Drug interaction and event prediction method and model based on attention neural network

Abstract

The present invention provides a drug interaction and event prediction method and model based on an attention neural network. Based on a deep attention neural network framework, a more comprehensive and integrated feature vector representation is obtained by integrating heterogeneous drug feature networks. Combined with a graph representation learning method, the differences in the effects of drugs on other drugs in different features and dimensions are analyzed. By using multiple data sources, the accuracy of drug interaction prediction and the robustness of the model are improved, the model is made interpretable, and the function of predicting drug interaction events is realized, which helps to understand the hidden mechanisms behind adverse drug reactions. At the same time, it can guide the rational combination of drugs and promote drug safety.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning and drug design and medical technology, and specifically relates to a drug interaction and event prediction method and model based on an attention neural network. Background Art

[0002] Drug interaction is the phenomenon that the effect of a drug is changed due to the recent or simultaneous use of other drugs. With the rapid increase in the number of approved drugs, combination drug prescriptions have become a common phenomenon for doctors to treat patients' diseases. However, combination drug use may lead to toxic side effects. Therefore, the detection of drug interactions is an important task of drug safety supervision, which can provide effective and safe co-prescriptions for multiple drugs and avoid adverse reactions and dangerous reactions. Since the traditional clinical trial process is complicated, costly and time-consuming, there is an urgent need to develop computer-assisted calculation methods to detect drug-drug interactions. In recent years, machine learning algorithms have been successfully applied in the field of biomedicine, such as gene expression, mass spectrometry-based proteomics, identification of gene-gene interactions in whole-genome association studies, and prediction of regulatory elements from DNA and protein sequences.

[0003] The importance of integrating heterogeneous drug features for predicting drug interactions has been emphasized in many studies. The large amount of information collected on drugs can be constructed into feature networks of different properties. A large number of potential connections and features of drugs are hidden in the network structure. Studies have shown that representation learning from multiple feature networks of different properties performs better than a single network, and integrating multiple drug networks may result in a more comprehensive and integrated feature vector representation.

[0004] However, integrating diverse drug features in drug interaction prediction is a challenging task. There are correlations between heterogeneous drug features and redundant information, which may affect the performance of traditional classifiers, so an effective heterogeneous feature integration framework is needed. Deep Neural Network (DNN) is a representative deep learning algorithm that aims to model high-level features of drugs using multiple nonlinear and complex processing layers, so it is a suitable model for learning high-level representations from multiple drug features.

[0005] Most existing methods assume that a drug has the same effect on other drugs, but a drug may have different effects on other drugs across different features and dimensions. Therefore, an attention mechanism network can be used to capture the different attention vectors of each pair of drugs and automatically assign different weights to different features and dimensions of the drugs, thereby improving the prediction performance.

[0006] In addition to predicting the existence of drug interactions, drug interaction events are also worthy of attention. Drug interaction events are a more refined description of known drug interactions. Predicting drug interaction events helps to understand the mechanisms behind adverse reactions. Most existing methods are used to predict whether two drugs interact with each other. However, drug interactions can lead to different biological consequences or events, so it is necessary to conduct a deeper event analysis of drug interactions. Summary of the invention

[0007] The technical problem to be solved by the present invention is to provide a drug interaction and event prediction method and model based on an attention neural network, which is used to improve the accuracy of drug interaction prediction and predict drug interaction events while using multiple data sources.

[0008] The technical solution adopted by the present invention to solve the above technical problems is: a drug interaction and event prediction method based on an attention neural network, characterized in that it includes the following steps:

[0009] S1: Build a drug interaction and event prediction model based on attention neural network, including drug feature representation learning module, drug pair feature representation learning module, drug interaction and event prediction module; drug feature representation learning module includes multi-source feature acquisition module, multiple feature network construction module and SDNE module; drug pair feature representation learning module includes drug comprehensive feature representation splicing module, drug pair feature representation splicing module and ANN module; drug interaction and event prediction module includes DNN module; source feature acquisition module, multiple feature network construction module, SDNE module, drug comprehensive feature representation splicing module, drug pair feature representation splicing module, ANN module, DNN module are connected in sequence according to signal flow direction;

[0010] S2: construct multiple drug feature networks and use graph representation learning methods to learn drug feature representations from multiple drug feature networks;

[0011] S3: Concatenate the feature representations of drugs learned from multiple networks to obtain a comprehensive feature representation vector for each drug, and learn the feature representation of drug pairs through the attention neural network ANN;

[0012] S4: Input the feature representation of drug pairs into a deep neural network (DNN), and use the deep neural network to predict potential drug interactions and events;

[0013] S5: Training and optimizing attention neural network-based drug interaction and event prediction models.

