CN116206775A - Multi-dimensional characteristic fusion medicine-target interaction prediction method - Google Patents

Abstract

The invention provides a drug-target interaction prediction method that integrates multi-dimensional features; extracts information on interacting drugs and targets, as well as related diseases and side effects of drugs from medical databases, and constructs a heterogeneous network; uses The heterogeneous graph attention neural network extracts the topological structure features in the heterogeneous network; expresses the drug SMILES molecular sequence as a molecular graph structure, and extracts the drug feature information; simultaneously extracts the target feature information; the extracted drug and target feature information A message passing process incorporating a heterogeneous graph attention neural network is trained, the model is saved, and the relational predictions for drugs and targets are made. The present invention effectively utilizes the biometric information of the drug and the target, has a higher accuracy rate when predicting the drug-target relationship, improves the efficiency and accuracy of drug-target relationship verification, and effectively shortens the drug research and development cycle. Greatly reduce the cost of new drug development.

Description

A drug-target interaction prediction method integrating multi-dimensional features

Technical Field

The present invention relates to the field of medical artificial intelligence technology, and in particular to a drug-target interaction prediction method integrating multi-dimensional features.

Background Art

New drug development is a long and expensive process. It usually takes 10 to 17 years from the determination of ideas to the launch of drugs, and the capital investment will be between US$700 million and US$2.7 billion. The discovery method of using existing drugs to treat new indications has the advantages of low R&D costs and short development time. Therefore, reusing existing drugs to treat common and rare diseases has become increasingly attractive. In the process of identifying new candidate compounds with potential therapeutic effects, predicting drug-target interactions is an essential step. Drugs play an important role in the human body by interacting with various targets, which can enhance or inhibit their functions and play a regulatory role to achieve the purpose of treating a certain disease. Therefore, identifying drug-target interactions can help understand the mechanism of action of drugs and play a vital role in the discovery of new targets and drug repositioning.

At present, structure-based methods, ligand similarity-based methods and network-based methods are the main ways to predict drug-target interactions. Among them, structure-based methods usually require the understanding of the three-dimensional structure of proteins, so their performance is often poor for proteins with unknown structures. Ligand similarity-based methods use the common sense of known ligands for prediction. If the target compound is not marked in the target ligand library, such methods will not be able to produce reliable prediction results. Network-based methods make full use of the potential correlation between drugs and targets and have become the mainstream technology for analyzing and solving problems related to drug-target interaction prediction. Inspired by information transfer and clustering tasks in deep learning, drug target prediction can mine large-scale data on graph neural networks. Among them, methods based on graph convolutional networks are particularly outstanding. A large amount of effective hidden information is stored in the huge heterogeneous network composed of drugs and their related data. Using graph convolutional networks to process this information can effectively mine the potential associations in the network, which is conducive to drug discovery research. However, these methods generally ignore the use of biological knowledge, such as the biological structure characteristics in the compound sequence, so they cannot obtain the potential features in the data, and there is still a lot of room for improvement in model performance.

Summary of the invention

The purpose of the present invention is to propose a graph neural network model that incorporates drug molecular structure and protein biological structure information, which automatically predicts the interaction relationship between drugs and targets, improves verification efficiency and reduces verification costs.

To achieve the above objectives, the technical solution of the present application is: a drug-target interaction prediction method integrating multi-dimensional features, comprising:

Step 1: Extract information about interacting drugs and targets, as well as related diseases and drug side effects from the medical database, preprocess them, and construct a heterogeneous network;

Step 2: Using a heterogeneous graph attention neural network to extract network topology features in the heterogeneous network;

Step 3: Represent the drug SMILES molecular sequence as a molecular graph structure and use the molecular attention Transformer network to extract drug structural feature information;

Step 4: Embed the target sequence information and process it using a convolutional neural network and a bidirectional long short-term memory network to extract the target structure feature information;

Step 5: Integrate the extracted drug structure feature information and target structure feature information into the message passing process of the heterogeneous graph attention neural network;

Step 6: Use the cross entropy loss function to optimize the prediction model and then save the prediction model;

Step 7: Load the prediction model, input the drug and target information to be predicted, predict the relationship between the drug and the target, and output the prediction result.

