Drug-target binding affinity prediction model and method based on graph dilation convolution strategy

Abstract

The present invention discloses a drug-target binding affinity prediction model and method based on a graph dilation and convolution strategy. The present invention performs feature encoding on drug molecules, local information, and targets through a feature encoding module. Feature extraction is performed on the drug molecular structure, local chemical information of the drug molecule, and target structure through a multi-channel general aggregation network module using a graph dilation and convolution strategy, a drug sequence representation learning module using a multi-layer residual convolution network, and a target sequence representation learning module using a bidirectional long-short cycle memory network. Finally, the extracted global drug structural features, local drug chemical features, and target sequence features are concatenated and passed through a DTA prediction module to predict the drug-target affinity value. This effectively improves the drug-target binding affinity prediction accuracy and the success rate of the drug redirection process, solving the problems of low drug-target binding affinity prediction accuracy and low drug redirection process success rate.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and more particularly to a drug target binding affinity prediction model and method based on a graph dilation and convolution strategy. Background Art

[0002] In the field of bioinformatics, the traditional drug discovery process requires significant time and cost, and carries a high risk of failure. Currently, most drugs achieve their therapeutic effects through in vivo interactions with specific target molecules such as enzymes, nuclear receptors, G-protein-coupled receptors, and ion channels. Therefore, accurately predicting drug-target interactions (DTI) can significantly shorten the drug development process.

[0003] The strength of a drug-target interaction is typically measured by the binding affinity between the two. Therefore, DTI prediction is also called drug-target binding affinity (DTA) prediction. Researchers hope to discover potential DTIs from existing drugs to accelerate the drug development process, a process known as drug repositioning. Drug repositioning aims to identify new pharmacological indications for existing drugs. It allows newly developed drugs to treat diseases beyond their intended therapeutic use, while also establishing new therapeutic uses for known drugs. With the advancement of computing technology and artificial intelligence algorithms, various computational models have emerged to predict large-scale potential drug-target relationships. For example, the prediction method based on machine learning mainly makes predictions through multiple similarity networks between drugs and targets, interaction networks between drugs, and interaction networks between targets; the prediction method based on convolutional neural networks mainly makes predictions through the chemical structure sequence of drug small molecules, the amino acid sequence of target proteins, and the binding affinity values ​​of known drug-target pairs; the prediction method based on graph neural networks regards the structure of the drug as a molecular graph, uses the graph neural network model to extract the global chemical structure feature information in the molecular graph, and combines the temporal features in the amino acid sequence of the target protein to predict the drug-target binding affinity.

[0004] In CN202211004610.X, a drug target prediction method based on a graph convolutional neural network is disclosed. The method adaptively selects appropriate meta-paths based on different path distances, allowing the selected meta-paths to cover more valid information. The method uses an attention mechanism to fuse neighborhood features obtained under different meta-paths, differentially integrating features of different types of nodes. Features of different near and far neighborhoods are obtained hierarchically based on distance, which can better capture high-order topological structural features. The method uses an MLP network that considers inter-class confidence, path confidence, and degree confidence to perform final drug target prediction. However, this method ignores the inherent structural information of the drug and target, resulting in low drug target prediction accuracy and affecting the success rate of the drug redirection process. Summary of the Invention

[0005] The main technical problem to be solved by the present invention is to provide a drug-target binding affinity prediction model based on graph dilation convolution strategy to address the shortcomings of the existing technology in that the prediction accuracy of drug-target binding affinity is not high and the success rate of the drug redirection process is low.

[0006] Another technical problem solved by the present invention is to provide a drug target binding affinity prediction method based on graph dilation convolution strategy.

[0007] The purpose of the present invention is achieved through the following technical solutions:

[0008] A drug-target binding affinity prediction model based on graph dilation convolution strategy includes a feature encoding module, a representation learning module and a DTA prediction module.

