Drug relocation method based on multi-source biological information and graph neural network

Abstract

The invention discloses a drug relocation method based on multi-source biological information and a graph neural network, and the method comprises the steps: firstly integrating drug and disease information, and constructing a relation network containing homogeneous and heterogeneous information; on the basis, local and global features are extracted at the same time by using a graph convolution operation and a multi-head attention mechanism, a deep residual network is fused, model training is accelerated, a feature-oriented down-sampling strategy and a weighted cross entropy loss function are adopted, and more negative sample information is incorporated into a model training process. Experimental results show that the GADRC model is superior to six latest baseline methods. Case research and molecular docking experiments further verify the effectiveness of the method in predicting the potential incidence relation of the drug diseases, and the method has practical application value.

Description

Technical Field

[0001] The present invention relates to the technical field of drug repositioning, and in particular to a drug repositioning technology based on multi-source biological information and graph neural networks. Background Art

[0002] Drug repositioning is an innovative drug development strategy that aims to discover new uses for existing drugs. This process significantly shortens the drug development cycle and reduces development costs. Studies have shown that drug repositioning can save up to 85% of new drug development costs while contributing 25% to the pharmaceutical industry's annual revenue. With the increase in clinical cases, such as the successful application of thalidomide in the treatment of leprosy complications and multiple myeloma, the clinical value of drug repositioning has become increasingly prominent. By leveraging the safety and efficacy of existing drugs, drug repositioning not only increases the success rate of new treatment options, but also provides an effective response plan for responding to public health emergencies such as the COVID-19 epidemic, and many candidate drugs have quickly entered the clinical trial stage.

[0003] Although significant progress has been made in the research of drug repositioning, it still faces many challenges at the technical level. Currently, computational methods are mostly divided into two categories: machine learning and deep learning. Machine learning models rely on a large amount of high-quality data and require a lot of manual intervention in feature selection and processing, which limits their application effect. In addition, many existing models focus on the extraction of local features and fail to fully reflect the complex interactions between drugs and diseases. Although deep learning models have shown stronger feature learning capabilities, they are still affected by data noise and feature learning limitations. These problems lead to deficiencies in the predictive performance and generalization capabilities of drug repositioning. Summary of the invention

[0004] In view of the defects of the prior art, the present invention provides a drug repositioning method based on multi-source biological information and graph neural network.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] The present invention is a drug repositioning method based on multi-source biological information and graph neural network, comprising the following steps:

[0007] S1. Data processing stage: construct a relationship network, which covers three different relationships: the association between drugs and diseases, the similarity between drugs, and the similarity between diseases.

[0008] S2, information aggregation stage: According to the relationship network constructed in step 1, the first features of drugs and diseases are obtained respectively through graph convolution operation, and the second features of drugs and diseases are obtained through residual learning mechanism. The first and second features of drugs and diseases are fused to obtain the third feature, and the fourth feature is obtained by multi-head attention mechanism.

[0009] S3, prediction stage: multiply the fourth feature of the drug constructed in step 2 with the fourth feature of the disease, and use the dot product decoder to obtain the correlation prediction between the drug and the disease, and reposition the drug.

[0010] S4. Method performance evaluation: Ten-fold cross validation is used to comprehensively evaluate the performance of the proposed method. The specific evaluation indicators include Accuracy (Acc), precision (Pre), recall (Recall), F1-score (F1), AUC and AUPR, and a comparative analysis is performed with the classic model.

[0011] Furthermore, the specific sub-steps of step S1 are as follows:

[0012] S1-1. Data collection: Drug data comes from the DrugBank database, using the SMILES sequence of the drug to obtain the chemical structure information of the drug. Disease data mainly comes from the OMIM database, using the clinical manifestations and phenotype information of the disease. The correlation between the two comes from the CTD database.

[0013] S1-2. Similarity calculation: Use the RDKit tool to obtain drug fingerprint similarity and obtain a drug similarity matrix with a matrix size of R X×X , X represents the total number of drugs. Use MinMiner tool to calculate the disease phenotype similarity and obtain the disease similarity matrix. The matrix size is R Y×Y , Y represents the total number of diseases. Since the single matrix needs to be completed, Gaussian interaction spectrum kernel similarity (GIP) is introduced.

[0014] S1-3, Similarity fusion: In order to solve the problem that the single similarity matrix of drugs and diseases is too sparse, the SMILES sequence similarity and GIP similarity of drugs are fused. According to the similarity between every two nodes in the similarity matrix, a threshold is set. Values ​​above the threshold represent association, and values ​​below the threshold represent no association, thereby constructing a relationship network.