Furthermore, the specific implementation process of step 1 includes:

Step 1.1: Screen the drug information and target information from the medical database and delete the drug information and target information that have no interaction relationship;

Step 1.2: Obtain the SMILES molecular sequence corresponding to the drug and the sequence information corresponding to the target from the medical database, which are used as the biological feature representation information of the drug and the target respectively;

Step 1.3: Extract disease and drug side effect information related to the above drugs and targets;

Step 1.4: As shown in Figure 2(a), the extracted drugs, targets, diseases and drug side effects are taken as nodes, and the association information between them is represented as edges to construct a heterogeneous network G = (V, E), where V represents the node set and E represents the edge set;

Step 1.5: Integrate the drugs and targets with interaction relationships and construct them into the form of <drug number, target number, label>, and mark the label as 1;

Step 1.6: Randomly construct unknown drug-target relationships as negative examples at a ratio of 1:10, and mark them as 0.

Furthermore, the specific implementation process of step 2 includes:

In the heterogeneous network, the embedding formula of the initialization node is f ⁰ :V→R ^d , where f ⁰ (v) represents the d-dimensional mapping of each node v; the neighbor node information aggregation of node v is defined as:

where σ(·) represents the nonlinear activation function in the propagation process of a layer of neural network, K is the number of attention layers, _Nv represents all the neighboring nodes of node v, W is a shared weight parameter, a represents the weight vector of the attention mechanism, and AconC is a new activation function that can be adaptively learned.

Furthermore, the specific implementation process of step 3 includes:

Step 3.1: By calling the RDKit function library in the Python library, the SMILES molecular sequence of each drug is represented as a molecular graph, where the vertices and edges of the graph represent the atoms and chemical bonds of the drug, respectively. Each drug molecule is represented by a feature matrix and an adjacency matrix, and each row of the feature matrix corresponds to the attributes of each atom.

Step 3.2: Since each SMILES sequence has a different length, in order to create an efficient representation, a SMILES sequence with a maximum length of 100 characters is selected, which can cover at least 90% of the compounds in the dataset. Sequences longer than the maximum character length are truncated, while sequences shorter than the maximum character length are padded with 0;

Step 3.3: Referring to Figure 2(b), the molecular attention Transformer network is used to extract the drug feature representation S _drug ; the calculation formula of the molecular multi-head self-attention layer is as follows:

表示分子图的邻接矩阵，

表示原子间的距离；

分别是查询向量矩阵、键向量矩阵和值向量矩阵，其中W是可学习的参数，i∈(1,...,h)，h是多头注意力的头数；λ_a、λ_d和λ_g表示加权自注意、距离和邻接矩阵的标量参数。in

represents the adjacency matrix of the molecular graph,

represents the distance between atoms;

are the query vector matrix, key vector matrix, and value vector matrix, respectively, where W is a learnable parameter, i∈(1,...,h), h is the number of heads of multi-head attention; _λa , _λd , and _λg represent scalar parameters of weighted self-attention, distance, and adjacency matrices.

Furthermore, the specific implementation process of step 4 includes:

Step 4.1: Randomly initialize an index table corresponding to all amino acids appearing in the target sequence, with a size of 26×100; correspond the amino acids in each target sequence to the index table to construct an embedding matrix of the target sequence; the length of the embedding matrix is the maximum length in the target sequence, which is set to 1000; during the model training process, the embedding vector is continuously optimized, so the relevant information in the index table will continue to change with the optimization of the model;

Step 4.2: As shown in Figure 2(c), a convolutional neural network and a bidirectional long short-term memory network are used to extract feature information from the target sequence.

