network of diseases.

3. The drug repositioning prediction method based on multi-similarity network and knowledge graph according to claim 1, characterized in that: The method of fusing the multi-similarity networks of the drugs and diseases by using the graph regularization integration method and extracting the common attribute feature representation of the drugs and diseases based on the fused results includes: Based on the multi-similarity network of drugs or diseases, several candidate similarity matrices are obtained, and all the candidate similarity matrices are filtered using a heuristic similarity selection process to obtain multiple first similarity matrices after screening; Using a diffusion denoising method, further performing denoising processing on the plurality of first similarity matrices to obtain a plurality of second similarity matrices; By using a joint decomposition method with graph regularization, a plurality of the second similarity matrices corresponding to drugs and diseases are uniformly analyzed to learn the common attribute feature representation of drugs and diseases.

4. A drug repositioning prediction method based on multi-similarity network and knowledge graph according to claim 3, characterized in that: The expression for obtaining the second similarity matrix is: In the formula, represents the second similarity matrix after the denoising of the u-th first similarity matrix, u represents the sequence number of the first similarity matrix, α represents the hyperparameter used to increase the self-similarity strength, Q (u) represents the transfer matrix corresponding to the u-th first similarity matrix, and the elements of the transfer matrix are calculated by normalizing and symmetrizing the elements in the first similarity matrix, S (u) represents the u-th first similarity matrix.

5. The drug repositioning prediction method based on multi-similarity network and knowledge graph according to claim 3, characterized in that: The joint decomposition method with graph regularization is used to uniformly analyze multiple second similarity matrices corresponding to drugs and diseases, and learn the common attribute feature representation of drugs and diseases, including: The joint decomposition method with graph regularization includes two parts: joint decomposition and graph regularization processing; First, a joint decomposition method is used to uniformly analyze multiple second similarity matrices, separate inconsistencies and explore the correlations between the second similarity matrices, and obtain a result after joint decomposition; Then, a graph regularization process is used to perform consistency constraints on the plurality of second similarity matrices to obtain a graph regularized result; Combining the results after joint decomposition and graph regularization, we obtain the common attribute feature representation of drugs and diseases.

6. A drug repositioning prediction method based on multi-similarity network and knowledge graph according to claim 5, characterized in that: The expression of the result after the joint decomposition is: In the formula, represents the result after joint decomposition, V represents the number of the second similarity matrix, v represents the sequence number of the second similarity matrix, KL(·) represents the KL divergence function, represents the second similarity matrix, Represents the reconstruction matrix, which is obtained by converting the common feature matrix X and the specific feature matrix W of the second similarity matrix v The product of is normalized, N represents the number of nodes in the similarity matrix, i represents the row number of the second similarity matrix, j represents the column number of the second similarity matrix, represents the i-th row and j-th column element of the v-th second similarity matrix, log represents the logarithmic function, x i represents the column vector in the common feature matrix X, represents the column vector in a specific feature matrix, j′ and With j and Similar, j ′ Responsible for enumerating all columns of the similarity matrix through the inner product and exponential operations to approximate the reconstructed matrix x i T represents the transpose of the column vector in the common feature matrix X, and exp represents the exponential function with a natural constant as the base; The expression of the result after graph regularization is: In the formula, represents the result after graph regularization, x i and x j represents the column vector in the common feature matrix X, Tr represents the trace of the matrix, Represents the second similarity matrix The diagonal matrix of Represents the Laplacian matrix of a graph.

7. The drug repositioning prediction method based on multi-similarity network and knowledge graph according to claim 1, characterized in that: The fusing of the attribute feature representation of the drug and the disease with the topological feature representation includes: The attribute feature representation of the drug or disease and the topological feature representation are fused by using an attention feature fusion method. The calculation formula of the attention feature fusion is: Where Z represents the fusion feature vector of drug or disease, represents the disjoint union operation. represents the element-by-element product operation, M represents the attention weight generated using the multi-scale channel attention module, G represents the attribute feature representation of the drug or disease, and Y represents the topological feature representation of the drug or disease; 8. The drug repositioning prediction method based on multi-similarity network and knowledge graph according to claim 1, characterized in that: The training of drug-disease heterogeneous network comprises: The training is performed in the graph convolutional neural network. The training process includes intra-domain feature extraction and inter-domain feature extraction. The extracted intra-domain features and inter-domain features are merged to obtain the output embedding of each layer node. The expression of the output embedding is: In the formula, H l+1 represents the output embedding of the l-th layer node, represents the in-domain output features of the lth layer, represents the inter-domain output feature of the lth layer, H l represents the input features of the lth layer, represents the in-domain output features of the l-th layer of drugs, represents the inter-domain output features of the l-th layer of drugs, represents the input features of the l-th layer of drugs, represents the in-domain output features of the disease at layer l, represents the inter-domain output features of the disease at layer l, Represents the input features of the disease at layer l; Then, different attentions are used to integrate the output embeddings to obtain the embedding information of drugs and diseases. The expression of the embedding information is: In the formula, H R ∈R N×k represents the final embedding of the drug, R N×k represents the final embedding size of N drugs with k nearest neighbor similar nodes, H D ∈R M×k represents the final embedding of the disease, R M×k represents the final embedding size of M drugs with k nearest neighbor similar nodes, L represents the hyperparameter, β l ∈R is automatically initialized by the neural network learning It represents the contribution of the embedded data of different convolutional layers to the final embedded data.

9. The drug repositioning prediction method based on multi-similarity network and knowledge graph according to claim 1, characterized in that: The reconstruction of the relationship between drugs and diseases includes: The definition of the decoder function used in reconstructing the association between drugs and diseases is as follows: In the formula, represents the predicted probability score matrix, H R represents the final embedding of the drug, H D represents the final embedding of the disease, sigmoid represents the activation function, and H D T represents the transpose of the final embedding of disease, drug r i With disease j The correlation prediction score between The corresponding (i,j) term of is given; At the same time, the parameters are learned by minimizing the weighted binary cross entropy loss. The loss function is defined as follows: In the formula, loss represents the loss function, N represents the number of drug nodes, and M represents the number of disease nodes. Indicates drug r i and diseases j The association prediction score of , log represents the logarithmic function, (i, j) represents the drug-disease pair, S + represents the set of all known drug-disease association information, S - represents the set of all unknown or unobserved drug-disease association information, and the balance factor Used to reduce the impact of data imbalance, |S - | and |S + |It is S + With S - The number of pairs in ; Finally, the Adam optimizer is used to optimize the drug-disease heterogeneous network and initialize the weights of each parameter in the drug-disease heterogeneous network.

Images