[0014] According to the above scheme, in step S2, the specific steps are:

[0015] S21: Collect multiple common drug features including drug chemical substructures, targets, enzymes, pathways and drug interactions;

[0016] S22: construct multiple feature networks for each drug, including drug-substructure network, drug-target network, drug-enzyme network, drug-reaction pathway network and drug-interaction network. The nodes of the feature network are drugs and features, and the links between nodes are the connections between drug nodes and feature nodes.

[0017] S23: Learning feature representations of drug nodes from feature networks via structured deep network embedding (SDNE) while maintaining first-order and second-order proximities.

[0018] Furthermore, in step S23, the specific steps are as follows: the original features of the nodes are input into an L-layer autoencoder network, the structured deep network is embedded in SDNE, and the first-order and second-order proximity of the drug pair are optimized simultaneously, and the first-order proximity constrains the potential representation vector of a pair of nodes i and j. and vector The second-order proximity is calculated by minimizing the output vector and input x i The reconstruction error of is reconstructed; let ⊙ be the Hadamard product, is the reconstruction data of the automatic encoder, b i is the input x i The penalty weight of the reconstruction error is:

[0019]

[0020] If there is a link between node i and node j, then the adjacency matrix M corresponding to the feature network is i,j =1, b i,j =β>1; if there is no link between node i and node j, then M i,j =0, b i,j =1;

[0021] Let node v i The output of the Lth layer is is the learned node feature representation; W (l) and is the weight matrix of the Lth layer, α is the parameter controlling the first-order loss, and λ is the parameter of the regularization term; then the optimization objective function of the drug feature representation learning module embedded in SDNE through the structured deep network is:

[0022]

[0023] The first term of the objective function L controls the second-order proximity to extract global features, the second term controls the first-order proximity to extract local features, and the third term is the L2 norm adjustment term; the learned feature representation maintains local and global structures; each drug learns 5 drug feature representations, which are: and

[0024] Furthermore, in step S3, the specific steps are:

[0025] S31: Five feature representations of spliced ​​drugs and Get the comprehensive feature representation vector for each drug

[0026] S32: Learning feature representations of drug pairs by fusing comprehensive feature representations of drugs through attention neural network ANN;

[0027] S33: The attention neural network ANN captures different attention vectors for each drug pair and assigns different weights to different features and dimensions of the drugs.

[0028] Furthermore, in step S32, the specific steps are: let the characteristic representation of the drug pair consisting of drug i and drug j be F i,j ⊙ is a bitwise multiplication operation; the comprehensive characteristics of each drug in the drug pair are represented by E i and E j ; K is the dimension of the drug comprehensive feature representation vector; a i,j =(a i,j,1 ,a i,j,2 ,…,a i,j,K ) is a K-dimensional attention vector, which is used to reflect the importance of different dimensions when fusing the representation vectors of drug pairs; the feature representation of drug pairs is defined as:

[0029] F i，j =a i，j ⊙(E i ⊙E j ).

[0030] Furthermore, in step S33, the specific steps are as follows: RELU is set as the activation function; b and W are the bias vector and the weight matrix respectively; V T is the weight vector; [E i ,E j ] is the drug characteristic representation E i and E j The connection vector of ; then:

[0031]

[0032] The k-th dimension attention vector a i,j,k for:

[0033]

[0034] Furthermore, in step S4, the specific steps are:

[0035] S41: The feature representation of the drug pair is used as the input of a deep neural network, and the dimension of the input layer of the deep neural network is the same as the dimension of the drug pair feature representation vector; through multiple fully connected hidden layers, the output layer of the deep neural network contains 2 neurons indicating whether the drug interaction occurs or not or 65 neurons indicating 65 drug interaction events;

[0036] S42: The softmax function is used to generate the probability of the output node, and the linear rectification function RELU is used as the activation function of all hidden layers.

[0037] According to the above scheme, in step S5, the specific steps are:

[0038] S51: training and optimizing drug feature representation learning module;

[0039] S52: End-to-end training and optimization of the drug pair feature representation learning module and the drug interaction and event prediction module; select the binary cross entropy loss function to optimize the model and use the Adam optimizer to optimize the drug interaction prediction module; between hidden layers, use batch normalization layers to accelerate convergence and random dropout layers to avoid overfitting and improve generalization ability.