[0009] The feature encoding module performs initial feature encoding on the drug molecular structure, local chemical information of the drug molecule and the target structure;

[0010] The representation learning module includes a multi-channel general aggregation network module, a drug sequence representation learning module and a target sequence representation learning module:

[0011] The multi-channel general aggregation network module (McGEN module) includes two channels, each of which includes three GENs. It aggregates node feature information based on the atomic node association feature matrix of the drug molecular structure in the feature encoding module, learns the potential feature expression of each atom, and improves prediction accuracy.

[0012] The drug sequence representation learning module (MLRCN module) includes a three-layer feature extraction network constructed by 6 residual connection blocks. Each residual connection block has two CNN modules, and each CNN module is followed by a ReLU activation function, which outputs the potential feature expression of the local chemical information of the drug based on the initial features of the local chemical information of the drug molecules, thereby enriching the characteristic information of the drug.

[0013] The target sequence representation learning module (BiLSTM module) consists of two long short-term memory network modules (LSTM modules), which extract structural features from the target sequence from the forward and backward directions respectively to obtain the deep feature expression of the target.

[0014] The DTA prediction module includes a multi-layer perceptron, which integrates the outputs of the multi-channel general aggregation network module, the drug sequence representation learning module and the target sequence representation learning module and then feeds them into the MLP network for DTA prediction.

[0015] Furthermore, the number of GEN feature channels is 78, 312 and 624 respectively.

[0016] Furthermore, the number of residual connection blocks of the three-layer feature extraction network is 3, 2 and 1 respectively.

[0017] A method for predicting drug-target binding affinity based on a graph dilation convolution strategy comprises the following steps:

[0018] S1. Representing the drug molecule sequence and the target amino acid sequence as two-dimensional vector structures using a feature encoding module, and converting the drug into a two-dimensional structure to obtain initial features of the drug molecular structure, initial features of the drug's local chemical information, and initial features of the target;

[0019] S2. Extract the initial features of the drug molecule, the initial features of the drug's local chemical information, and the initial features of the target through the multi-channel general aggregation network module, the drug sequence representation learning module, and the target sequence representation learning module, respectively, to obtain the drug's global structural features, the drug's local chemical features, and the target sequence features;

[0020] S3. The extracted global structural features, local chemical features and target sequence features are concatenated and then used in the DTA prediction module to predict the drug-target affinity value.

[0021] Furthermore, the initial characteristics of the drug molecule Expressed as:

[0022] Where v i is the i-th atomic node of the drug molecule, is the initial feature matrix The size of N atom is the number of atomic nodes of the drug molecule, N fea For node v i The number of neighbor nodes;

[0023] Furthermore, the atomic node association characteristics of the drug molecule

[0024] Furthermore, the initial feature H of the local chemical information of the drug s Expressed as

[0025] Furthermore, the initial structural characteristics of the target H l Expressed as

[0026] Furthermore, the step of obtaining the initial characteristics of the drug molecule through a multi-channel universal aggregation network module includes:

[0027] Initial characterization of drug molecules Associate features with atomic nodes Channel 1 of the McGEN module is input. Channel 1 aggregates and updates the feature information of the neighboring nodes of the central node by stacking three layers of GEN to obtain the feature expression of the central node. Channel 2 uses the graph expansion convolution strategy to construct the indirect neighbor node association features of atomic nodes Indicates Jn Multiply, Jn is the number of jumps. By stacking GEN, the feature information of the indirect neighbor nodes of the central node is aggregated and updated to obtain the feature expression of the central node Finally, the McGEN module cascades the outputs of the two channels to obtain D k The potential feature expression

[0028] Furthermore, the steps of the initial features of the drug local chemical information through the drug sequence representation learning module are:

[0029] The input is the initial feature matrix H of the local chemical information of the drug encoded by the feature encoding module s After the drug sequence representation learning module, the final output is the potential feature expression of the local chemical information of the drug

[0030] Furthermore, the initial features of the target are represented by the target sequence learning module in the following steps:

