Drug target interaction prediction method of bipartite attribute graph based on feature enhancement

Abstract

The invention discloses a drug-target interaction prediction method based on feature enhancement and a bipartite graph, the method aims to predict whether a drug and a target interact or not, and a model mainly comprises the steps of bipartite graph construction, feature enhancement processing, model definition, model training and model evaluation. According to the drug target interaction prediction method based on the feature enhanced bipartite attribute graph, prediction of drug target interaction is achieved, and compared with an existing algorithm, the drug target interaction prediction method has the following advantages that the innovative method combines pharmacological attributes of drugs with multi-source bioinformatics attributes of the targets, the similarity information is fused, and the accuracy of drug target interaction prediction is improved. And the features are enhanced. The model adopts an advanced feature fusion strategy, drugs and targets are complexly fused with wide biological and pharmacological attributes and similarity information, an advanced graph neural network is used for feature extraction, and a new perspective is provided for DTI prediction.

Description

Technical Field

[0001] The invention belongs to the field of drug-target interaction and is a drug-target interaction prediction method based on a feature-enhanced bipartite attribute graph. Background Art

[0002] Drug repositioning, drug repurposing, drug reanalysis, drug redirection and drug rediscovery are all important research topics in today's society. However, the discovery of new drugs is a very expensive and time-consuming process, and involves health and safety issues. In drug discovery and drug combination design, verifying the interaction between drugs and specific targets (DTI) is a key step in proving whether the drug is effective. Computer-aided drug design is an effective drug discovery method that can significantly save time and cost. In the process of new drug development, studying the interaction between drug molecules with complex chemical structures and a large number of protein targets is the key to drug treatment of diseases, and the prediction of drug-target interaction is an important direction of current research. Current drug-target interaction methods are divided into the following categories: docking-based methods, network-based methods, matrix decomposition-based methods and deep learning-based methods. Docking-based methods can enhance the possibility of successfully identifying DTIs by discovering more interactions or binding sites between drugs and their targets. Network-based methods calculate the similarity between drugs and targets based on network topology, including building a network of drugs, proteins or both to identify new DTIs. Matrix factorization-based methods aim to find two matrices whose multiplication results in an interaction matrix, assuming that the drug and target are in the same distance space, so the distance between the drug and the target can be used to measure the strength of their interaction, and the drug and target can be embedded in a common low-dimensional subspace, thereby achieving drug-target interaction prediction. Deep learning-based methods provide powerful expressive power in identifying, processing, and determining complex patterns in molecular data. Neural networks can fuse drug and target features, changing the status quo of simple concatenation or tensor product of drug and target features. Despite these advances, existing models often face challenges. Extracting a large number of informative features may lead to over-smoothing problems and fail to fully utilize the rich topological information and complex interactions of heterogeneous data. Summary of the invention

[0003] To address the current shortcomings of drug-target interaction prediction methods, we proposed a feature-enhanced drug target representation learning that integrates the pharmacological and bioinformatic properties of drug targets to enhance drug-target interaction prediction.

[0004] The present invention provides a method for predicting drug-target interaction based on a feature-enhanced bipartite attribute graph, the method comprising the following steps: Step S1: Construction of bipartite graph The drug-target interaction relationships are collected from the database, and a bipartite graph of drug-target interactions is constructed.

[0005] Step S2: Feature enhancement processing The drug SMILES are used to generate molecular graphs using RDKit. The drug pharmacological properties, target protein sequences, pathway and chromosome information are converted into one-hot encoding, and the similarity matrix of the drug and target is calculated respectively. The attribute matrix and similarity matrix of the drug and target are fused, and the dimension is reduced by splicing and principal component analysis to obtain a feature-enhanced matrix representation.

[0006] Step S3: Define the model Define the model structure and set training parameters.

[0007] Step S4: Training the model It consists of ten iterations, where each subset is used alternately as a test set, and the remaining nine subsets constitute the training set and are input into the model for training. The final performance indicator is the average of the ten test results.

[0008] Step S5: Evaluate the model AUROC, AUPR and F1 scores are obtained based on the training results. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 is a flow chart of the present invention; DETAILED DESCRIPTION

[0010] The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention. The present invention will now be described in further detail by way of the accompanying drawings and examples.

[0011] Figure 1 The present invention provides a schematic diagram of a drug-target interaction prediction model process based on a feature-enhanced bipartite attribute graph. Figure 1 As shown, this embodiment mainly includes the following steps: Step S1: Bipartite graph construction The drug-target interaction relationships are collected from the database, and a bipartite graph of drug-target interactions is constructed.

[0012] Step S2: Feature enhancement processing The drug SMILES are used to generate molecular graphs using RDKit. The drug pharmacological properties, target protein sequences, pathway and chromosome information are converted into one-hot encoding, and the similarity matrix of the drug and target is calculated respectively. The attribute matrix and similarity matrix of the drug and target are fused, and the dimension is reduced by splicing and principal component analysis to obtain a feature-enhanced matrix representation.