[0040] The drug interaction and event prediction model based on attention neural network includes a drug feature representation learning module, a drug pair feature representation learning module, and a drug interaction and event prediction module; the drug feature representation learning module includes a multi-source feature acquisition module, a multiple feature network construction module and a SDNE module; the drug pair feature representation learning module includes a drug comprehensive feature representation splicing module, a drug pair feature representation splicing module and an ANN module; the drug interaction and event prediction module includes a DNN module; the source feature acquisition module, the multiple feature network construction module, the SDNE module, the drug comprehensive feature representation splicing module, the drug pair feature representation splicing module, the ANN module, and the DNN module are connected in sequence according to the signal flow direction.

[0041] A computer storage medium stores a computer program executable by a computer processor, wherein the computer program executes a drug interaction and event prediction method based on an attention neural network.

[0042] The beneficial effects of the present invention are:

[0043] 1. The drug interaction and event prediction method and model based on the attention neural network of the present invention is based on the deep attention neural network framework, by integrating heterogeneous drug feature networks and combining graph representation learning methods, to analyze the differences in the effects of drugs on other drugs in different features and dimensions. By using multiple data sources, the accuracy of drug interaction prediction and the robustness of the model are improved, the model is interpretable, and the function of predicting drug interaction events is realized.

[0044] 2. The present invention can not only predict the interactions between traditional drug pairs, but also predict drug interactions and their events through multiple heterogeneous data sources and deep attention neural networks, and further predict 65 drug interaction events, which helps to understand the mechanisms behind adverse drug reactions. At the same time, it can guide the rational combination of drugs and promote drug safety.

[0045] 3. Since the representation learning obtained from multiple networks with different properties performs better than that of a single network, the present invention uses existing data sets to provide a large amount of information on drugs, which is used to construct networks with different properties and containing a large number of potential connections and features. Multiple drug networks are integrated to obtain a more comprehensive and integrated feature vector representation.

[0046] 4. A drug may have different degrees of influence on other drugs across different features and dimensions. The present invention proposes a network with an attention mechanism to capture the different attention vectors of each pair of drugs. The attention neural network automatically assigns different weights to the different features and dimensions of the drugs.

[0047] 5. Compared with traditional classifiers that rely heavily on feature engineering and redundant feature processing, the deep neural network used in the present invention has a strong advantage in automatically extracting high-quality feature sets, thereby improving the accuracy of classification and prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a functional block diagram of an embodiment of the present invention.

[0049] Figure 2 is a flow chart of an embodiment of the present invention.

[0050] Figure 3 It is a visualization diagram of the average attention weights of five drug features in an embodiment of the present invention.

[0051] Figure 4 A diagram showing the types of drug interaction events predicted by an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0053] See also Figure 1 The drug interaction and event prediction model based on attention neural network of the present invention includes a drug feature representation learning module, a drug pair feature representation learning module, and a drug interaction and event prediction module; the drug feature representation learning module includes a multi-source feature acquisition module, a multiple feature network construction module and an SDNE module; the drug pair feature representation learning module includes a drug comprehensive feature representation splicing module, a drug pair feature representation splicing module and an ANN module; the drug interaction and event prediction module includes a DNN module; the source feature acquisition module, the multiple feature network construction module, the SDNE module, the drug comprehensive feature representation splicing module, the drug pair feature representation splicing module, the ANN module, and the DNN module are connected in sequence according to the signal flow direction.

[0054] See also Figure 2 The drug interaction and event prediction method based on attention neural network of the present invention comprises the following steps:

[0055] S1: Build a drug interaction and event prediction model based on attention neural network;

[0056] S2: Construct multiple drug feature networks and use graph representation learning methods to learn drug feature representations from multiple drug feature networks:

[0057] S21: Collect multiple common drug features including drug chemical substructures, targets, enzymes, pathways and drug interactions;

[0058] See Table 1. Five common drug feature data, namely drug chemical substructure, target, enzyme, pathway and drug interaction, were collected from the DrugBank database (version 5.1.0) released in April 2018. The drug dataset constructed by these drug features contains 841 drugs, 619 chemical substructures, 1333 targets, 214 enzymes and 307 reaction pathways. The target protein of the drug is mapped to the KEGG drug database using the ID mapping server to obtain the drug reaction pathway. There are 353,220 drug pairs for the 841 drugs, including 82,620 known pairwise drug interactions.