[0031] (1) The LSTM model first calculates the input feature x at time t. t and the hidden layer state h at the previous moment t-1 Calculate the forget gate and memory gate of the cell and select the information to be forgotten f t and the information that needs to be remembered t , while recording the temporary cell state in:

[0032] f t =σ(W f ·[h t-1 ,x t]+b f ),

[0033] i t =σ(W i ·[h t-1 ,x t ]+b i ),

[0034]

[0035] Where σ is the Sigmoid activation function, W is the learnable weight matrix, and b is the bias term;

[0036] (2) LSTM will be based on the cell state C at the previous moment t-1 Calculate the current cell state C t , its calculation expression is as follows:

[0037]

[0038] (3) After comprehensively considering the cell information at the previous moment, LSTM calculates the hidden layer state h at the current moment. t Calculate and get

[0039] O t =σ(W O ·[h t-1 ,x t ]+b O ),

[0040] h t =O t ×tanh(C t ).

[0041] (4) BiLSTM will output hidden layer state information with the same length as the target encoding sequence

[0042]

[0043] Furthermore, the multi-layer perceptron performs the following steps to predict the drug-target affinity value:

[0044] (1) and After integration, it is sent to the MLP network for DTA prediction. The integrated drug-target pair features are

[0045] (2) The drug-target pair features are used as the input of the MLP, and the binding affinity value is calculated after forward propagation as the output of the MLP. Subsequently, the MLP establishes a loss function by comparing the predicted binding affinity value with the true binding affinity value, and uses backpropagation to update the hidden layer.

[0046] Furthermore, the loss function used by the MLP is:

[0047]

[0048] where Y i is the true binding affinity value, To predict the binding affinity value, n is the number of drug-target pairs.

[0049] Compared with the prior art, the beneficial effects are:

[0050] When extracting the global structural features of the drug molecule graph, the present invention adopts a graph dilation convolution strategy to construct an atomic node association matrix for the atomic nodes in the indirect neighborhood of each atomic node in the drug molecule graph, thereby alleviating the problem of decreased prediction accuracy caused by the same or similar topological structures of the subgraphs during the graph convolution process. In the process of extracting local chemical information of drugs, the present invention utilizes a multi-layer residual convolution module to effectively extract the potential structural features in the SMILES sequence by stacking residual convolution blocks, thereby enriching the characteristic information of the drug molecule graph. When extracting target sequence features, a bidirectional long-short cycle memory network is used to model the amino acid sequence of the target protein in the forward and backward directions, respectively, comprehensively considering the temporal features of the sequence, obtaining the deep characteristic information of the target, and finally integrating these features and predicting the affinity between the two. By combining the above-mentioned means, the present invention effectively improves the prediction accuracy of drug-target binding affinity and the success rate of the drug redirection process, realizes the drug repositioning process, and accelerates drug development. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is the McGEN module architecture diagram in Example 1;

[0052] Figure 2 This is the MLRCN module architecture diagram in Example 1;

[0053] Figure 3 This is the BiLSTM module architecture diagram in Example 1;

[0054] Figure 4 This is the DTA prediction module architecture diagram in Example 1;

[0055] Figure 5 This is a flow chart in Example 2;

[0056] Figure 6 This is a diagram of the drug molecule and protein structure in Example 4. DETAILED DESCRIPTION

[0057] The present invention will be further explained and illustrated below with reference to the embodiments, but the specific embodiments do not limit the present invention in any form.

[0058] Example 1

[0059] This embodiment provides a drug-target binding affinity prediction model based on a graph dilation convolution strategy, comprising a feature encoding module, a representation learning module, and a DTA prediction module;

[0060] The feature encoding module performs initial feature encoding on the drug molecular structure, local chemical information of the drug molecule and the target structure.