[0013] Step S3: Define the model Utilize aggregated graph convolutional networks and fully connected layers for feature extraction and prediction tasks.

[0014] Step S4: Training the model It consists of ten iterations, where each subset is used alternately as a test set, and the remaining nine subsets constitute the training set and are input into the model for training. The final performance indicator is the average of the ten test results.

[0015] Step S5: Evaluate the model AUROC, AUPR and F1 scores are obtained based on the training results.

[0016] Example 1 The present invention is tested using a dataset of a drug-target interaction prediction task This task includes two data sets. Our self-built TCM data set includes 17 traditional Chinese medicines, 222 compounds, 270 targets, and 2211 pairs of interaction relationships. Wang's DrugBank data set includes 708 drugs, 1512 targets, and 1923 pairs of interaction relationships. The traditional Chinese medicine data set integrates two attributes of drugs (SMILES and pharmacological attributes) and three attributes of targets (protein sequence, pathway, and chromosome). The DrugBank data set contains the chemical structure attributes of drugs and the primary protein sequence attributes of targets. At the same time, it was compared with 5 existing methods, namely DTI-GCN, DTI-GAT, DTI-SuperGAT, DTI-MFC, and DTI-Graph. The existing 5 methods all run under their optimal parameters. The relevant parameters of the method of the present invention are set as follows: the number of epochs is 150, and the learning rate is 0.001.

[0017] Table 1 Experimental comparison on TCM dataset: Models AUROC AUPR F1 DTI-BGCGCN 0.9505 0.9601 0.8915 DTI-GCN 0.8979 0.9296 0.8431 DTI-GAT 0.9068 0.9049 0.8475 DTI-SuperGAT 0.9027 0.8944 0.8343 DTI-MFC 0.9149 0.9272 0.8813 DTI-Graph 0.9253 0.9316 0.8849 Table 1 shows the various indexes of the present invention compared with the other five algorithms on the TCM data set. It can be seen from the table that the AUROC value of the present invention is 0.9505, the AUPR value is 0.9601, and the F1 value is 0.8915. Compared with the other five algorithms, it is obvious that the effect of the present invention is the best.

[0018] Table 2 Experimental comparison on DrugBank dataset: Models AUROC AUPR F1 DTI-BGCGCN 0.9493 0.9524 0.8314 DTI-GCN 0.9184 0.9339 0.8174 DTI-GAT 0.9282 0.9331 0.8217 DTI-SuperGAT 0.9335 0.9352 0.8273 DTI-MFC 0.8152 0.8777 0.7925 DTI-Graph 0.8533 0.8867 0.8213 Table 2 shows the various indexes of the present invention compared with the other five algorithms on the DrugBank data set. It can be seen from the table that the AUROC value of the present invention is 0.9493, the AUPR value is 0.9524, and the F1 value is 0.8314. Compared with the other five algorithms, it is obvious that the effect of the present invention is the best.

Claims

1. A drug-target interaction prediction method based on a feature-enhanced bipartite attribute graph, which is characterized by multivariate attribute information and a unique feature enhancement method, comprising the following steps: Step S1: Construction of bipartite graph The drug-target interaction relationships are collected from the database, and a bipartite graph of drug-target interactions is constructed.

2. Step S2: Feature enhancement processing The drug SMILES are used to generate molecular graphs using RDKit. The drug pharmacological properties, target protein sequences, pathway and chromosome information are converted into one-hot encoding, and the similarity matrix of the drug and target is calculated respectively. The attribute matrix and similarity matrix of the drug and target are fused, and the dimension is reduced by splicing and principal component analysis to obtain a feature-enhanced matrix representation.

3. Step S3: Define the model Define the model structure and set training parameters.

4. Step S4: Training the model It consists of ten iterations, where each subset is used alternately as a test set, and the remaining nine subsets constitute the training set and are input into the model for training. The final performance indicator is the average of the ten test results.

5. Step S5: Evaluate the model AUROC, AUPR and F1 scores are obtained based on the training results.

6. The method for predicting drug-target interaction based on feature-enhanced bipartite attribute graph according to claim 1, characterized in that: In step S1, the drug-target interaction relationship is collected from the TCMSP database and the DrugBank database.

7. The method for predicting drug-target interaction based on feature-enhanced bipartite attribute graph according to claim 1, characterized in that: In step S2, according to the drug SMILES formula, it is first converted into a molecular graph through RDKit, and then the drug pharmacological properties, target protein sequence, pathway and chromosome information are converted into one-hot encoding, and the Jaccard similarity is calculated. Feature enhancement processing is performed through splicing and principal component analysis dimensionality reduction to obtain the feature representation of the drug target.

8. The method for predicting drug-target interaction based on feature-enhanced bipartite attribute graph according to claim 1, characterized in that: Step S3 will use a clustering graph convolutional network to divide the nodes in the bipartite graph into several closely connected subgraphs, effectively integrating neighborhood information from the nodes, thereby enhancing the representation ability of the nodes to achieve feature extraction.

Images