Network pharmacology method based on drug-target interaction prediction model
By using a deep learning-based drug-target interaction prediction model that combines GCN, CNN, and Bi-LSTM, the problem of incomplete data in network pharmacology is solved, and the accuracy of analysis and prediction is improved.
Patent Information
- Application Number
- CN202211079616.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-09-05
- Publication Date
- 2026-01-02
- Estimated Expiration
- 2042-09-05
AI Technical Summary
Existing network pharmacology methods lack comprehensive drug, gene, and protein data, affecting the accuracy of analysis and prediction.
A deep learning-based drug-target interaction prediction model is designed, combining graph convolutional neural networks (GCN), convolutional neural networks (CNN), and bidirectional long short-term memory networks (Bi-LSTM), and introducing an attention mechanism to construct a drug-target interaction prediction model (DTI) for network pharmacology analysis.
It improves the accuracy of network pharmacology analysis and prediction by acquiring structural and sequence information of drugs and targets and enhancing feature extraction through attention mechanisms, thereby improving prediction performance.
Smart Images
Figure CN115719611B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of computer-aided drug design, in particular to a network pharmacology method based on a drug-target interaction prediction model. BACKGROUND
[0002] In recent years, with the rapid development of social economy, people pay more and more attention to health problems, so improving the medical level can increase the social happiness index, and the improvement of the medical level cannot be separated from the research and development of drugs. However, the cycle of drug research and development is long and expensive, and the number of drugs approved by the US FDA for listing has been declining in the past decade. Therefore, the drug research and development mode of "one drug, one target and one disease" is no longer applicable.
[0003] With the development of systems biology, omics, multidirectional pharmacology and bioinformatics methods, the concept of network pharmacology has been put forward. Its emergence has ended the traditional drug research and development mode of single target for single chemical structure formula to treat diseases, and opened up a new research mode of complex network relationship between multi-target and multiple diseases. Network pharmacology mainly reveals the complex relationship among drugs, targets and diseases by using high-throughput screening, network analysis and visualization technology, and analyzes and predicts the pharmacological mechanism of drugs on this basis, and verifies it through experiments. However, there are still some problems in the traditional network pharmacology method, such as the data of drugs, genes and proteins obtained from the database is not comprehensive enough. These will affect the accuracy of network pharmacology analysis and prediction.
[0004] As the most effective and direct method to discover new drugs and new targets, the drug-target interaction (DTI) prediction method can discover the interaction relationship between new chemical structure and target from known drugs and diseases. To some extent, the use of DTI method can help to solve the problem of incomplete data in network pharmacology method, which has great significance.
[0005] Based on this, the present application designs a DTI model based on deep learning, which is integrated into the process of network pharmacology method, so as to improve the accuracy of network pharmacology analysis and prediction. SUMMARY
[0006] The first purpose of the present application is to solve the problem of incomplete related data in network pharmacology method, and a drug-target interaction prediction model based on deep learning is proposed, which can be used in network pharmacology analysis method to improve the accuracy of network pharmacology analysis and prediction.
[0007] A network pharmacology method based on a drug-target interaction prediction model, comprising the steps of:
[0008] Step (1): Obtain the chemical structure of the drug and the information of the target and the information of the disease target, construct the drug compound database and the disease target database;
[0009] Step (2): According to Davis and KIBA data sets, obtain the interaction relationship between each drug target and each disease target as the label of the training set; construct the training set from the chemical structure of each drug, disease target data and label, and construct the test set from the chemical structure of each drug and disease target data;
[0010] Step (3): Construct the drug-target interaction prediction model DTI and train it with the training set;
[0011] The drug-target interaction prediction model DTI includes parallel drug feature extractor, disease target feature extractor, and attention module, fusion module;
[0012] The drug feature extractor includes parallel drug structure feature extraction module and drug sequence feature extraction module;
[0013] The drug structure feature extraction module includes a first embedding layer and a graph convolutional neural network (GCN); the first embedding layer uses RDKit to convert the SMILES chemical structure of the drug into a molecular graph structure and obtains the corresponding structure information. The graph convolutional neural network (GCN) is used to receive the output of the first embedding layer to extract the structure features of the drug chemical structure;
[0014] The drug sequence feature extraction module includes a second embedding layer and a convolutional neural network (CNN); the second embedding layer uses Embedding to convert the SMILES chemical structure of the drug into a vector containing sequence information of the drug chemical structure. The convolutional neural network (CNN) is used to receive the output of the second embedding layer to extract the sequence features of the drug chemical structure;
[0015] The disease target feature extractor includes a third embedding layer and a bidirectional long short-term memory network (Bi-LSTM); the third embedding layer uses Embedding to convert the amino acid sequence information of the disease target into a vector containing sequence information of the disease target. The bidirectional long short-term memory network (Bi-LSTM) is used to receive the output of the third embedding layer to extract the disease target features;
[0016] The attention module uses attention mechanism Attention to extract more important subsequences in the disease target protein and more important subsequences in the drug compound.