[0061] The representation learning module includes a multi-channel general aggregation network module (McGEN module), a drug sequence representation learning module, and a target sequence representation learning module:

[0062] like Figure 1 The multi-channel general aggregation network module (McGEN module) mainly consists of two channels, each of which includes three GEN models with 78, 312, and 624 feature channels, respectively. It aggregates node feature information based on the atomic node association feature matrix of the drug molecular structure in the feature encoding module and learns the potential feature expression of each node.

[0063] The general aggregation function used in the GEN model is SoftMax_Agg μ (·), its expression is:

[0064]

[0065] μ is a learnable continuous variable with an initial value of 1.

[0066] like Figure 2 The drug sequence representation learning module (MLRCN module) includes a three-layer feature extraction network constructed from six residual connection blocks. The number of residual connection blocks in the three-layer feature extraction network is 3, 2, and 1, respectively. Each residual connection block contains two convolutional neural network (CNN) modules, each followed by a ReLU activation function. It outputs latent features of the local chemical information of the drug molecules based on the initial features of the local chemical information of the drug molecules.

[0067] like Figure 3 The target sequence representation learning module (BiLSTM module) consists of two long short-term memory network modules (Long Short-Term Memory, LSTM), which extract the amino acid temporal features in the sequence from the forward and backward directions of the target sequence respectively.

[0068] like Figure 4,The DTA prediction module includes a multi-layer perceptron, which ,integrates the outputs of the multi-channel general aggregation network module, ,drug sequence representation learning module and target sequence representation learning ,module and then feeds it into the MLP network for DTA prediction.

[0069] Example 2

[0070] like Figure 5 This embodiment provides a prediction method based on the drug-target binding affinity prediction model based on the graph dilation convolution strategy described in Example 1, and the steps include:

[0071] S1. Representing the drug molecule sequence and the target amino acid sequence as two-dimensional vector structures using a feature encoding module, and converting the drug into a two-dimensional structure to obtain initial features of the drug molecule, initial features of the drug's local chemical information, and initial features of the target;

[0072] S2. Extract the initial features of the drug molecule using a multi-channel general aggregation network module to obtain the drug's global structural features, extract the initial features of the drug's local chemical information using a drug sequence representation learning module to obtain the drug's local chemical features, and extract the initial features of the target using a target sequence representation learning module to obtain the target sequence features;

[0073] S3. The extracted global structural features, local chemical features and target sequence features are concatenated and passed through a fully connected layer to predict the drug-target affinity value.

[0074] Example 3

[0075] Drug molecules are typically represented as SMILES sequences, which use ASCII characters to describe the molecular structure, representing the drug's 3-D chemical structure as a string of characters. SMILES sequences consist of symbols representing atomic nodes in the molecular graph and symbols representing the chemical structure ('S' represents a sulfur atom, '=' represents a carbon-carbon double bond, and '#' represents a carbon-carbon triple bond). In the drug molecule graph, atoms form various functional groups through different valence bonds, such as aromatic hydrocarbons, hydroxyl groups, and carbonyl groups. This functional group information is called the local chemical information of the drug molecule.

[0076] Proteins are structurally distinct biological macromolecules composed of multiple amino acids linked by peptide bonds. Because representing protein molecules using molecular graphs is difficult and unreliable, proteins are typically represented as sequences of ASCII characters, with each character representing an amino acid (e.g., A for alanine and R for arginine).

[0077] This embodiment provides specific prediction steps based on the method described in Example 2, specifically:

[0078] S1. Use Represents a set of SMILES strings for drug molecules, where D k represents the SMILES sequence expression of the kth drug molecule, N d Represents the total number of drug molecules. k The lengths of D k The length of the feature matrix is ​​set to 1×200. k If the sequence length is less than 200, it will be padded with 0; if it exceeds 200, the excess part will be truncated.

