Protein three-mode and multi-feature fusion-based drug target affinity prediction algorithm

Abstract

The invention provides a drug target affinity prediction algorithm based on protein three-mode and multi-feature fusion, and the algorithm makes full use of sequence information and two-dimensional and three-dimensional structural features of drugs and proteins, and realizes deep integration of data. The technical innovation point lies in that a multi-modal data fusion mechanism and an advanced graph neural network technology are introduced, and the sensitivity and recognition capability of the model to key biomarkers are enhanced through a dynamic attention mechanism. According to the technical advantage, the method has higher accuracy and stability when processing complex biological data, and compared with the prior art, the generalization ability and the interpretability are remarkably improved. In the aspect of technical maturity, the algorithm has passed multiple rounds of verification and shows good prediction performance and adaptability. The method is simple in operation and wide in market application prospect, can assist in new drug discovery and accelerate the drug design process, can also promote the development of the precision medical field, and particularly has important value in targeted therapy and disease prevention strategy formulation.

Description

Technical Field

[0001] The present invention relates to the fields of bioinformatics and drug design. Specifically, an algorithm for predicting drug target affinity based on protein trimodality and multi-feature fusion is designed. Background Art

[0002] In the process of drug design and development, accurately predicting the affinity between drug molecules and their target proteins is a crucial step. This prediction is not only related to the efficacy of the drug, but also directly affects the selectivity and toxicity of the drug. Traditional drug target affinity prediction methods are mainly based on empirical rules or basic molecular docking techniques. These methods often fail to fully consider the complex interactions between proteins and drug molecules, resulting in limited accuracy and reliability of the prediction results.

[0003] With the rapid development of bioinformatics, computational biology and artificial intelligence technologies, machine learning-based methods have begun to be used to improve the prediction of drug target affinity. These methods are able to process large amounts of biological data and extract deep feature information. However, existing prediction models still show certain limitations when dealing with the multidimensional properties of proteins and drug molecules, especially when dealing with multidimensional data of proteins (such as sequence information, two-dimensional and three-dimensional structures). This information is often not effectively used for in-depth analysis, limiting the accuracy and generalization ability of the prediction models.

[0004] In addition, the splicing and integration of multimodal data often lacks efficiency and accuracy. Therefore, there is an urgent need to develop a new algorithm that can effectively integrate multimodal data in proteins, thereby significantly improving the scientific nature of drug target affinity prediction.

[0005] This paper proposes a drug target affinity prediction algorithm based on protein trimodality and multi-feature fusion. This method fully extracts drug molecular graph features by using the graph agent attention module, and combines protein sequence information, protein two-dimensional structure and protein pocket three-dimensional structure to fully obtain protein potential features, thus overcoming the limitations of the existing technology. It provides an efficient and accurate solution for drug research and development. Summary of the invention

[0006] The purpose of the present invention is to provide a drug target affinity prediction algorithm based on protein trimodality and multi-feature fusion to solve the above-mentioned problem of improving the accuracy and generalization ability of the prediction model.

[0007] In order to achieve the above object, the present invention provides the following scheme:

[0008] The present invention provides a drug target affinity prediction algorithm based on protein trimodality and multi-feature fusion, comprising the following steps:

[0009] S1: RDKit is responsible for parsing the drug SMILES (Simplified Molecular Input Linear Expression) string, building an accurate molecular structure model, and inputting it into the graph residual agent attention module to extract its potential features D;

[0010] S2: Use AlphaFold3 to obtain the high-precision three-dimensional structure of the target protein.

[0011] S3: The three-dimensional structure data of proteins are converted into residue-level graphical representations using BioPython tools and then input into the deep graph isomorphism network to extract and analyze the two-dimensional structure information of proteins. G1 ;

[0012] S4: Use the DoGSiteScorer tool to identify the binding pocket position on the protein surface and input it into the deep graph convolutional network to extract the protein three-dimensional information P G2 ;

[0013] S5: The protein amino acid sequence is first input into an embedding layer to obtain the one-dimensional sequence information of the protein. Subsequently, this one-dimensional sequence information is further input into a convolutional interactive attention module to effectively extract and enhance the key features of the protein sequence;

[0014] S6: Build a multi-feature fusion engine to integrate the sequence information, two-dimensional structure information, three-dimensional structure information and drug molecule structure information of proteins. After inputting these multi-dimensional data into the fusion engine, the system can process and output the corresponding binding affinity value.