[0015] Furthermore, the specific sub-steps of step S2 are as follows:

[0016] S2-1, Feature learning: First, the three relationship networks obtained in steps 1-3 are processed through convolution operations to aggregate the neighbor features of each node to enhance the node representation ability by utilizing the topological structure of the graph. In the feature update process, not only the neighbor node information is considered, but also self-connection is introduced to ensure that each node retains its own characteristics during feature propagation. Combined with the learnable parameter matrix, the first features of drugs and diseases are finally obtained.

[0017] S2-2. Improve prediction stability: Introduce the short-circuit link and residual learning mechanism of the deep residual network to obtain the second characteristics of drugs and diseases. The residual learning mechanism effectively captures the complex nonlinear relationship between the first characteristic nodes, and the multi-layer feature extraction capability significantly reduces the information loss in high-dimensional sparse data processing, alleviates the risk of gradient vanishing and explosion, and thus ensures the stability and consistency of the prediction results.

[0018] S2-3, optimize node information aggregation: First, the first feature and the second feature are fused in equal proportion to obtain the third feature. Then, the third feature is mapped to a new representation space through a linear transformation, and the activation function LeakReLU is used to process these representations. The attention coefficient is calculated through a learnable parameter vector to determine the weight of each neighbor node. According to these weights, the features of the neighbor nodes are weighted and summed to obtain the output of each head. Finally, the outputs of all heads are integrated by connection or averaging to obtain the fourth feature representation of drugs and diseases.

[0019] Furthermore, the specific sub-steps of step S3 are as follows:

[0020] S3-1, association prediction probability: multiply the fourth characteristic matrix of the drug constructed in step 2-3 with the fourth characteristic matrix of the disease, and use the dot product decoder to obtain the prediction probability y′ between the drug and the disease ij .

[0021] S3-2. Loss function: This method uses the weighted cross entropy loss function L to quantify the gap between the predicted value and the true label.

[0022] Furthermore, the specific sub-steps of step S4 are as follows:

[0023] S4-1. Feature-guided downsampling: In order to effectively solve the evaluation bias caused by unbalanced data distribution, the present invention adopts a ten-fold cross-validation strategy of feature-guided downsampling to comprehensively evaluate the model performance. Specifically, negative samples that are ten times the number of positive samples are randomly selected and combined with positive samples to form a small data set, which is then divided into ten mutually exclusive subsets. In each iteration, nine subsets are used as training sets, and the remaining subset is used as a validation set, and the cycle is repeated ten times to obtain the average result.

[0024] S4-2. Evaluation indicators: Specific evaluation indicators include AUC (the percentage of the area under the ROC curve to the total area), AUPR (the percentage of the area under the PR curve to the total area), and F1 score, which can indicate the performance of the model by counting each predicted answer in the entire sample and is calculated as: Where Pre is the accuracy of the sample, which represents the percentage of correct predictions of the model, calculated as Rec is the regression value of the sample, indicating how many of all standard label answers the model correctly identified, calculated as Acc is one of the commonly used evaluation indicators and is calculated as Where TP, TN, FP and FN represent the true positive rate, true negative rate, false positive rate and false negative rate respectively.

[0025] S4-3. Case study: This method sets all drug-disease samples as unknown, and then uses this method to generate association scores for two common specific clinical manifestations, Alzheimer's disease and breast cancer. In the generated prediction results, these scores are sorted in descending order, and it is proved that most of these drugs are related to the disease based on the literature in PubMed.

[0026] S4-4, Molecular docking: This method conducted molecular docking experiments for Alzheimer's disease to evaluate the interaction between candidate drugs and related target proteins. Five target proteins related to Alzheimer's disease were selected, and molecular docking analysis was performed using AutoDockVina software to calculate the binding energy between these target proteins and the top ten predicted potential drugs. The results showed that oxcarbazepine, risperidone, and tocapruone showed high binding energy values, indicating that they have significant binding ability with the target protein.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] 1. Dual-image feature extraction: This paper uses a dual-image encoder to extract features from homogeneous and heterogeneous networks, capturing neighbor and global embeddings, which promotes a comprehensive and in-depth study of drug- and disease-related attention-aware representations.

[0029] 2. Integrate more sample information: Innovatively introduce a "feature-guided downsampling method" to construct training and test data sets, and combine it with a weighted cross entropy loss function to capture sample features more comprehensively.