[0079] S2.D k Perform label encoding to obtain the initial label h of the drug SMILES sequence k =[x1,x2,...,x 200 ] T , and then the one-hot encoding method is used to obtain the initial features of the local chemical information of the drug. The process can be formalized as follows:

[0080]

[0081] Among them, the one-hot encoding function One-Hot(·) is obtained by k Each label is represented as a 128-dimensional vector, and the initial feature expression of the local chemical information of the drug is obtained

[0082] S3. Using the open source tool RDKit, the drug structure coding module will convert D k Convert to the corresponding molecular graph G k =(V k ,E k ), where v i ∈V k (i=1,2,...,N atom ) represents the i-th atomic node of the molecule, e i,j ∈E k (j∈{1,2,...,N N}) represents the atomic node v i With v j In the drug molecule diagram, GDilatedDTA does not consider the strength of chemical bonds, so G k is an undirected graph. N atom and N N Represents the number of atoms and node v respectively i The number of first-order neighbor nodes. GDilatedDTA uses the DeepChem method to represent each node as a 78-dimensional binary feature vector. The feature vector contains five pieces of information: atomic symbol h sym , the number of adjacent atoms hnei 、Number of adjacent hydrogen atoms h H 、Atomic implicit valence h hid And whether the atom belongs to the aromatic hydrocarbon structure h isaro By v i The one-hot encoding method can be used to obtain the initial feature representation of each atomic node in the k-th drug molecule:

[0083]

[0084] Where || represents vector concatenation, N fea represents the feature vector dimension, which is 78. At this time, the drug molecule D k The initial feature H Dk Expressed as:

[0085]

[0086] For atomic node association features, GDilatedDTA uses the adjacency matrix Indicates that when the atomic node v i With v j When there is a bond connection, Otherwise 0.

[0087] S4. Use Target sequence set, where T l Indicates the amino acid sequence of the first target, N t Indicates the total number of target molecules. GDilatedDTA sets the target feature matrix dimension to 1×1000. When the target sequence length is less than 1000, it is padded with 0; if the length exceeds 1000, the excess part is truncated. The target structure encoding module uses alphabetical order to encode T l Perform label encoding to obtain the target label h l =[x1,x2,...,x 1000 ] T Then the initial structural characteristics of the target are obtained by formula (1):

[0088]

[0089] Represents the initial structural features of the lth target molecule. Each amino acid in the target is mapped to a 128-dimensional vector space through the one-hot encoding function One-Hot(·).

[0090] S5. Obtained drug molecule D k The initial characteristics Associate features with atomic nodes Input to the McGEN module, channel 1 input drug molecule Dk The initial characteristics and atomic node associated feature matrix Then, by stacking three layers of GEN, channel 1 aggregates and updates the feature information of the neighboring nodes of the central node to obtain the feature expression of the central node. The atomic node correlation feature matrix of channel 2 is Similarly, the feature information of the indirect neighbor nodes of the central node is aggregated and updated by stacking GEN to obtain the feature expression of the central node Finally, McGEN cascades the outputs of the two channels to obtain D k The potential feature expression

[0091] S6. Initial features H of the obtained drug local chemical information s Input the MLRCN module and output the potential feature expression of drug local chemical information

[0092] S7. The amino acid sequence features of the target are fed into the LSTM module from two directions. The LSTM model first calculates the target amino acid sequence features x at time t. t and the hidden layer state h at the previous moment t-1 Calculate the forget gate and memory gate of the cell and select the information to be forgotten f t and the information that needs to be remembered t , while recording the temporary cell state

[0093] f t =σ(W f ·[h t-1 ,x t ]+b f ),

[0094] i t =σ(W i ·[h t-1 ,x t ]+b i ),

[0095]

[0096] Among them, σ is the Sigmoid activation function, W is the learnable weight matrix, and b is the bias term. Then, LSTM will be based on the cell state C at the previous moment. t-1 Calculate the current cell state C t , its calculation expression is as follows:

[0097]

[0098] After comprehensively considering the cell information at the previous moment, LSTM calculates the hidden layer state h at the current moment. t Perform calculations.

[0099] O t =σ(W O ·[h t-1 ,x t ]+b O ),

[0100] h t =O t ×tanh(C t ).

