Drug target binding affinity prediction method based on graph neural network

Abstract

The invention discloses a drug target binding affinity prediction method based on a graph neural network, and belongs to the technical field of biological information, and the method comprises the following steps: S1, sample preprocessing; s2, constructing a model; s3, model training; and S4, DTA prediction is carried out. The method comprises the following steps: firstly, obtaining related data through a Davis data set and a KIBA data set, respectively obtaining a drug molecular map, a drug Morgan fingerprint and protein sequence information through data preprocessing, and respectively inputting the drug molecular map, the drug Morgan fingerprint and the protein sequence information into a GATv2 network, a multi-layer perceptron network and a multi-scale convolutional neural network for feature extraction; and performing feature splicing on the Morgan fingerprints and molecular map features through layer attention to obtain drug features, splicing the drug features with protein features, and inputting the spliced drug features into a prediction network to perform drug target binding affinity prediction.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a method for predicting drug target binding affinity based on graph neural network. Background Art

[0002] Proteins are involved in numerous cellular processes, and many diseases are caused by abnormal proteins. Therefore, a large number of drugs have been developed to act on proteins. Drug-target binding affinity refers to the strength or tightness of the binding of a drug to its target. In the process of drug design and discovery, understanding the binding affinity between a drug and its target is crucial for evaluating the efficacy of the drug and selecting appropriate drug candidate compounds. Structural changes in drugs can have a huge impact on the strength of protein binding affinity, so it is very important to find specific drug-protein pairs.

[0003] Although experimental methods have been applied to drug discovery and drug repurposing, in order to identify effective and safe drugs for a given protein, pharmacologists must test thousands of chemical compounds, which is very time-consuming and laborious. The rise of computers has brought computational methods into drug discovery. In the biomedical field, computational methods have become a new hot spot as a supplement to experimental methods. In drug discovery and drug repositioning, drug-target interaction (DTI) plays a vital role. Drug-target binding affinity (DTA) is a regression problem. Drug-target binding affinity provides information about the strength of the interaction between drug-target pairs, usually expressed as dissociation constants, inhibition constants, and half-maximal inhibition concentrations. With the development of machine learning and artificial intelligence technologies, more and more studies have begun to explore the use of these technologies to predict the binding affinity between drugs and targets. Machine learning algorithms can learn patterns from a large amount of known drug-target interaction data, thereby realizing the prediction of unknown drug-target interactions. This data-driven approach usually has certain advantages in prediction accuracy and efficiency.

[0004] In short, as diseases are constantly being conquered, the speed of drug discovery has also become extremely high. Technical means relying on deep learning have greatly improved the efficiency and success rate in actual drug research and development. Summary of the invention

[0005] The technical problem to be solved by the present invention is that the accuracy of drug target binding affinity prediction may be improved. In terms of drug feature extraction, the global information of the drug is difficult to capture, and a simple two-layer graph neural network cannot obtain its specific structural information, such as ring structure; in terms of protein feature extraction, protein feature extraction is not comprehensive. A drug target binding affinity prediction method based on graph neural network is provided.

[0006] The present invention solves the above technical problems through the following technical solutions, and the present invention comprises the following steps:

[0007] S1: Sample pretreatment

[0008] The Davis and KIBA datasets are divided into training sets and test sets according to the set ratio, and the drug protein information in the training sets and test sets is processed into drug molecular graphs, Morgan fingerprints and protein sequence information;

[0009] S2: Model construction

[0010] Construct an initial drug-target binding affinity prediction model, including a drug protein feature extraction model based on graph neural network and multi-layer perceptron network, and a prediction model based on prediction network;

[0011] S3: Model training

[0012] The initial drug-target binding affinity prediction model is trained using the training set, and the final drug-target binding affinity prediction model is obtained after the training;

[0013] S4: DTA Forecast

[0014] The final drug-target binding affinity prediction model is tested on the test set and the binding affinity index is output.

[0015] Furthermore, in step S1, the specific processing process is as follows:

[0016] S11: Divide the Davis and KIBA datasets into training sets and test sets according to the set ratio;

[0017] S12: Use RDkit to preprocess the drug SMILES sequences of samples in the dataset into a graph with nodes and adjacency matrix, that is, to obtain drug molecular graph information;

[0018] S13: Generate a Morgan fingerprint with a radius of 2 and a bit number of 1024 using RDkit according to the drug SMILES sequence, and obtain the Morgan fingerprint information;

[0019] S14: The protein FASTA sequence of the sample in the data set is converted into a corresponding integer sequence through the corresponding vocabulary, that is, the protein sequence information is obtained.