[0030] 3. Residual acceleration: This method uses deep residual neural networks to accelerate model convergence, alleviate the risks of gradient explosion and gradient vanishing, and enhance the predictive performance of the model.

[0031] 4. It can better guide biological experiments and drug development: Case analysis of specific diseases and molecular docking experiments show that this method can not only reveal the potential relationship between drugs and diseases, but also deeply explore the mechanism of action and toxicity of drugs, which provides a more important basis and reference for the clinical testing of new drugs. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be 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 labor.

[0033] Figure 1 This is the overall flow chart of the drug repositioning method based on multi-source biological information and graph neural network.

[0034] Figure 2 This is a result diagram of a molecular docking experiment of oxcarbazepine and 4EY5 targeting protein using the method of the present invention. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] like Figure 1 , an embodiment of the present invention provides a drug repositioning method based on multi-source biological information and graph neural network:

[0038] Drug repositioning methods based on multi-source biological information and graph neural networks include:

[0039] S1. Data processing stage: construct a relationship network, which covers three different relationships: the association between drugs and diseases, the similarity between drugs, and the similarity between diseases.

[0040] Specifically, in the embodiment of the present invention, step S1 includes the following steps:

[0041] S1-1. Data collection: Drug data comes from the DrugBank database, using the SMILES sequence of the drug to obtain the chemical structure information of the drug. Disease data mainly comes from the OMIM database, using the clinical manifestations and phenotype information of the disease. The correlation between the two comes from the CTD database.

[0042] S1-2. Similarity calculation: Use the RDKit tool to obtain drug fingerprint similarity and obtain a drug similarity matrix with a matrix size of R X×X , X represents the total number of drugs. Use MinMiner tool to calculate the disease phenotype similarity and obtain the disease similarity matrix. The matrix size is R Y×Y , Y represents the total number of diseases. Due to the completion of the single matrix, the Gaussian interaction spectrum kernel similarity (GIP) is introduced, and the drug r i and r j The GIP between is calculated as:

[0043] D rg (r i ,r j )=exp(-φ r ||IP(r i )-IP(r j )|| 2 )

[0044]

[0045] Among them, IP(r i ) represents the drug-disease association matrix, drug r i The corresponding column, φ′ r Equal to 1, similarly, disease s i and j The GIP between can be expressed as:

[0046] D sg (s i ,s j )=exp(-φ s ∥IP(s i )-IP(s j )∥ 2 )

[0047]

[0048] Among them, IP(s i ) represents the disease s in the drug-disease association matrix i The corresponding column, φ′ s Equal to 1.

[0049] S1-3. Similarity fusion: In order to solve the problem that the single similarity matrix of drugs and diseases is too sparse, the SMILES sequence similarity and GIP similarity of drugs are fused. The formula is as follows:

[0050]

[0051]

[0052] According to the similarity between every two nodes in the similarity matrix, a threshold is set. Nodes above the threshold represent a relationship, while nodes below the threshold represent no relationship, thus constructing a relationship network.

[0053] S2, information aggregation stage: According to the relationship network constructed in step 1, the first features of drugs and diseases are obtained respectively through graph convolution operation, and the second features of drugs and diseases are obtained through residual learning mechanism. The first and second features of drugs and diseases are fused to obtain the third feature, and the fourth feature is obtained by multi-head attention mechanism.

[0054] Specifically, in the embodiment of the present invention, step S2 includes the following steps:

[0055] S2-1. Feature learning: First, the convolution operation is used to process the three relationship networks obtained in step 1. By aggregating the neighbor features of each node, the topological structure of the graph is used to enhance the representation ability of the node. Specifically, in the feature update process, not only the information of the neighbor nodes is considered, but also self-connection is introduced to ensure that each node retains its own features during feature propagation. Combined with the learnable parameter matrix, the first features of drugs and diseases are obtained. The node feature update is expressed as:

[0056]

[0057] Among them, N(i) represents the neighbor set of node i, x′ i is the updated feature vector of node i, x j represents the feature vector of node j. j,i is the edge weight from node j to node i, and are the degrees of node i and node j respectively.

[0058] S2-2. Improve prediction stability: Introduce the short-circuit link and residual learning mechanism of the deep residual network to obtain the second feature of drugs and diseases. The residual learning mechanism effectively captures the complex nonlinear relationship between the first feature nodes, and the multi-layer feature extraction capability significantly reduces the information loss of processing high-dimensional sparse data, alleviates the risk of gradient vanishing and gradient explosion, and ensures the stability and consistency of the prediction results. The processing steps are expressed as:

[0059] Y=F(X′)+X′

[0060] F(X′)=σ(B(G(X′)))

[0061] Among them, X′ represents the first feature of the input, σ represents the sigmoid activation function, B represents batch normalization, G represents the convolution operation, and Y represents the obtained second feature.