[0017]
[0018]
[0019] D d = relu(DW + b)
[0020] P p = relu(PW + b)
[0021]
[0022]
[0023]
[0024] wherein D is a set of representations of chemical structures of drugs; P is a representation of a disease target protein; N is a sequence length of a drug SMILES structure; M is a length of an amino acid sequence of a target protein; F is a fixed width; D d and P p are attention vectors converted by using a multi-layer perceptron and a ReLU function; W is a weight matrix; b is a bias; a is a calculated attention score, representing the strength of interaction between a drug compound sub-sequence and a disease target protein sub-sequence; a d is an attention vector of a drug, being an attention score of each position in a drug sequence for the entire protein sequence; a p is an attention vector of a disease target protein, being an attention score of each position in a protein sequence for the entire drug sequence; a i,j is an attention score of the i-th position of a drug sequence for the j-th position of a disease target sequence or an attention score of the j-th position of a disease target sequence for the i-th position of a drug sequence.
[0025] Step (4): using a test set to predict the trained drug-target interaction prediction model DTI, obtaining the action target of the chemical structure of the drug;
[0026] Step (5): extracting the intersection target according to the action target of the drug, the disease target and the action target predicted by the drug-target interaction prediction model DTI; the extraction operation is a conventional technology, and therefore is not explained in detail.
[0027] Step (6): based on the intersection target, performing protein interaction analysis, network topology analysis and enrichment analysis to obtain the core drug compound and the core disease target required by the disease;
[0028] Step (7): verifying the molecular docking according to the core drug compound and the core disease target required by the disease.
[0029] Further, in step (1), the drug chemical structure formula of the drug compound database is described by SMILES descriptor (a standard for explicitly describing molecular structure by ASCII string), and the drug target is calculated by SwissTargetPrediction tool; in step (A), the SMILES descriptor information of the chemical structure formula is collected, and then the SwissTargetPrediction tool is used to calculate the target of each chemical structure formula, and the UniProt database is used to convert the amino acid sequence of the protein; the amino acid sequence of the corresponding disease target is also calculated when collecting the disease target information.
[0030] Further, in step (1), the disease target information in the disease target database is converted into the amino acid sequence of the disease target protein by using the UniProt database. The amino acid sequence of the corresponding disease target is also calculated when collecting the disease target information.
[0031] Further, in step (2), each sample of the training set is composed of a drug chemical structure formula (SMILES) and a target protein (amino acid sequence), and if the chemical structure formula and the target protein have an interaction relationship, it is a positive sample (1), otherwise it is a negative sample (0); the sample of the test set is composed of a drug chemical structure formula and a disease target.
[0032] Further, in step (3), the model is trained by using the cross-validation method.
[0033] Further, in step (5), the drug target and the disease target are the drug target and the disease target output by the drug-target interaction prediction model DTI.
[0034] Further, in step (5), the drug target and the disease target are obtained by using the Venn diagram.
[0035] Further, in step (6), the compound-disease target-disease network obtained by using Cytoscape software is analyzed, the key target is obtained by analyzing the intersection target protein interaction, and then the core drug compound and the core disease target are obtained by analyzing the gene ontology (Gene Ontology, GO) and the kyoto gene and genome encyclopedia (Kyoto Encyclopedia of Genes and Genomes, KEGG).
[0036] The second object of the present application is to provide a precise matching disease drug screening system, which comprises a drug-target interaction prediction model DTI which has been trained and tested.
[0037] A third object of the present application is to provide an electronic device comprising a processor and a memory having stored machine executable instructions executable by the processor to implement the system.
[0038] The beneficial effects of the present application are embodied in:
[0039] (1) The structure and sequence information of the chemical structural formula are obtained using GCN and CNN, and the attention mechanism is added so that the DTI model can fully utilize the relevant feature information of the chemical structural formula and the target protein, and obtain better prediction performance.