[0020] Furthermore, in step S2, the specific processing process of the drug protein feature extraction model is as follows:

[0021] S21: Send the drug molecule graph to the GATv2 network to obtain 96-dimensional drug molecule graph features. The GATv2 network is a graph neural network.

[0022] S22: Send the drug Morgan fingerprint into the multi-layer perceptron network to obtain the 96-dimensional drug Morgan fingerprint feature;

[0023] S23: The drug Morgan fingerprint features and drug molecular graph features are concatenated and reduced in dimension after passing through the layer attention module to obtain a 96-dimensional drug feature;

[0024] S24: The protein integer sequence is passed through a multi-scale convolutional neural network to obtain 96-dimensional protein features.

[0025] Furthermore, in step S2, the processing process of the prediction model is as follows:

[0026] S25: concatenate the drug feature with the protein feature to obtain a concatenated feature;

[0027] S26: Send the spliced ​​features into the prediction network to obtain the binding affinity prediction results.

[0028] Furthermore, in step S21, the drug molecule graph is represented as G=(V,E), wherein V represents the node of the drug atom, and E represents the set of edges connecting the drug nodes. Each node obtains an n-dimensional representation through Deepchem including extracted atomic features, bond features, substructure features, and topological structure features.

[0029] Furthermore, in step S21, the GATv2 network adopts a dynamic attention mechanism to aggregate the corresponding graph representations obtained through the drug SMILES sequence to obtain drug molecular graph features, and the aggregation method is as follows:

[0030] e(h i ,h j )=a T LeakyRelu(W 0 [h i ||h j ])j∈N i

[0031] Among them, e(h i ,h j ) represents the importance of neighbor node j to the original node i, N i Represented as the set of all neighbor nodes of node i is a learnable matrix.

[0032] Furthermore, in step S22, the molecular structure is traversed through the drug Morgan fingerprint, and the environmental information around each atom is recorded to obtain a 1024-dimensional vector, and then the 1024-dimensional vector is reduced to a 96-dimensional vector through a multi-layer perceptron network, and the supplementary structural information of the drug is output, that is, the drug Morgan fingerprint feature is obtained.

[0033] Furthermore, in step S23, the drug molecular graph features and the drug Morgan fingerprint features are spliced ​​by the layer attention module, and the splicing method is as follows:

[0034] h k =[W 1 h j ∥(1-W 1 )h i ].

[0035] Furthermore, in step S24, the multi-scale convolutional neural network includes three convolutional branches, different convolutional branches have different numbers of convolutional layers, and the corresponding convolutional kernel sizes are also different. The processing method of the multi-scale convolutional neural network is as follows:

[0036] y=W 2 (m(F 1 (Q))∥m(F 2 (Q))∥m(F 3 (Q)

[0037] Among them, y represents protein features, Q is the input matrix obtained by protein integer sequence, and F i represents the i-th convolution branch, m represents the maximum pooling process, and W 2 is a learnable matrix.

[0038] Furthermore, in step S26, the prediction network includes three linear transformation layers and a fully connected layer. A Relu activation function layer and a dropout layer are respectively connected after each linear transformation layer. The binding index is obtained through the three linear transformation layers, and then processed through the fully connected layer to obtain the final drug protein binding affinity index.

[0039] Compared with the prior art, the present invention has the following advantages:

[0040] 1. The information of Morgan fingerprint is integrated into the drug feature extraction part, which can supplement the missing structural information of the drug molecular graph when it passes through the graph neural network. After obtaining the molecular graph features and Morgan fingerprint features, layer attention is used to obtain their corresponding contributions, and finally obtain more comprehensive drug feature information;

[0041] 2. For the entire protein sequence, only a part of it plays a role in the reaction. For example, the action site in the enzyme protein is only a part of the area involved in the catalytic reaction. Therefore, a multi-scale convolutional neural network is used in the protein feature extraction part. It is a network composed of 3 convolution branches. Different convolution branches have different numbers of convolution layers, which means that the size of their convolution kernels is different, which enables the convolutional neural network to extract more local information at more scales. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 is a schematic diagram of a process of predicting drug target binding affinity based on a graph neural network in an embodiment of the present invention;

[0043] Figure 2 Schematic diagram of the drug target binding affinity prediction model structure in an embodiment of the present invention;

[0044] Figure 3 is a schematic diagram of the structure of a multi-scale convolutional neural network in an embodiment of the present invention;

[0045] Figure 4 is a schematic diagram of the structure of a prediction network in a prediction model in an embodiment of the present invention;

[0046] Figure 5 : is the attention heat map of different graph neural networks in the embodiments of the present invention, where a is the GAT network and b is the GATv2 network;

[0047] Figure 6 Schematic diagram of fitting curves of the final drug target binding affinity prediction model in the embodiment of the present invention on different data sets, wherein a is the Davis data set and b is the KIBA data set. DETAILED DESCRIPTION

[0048] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given, but the protection scope of the present invention is not limited to the following embodiment.