Furthermore, the specific implementation process of step 4.2 includes:

Step 4.2.1: Use the embedding matrix obtained in step 4.1 as the input of the convolutional neural network; empty labels are automatically filled for target sequences that are shorter than the length of the embedding matrix; each CNN block uses three consecutive one-dimensional convolutional layers, and the number of convolutional kernels increases with the number of layers. The second layer uses twice the convolutional kernels of the first layer, and the third layer uses three times the convolutional kernels of the first layer;

Step 4.2.2: Use the BiLSTM layer to receive the output of the convolutional layer. The final output is the protein structure feature, expressed as S _protein , and the formula is as follows:

Among them, w and m represent the weight matrix and convolution window size respectively, h is the LSTM hidden layer state, and x is the feature representation of the protein sequence.

Furthermore, the specific implementation process of step 5 includes:

Referring to Figure 2(d), the drug structure feature vector S _drug obtained in step 3 and the target structure feature vector S _protein obtained in step 4 are concatenated in the heterogeneous graph attention neural network message passing stage, and the formula for updating the node embedding in equation (1) is as follows:

Furthermore, the specific implementation process of step 6 includes:

Step 6.1: See Figure 2(e). After obtaining the feature representation of the drug and the target, the inner product method is used to predict the drug-target interaction. Given a drug node u and a protein node v, _fu and _fv represent their features. The probability of interaction between u and v is:

P＝σ(( _fu ) ^Tfv ₎ (5)

为s型函数，P表示u和v之间的相互作用预测得分；in

is a sigmoid function, P represents the interaction prediction score between u and v;

Step 6.2: Use the cross entropy loss function to optimize the prediction model training, use 10-fold cross validation to test the prediction model performance, and save the best prediction model Model _best .

Furthermore, the specific implementation process of step 7 includes:

Load the prediction model Model _best in step 6.2, input the drug-target information in the validation data into the prediction model, determine whether there is an interaction relationship between the drug and the target, and output the corresponding evaluation index.

Due to the adoption of the above technical scheme, the present invention can achieve the following technical effects: the present invention adopts a deep learning model, utilizes the information of drugs, targets, diseases and drug side effects in the medical database, combines the structural characteristics of drugs and targets, and automatically predicts the drug-target interaction information through the model. It effectively extracts the characteristic information in the drug molecule and protein structure, has a higher accuracy rate and robustness when predicting the drug-target relationship, improves the efficiency and accuracy of drug-target relationship verification, effectively shortens the drug development cycle, greatly reduces the cost of new drug development, and provides an important foundation and guarantee for new drug development and drug reuse.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a drug-target interaction prediction method integrating multi-dimensional features;

FIG2 is a model structure diagram of a drug-target interaction prediction method integrating multi-dimensional features.

DETAILED DESCRIPTION

The embodiments of the present invention are implemented on the premise of the technical solution of the present invention, and detailed implementation methods and specific operation processes are given, but the protection scope of the present invention is not limited to the following embodiments.

The present invention is described in detail below in conjunction with embodiments so that those skilled in the art can implement the invention according to the description.

Example 1

This embodiment uses Windows system as the development environment, Pycharm as the development platform, Python as the development language, and adopts the drug-target interaction prediction method integrating multi-dimensional features of the present invention to predict the drug-target interaction relationship.

In this embodiment, a drug-target interaction prediction method integrating multi-dimensional features includes the following steps:

708 drugs, 1512 proteins, 5603 diseases and 4192 drug side effects were extracted from DrugBank, PubChem, HPRD, Comparative Toxicogenomics and SIDER databases. The existing drug-target interaction relationships were marked as positive examples, and the data labels were set to 1, totaling 1923 cases. 19230 cases were randomly selected from drug-target pairs that were not marked as positive examples to construct negative examples, and the data labels were set to 0. A heterogeneous network was constructed using the above data.

The heterogeneous network, drug SMILES sequence, and protein sequence are used as input to train and save the prediction model to obtain the prediction score of the evaluation indicators of the interaction between drugs and targets. The evaluation indicators include the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPR).

According to the above steps, the present invention compares the drug-target relationship prediction effect with the EEG-DTI model, NeoDTI model, DTINet model, MSCMF model and HNM model. As can be seen from Table 1, the method proposed in this invention is significantly better than other methods in both AUROC and AUPR.