[0062] S2-3, optimize node information aggregation: first, fuse the first and second features in equal proportion to obtain the third feature, map the third feature to a new representation space through linear transformation, use the activation function LeakReLU to process these representations, and use a learnable parameter vector to calculate the attention coefficient and determine the weight of each neighbor node. Then, based on these weights, perform weighted summation on the features of the neighbor nodes to obtain the output of each head. Finally, the outputs of all heads are integrated by concatenation or averaging to obtain the fourth feature representation of drugs and diseases. The node feature update is expressed as:

[0063]

[0064] In the formula, H i new represents the updated feature of node i, which combines the information of its neighbors and its own features. is the set of neighbor nodes of node i. i,j is the attention coefficient of node i to neighbor node j, reflecting the importance of neighbor nodes in the update process. j H represents the characteristics of neighbor node j, and W is the linear transformation matrix. i is the original feature of node i, which is combined with the updated feature through the residual connection to retain the historical information of the node.

[0065] S3, prediction stage: multiply the fourth feature of the drug constructed in step 2 with the fourth feature of the disease, and use the dot product decoder to obtain the correlation prediction between the drug and the disease, and reposition the drug.

[0066] Specifically, in the embodiment of the present invention, step S3 includes the following steps:

[0067] S3-1, association prediction probability: multiply the fourth characteristic matrix of the drug constructed in step 2-3 with the fourth characteristic matrix of the disease, and use the dot product decoder to obtain the prediction probability y′ between the drug and the disease ij

[0068] S3-2. Loss function: This method uses the weighted cross entropy loss function L to quantify the gap between the predicted value and the true label.

[0069] S4. Method performance evaluation: Ten-fold cross validation is used to comprehensively evaluate the performance of the proposed method. The specific evaluation indicators include Accuracy (Acc), precision (Pre), recall (Recall), F1-score (F1), AUC and AUPR, and a comparative analysis is performed with the classic model.

[0070] Specifically, in the embodiment of the present invention, step S4 includes the following steps:

[0071] S4-1. Feature-guided downsampling: In order to effectively solve the evaluation bias caused by unbalanced data distribution, the present invention uses a ten-fold cross-validation strategy of feature-guided downsampling to comprehensively evaluate the model performance. Specifically, negative samples that are ten times the number of positive samples are randomly selected and combined with the original positive samples into a small data set, which is then divided into ten mutually exclusive subsets. In each iteration, nine subsets are used as training sets, and the remaining subset is used as a validation set. The average result is obtained by iterating ten times.

[0072] S4-2. Evaluation indicators: Specific evaluation indicators include AUC (the percentage of the area under the ROC curve to the total area), AUPR (the percentage of the area under the PR curve to the total area), and F1 score, which can indicate the performance of the model by counting each predicted answer in the entire sample and is calculated as: Where Pre is the accuracy of the sample, which represents the percentage of correct predictions of the model, calculated as Rec is the regression value of the sample, indicating how many of all standard label answers the model correctly identified, calculated as Acc is one of the commonly used evaluation indicators and is calculated as Where TP, TN, FP and FN represent the true positive rate, true negative rate, false positive rate and false negative rate respectively.

[0073] S4-3. Case study: This method sets all drug-disease samples as unknown, and then uses this method to generate association scores for two common specific clinical manifestations, Alzheimer's disease and breast cancer. In the generated prediction results, these scores are sorted in descending order, and it is proved that most of these drugs are related to the disease based on the literature in PubMed.

[0074] S4-4, Molecular docking: This method conducted molecular docking experiments for Alzheimer's disease to evaluate the interaction between candidate drugs and related target proteins. We selected five target proteins related to Alzheimer's disease and used AutoDock Vina software for molecular docking analysis to calculate the binding energy between these target proteins and the top ten predicted potential drugs. By summarizing the results, we found that oxcarbazepine, risperidone and tocapruone showed relatively high binding energy values, indicating that they have significant binding ability with the target protein.

[0075] Table 1 and Table 2 are the results of case studies on Alzheimer's disease and breast cancer using the method of the present invention.