[0049] This embodiment provides a technical solution: a method for predicting drug target binding affinity based on a graph neural network, comprising the following steps:

[0050] S1: Sample pretreatment

[0051] The Davis and KIBA datasets are divided into training sets and test sets according to the set ratio, and the drug protein information in the training sets and test sets is processed into drug molecular graphs, Morgan fingerprints and protein sequence information;

[0052] S2: Model construction

[0053] Construct an initial drug-target binding affinity prediction model, including a drug protein feature extraction model based on graph neural network and multi-layer perceptron network, and a prediction model based on prediction network;

[0054] S3: Model training

[0055] The initial drug-target binding affinity prediction model is trained using the training set, and the final drug-target binding affinity prediction model is obtained after the training;

[0056] S4: DTA Forecast

[0057] The final drug-target binding affinity prediction model is used to predict the binding affinity index on the test set.

[0058] In this embodiment, step S1 includes the following sub-steps:

[0059] S11: The data sets used in the present invention are Davis and KIBA data sets, wherein the Davis data set includes interactions between 442 proteins and 68 drugs. The KIBA data set initially contained 467 proteins, 52,498 drugs, and KIBA scores between proteins and drugs. The KIBA score is used to measure the biological activity of kinase inhibitors and is considered to be related to the binding affinity of the kinase. In order to make the data set more balanced, a data set containing 229 proteins and 2,111 drugs was finally formed after screening. Among them, all drugs in the Davis data set interact with proteins, and there are 118,254 interactions in the KIBA data set.

[0060] Table 1 The number and proportion of various types of data in the dataset

[0061] Davis KIBA protein 442 299 drug 68 2111 interaction 30056 118254

[0062] S12: The drug SMILES sequence was preprocessed into a graph with nodes and adjacency matrix and Morgan fingerprint with radius 2 and bit number 1024 by RDkit. The protein FASTA sequence was converted into an integer sequence by dictionary. SMILES is a short ASCII string used to describe chemical substances.

[0063] In this example, the specific processing of the initial drug target binding affinity prediction model is as follows:

[0064] S21: Send the drug molecule graph to the GATv2 network to obtain 96-dimensional drug molecule graph features. The GATv2 network is a graph neural network.

[0065] S22: Send the drug Morgan fingerprint into the multi-layer perceptron network to obtain the 96-dimensional drug Morgan fingerprint feature;

[0066] S23: The drug Morgan fingerprint features and drug molecular graph features are concatenated and reduced in dimension after passing through the layer attention module to obtain a 96-dimensional drug feature;

[0067] S24: Send the protein integer sequence to the multi-scale convolutional neural network to obtain 96-dimensional protein features;

[0068] S25: splicing the drug feature with the protein feature to obtain a spliced ​​feature;

[0069] S26: Send the spliced ​​features into the prediction network to obtain the binding affinity prediction results.

[0070] In this embodiment, the GATv2 network adopts a dynamic attention mechanism to aggregate the corresponding graph representations obtained through the drug SMILES sequence to obtain the drug molecule graph features, that is, the graph G = (V, E), where V represents the nodes of drug atoms and E represents the set of edges connecting the drug nodes. The n-dimensional representation of each node obtained by Deepchem includes the extracted atomic features (atom type, charge, number of bonds, number of hydrogen bonds, etc.), bond features (bond type (single bond, double bond, triple bond), bond length, bond angle, etc.), substructure features (such as the number and size of rings, connections between rings, etc.) and topological structure features (such as the topological structure of the molecule, connection patterns between atoms, etc.). Among them, e(h i ,h j ) represents the importance of neighbor node j to the original node i, N i Represented as the set of all neighbor nodes of node i is a learnable matrix that is aggregated as follows:

[0071] e(h i ,h j )=a T LeakyRelu(W 0 [h i ||h j ])j∈N i

[0072] In this embodiment, the molecular structure is traversed through the drug Morgan fingerprint, and the environmental information around each atom is recorded to obtain a 1024-dimensional vector. The 1024-dimensional vector is then reduced to a 96-dimensional vector through a multi-layer perceptron network, and the supplementary structural information of the drug is output, that is, the drug Morgan fingerprint feature is obtained.