[0015] Furthermore, in S1, through the design of the composite network structure, not only can the basic information of the drug molecular structure be captured, but also the deep biological activity characteristics of the molecule can be learned through the attention mechanism and multi-layer neural network. The specific steps are as follows:

[0016] The SMILES format of the drug is converted into the corresponding molecular graph representation G(X,E) through the chemical information tool library RDKit, where X is the atomic attribute matrix, reflecting the chemical properties of each atom in the molecular graph; E is the edge attribute matrix, representing the type of chemical bonds between atoms and their characteristics;

[0017] In order to fully extract the structural information of the drug molecular graph, the converted molecular graph G is input into the convolutional neural network (GCN) to learn the deep feature representation of the molecular graph: 3 =graph 3 (G) and M 4 =M 3 +graph(M 3 );M iRepresents the features of the i-th layer, graph 3 (·) represents a three-layer graph neural network, and graph(·) represents a single-layer graph neural network. i , the node feature update can be expressed as: in represents the adjacency matrix of the drug molecule graph, express The degree matrix, W (i) represents the weight parameter of the i-th layer, and σ represents the activation function.

[0018] The output feature M after processing by the graph convolutional neural network 4 , input to the agent attention network, which strengthens the ability to capture key information by learning more critical atomic nodes in the drug molecule graph. Calculate the attention score A and update the node representation through the attention mechanism: A = Wpooling (Q) + b and D proxy =Attn(Q,A,Attn(A,K,V)); where Q, K, V are determined by M 4 The mappings represent query vector, key vector and value vector respectively, W represents the weight of the aggregation layer, and b represents the bias term;

[0019] By aggregating features and the output of the proxy attention network, this information is fed into a feed-forward neural network (FNN) to further learn high-level representations of nonlinear relationships and features: D = pooling (W2 (W1 (α (G + D proxy ))+b1)+b2); where W1, W2 represent weight matrices, b1, b2 represent bias vectors, and α represents an activation function, which is used to introduce nonlinearity and enhance the expressiveness of the model.

[0020] Furthermore, in S2, the protein three-dimensional structure P = {A1, A2, ..., A n}, where A i =(x i ,y i ,z i ) represents the three-dimensional coordinates of the ith atom in the protein, and n represents the total number of atoms in the protein.

[0021] Furthermore, in S3, using the bioinformatics tool BioPython, protein three-dimensional structure data is accurately converted into a residue-level graphical representation P G1 =(V,E), where v iIndicates that each node corresponds to a residue, and E represents the interaction relationship between residues. Finally, it is input into the deep graph isomorphism network to extract the two-dimensional structural features. The node information update rule in the graph isomorphism neural network is: in represents the feature of node v after the i-th iteration, N (V) Represents the set of neighbor nodes of node v. ∈ (i) represents a trainable parameter, MLP i Represents the multilayer perceptron of the i-th layer.

[0022] Furthermore, in S4, DoGSiteScorer is used to scan protein structures to identify and output potential binding pockets, and the protein pocket features P are output. G2 =(V,E). V represents the amino acid residues of the protein, and E represents the interaction between the residues. The deep convolutional network and pooling are then used to extract the three-dimensional structural features of the protein to enhance the model's understanding of the complex three-dimensional structure of the protein. The specific implementation is as follows:

[0023] The graph convolutional network is used to extract features from the pocket graph, which includes multi-layer GNN processing, and then the features are further optimized by using a hybrid pooling strategy. The deep convolutional network and pooling formula are expressed as: G2-GNN =WGNN(σ(GNN(GNN(P G2 ))))+b and P G2′ =σ(W1(meanpool(P G2-GNN )+maxpool(P G2-GNN ))+b2). ​​Where σ represents the activation function, W and W1 represent trainable weights, b and b2 represent bias terms, meanpool() and maxpool() represent average pooling and maximum pooling, respectively. Deep convolutional network methods can improve the accuracy and efficiency of protein three-dimensional structure analysis. The comprehensive pooling strategy aims to enhance the generalization ability of the network and optimize feature expression, thereby improving the accuracy and efficiency of the model in subsequent applications.