[0076]

[0077] Table 2:

[0078]

[0079] Tables 3 and 4 are performance comparisons of the proposed method and six benchmark methods: deepDR, GCNGAT, HNet-DNN, HINGRL, DRWBNCF, and DDAGDL on two benchmark datasets.

[0080] Table 3: shows the data comparison between the method of the present invention and six drug repositioning methods on the C dataset:

[0081]

[0082] Table 4: shows the data comparison between the method of the present invention and six drug repositioning methods on the F dataset:

[0083]

[0084] The present invention proposes a drug repositioning method based on multi-source biological information and graph neural network. First, the drug and disease information are integrated and a relationship network containing homogeneous and heterogeneous information is constructed. On this basis, the graph convolution operation and multi-head attention mechanism are used to simultaneously extract local and global features, and the deep residual network is integrated to accelerate model training. The feature-oriented downsampling strategy and weighted cross entropy loss function are adopted to incorporate more negative sample information into the model training process. Experimental results show that the GADRC model outperforms the six latest baseline methods. Case studies and molecular docking experiments further verify the effectiveness of this method in predicting potential drug-disease associations, and it has practical application value.

[0085] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0086] The above are only preferred implementations of the present invention. The protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A drug repositioning method based on multi-source biological information and graph neural network first integrates drug and disease information and constructs a relational network containing homogeneous and heterogeneous information. On this basis, graph convolution operations and multi-head attention mechanisms are used to simultaneously extract local and global features, integrate deep residual networks, accelerate model training, adopt feature-oriented downsampling strategies and weighted cross entropy loss functions, incorporate more negative sample information into the model training process, and finally use dot product decoders to complete the prediction task. The method includes four stages: data processing, information aggregation, prediction, and method performance evaluation. The steps are as follows: S1, data processing stage: construct a relationship network, which covers three different relationships, namely the association relationship between drugs and diseases, the similarity relationship between drugs and drugs, and the similarity relationship between diseases; The detailed sub-steps of step S1 include the following steps: S1-1. Data collection: Drug data mainly comes from the DrugBank database, using the SMILES sequence of the drug to obtain the chemical structure information of the drug. Disease data mainly comes from the OMIM database, using the clinical manifestations and phenotype information of the disease. The correlation between the two comes from the CTD database; S1-2. Similarity calculation: Use the RDKit tool to obtain drug fingerprint similarity and obtain a drug similarity matrix with a matrix size of R X×X , X represents the total number of drugs, and the MinMiner tool is used to calculate the disease phenotype similarity to obtain the disease similarity matrix, the matrix size is R Y×Y , Y represents the total number of diseases. Due to the completion of the single matrix, the Gaussian interaction spectrum kernel similarity (GIP) is introduced. The drug r i and r j The GIP between is calculated as: D rg (r i ,r j )=exp(-φ r ||IP(r i )-IP(r j )|| 2 ) Among them, IP(r i ) represents the drug-disease association matrix, drug r i The corresponding column, φ′ r Equal to 1, similarly, disease s i and j The GIP between can be expressed as: D sg (s i ,s j )=exp(-φ s ||IP(s i )-IP(s j )|| 2 ) Among them, IP(s i ) represents the disease s in the drug-disease association matrix i The corresponding column, φ′ s is equal to 1; S1-3. Similarity fusion: In order to solve the problem that the single similarity matrix of drugs and diseases is too sparse, the SMILES sequence similarity and GIP similarity of drugs are fused. The formula is as follows: According to the similarity between every two nodes in the similarity matrix, a threshold is set. If it is higher than the threshold, it means there is a relationship, and if it is lower than the threshold, it means there is no relationship, and a relationship network is constructed; S2, information aggregation stage: According to the relationship network constructed in step 1, the first features of drugs and diseases are obtained through graph convolution operation, the second features of drugs and diseases are obtained through residual learning mechanism, the first and second features of drugs and diseases are fused to obtain the third feature, and the fourth feature is obtained by multi-head attention mechanism. The detailed sub-steps of step S2 include the following steps: S2-1. Feature learning: First, the convolution operation is used to process the three relationship networks obtained in step 1. By aggregating the neighbor features of each node, the topological structure of the graph is used to enhance the representation ability of the node. That is, in the feature update process, not only the information of the neighbor nodes is considered, but also self-connection is introduced to ensure that each node retains its own features during feature propagation. Combined with the learnable parameter matrix, the first features of drugs and diseases are obtained. The node feature update is expressed as: Among them, N(i) represents the neighbor set of node i, x′ i is the updated feature vector of node i, x j represents the feature vector of node j, e