Drug-drug interaction prediction method based on the global structure of biological networks

Abstract

The present invention discloses a method for predicting drug-drug interactions based on the global structure of a biological network, comprising the following steps: S1, constructing a biological heterogeneous network containing drugs, proteins, and diseases; S2, selecting a set number of node pairs in the biological heterogeneous network, and labeling the node pairs into four different categories based on the length of the shortest path between the node pairs; S3, calculating the number of neighbors of each node, and converting the number of neighbors of all nodes into a one-hot encoding vector as the initial feature of the node; S4, inputting the adjacency matrix of the biological network and the initial features of the nodes into a graph attention neural network to predict the shortest path category of the node pair, and comparing it with the original path label of the node pair; S5, using a trained drug-drug interaction prediction model to predict drug-drug interactions. The present invention reduces the dependence of deep neural network models on labeled data.

Description

Technical Field

[0001] The present invention relates to the field of pharmaceutical technology, and more particularly to a method for predicting drug-drug interactions based on the global structure of a biological network. Background Art

[0002] Drug-drug interactions increase the risk of death or medication discontinuation. Studies have shown that 6.7% of hospitalized patients in the United States experience serious adverse drug reactions, resulting in a 0.32% mortality rate. However, the simultaneous use of multiple medications is essential for treating complex diseases. Therefore, predicting drug-drug interactions is a critical task in pharmacovigilance. However, traditional wet lab experiments are expensive and time-consuming, making them incapable of large-scale prediction of drug-drug interactions. Furthermore, the rapid and effective use of computer technology to predict drug-drug interactions has important scientific significance and value.

[0003] With the rapid development of deep learning technology in fields such as computer vision, natural language, and speech recognition, more and more research has begun to utilize deep learning for drug discovery in recent years, including drug-target interaction prediction and drug-drug interaction prediction. Graph deep learning technology, in particular, has been widely applied to drug-drug interaction prediction. This is primarily due to the massive amount of biomedical network data generated by high-throughput technologies, such as drug-drug interaction networks, drug-protein interaction networks, and drug-disease interaction networks. These networks not only simulate complex biological systems but are also highly suitable for graph neural network models. Simultaneously, the rapid development of artificial intelligence technology has led to the emergence of many advanced graph neural network models, such as graph convolutional networks, graph attention networks, and graph isomorphism networks. These graph network models possess powerful representational capabilities and can naturally integrate the connectivity patterns and characteristic attributes of graph-structured data.

[0004] However, it is well known that deep neural networks, including graph neural networks, rely heavily on large amounts of labeled data. For graphs with sparsely labeled nodes, large-scale data requirements may not be met in many practical scenarios. Similarly, biomedical networks are highly sparse and have very little labeled data, making it difficult to directly use graph neural networks for drug-drug interaction prediction. Secondly, existing graph neural networks are limited to aggregating information in local areas and ignore the associations between distant nodes, which can easily lead to oversmoothing. To this end, it is necessary to develop a drug-drug interaction prediction method based on the global structure of biological networks. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for predicting drug-drug interactions based on the global structure of biological networks to overcome the defects of the existing technology.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] A method for predicting drug-drug interactions based on the global structure of a biological network comprises the following steps:

[0008] S1. Extract drug-drug interaction networks, drug-target interaction networks, protein-protein interaction networks, drug-disease association networks, and disease-protein association networks, and compare and fuse them to obtain a biologically heterogeneous network containing drugs, proteins, and diseases;

[0009] S2. Select a set number of node pairs in the biological heterogeneous network, label the node pairs into four different categories based on the length of the shortest path between the node pairs, and use the selected node pairs and corresponding labels as training samples for self-supervised learning;

[0010] S3. Calculate the number of neighbors of each node and convert the number of neighbors of all nodes into a one-hot encoding vector as the initial feature of the node;

[0011] S4, inputting the adjacency matrix of the biological network and the initial features of the nodes into the graph attention neural network to predict the shortest path category of the node pair of the training sample in step S2, and comparing it with the original path label of the node pair;

[0012] S5. All drug pairs with interactions in the drug-drug interaction network are regarded as positive samples, and an equal number of unknown drug pairs are randomly selected from the drug-drug interaction network as negative samples. All positive samples and negative samples are mixed together and divided into training sets and test sets according to a certain ratio. The sample representations and labels in the training set are input into the deep neural network for model training. After the maximum number of iterations is reached, the drug-drug interaction prediction model is obtained, and the trained drug-drug interaction prediction model is used to predict drug-drug interactions.

[0013] Furthermore, in step S1, based on the biomedical big data of DrugBank, UniProt, HPRD, SIDER, CTD, NDFRT, repoDB, HuRI, DrugCentral, BioGRID, OMIM and STRING, drug-target interaction networks, protein-protein interaction networks, drug-disease association networks and disease-protein association networks are extracted.