[0024] Furthermore, in S5, the specific method is:

[0025] The protein amino acid sequence is taken as input and passed to the embedding layer. The role of the embedding layer is to transform the discrete amino acid sequence P∈R b×n Mapped into a high-dimensional continuous vector P∈R b×n×d b represents the number of samples in the batch, n represents the protein length and d represents the dimension of each amino acid in the protein sequence;

[0026] The target feature P is split into two independent sequence features P1 and P2, and a convolutional neural network (CNN) is used for separate training to extract local features of each sequence:

[0027] Use pooling technology to process The maximum pooling and average pooling methods are combined to retain the key information in the sequence. By integrating the advantages of the two pooling methods, the information retention effect is optimized and the model's ability to capture key biological features is improved:

[0028] Using interactive attention networks to locate and emphasize important binding sites in protein sequences: Where d is the normalization factor of the dimension. By scaling the dot product attention mechanism, the model can effectively distinguish the importance of different sequence features and assign weights;

[0029] The two attention vectors CA1 and CA2 are concatenated and the nonlinear relationship in the protein sequence is learned through full connection. This step enables the model to not only learn the influence of a single feature, but also capture and model the complex patterns generated by the interaction of multiple features: P = W([CA1; CA2]) + b;

[0030] Furthermore, in S6, in order to integrate global and local information and more effectively simulate the complex interactions between drugs and targets, a multimodal feature fusion is used to fuse protein sequence information P and protein two-dimensional structure information P G1′ , protein pocket map information P G2′ The drug molecule graph is input into the multimodal fusion engine to predict the affinity value. The specific method is:

[0031] First, the protein two-dimensional structure information P G1′ , protein pocket map information P G2′ :P G =P G1′ +P G2′ , α=Tanh(WP G +b), T G =(α)P G1′ +(1-α)P G2′ ; Where Tanh() represents the activation function.

[0032] Using a weighted hybrid strategy, protein structure representation T G and protein sequence representation P in a weighted manner: w = σ(Tanh(W[T G ; P] + b)); Next, the calculated weight w is used to adjust the synthetic representation W P =Relu(w·[TG ; P]); where Relu() represents the activation function. The fused representation is further processed using a fully connected layer: T = (σ(W1W P +b1);

[0033] After obtaining the refined protein representation T, it is spliced ​​with the drug representation D and the interaction between the drug and the target is captured by the deep predictor: X con =[T;D];y=Relu((W2(Relu(W1X con +b1))+b2)).

[0034] Compared with the prior art, the present invention has achieved the following beneficial technical effects:

[0035] 1. By comprehensively utilizing protein sequence information, two-dimensional and three-dimensional structural information, and the structural characteristics of drug molecules, this algorithm can more accurately simulate the complex interactions between drugs and target proteins. This deep feature fusion significantly improves the accuracy of affinity prediction, which is better than traditional methods that rely on a single data source or simplified models;

[0036] 2. By introducing a multimodal data fusion mechanism, the present invention not only processes data from different biological information channels, but also optimizes feature selection and data processing through intelligent algorithms. This enables the model to have better generalization capabilities and adapt to changing biological data and complex practical application scenarios;

[0037] 3. This algorithm is designed with a specific structure so that the features extracted from the model not only contribute to the prediction results, but also have biological significance. This enhances the interpretability of the model and helps researchers understand the mechanism of action between drugs and proteins, thereby guiding experimental design and drug improvement;

[0038] 4. By providing more accurate and reliable predictions of drug target affinity, the present invention provides a powerful tool for the formulation of precision medicine and personalized treatment strategies, helping to achieve more precise disease treatment and management. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is an overall flow chart of a drug target affinity prediction method based on a protein grid of the present invention;

[0040] Figure 2 It is an architecture diagram of the graph residual proxy attention module in the present invention;

[0041] Figure 3 It is the architecture diagram of the convolutional interactive attention module in the present invention. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] The core of the present invention is to provide a drug target affinity prediction algorithm based on protein trimodality and multi-feature fusion to solve the above-mentioned problem of improving the accuracy and generalization ability of the prediction model.