[0073] In this embodiment, the drug molecular graph features and the drug Morgan fingerprint features are spliced ​​by layer attention, and the splicing method is as follows:

[0074] h k =[W 1 h j ∥(1-W 1 )h i ].

[0075] In this embodiment, the multi-scale convolutional neural network includes three convolutional branches, where different convolutional branches have different numbers of convolutional layers, which means that the sizes of their convolutional kernels are different, which enables the convolutional neural network to extract local information at more scales. The protein integer sequence is embedded to obtain the input matrix The input matrix Q passes through F i Mapping to Matrix After activation by the activation function Relu, the processing method of the multi-scale convolutional neural network is as follows:

[0076] y=W 2 (m(F 1 (Q))∥m(F 2 (Q))∥m(F 3 (Q)

[0077] Among them, F i It represents a branch among multiple branches, m represents the maximum pooling, is a learnable matrix.

[0078] In this embodiment, after obtaining the drug and protein feature vectors, they are concatenated and sent to the prediction network, which includes three linear transformation layers and a fully connected layer. After each linear transformation layer, a Relu activation function layer and a dropout layer are connected respectively, and then the final drug-protein binding affinity index is obtained through processing in the fully connected layer. The prediction network maps the high-dimensional features obtained after the fusion of drug-protein features to the final output, i.e., the drug-target binding affinity index.

[0079] In this embodiment, Table 2 shows the impact of adjusting different modules on the experimental results, where MCNN means replacing with a multi-scale convolution module, ECFP means adding Morgan fingerprint information and feature extraction modules, and all performances are improved in Davis and KIBA datasets compared to adding only multi-scale convolution. LA (layer) means adding layer attention, and all indicators are improved in both datasets compared to adding ECFP and MCNN modules, and the final results are obtained in the ablation experiment. The lower the MSE, the better, CI and The higher the better.

[0080] Table 2 Comparison of experimental results of adding different modules

[0081]

[0082] Table 3 shows the results of model comparison of different graph neural networks, including the mainstream models of graph neural networks, including GCN, GAT, GIN, GCN-GAT, and GATv2 networks. Experimental data show that the GATv2 network has achieved good results in all indicators of the two datasets. Compared with other graph neural networks, only the MSE of the KIBA dataset is slightly lower than GIN by 0.1%, and all other indicators are the best in the comparison of graph neural networks. This is because the GATv2 network uses dynamic attention, and in the Davis and KIBA datasets, the Davis dataset has a small number of samples and an unbalanced label distribution, while the KIBA dataset has a large number of single-label samples and is very concentrated. The GATv2 network can better distinguish the similarities and differences between samples.

[0083] Table 3 Comparison of experimental results of different graph neural networks

[0084]

[0085] Table 4 shows the experimental results of different networks on Davis and KIBA datasets, including traditional machine learning algorithms and the latest deep learning algorithms. Traditional machine learning methods include KronRLS and SimBoost, and deep learning methods include DeepDTA, GraphDTA, MATT DTI, DeepGLSTM, and DGDTA. As can be seen from Table 4, the drug target binding affinity prediction method involved in the present invention is ahead of other models in many indicators in the test set. The CIs are 0.903 and 0.904 in the two datasets, respectively, which are the best among all models.

[0086] Table 4 Experimental results of different networks on Davis and KIBA datasets

[0087]

[0088]

[0089] In summary, the drug-target binding affinity prediction method based on graph neural network in the above embodiment can supplement the structural information of drug characteristics by adding the Morgan fingerprint information of the drug to the drug protein sequence data. For example, the graph neural network cannot know the specific structural information, such as ring structure, etc.; an optimization scheme is proposed to aggregate the drug molecular graph features and the drug Morgan fingerprint features by adding LA attention; since the protein participates in the drug action only at a certain point, the feature extraction of the protein is also improved in the present invention, and multi-scale convolution is used to obtain local features of different scales to simulate drug target binding; finally, the effectiveness of the improvement is proved through various comparative experiments, and the best results are obtained in multiple evaluation indicators, proving that it has the best effect.

[0090] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A drug target binding affinity prediction method based on graph neural network, It is characterized in that The following steps are involved: S1: Sample pretreatment The Davis and KIBA datasets are divided into training sets and test sets according to the set ratio, and the drug protein information in the training sets and test sets is processed into drug molecular graphs, Morgan fingerprints and protein sequence information; S2: Model construction Construct an initial drug-target binding affinity prediction model, including a drug protein feature extraction model based on graph neural network and multi-layer perceptron network, and a prediction model based on prediction network; S3: Model training The initial drug-target binding affinity prediction model is trained using the training set, and the final drug-target binding affinity prediction model is obtained after the training; S4: DTA Forecast The final drug-target binding affinity prediction model is tested on the test set and the binding affinity index is output.