[0040] (2) The drug-target interaction prediction model is integrated, so that the data amount of the action target of the drug chemical structural formula for the disease is increased, and the accuracy of the analysis and prediction of the network pharmacology method is improved to a certain extent. BRIEF DESCRIPTION OF DRAWINGS
[0041] Figure 1 A schematic diagram of the network pharmacology method based on drug-target interaction prediction of the present application. Figure 2 A structural diagram of the drug-target interaction prediction model. DETAILED DESCRIPTION
[0042] The present application will be further described below in conjunction with the accompanying drawings and specific examples. It should be understood that these examples are only used to illustrate the present application and not to limit the scope of the present application. The operation methods not specified in the following examples are usually performed under conventional conditions or under the conditions recommended by the manufacturers.
[0043] The network pharmacology method based on the drug-target interaction prediction model of the present application is shown in Figure 1 Specifically, the method comprises the following steps:
[0044] (A) The drug chemical structural formula information downloaded from the databases such as TCMSP, DrugBank, PubChem, etc. includes SMILES descriptor and SDF format molecular structural formula. The target data information of the disease is obtained from the databases such as OMIM, TTD, GeneCards, etc. Then the SwissTargetPrediction software is used to calculate the action target according to the molecular structure information of the chemical structural formula, and the UniProt database is used to convert to obtain the amino acid sequence of the target. Thus, the drug chemical structural formula database and the disease target database are constructed.
[0045] (B) Each sample of the training set consists of a chemical structure formula (SMILES) and a target (amino acid sequence), and is a positive sample 1 if the chemical structure formula and the target have an interaction relationship, otherwise it is a negative sample 0. The test set is constructed by one-to-one correspondence between the chemical structure formula and all the collected disease targets, and their interaction relationship is unknown.
[0046] (C) Design a drug-target interaction model DTI that can predict the interaction results of chemical structure formula and target. The specific steps are as follows:
[0047] (1) Drug chemical structure formula feature extraction. Use the RDKit open source package of Python to convert the SMILES of the chemical structure formula into a graph structure, and input it into the GCN layer to obtain the feature vector representation. On the other hand, the SMILES descriptor is converted into a vector containing the sequence information of the drug chemical structure formula by Embedding, and then input into the CNN layer to obtain the sequence information feature vector representation. Then the structure feature vector and the sequence feature vector are fused to obtain the total feature vector representation.
[0048] (2) Target protein feature extraction. The amino acid sequence of the target protein is converted into a vector containing the sequence information of the disease target by Embedding, and then input into the Bi-LSTM layer to extract the feature vector.
[0049] (3) Attention mechanism. Add an attention vector to the feature vectors of the drug chemical structure formula and the disease target protein to enable the model to determine which subsequence in the protein is more important to the drug, and which subsequence in the drug is more important to the protein.
[0050] (4) Full connection layer output. After combining the feature vectors of the chemical structure formula and the target protein, input them into the full connection layer, and output 0 or 1 through the Sigmoid function.
[0051] (D) Train the DTI model based on the training set. In order to obtain better training effect, five-fold cross-validation is used during training. The training parameters are as follows: Learning rate: 0.0005, Optimizer: Adam, Batch_size: 32.
[0052] (E) Use the model trained in step (D) to predict the test set obtained in step (B), record all the prediction results, and select the samples with the result of 1 to add to the drug chemical structure formula database.
[0053] (F) First, remove the duplicate drug chemical structure formula target points obtained in steps (B) and (E), and then import them into the Venn diagram together with the disease targets to extract the intersection targets.
[0054] (G) The intersection target points obtained in step (F) are introduced into the STRING database for protein interaction analysis, and then the TSV format file obtained is opened in Cytoscape software, and the values of Closeness, Betweenness and Degree are calculated by Centiscape to screen the key target points. Then the key target points are introduced into Metascape for GO and KEGG analysis, and the network of chemical structure-target-pathway-disease is obtained, and then the network topology analysis is performed by Cytoscape to calculate the core drug compounds and core disease targets.
[0055] (H) First, the 3D structure of the drug chemical structure is obtained by using the PubChem database and OpenBabel software, and then hydrogen is added, the rotation key is set, and other operations are performed to convert it into a PDBQT format file by using AutoDockTools software. Then the PDB database and Aphafold2 are used to obtain the PDB format file of the three-dimensional structure of the disease target protein. Finally, the global docking box is generated by AutoDockTools and the blind docking is performed by using qvina-w. If the binding energy is <-4.25 kcal / mol, it indicates that the drug chemical structure can spontaneously bind to the disease target protein, if the binding energy is <-5.0 kcal / mol, it indicates that the two have good binding, and if the binding energy is <-7.0 kcal / mol, it indicates that the two have strong binding activity. After obtaining the drug chemical structure and disease target with good binding performance, zebrafish experiments, mouse experiments and other medical experiments are performed for further verification.