[0101] Finally, the BiLSTM module outputs the hidden layer state information with the same length as the target encoding sequence

[0102] S8.GDilatedDTA obtains the feature extraction module and After integration, it is sent to the MLP network for DTA prediction. The integrated drug-target pair features are The drug-target pair features are used as the input of the MLP. After forward propagation, the binding affinity value is calculated as the output of the MLP. Subsequently, the MLP establishes a loss function by comparing the predicted binding affinity value with the actual binding affinity value, and uses backpropagation to update the hidden layer. The loss function used by the MLP is shown in Equation (2):

[0103]

[0104] Among them, Y i is the true binding affinity value, To predict the binding affinity value, n is the number of drug-target pairs.

[0105] Example 4

[0106] To further validate the model's interpretability, this example selected PI3Kdelta (PDB ID: 5I6U) and CHEMBL3806195 as examples for prediction. To highlight the interaction regions, regions with weights exceeding 1.07E-2 in the protein and all atomic nodes in the drug molecule were colored.

[0107] like Figure 6The blue in the protein structure and the orange in the drug molecule image represent the locations within the binding pocket where the target interacts with the drug molecule. Darker orange in the drug molecule image indicates a higher weight, while darker blue indicates a lower weight. The amino acid weights in PI3Kdelta are mostly concentrated between 6.54E-4 and 5.76E-3. The GDilatedDTA model primarily captures the amino acid region: residues 812-919. Residues 907-914 have relatively high weights, with a peak at ASP-911, which falls squarely within the binding pocket.

[0108] from Figure 6 As can be seen in the drug molecule diagram, the drug atom node bound to residue ASP-911 also has a large weight. The weight visualization analysis results show that the proposed model and method accurately predict the binding site of the drug-target interaction and improve the model's interpretability.

[0109] Obviously, the above embodiments of the present invention are merely examples for the purpose of clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A drug-target binding affinity prediction model based on graph dilation and convolution strategy, characterized by: The model includes a feature encoding module, a representation learning module and a DTA prediction module; The feature encoding module performs initial feature encoding on the drug molecular structure, local chemical information of the drug molecule and the target structure; The representation learning module includes a multi-channel general aggregation network module, a drug sequence representation learning module and a target sequence representation learning module: The multi-channel general aggregation network module includes two channels, each of which includes three GENs, which aggregate node feature information based on the atomic node association feature matrix of the drug molecules in the feature encoding module and learn the potential feature expression of each atom; The drug sequence representation learning module includes a three-layer feature extraction network constructed by six residual connection blocks. Each residual connection block has two convolutional neural network modules, and each convolutional neural network module is followed by a ReLU activation function, which outputs the latent feature expression of the local chemical information of the drug molecules based on the initial features of the local chemical information of the drug molecules. The target sequence representation learning module consists of two long short-term memory network modules, which extract structural features from the target sequence from the forward and backward directions respectively; The DTA prediction module includes a multi-layer perceptron, which integrates the outputs of the multi-channel general aggregation network module, the drug sequence representation learning module and the target sequence representation learning module and then feeds them into the MLP network for DTA prediction.

2. The drug-target binding affinity prediction model based on graph dilation and convolution strategy according to claim 1, characterized in that: The number of GEN feature channels is 78, 312, and 624 respectively.

3. The drug-target binding affinity prediction model based on graph dilation and convolution strategy according to claim 1, characterized in that: The number of residual connection blocks of the three-layer feature extraction network is 3, 2 and 1 respectively.

4. The prediction method of the drug-target binding affinity prediction model based on the graph dilation convolution strategy according to claim 1, characterized in that: The following steps are involved: S1. Representing the drug molecule sequence and the target amino acid sequence as two-dimensional vector structures using a feature encoding module, and converting the drug into a two-dimensional structure to obtain initial features of the drug molecular structure, initial features of the drug's local chemical information, and initial features of the target; S2. Extract the initial features of the drug's molecular structure, the initial features of the drug's local chemical information, and the initial features of the target through a multi-channel general aggregation network module, a drug sequence representation learning module, and a target sequence representation learning module, respectively, to obtain the drug's global structural features, the drug's local chemical features, and the target sequence features. S3. The extracted global structural features, local chemical features and target sequence features are concatenated and then used in the DTA prediction module to predict the drug-target affinity value.