j,i is the edge weight from node j to node i, and are the degrees of node i and node j respectively; S2-2. Improve prediction stability: Introduce the short-circuit link and residual learning mechanism of the deep residual network to obtain the second features of drugs and diseases. The residual learning mechanism effectively captures the complex nonlinear relationship between the first feature nodes, and the multi-layer feature extraction capability significantly reduces the information loss of processing high-dimensional sparse data, alleviates the risk of gradient vanishing and gradient explosion, and ensures the stability and consistency of the prediction results. The processing steps are expressed as follows: Y=F(X′)+X′ F(X′)=σB(G(X′))) Among them, X′ represents the first feature of the input, σ represents the sigmoid activation function, B represents batch normalization, G represents the convolution operation, and Y represents the obtained second feature; S2-3, optimize node information aggregation: First, the first feature and the second feature are fused in equal proportion to generate the third feature. Then, the third feature is mapped to a new representation space through linear transformation, and the LeakReLU activation function is used to process these representations. Then, the attention coefficient is calculated using the learnable parameter vector to determine the weight of each neighbor node. According to these weights, the features of the neighbor nodes are weighted summed to obtain the output of each head. Finally, the outputs of all heads are integrated by concatenation or averaging to form the fourth feature representation of drugs and diseases. The node feature update is expressed as: In the formula, represents the updated features of node i, which combines the information of its neighbors and its own features; is the set of neighbor nodes of node i; α i,j is the attention coefficient of node i to neighbor node j, reflecting the importance of neighbor nodes in the update process, H j represents the characteristics of neighbor node j, W is the linear transformation matrix; H i is the original feature of node i, which is combined with the updated feature through residual connection to preserve the historical information of the node; S3, prediction stage: multiply the fourth feature of the drug constructed in step 2 with the fourth feature of the disease, and use the dot product decoder to obtain the correlation prediction between the drug and the disease, and reposition the drug; The detailed sub-steps of step S3 include the following steps: S3-1, association prediction probability: multiply the fourth characteristic matrix of the drug constructed in step 2-3 with the fourth characteristic matrix of the disease, and use the dot product decoder to obtain the prediction probability y′ between the drug and the disease ij ; S3-2, loss function: This method uses the weighted cross entropy loss function L to quantify the gap between the predicted value and the true label; S4. Method performance evaluation: Ten-fold cross validation is used to comprehensively evaluate the performance of the proposed method. The specific evaluation indicators include Accuracy (Acc), precision (Pre), recall (Recall), F1-score (F1), AUC and AUPR, and a comparative analysis is performed with the classic model; The detailed sub-steps of step S4 include the following steps: S4-1. Feature-guided downsampling: In order to effectively solve the evaluation bias caused by imbalanced data distribution, the present invention uses a ten-fold cross-validation strategy of feature-guided downsampling to comprehensively evaluate the model performance. Specifically, negative samples that are ten times the number of positive samples are randomly selected and combined with the original positive samples into a small data set, which is then divided into ten mutually exclusive subsets. In each iteration, nine subsets are used as training sets and the remaining subset is used as a validation set. The average result is obtained by iterating ten times in a cycle. S4-2. Evaluation indicators: Specific evaluation indicators include AUC (the percentage of the area under the ROC curve to the total area), AUPR (the percentage of the area under the PR curve to the total area), and F1 score, which can indicate the performance of the model by counting each predicted answer in the entire sample and is calculated as: Where Pre is the accuracy of the sample, which represents the percentage of correct predictions of the model, calculated as Rec is the regression value of the sample, indicating how many of all standard label answers the model correctly identified, calculated as Acc is one of the commonly used evaluation indicators and is calculated as Where TP, TN, FP and FN represent the true positive rate, true negative rate, false positive rate and false negative rate respectively; S4-3. Case study: This method sets all drug-disease samples as unknown, and then uses this method to generate association scores for two common specific clinical manifestations, Alzheimer's disease and breast cancer. In the generated prediction results, these scores are sorted in descending order, and it is proved that most of these drugs are related to the disease based on the literature in PubMed; S4-4. Molecular docking: This method conducted molecular docking experiments for Alzheimer's disease to evaluate the interaction between candidate drugs and related target proteins. We selected five target proteins related to Alzheimer's disease and used AutoDockVina software for molecular docking analysis to calculate the binding energy between these target proteins and the top ten predicted potential drugs. The results showed that the binding energy values ​​of oxcarbazepine, risperidone and tocapruone were relatively high, indicating that they had significant binding ability with the target proteins.

Images