[0014] Furthermore, in step S2, the node pairs are marked into four different categories according to the length of the shortest path between the node pairs, specifically: the shortest path between the node pairs is calculated, if the shortest path is less than the set value, the path length is marked as the class label of the corresponding node, if the shortest path is greater than or equal to the set value, the corresponding node is classified into a category with a path of 4.

[0015] Furthermore, in step S4, a multi-head attention mechanism is used in the graph attention network to aggregate the features of the node itself and its neighboring nodes, and the formula is:

[0016]

[0017] Where δ represents the nonlinear activation ReLU function, K is the number of multi-head attention, || represents the concatenation of feature vectors, and W k represents the learnable parameter matrix in the k-th attention mechanism, represents the input features of node j, N i are all neighbors of node i, is the attention coefficient learned in the k-th attention mechanism, and the calculation formula is:

[0018]

[0019] Where, is a hyperparameter, ε(·) represents the nonlinear activation function LeakyReLU, which is calculated as:

[0020]

[0021] Where θ is a hyperparameter.

[0022] Furthermore, in step S4, the cross entropy function is used to calculate the loss, and the Adam optimizer is used for training and updating parameters.

[0023] Furthermore, the deep neural network in step S5 includes an input layer, a hidden layer and an output layer, the hidden layer adopts a ReLU activation function, and the output layer adopts a sigmoid activation function.

[0024] Compared with the existing technology, the advantages of the present invention are: the present invention provides a drug-drug interaction prediction method based on the global structure of the biological network, which first integrates multiple network data and can capture multimodal network information; secondly, it designs a representation learning task based on biological network path length classification to guide the graph attention network model to generate high-quality drug representation vectors, thereby reducing the dependence of the deep neural network model on label data; at the same time, the representation learning sample integrates biological entity pairs from the global structure of the network, so that the generated representation vector captures the global structural information of the network, further improving the prediction performance of drug-drug interactions. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. 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 any creative work.

[0026] Figure 1 It is a framework diagram of the drug-drug interaction prediction method based on the global structure of biological networks of the present invention.

[0027] Figure 2 It is the experimental result figure of the present invention. DETAILED DESCRIPTION

[0028] The preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more precise definition of the protection scope of the present invention.

[0029] See Figure 1 and Figure 2 As shown, this embodiment discloses a method for predicting drug-drug interactions based on the global structure of a biological network, comprising the following steps:

[0030] Step S1, constructing a biological heterogeneous network: Based on the biomedical big data of DrugBank, UniProt, HPRD, SIDER, CTD, NDFRT, repoDB, HuRI, DrugCentral, BioGRID, OMIM and STRING, extract the drug-target interaction network, protein-protein interaction network, drug-disease association network and disease-protein association network, and compare and fuse them to obtain a biological heterogeneous network including drugs, proteins and diseases.

[0031] Step S2: Construct training samples for self-supervised representation learning: Select a set number of node pairs in the biological heterogeneous network, label the node pairs into four different categories based on the length of the shortest path between the node pairs, and use the selected node pairs and corresponding labels as training samples for self-supervised learning.

[0032] Specifically, node pairs are labeled into four different categories based on the length of the shortest path between them: the shortest path between the node pairs is calculated. If the shortest path is less than a set value of 4, the path length is marked as the class label of the corresponding node. If the shortest path is greater than or equal to the set value of 4, the corresponding node is classified into the category with a path length of 4. For example, when the path length is less than 4, it is marked as 1, 2, or 3, and when it is greater than or equal to 4, it is marked as 4. (There are 4 categories in total, which is consistent with the four categories of S2.

[0033] Step S3, initial feature setting: Calculate the number of neighbors of each node and convert the number of neighbors of all nodes into a one-hot encoding vector as the initial feature of the node.

[0034] Step S4: Input the adjacency matrix of the biological network and the initial features of the nodes into the graph attention neural network to predict the shortest path category of the node pairs of the training samples in step S2, and compare them with the original path labels of the node pairs.

[0035] In this embodiment, a multi-head attention mechanism is used in the graph attention network to aggregate the features of the node itself and its neighboring nodes. The formula is:

[0036]

[0037] Where δ represents the nonlinear activation ReLU function, K is the number of multi-head attention, || represents the concatenation of feature vectors, and W k represents the learnable parameter matrix in the k-th attention mechanism, represents the input features of node j, N i are all neighbors of node i, is the attention coefficient learned in the k-th attention mechanism, and the calculation formula is:

[0038]

[0039] Where, is a hyperparameter, ε(·) represents the nonlinear activation function LeakyReLU, which is calculated as:

[0040]

[0041] Where θ is a hyperparameter.

[0042] In this embodiment, the cross entropy function is used to calculate the loss, and the Adam optimizer is used for training and parameter update.