Table 1 Comparison of prediction results of drug-target relationship by different models

Example 2

In this example, Windows system is used as the development environment, Pycharm is used as the development platform, Python is used as the development language, and the drug-target interaction prediction method integrating multi-dimensional features of the present invention is used to predict potential therapeutic drugs for COVID-19.

In this embodiment, a drug-target interaction prediction method integrating multi-dimensional features includes the following steps:

146 targets closely related to COVID-19 and 708 candidate drugs were extracted from the Comparative Toxicogenomics Database and DrugBank Database; SMILES sequences of drugs and sequence structures of targets were obtained from the PubChem Database; 1456 diseases and 4192 drug side effects related to drugs and targets were extracted from the HPRD Database and SIDER Database; heterogeneous networks were constructed using the acquired data;

The heterogeneous network and the sequence data of drugs and proteins are used as input, and the saved prediction model is loaded to obtain the prediction score of the interaction relationship between drugs and different targets; the prediction score is sorted in descending order, and the top 10 candidate drugs with confidence for each target are extracted, and the confidence scores of these drugs are required to be greater than 0.5. After such processing, only 15 targets meet the requirements, and 150 candidate drugs are finally screened.

In the experimental results, 54 of the 150 drugs screened have appeared in COVID-19 clinical studies, and some of the data are shown in Table 2. This method can quickly and more specifically find candidate drugs for subsequent wet experiments.

Table 2 The present invention screened out therapeutic drugs related to COVID-19

The foregoing description of specific exemplary embodiments of the present invention is for the purpose of illustration and demonstration. These descriptions are not intended to limit the present invention to the precise form disclosed, and it is clear that many changes and variations can be made based on the above teachings. The purpose of selecting and describing exemplary embodiments is to explain the specific principles of the present invention and its practical application, so that those skilled in the art can realize and utilize various different exemplary embodiments of the present invention and various different selections and changes. The scope of the present invention is intended to be limited by the claims and their equivalents.

Claims (9)

1. A drug-target interaction prediction method integrating multi-dimensional features, characterized by comprising: Step 1: Extract information about interacting drugs and targets, as well as related diseases and drug side effects from the medical database, preprocess them, and construct a heterogeneous network; Step 2: Using a heterogeneous graph attention neural network to extract network topology features in the heterogeneous network; Step 3: Represent the drug SMILES molecular sequence as a molecular graph structure and use the molecular attention Transformer network to extract drug structural feature information; Step 4: Embed the target sequence information and process it using a convolutional neural network and a bidirectional long short-term memory network to extract the target structure feature information; Step 5: Integrate the extracted drug structure feature information and target structure feature information into the message passing process of the heterogeneous graph attention neural network; Step 6: Use the cross entropy loss function to optimize the prediction model and then save the prediction model; Step 7: Load the prediction model, input the drug and target information to be predicted, predict the relationship between the drug and the target, and output the prediction result. 2. According to the method for predicting drug-target interaction integrating multi-dimensional features in claim 1, the specific implementation process of step 1 comprises: Step 1.1: Screen the drug information and target information from the medical database and delete the drug information and target information that have no interaction relationship; Step 1.2: Obtain the SMILES molecular sequence corresponding to the drug and the sequence information corresponding to the target from the medical database, which are used as the biological feature representation information of the drug and the target respectively; Step 1.3: Extract disease and drug side effect information related to the above drugs and targets; Step 1.4: Take the extracted drugs, targets, diseases and drug side effects as nodes, and the association information between them as edges to construct a heterogeneous network G = (V, E), where V represents the node set and E represents the edge set; Step 1.5: Integrate the drugs and targets with interaction relationships and construct them into the form of <drug number, target number, label>, and mark the label as 1; Step 1.6: Randomly construct unknown drug-target relationships as negative examples in a certain proportion and mark them as 0. 3. According to the method for predicting drug-target interaction integrating multi-dimensional features in claim 1, the specific implementation process of step 2 comprises: In the heterogeneous network, the embedding formula of the initialization node is f ⁰ :V→R ^d , where f ⁰ (v) represents the d-dimensional mapping of each node v; the neighbor node information aggregation of node v is defined as:

where σ(·) represents the nonlinear activation function in the propagation process of a layer of neural network, K is the number of attention layers, _Nv represents all the neighboring nodes of node v, W is a shared weight parameter, a represents the weight vector of the attention mechanism, and AconC is a new activation function that can be adaptively learned. 4. According to the method for predicting drug-target interaction integrating multi-dimensional features in claim 1, the specific implementation process of step 3 comprises: Step 3.1: By calling the RDKit function library in the Python library, the SMILES molecular sequence of each drug is represented as a molecular graph, where the vertices and edges of the graph represent the atoms and chemical bonds of the drug, respectively. Each drug molecule is represented by a feature matrix and an adjacency matrix, and each row of the feature matrix corresponds to the attributes of each atom. Step 3.2: Select a SMILES sequence with a maximum length of 100 characters. Sequences longer than the maximum length are truncated, while sequences shorter than the maximum length are padded with 0s. Step 3.3: Use the molecular attention Transformer network to extract the drug feature representation S _drug ; the calculation formula of the molecular multi-head self-attention layer is as follows:

in

represents the adjacency matrix of the molecular graph,

represents the distance between atoms; _Qi = _XWiq , _Ki = _XWik , _Vi = _XWiv are the query vector matrix ^, key vector matrix and value vector matrix respectively, where ^W is a learnable parameter, i∈(1,...,h) ^, h is the number of heads of multi-head attention; _λa , _λd and _λg represent scalar parameters of weighted self-attention, distance and adjacency matrices. 5. According to the method for predicting drug-target interaction integrating multi-dimensional features in claim 1, the specific implementation process of step 4 comprises: Step 4.1: randomly initialize an index table corresponding to all amino acids appearing in the target sequence; correspond each amino acid in the target sequence to the index table to construct an embedding matrix of the target sequence; the length of the embedding matrix is the maximum length in the target sequence; Step 4.2: Use convolutional neural network and bidirectional long short-term memory network to extract feature information from the target sequence. 6. According to the method for predicting drug-target interaction integrating multi-dimensional features in claim 5, it is characterized in that the specific implementation process of step 4.2 includes: Step 4.2.1: Use the embedding matrix obtained in step 4.1 as the input of the convolutional neural network; for target sequences that are shorter than the length of the embedding matrix, empty labels are automatically filled; each CNN block uses three consecutive one-dimensional convolutional layers, and the number of convolutional kernels increases with the number of layers, that is, the second layer uses twice the convolutional kernels of the first layer, and the third layer uses three times the convolutional kernels of the first layer; Step 4.2.2: Use the BiLSTM layer to receive the output of the convolutional layer. The final output is the protein structure feature, expressed as S _protein , and the formula is as follows:

Among them, w and m represent the weight matrix and convolution window size respectively, h is the LSTM hidden layer state, and x is the feature representation of the protein sequence. 7. According to the method for predicting drug-target interaction integrating multi-dimensional features in claim 1, the specific implementation process of step 5 comprises: The drug structure feature vector S _drug obtained in step 3 and the target structure feature vector S _protein obtained in step 4 are concatenated in the heterogeneous graph attention neural network message passing stage, and the formula for updating the node embedding in formula (1) is as follows:

8. According to the method for predicting drug-target interaction integrating multi-dimensional features in claim 1, the specific implementation process of step 6 comprises: Step 6.1: After obtaining the feature representation of drugs and targets, the inner product method is used to predict drug-target interactions; given a drug node u and a protein node v, _fu and _fv represent their features; the probability of interaction between u and v is: P＝σ(( _fu ) ^Tfv ₎ (5) in

is a sigmoid function, P represents the interaction prediction score between u and v; Step 6.2: Use the cross entropy loss function to optimize the prediction model training, use 10-fold cross validation to test the prediction model performance, and save the best prediction model Model _best . 9. According to the method for predicting drug-target interaction integrating multi-dimensional features in claim 8, the specific implementation process of step 7 comprises: Load the prediction model Model _best in step 6.2, input the drug-target information in the validation data into the prediction model, determine whether there is an interaction relationship between the drug and the target, and output the corresponding evaluation index.

Images