[0059] Table 1 Drug characteristics and source description

[0060]

[0061] S22: construct multiple feature networks for each drug, including drug-substructure network, drug-target network, drug-enzyme network, drug-reaction pathway network and drug-interaction network. The nodes of the feature network are drugs and features, and the links between nodes are the connections between drug nodes and feature nodes.

[0062] S23: Learn the feature representation of drug nodes from the feature network by simultaneously preserving the first - order and second - order proximities through SDNE (Structural Deep Network Embedding).

[0063] See Figure 2 , input the original features of the nodes into an L - layer auto - encoder network. SDNE simultaneously optimizes the first - order and second - order proximities of drug pairs. The first - order proximity constrains the similarity of the latent representation vectors and vector . The second - order proximity is reconstructed by minimizing the reconstruction error between the output vector and the input x i . Let ⊙ be the Hadamard product, be the reconstructed data of the auto - encoder, and b i be the penalty weight of the reconstruction error of the input x i . Then:[[]]

[0064]

[0065] If there is a link between node i and node j, then the corresponding adjacency matrix M i,j = 1, b i,j = β>1; if there is no link between node i and node j, then M i,j = 0, b i,j = 1;

[0066] Let the output of the L - th layer of node v i be which is the learned node feature representation; W (l) and are the weight matrices of the L - th layer, α is the parameter controlling the first - order loss, and λ is the parameter of the regularization term. Then the optimization objective function of the drug feature representation learning module through SDNE is:[[]]

[0067]

[0068] The objective function L has three terms. The first term controls the second - order proximity to extract global features, the second term controls the first - order proximity to extract local features, and the third term is the L2 - norm regularization term. Therefore, the learned feature representation preserves the local and global structures. Finally, each drug learns 5 drug feature representations, which are respectively:[[]] and

[0069] S3: Concatenate the feature representations of drugs learned from multiple networks to obtain a comprehensive feature representation vector for each drug, and learn the feature representation of drug pairs through an attention neural network (ANN):

[0070] S31: Five feature representations of spliced ​​drugs and Get the comprehensive feature representation vector for each drug

[0071] S32: Learning feature representations of drug pairs by fusing comprehensive feature representations of drugs via attention neural networks;

[0072] Assume that the characteristic of the drug pair consisting of drug i and drug j is represented by F i,j ⊙ is a bitwise multiplication operation; the comprehensive characteristics of each drug in the drug pair are represented by E i and E j ; K is the dimension of the drug comprehensive feature representation vector; a i,j =(a i,j,1 ,a i,j,2 ,…,a i,j,K ) is a K-dimensional attention vector, which is used to reflect the importance of different dimensions when fusing the representation vectors of drug pairs; the feature representation of drug pairs is defined as:

[0073] F i，j =a i，j ⊙(E i ⊙E j );

[0074] S33: The attention neural network captures different attention vectors for each drug pair and assigns different weights to different features and dimensions of the drugs.

[0075] Let RELU be the activation function; b and W are the bias vector and weight matrix respectively; V T is the weight vector; [E i ,E j ] is the drug characteristic representation E i and E j The connection vector of ; then:

[0076]

[0077] The k-th dimension attention vector a i,j,k for:

[0078]

[0079] See also Figure 2 , the attention neural network outputs the feature representation vector of the drug pair. Figure 3 , the average attention weights of all drug pairs on the five drug features are visualized; the attention weights of drug pairs intuitively reflect the different contributions of the five drug features to the prediction of drug interactions.

[0080] S4: The feature representation of the drug pair is input into a deep neural network (DNN), and the deep neural network is used to predict potential drug interactions and events; the output layer of the deep neural network contains 2 neurons for predicting drug interactions or 65 neurons for predicting drug interaction events.

[0081] S41: Predicting whether a drug interaction occurs is considered a binary classification problem, and predicting the occurrence of a drug interaction is considered a multivariate classification problem; see Figure 2 , the feature representation of the drug pair is used as the input of the deep neural network, and the dimension of the input layer of the deep neural network is the same as the dimension of the drug pair feature representation vector; the drug pair feature representation vector is used as input and passes through multiple fully connected hidden layers; the output layer of the deep neural network contains 2 neurons indicating whether the drug interaction occurs or not, or 65 neurons indicating 65 drug interaction events;

[0082] S42: The softmax function is used to generate the probability of the output node, and ReLU (Rectified Linear Unit) is used as the activation function of all hidden layers.