Claims
1. A network pharmacology method based on a drug-target interaction prediction model, characterized in that, The method comprises the steps of: Step (1): obtaining the chemical structure of the drug and the information of the target and the information of the disease target, constructing a drug compound database and a disease target database; Step (2): according to Davis and KIBA data sets, the interaction relationship between each drug target and each disease target is obtained as the label of the training set; the training set is constructed by the chemical structure of each drug, the disease target data and the label, and the test set is constructed by the chemical structure of each drug and the disease target data; Step (3): constructing a drug-target interaction prediction model DTI and training it with the training set; The drug-target interaction prediction model DTI includes parallel drug feature extractors, disease target feature extractors, and attention modules and fusion modules; The drug feature extractor includes parallel drug structure feature extraction modules and drug sequence feature extraction modules; The drug structure feature extraction module includes a first embedding layer and a graph convolutional neural network GCN; the first embedding layer converts the SMILES chemical structure of the drug into a molecular graph structure and obtains the corresponding structure information; the graph convolutional neural network GCN is used to receive the output of the first embedding layer to extract the structure features of the drug chemical structure; The drug sequence feature extraction module includes a second embedding layer and a convolutional neural network CNN; the second embedding layer converts the SMILES chemical structure of the drug into a vector containing sequence information of the drug chemical structure; the convolutional neural network CNN is used to receive the output of the second embedding layer to extract the sequence features of the drug chemical structure; The disease target feature extractor includes a third embedding layer and a bidirectional long short-term memory network Bi-LSTM; The third embedding layer converts the amino acid sequence information of the disease target into a vector containing sequence information of the disease target; the bidirectional long short-term memory network Bi-LSTM is used to receive the output of the third embedding layer to extract the disease target features; Step (4): using the test set to predict the trained drug-target interaction prediction model DTI to obtain the target of the drug chemical structure; Step (5): according to the target of the drug, the disease target and the target predicted by the drug-target interaction prediction model DTI, the intersection target is extracted; Step (6): based on the intersection target, protein interaction analysis, network topology analysis and enrichment analysis are carried out to obtain the core drug compound and the core disease target required by the disease; Step (7): according to the core drug compound and the core disease target required by the disease, molecular docking verification is carried out.
2. The method of claim 1, wherein, The attention module in step (3) uses an attention mechanism Attention to extract more important subsequences in the disease target protein and more important subsequences in the drug compound.
3. The method of claim 2, wherein, The attention module is specifically: ； ； ； ； ； ； ； Where D is the set of drug chemical structure formula; P is the representation of disease target protein; N is the sequence length of drug SMILES structure; M is the length of target protein amino acid sequence; F is the fixed width; and is the attention vector converted by using multi-layer perceptron and ReLU function; W is the weight matrix; b is the bias; α is the calculated attention score, representing the strength of the interaction between the drug compound sub-sequence and the disease target protein sub-sequence; is the attention vector of the drug, which is the attention score of each position in the drug sequence to the entire protein sequence; is the attention vector of the disease target protein, which is the attention score of each position in the protein sequence to the entire drug sequence; is the attention score of the i-th position of the drug sequence to the j-th position of the disease target sequence or the attention score of the j-th position of the disease target sequence to the i-th position of the drug sequence. 4.The network pharmacology method based on a drug-target interaction prediction model according to claim 1, characterized in that, If the drug chemical structure and the target protein have an interaction relationship, the output of the drug-target interaction prediction model DTI is 1, otherwise it is 0. 5.The network pharmacology method based on a drug-target interaction prediction model according to claim 1, characterized in that, The step (3) uses a cross-validation method to train the model. 6.The network pharmacology method based on a drug-target interaction prediction model according to claim 5, characterized in that, In the step (5), the drug action target and the disease target are the drug action targets and the disease targets whose drug-target interaction prediction model DTI output results are 1.
7. The network pharmacology method based on a drug-target interaction prediction model according to claim 6, characterized in that, In the step (5), the obtained drug action targets and the disease targets are intersected by using a Venn diagram.
8. The network pharmacology method based on a drug-target interaction prediction model according to claim 7, characterized in that, In the step (6), the obtained compound-disease target-disease network is analyzed by using Cytoscape software, the key targets are obtained by performing protein interaction analysis on the intersection targets, and the core drug compounds and the core disease targets are obtained by performing gene ontology GO and Kyoto Encyclopedia of Genes and Genomes KEGG analysis.
9. An electronic device, comprising: The processor executes the machine executable instructions to implement the method in any one of claims 1-8.