2. A drug target binding affinity prediction method based on graph neural network according to claim 1, It is characterized in that In step S1, the specific processing process is as follows: S11: Divide the Davis and KIBA datasets into training sets and test sets according to the set ratio; S12: Use RDkit to preprocess the drug SMILES sequences of samples in the dataset into a graph with nodes and adjacency matrix, that is, to obtain drug molecular graph information; S13: Generate a Morgan fingerprint with a radius of 2 and a bit number of 1024 using RDkit according to the drug SMILES sequence, and obtain the Morgan fingerprint information; S14: The protein FASTA sequence of the sample in the data set is converted into a corresponding integer sequence through the corresponding vocabulary, that is, the protein sequence information is obtained.

3. A drug target binding affinity prediction method based on graph neural network according to claim 2, It is characterized in that In step S2, the specific processing process of the drug protein feature extraction model is as follows: S21: Send the drug molecule graph to the GATv2 network to obtain 96-dimensional drug molecule graph features. The GATv2 network is a graph neural network. S22: Send the drug Morgan fingerprint into the multi-layer perceptron network to obtain the 96-dimensional drug Morgan fingerprint feature; S23: The drug Morgan fingerprint features and drug molecular graph features are concatenated and reduced in dimension after passing through the layer attention module to obtain a 96-dimensional drug feature; S24: The protein integer sequence is passed through a multi-scale convolutional neural network to obtain 96-dimensional protein features.

4. A drug target binding affinity prediction method based on graph neural network according to claim 3, It is characterized in that In step S2, the processing process of the prediction model is as follows: S25: splicing the drug feature with the protein feature to obtain a spliced ​​feature; S26: Send the spliced ​​features into the prediction network to obtain the binding affinity prediction results.

5. A method for predicting drug target binding affinity based on graph neural network according to claim 4, It is characterized in that In step S21, the drug molecule graph is represented as G=(V,E), wherein V represents the node of the drug atom, and E represents the set of edges connecting the drug nodes. Each node obtains an n-dimensional representation through Deepchem including extracted atomic features, bond features, substructure features, and topological structure features.

6. A method for predicting drug target binding affinity based on graph neural network according to claim 4, It is characterized in that In step S21, the GATv2 network adopts a dynamic attention mechanism to aggregate the corresponding graph representations obtained through the drug SMILES sequence to obtain drug molecular graph features. The aggregation method is as follows: e(h i h j )=a T LeakyRelu(W 0 [h i ||h j ])j∈N i Among them, e(h i ,h j ) represents the importance of neighbor node j to the original node i, N i Represented as the set of all neighbor nodes of node i is a learnable matrix.

7. A method for predicting drug target binding affinity based on graph neural network according to claim 4, It is characterized in that In step S22, the molecular structure is traversed through the drug Morgan fingerprint, and the environmental information around each atom is recorded to obtain a 1024-dimensional vector, which is then reduced to a 96-dimensional vector through a multi-layer perceptron network, and the supplementary structural information of the drug is output, that is, the drug Morgan fingerprint feature is obtained.

8. A method for predicting drug target binding affinity based on graph neural network according to claim 4, It is characterized in that In step S23, the drug molecular graph features and the drug Morgan fingerprint features are spliced ​​together through the layer attention module, and the splicing method is as follows: h k =[W 1 h j ∥(1-W 1 )h i ]。 9. A method for predicting drug target binding affinity based on graph neural network according to claim 4, It is characterized in that In step S24, the multi-scale convolutional neural network includes three convolution branches. Different convolution branches have different numbers of convolutional layers, and the corresponding convolution kernel sizes are also different. The processing method of the multi-scale convolutional neural network is as follows: y=W 2 (m(F 1 (Q))∥m(F 2 (Q))∥m(F 3 (Q)) Among them, y represents protein features, Q is the input matrix obtained by protein integer sequence, and F i represents the i-th convolution branch, m represents the maximum pooling process, and W 2 is a learnable matrix.

10. A method for predicting drug target binding affinity based on graph neural network according to claim 4, It is characterized in that In step S26, the prediction network includes three linear transformation layers and a fully connected layer. A Relu activation function layer and a dropout layer are respectively connected after each linear transformation layer. The binding index is obtained through the three linear transformation layers, and then processed through the fully connected layer to obtain the final drug protein binding affinity index.

Images