[0083] S5: Training and optimizing the model:

[0084] S51: training and optimizing drug feature representation learning module;

[0085] S52: End-to-end training and optimization of the drug pair feature representation learning module and the drug interaction and event prediction module; select the binary cross entropy loss function to optimize the model and use the Adam optimizer to optimize the drug interaction prediction module; between hidden layers, use batch normalization layers to accelerate convergence and random inactivation (dropout) layers to avoid overfitting and improve generalization ability.

[0086] Referring to Table 2, the three modules (SDNE structured deep network embedding + ANN attention neural network + DNN deep neural network) in the model of the present invention are replaced or added or deleted to obtain different variants of the model, and compared with the model of the present invention. From the results:

[0087] The drug feature representation learning module using SDNE in the present invention has the best prediction performance compared with other graph representation learning methods (Node2vec, Metapath2vec and LINE); the drug pair feature representation learning module using ANN in the present invention can capture different attention vectors of each pair of drug pairs and automatically assign different weights to different features and dimensions of the drugs, thereby improving the prediction performance; the drug interaction and event prediction module using DNN in the present invention is better at learning advanced representation models from multiple drug features than other classifiers, and has better prediction performance. Therefore, the effectiveness of the method and model of the present invention has been verified experimentally.

[0088] Table 2 Prediction performance comparison table

[0089]

[0090] As shown in Table 3, the present invention also effectively predicts new drug interactions. There are 353,220 drug pairs for the 841 drugs collected, 82,620 known drug interactions and 270,600 unlabeled drug pairs, which may contain potential drug interactions between drugs. All known drug-drug interactions are used to train the model to predict drug interactions not included in the data set. Table 3 shows the top 20 drug interactions predicted by the present invention and 14 new drug interactions confirmed in the 5.1.7 version of the DrugBank database accessed on July 19, 2020:

[0091] Table 3 The top 20 deep drug interactions predicted by the present invention

[0092]

[0093]

[0094] Table 4 shows the effectiveness of the present invention in predicting drug interaction events. Table 4 lists the 10 most common events, numbered from #1 to #10; the present invention uses Figure 4 The model is trained on the drug interaction events in , and then the drug interaction events of other drug pairs are predicted, and the top 20 predictions related to each event are checked. The evidence of drug interaction events predicted by the model can be found in the latest version 5.1.7 of DrugBank database:

[0095] Table 4 List of verified drug interaction events

[0096]

[0097]

[0098] From the above analysis, the present invention adopts a deep attention neural network framework to build a model, which integrates multiple drug features to predict unknown drug interactions and their events, and has good prediction effect and good performance. At present, the present invention can predict 65 types of drug interaction events, and more and more types of interaction events will be included in the present invention.

[0099] The above embodiments are only used to illustrate the design ideas and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design ideas disclosed by the present invention are within the protection scope of the present invention.

Claims

1. Drug interaction and event prediction method based on attention neural network, Features: The following steps are involved: S1: Build a drug interaction and event prediction model based on attention neural network, including drug feature representation learning module, drug pair feature representation learning module, drug interaction and event prediction module; The drug feature representation learning module includes a multi-source feature acquisition module, a multi-feature network construction module and a SDNE module; the drug pair feature representation learning module includes a drug comprehensive feature representation splicing module, a drug pair feature representation splicing module and an attention neural network ANN module; the drug interaction and event prediction module includes a DNN module; the source feature acquisition module, the multi-feature network construction module, the SDNE module, the drug comprehensive feature representation splicing module, the drug pair feature representation splicing module, the ANN module and the DNN module are connected in sequence according to the signal flow direction; S2: construct multiple drug feature networks and use graph representation learning methods to learn drug feature representations from multiple drug feature networks; The specific steps are: S21: Collect multiple common drug features including drug chemical substructures, targets, enzymes, pathways and drug interactions; S22: construct multiple feature networks for each drug, including drug-substructure network, drug-target network, drug-enzyme network, drug-reaction pathway network and drug-interaction network. The nodes of the feature network are drugs and features, and the links between nodes are the connections between drug nodes and feature nodes. S23: Learning feature representations of drug nodes from feature networks by embedding SDNE with structured deep networks while maintaining first-order and second-order proximity; S3: Concatenate the feature representations of drugs learned from multiple networks to obtain a comprehensive feature representation vector for each drug, and learn the feature representation of drug pairs through the attention neural network ANN; The specific steps are: S31: Five feature representations of spliced ​​drugs and Get the comprehensive feature representation vector for each drug S32: Learning feature representations of drug pairs by fusing comprehensive feature representations of drugs through attention neural network ANN; S33: The attention neural network ANN captures different attention vectors for each drug pair and assigns different weights to different features and dimensions of the drugs; S4: Input the feature representation of the drug pair into the deep neural network DNN, and use the deep neural network to predict potential drug interactions and events; the specific steps are: S41: The feature representation of the drug pair is used as the input of a deep neural network, and the dimension of the input layer of the deep neural network is the same as the dimension of the drug pair feature representation vector; through multiple fully connected hidden layers, the output layer of the deep neural network contains 2 neurons indicating whether the drug interaction occurs or not or 65 neurons indicating 65 drug interaction events; S42: The softmax function is used to generate the probability of the output node, and the linear rectification function RELU is used as the activation function of all hidden layers; S5: Training and optimizing the drug interaction and event prediction model based on attention neural network; the specific steps are: S51: training and optimizing drug feature representation learning module; S52: End-to-end training and optimization of the drug pair feature representation learning module and the drug interaction and event prediction module; select the binary cross entropy loss function to optimize the model and use the Adam optimizer to optimize the drug interaction prediction module; between hidden layers, use batch normalization layers to accelerate convergence and random dropout layers to avoid overfitting and improve generalization ability.