[0043] Step S5: All drug pairs with interactions in the drug-drug interaction network are regarded as positive samples, and an equal number of unknown drug pairs are randomly selected from the drug-drug interaction network as negative samples. All positive samples and negative samples are mixed together and divided into training sets and test sets according to a certain ratio. The sample representations and labels in the training set are input into the deep neural network for model training. After the maximum number of iterations is reached, the drug-drug interaction prediction model is obtained, and the trained drug-drug interaction prediction model is used to predict drug-drug interactions and perform evaluation.

[0044] Among them, the deep neural network consists of an input layer, a hidden layer and an output layer. The hidden layer uses the ReLU activation function, and the output layer uses the sigmoid activation function.

[0045] This example uses 721 drugs in the DrugBank database as an example and uses the above method to predict drug-drug interactions. In the hot start test scenario, the area under the receiver operating curve (AUROC) is 0.817, and the area under the precision-recall curve (AUPR) is 0.851; in the cold start test scenario, the area under the receiver operating curve (AUROC) is 0.779, and the area under the precision-recall curve (AUPR) is 0.841. Figure 2 shown.

[0046] This embodiment first integrates multiple network data to capture multimodal network information; secondly, it designs a representation learning task based on biological network path length classification to guide the graph attention network model to generate high-quality drug representation vectors, thereby reducing the dependence of the deep neural network model on labeled data; at the same time, the representation learning sample integrates biological entity pairs from the global structure of the network, so that the generated representation vector captures the global structural information of the network, further improving the prediction performance of drug-drug interactions.

[0047] Although the embodiments of the present invention are described in conjunction with the accompanying drawings, the patent owner may make various changes or modifications within the scope of the appended claims. As long as they do not exceed the scope of protection described in the claims of the present invention, they should be within the scope of protection of the present invention.

Claims

1. A method for predicting drug-drug interactions based on the global structure of biological networks, characterized in that: The following steps are involved: S1. Extract drug-drug interaction networks, drug-target interaction networks, protein-protein interaction networks, drug-disease association networks, and disease-protein association networks, and compare and fuse them to obtain a biologically heterogeneous network containing drugs, proteins, and diseases; S2. Select a set number of node pairs in the biological heterogeneous network, label the node pairs into four different categories based on the length of the shortest path between the node pairs, and use the selected node pairs and corresponding labels as training samples for self-supervised learning; S3. Calculate the number of neighbors of each node and convert the number of neighbors of all nodes into a one-hot encoding vector as the initial feature of the node; S4, inputting the adjacency matrix of the biological network and the initial features of the nodes into the graph attention neural network to predict the shortest path category of the node pair of the training sample in step S2, and comparing it with the original path label of the node pair; S5. All drug pairs with interactions in the drug-drug interaction network are regarded as positive samples, and an equal number of unknown drug pairs are randomly selected from the drug-drug interaction network as negative samples. All positive samples and negative samples are mixed together and divided into training sets and test sets according to a certain ratio. The sample representations and labels in the training set are input into the deep neural network for model training. After the maximum number of iterations is reached, the drug-drug interaction prediction model is obtained, and the trained drug-drug interaction prediction model is used to predict drug-drug interactions.

2. The method for predicting drug-drug interactions based on the global structure of biological networks according to claim 1, characterized in that: In step S1, based on the biomedical big data of DrugBank, UniProt, HPRD, SIDER, CTD, NDFRT, repoDB, HuRI, DrugCentral, BioGRID, OMIM and STRING, drug-target interaction networks, protein-protein interaction networks, drug-disease association networks and disease-protein association networks are extracted.

3. The method for predicting drug-drug interactions based on the global structure of biological networks according to claim 1, characterized in that: In step S2, the node pairs are marked into four different categories according to the length of the shortest path between the node pairs. Specifically, the shortest path between the node pairs is calculated. If the shortest path is less than the set value 4, the path length is marked as the class label of the corresponding node. If the shortest path is greater than or equal to the set value 4, the corresponding node is classified into a category with a path length of 4.

4. The method for predicting drug-drug interactions based on the global structure of biological networks according to claim 1, characterized in that: In step S4, a multi-head attention mechanism is used in the graph attention network to aggregate the features of the node itself and its neighboring nodes, and the formula is: Where δ represents the nonlinear activation ReLU function, K is the number of multi-head attention, || represents the concatenation of feature vectors, and W k represents the learnable parameter matrix in the k-th attention mechanism, represents the input features of node j, N i are all neighbors of node i, is the attention coefficient learned in the k-th attention mechanism, and the calculation formula is: Where, is a hyperparameter, ε(·) represents the nonlinear activation function LeakyReLU, which is calculated as: Where θ is a hyperparameter.

5. The method for predicting drug-drug interactions based on the global structure of biological networks according to claim 1, characterized in that: In step S4, the cross entropy function is used to calculate the loss, and the Adam optimizer is used for training and parameter update.

6. The method for predicting drug-drug interactions based on the global structure of biological networks according to claim 1, characterized in that: The deep neural network in step S5 includes an input layer, a hidden layer and an output layer, the hidden layer uses a ReLU activation function, and the output layer uses a sigmoid activation function.

Images