5. The drug target binding affinity prediction method based on graph dilation and convolution strategy according to claim 4, characterized in that: Initial characterization of the drug molecule Expressed as: Where v i is the i-th atomic node of the drug molecule, is the initial feature matrix The size of N atom is the number of atomic nodes of the drug molecule, N fea For node v i The number of neighbor nodes; Atomic node association features of the drug molecule The initial feature H of the local chemical information of the drug s Expressed as: Initial structural characteristics of the target H l Expressed as:

6. The drug-target binding affinity prediction method based on graph dilation and convolution strategy according to claim 4, characterized in that: The steps of obtaining the initial characteristics of the drug molecule through a multi-channel universal aggregation network module include: Initial characterization of drug molecules Associate features with atomic nodes Input to channel 1 in the multi-channel general aggregation network module, channel 1 aggregates and updates the feature information of the neighbor nodes of the central node by stacking three layers of GEN to obtain the feature expression of the central node Channel 2 uses the graph expansion convolution strategy to construct the indirect neighbor node association features of atomic nodes Indicates Jn Multiply, Jn is the number of jumps, and the feature information of the indirect neighbor nodes of the central node is aggregated and updated by stacking GEN to obtain the feature expression of the central node Finally, the multi-channel general aggregation network module cascades the outputs of the two channels to obtain D k The potential feature expression 7. The method for drug-target binding affinity prediction based on graph dilation and convolution strategy according to claim 4, characterized in that: The steps of the initial features of the drug local chemical information through the drug sequence representation learning module are: Input the initial feature matrix H of the drug local chemical information encoded by the feature encoding module s After the drug sequence representation learning module, the potential feature expression of the local chemical information of the drug is output 8. The method for drug-target binding affinity prediction based on graph dilation and convolution strategy according to claim 4, characterized in that: The initial features of the target are represented by the target sequence learning module as follows: (1) According to the input feature x at time t t and the hidden layer state h at the previous moment t-1 Calculate the forget gate and memory gate of the cell and select the information to be forgotten f t and the information that needs to be remembered t , while recording the temporary cell state in: f t =σ(W f ·[h t-1 ,x t ]+b f ), i t =σ(W i ·[h t-1 ,x t ]+b i ), Where σ is the Sigmoid activation function, W is the learnable weight matrix, and b is the bias term; (2) The long short-term memory network is based on the cell state C at the previous moment. t-1 Calculate the current cell state C t , its calculation expression is as follows: (3) After comprehensively considering the cell information at the previous moment, the long short-term memory network calculates the hidden layer state h at the current moment. t Calculate and get The t =σ(W O ·[h t-1 ,x t ]+b O ), h t =O t ×tanh(C t ), (4) The target sequence representation learning module will output hidden layer state information with the same length as the target encoding sequence 9. The method for predicting drug-target binding affinity based on graph dilation and convolution strategy according to claim 4, characterized in that: The steps for predicting drug target affinity values ​​using a multi-layer perceptron are: (1) and After integration, it is sent to the MLP network for DTA prediction. The integrated drug-target pair features are (2) The drug-target pair features are used as the input of the MLP, and the binding affinity value is calculated after forward propagation as the output of the MLP. Subsequently, the MLP establishes a loss function by comparing the predicted binding affinity value with the true binding affinity value, and uses backpropagation to update the hidden layer.

10. The drug-target binding affinity prediction method based on graph dilation and convolution strategy according to claim 9, characterized in that: The loss function used by the MLP is: Among them, Y i is the true binding affinity value, To predict the binding affinity value, n is the number of drug-target pairs.

Images