2. The drug interaction and event prediction method based on attention neural network according to claim 1, Features: In the step S23, the specific steps are: inputting the original features of the nodes into an L-layer autoencoder network, embedding the structured deep network SDNE and optimizing the first-order and second-order proximity of the drug pair at the same time, and the first-order proximity constrains the potential representation vector of a pair of nodes i and j and vector The second-order proximity is calculated by minimizing the output vector and input x i The reconstruction error of is reconstructed; let ⊙ be the Hadamard product, is the reconstruction data of the automatic encoder, b i is the input x i The penalty weight of the reconstruction error is: If there is a link between node i and node j, then the adjacency matrix M corresponding to the feature network is i,j =1, b i,j =β>1; if there is no link between node i and node j, then M i,j =0, b i,j =1; Let node v i The output of the Lth layer is is the learned node feature representation; W (l) and is the weight matrix of the Lth layer, α is the parameter controlling the first-order loss, and λ is the parameter of the regularization term; then the optimization objective function of the drug feature representation learning module embedded in SDNE through the structured deep network is: The first term of the objective function L controls the second-order proximity to extract global features, the second term controls the first-order proximity to extract local features, and the third term is the L2 norm adjustment term; the learned feature representation maintains local and global structures; each drug learns 5 drug feature representations, which are: and 3. The drug interaction and event prediction method based on attention neural network according to claim 1, Features: In step S32, the specific steps are: let the characteristic representation of the drug pair consisting of drug i and drug j be F i,j ⊙ is a bitwise multiplication operation; the comprehensive characteristics of each drug in the drug pair are represented by Ei and E j ; K is the dimension of the drug comprehensive feature representation vector; a i,j =(a i,j,1 ,a i,j,2 ,…,a i,j,K ) is a K-dimensional attention vector, which is used to reflect the importance of different dimensions when fusing the representation vectors of drug pairs; the feature representation of drug pairs is defined as: F i，j =a i，j ⊙(E i ⊙E j )。 4. The drug interaction and event prediction method based on attention neural network according to claim 3, Features: In step S33, the specific steps are: set RELU as the activation function; b and W are the bias vector and weight matrix respectively; V T is the weight vector; [E i ,E j ] is the drug characteristic representation E i and E j The connection vector of , then: The k-th dimension attention vector ai,j,k is:

5. A prediction model for the drug interaction and event prediction method based on attention neural network according to any one of claims 1 to 4, Features: It includes drug feature representation learning module, drug pair feature representation learning module, and drug interaction and event prediction module; The drug feature representation learning module includes a multi-source feature acquisition module, a multi-feature network construction module, and a SDNE module; The drug pair feature representation learning module includes a drug comprehensive feature representation splicing module, a drug pair feature representation splicing module and an ANN module; The drug interaction and event prediction module includes a DNN module; The source feature acquisition module, the multiple feature network construction module, the SDNE module, the drug comprehensive feature representation splicing module, the drug pair feature representation splicing module, the ANN module, and the DNN module are connected in sequence according to the signal flow direction.

6. A computer storage medium, Features: A computer program executable by a computer processor is stored therein, and the computer program executes the drug interaction and event prediction method based on the attention neural network as described in any one of claims 1 to 4.

Images