[0044] In order to enable those skilled in the art to better understand the solution of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0045] Figure 1 A schematic diagram of the overall process of a protein grid-based drug target affinity prediction method, such as Figure 1 As shown, the following steps are included:

[0046] S1: Build an accurate molecular structure model based on the drug SMILES (Simplified Molecular Input Linear Expression) string. As an optional solution, RDKit can be responsible for parsing the drug SMILES string and input it into the graph residual agent attention module to extract its potential features D.

[0047] The drug-based SMILES format is converted into the corresponding molecular graph representation G(X,E), where X represents the atomic attribute matrix, reflecting the chemical properties of each atom in the molecular graph; E is the edge attribute matrix, representing the type of chemical bonds between atoms and their characteristics;

[0048] Specifically, Figure 2 As shown in Figure 1, in order to fully extract the structural information of the drug molecular graph, the converted molecular graph G is input into the convolutional neural network (GCN) to learn the deep feature representation of the molecular graph: 3 =graph 3 (G) and M 4 =M 3 +graph(M 3 );M i Represents the features of the i-th layer, graph 3 (·) represents a three-layer graph neural network, and graph(·) represents a single-layer graph neural network. i , the node feature update can be expressed as: in represents the adjacency matrix of the drug molecule graph, express The degree matrix, W(i) represents the weight parameter of the i-th layer, and σ represents the activation function.

[0049] The output feature M after processing by the graph convolutional neural network 4 , input to the agent attention network, which strengthens the ability to capture key information by learning more critical atomic nodes in the drug molecule graph. Calculate the attention score A and update the node representation through the attention mechanism: A = Wpooling (Q) + b and D proxy =Attn(Q,A,Attn(A,K,V)); where Q, K, V are determined by M 4 The mappings represent query vector, key vector and value vector respectively, W represents the weight of the aggregation layer, and b represents the bias term;

[0050] By aggregating features and the output of the proxy attention network, this information is fed into a feed-forward neural network (FNN) to further learn high-level representations of nonlinear relationships and features: D = pooling (W2 (W1 (α (G + D proxy ))+b1)+b2); where W1, W2 represent weight matrices, b1, b2 represent bias vectors, and α represents an activation function, which is used to introduce nonlinearity and enhance the expressiveness of the model;

[0051] S2: Use AlphaFold3 to obtain the high-precision three-dimensional structure of the target protein.

[0052] Obtaining protein three-dimensional structure based on protein amino acid sequence P = {A1, A2, ..., A n}, where A i =(x i ,y i ,z i ) represents the three-dimensional coordinates of the ith atom in the protein, and n represents the total number of atoms in the protein.

[0053] S3: The three-dimensional structure data of proteins are converted into residue-level graphical representations using BioPython tools and then input into the deep graph isomorphism network to extract and analyze the two-dimensional structure information of proteins. G1 .

[0054] Based on the protein three-dimensional structure data, it is accurately converted into a residue-level graphical representation P G1 =(V,E), where v i Indicates that each node corresponds to a residue, and E represents the interaction relationship between residues. Finally, it is input into the deep graph isomorphism network to extract the two-dimensional structural features. The node information update rule in the graph isomorphism neural network is: in represents the feature of node v after the i-th iteration, N (v) Represents the set of neighbor nodes of node v. ∈ (i) represents a trainable parameter, MLP i Represents the multilayer perceptron of the i-th layer.

[0055] S4: Use the DoGSiteScorer tool to identify the binding pocket position on the protein surface and input it into the deep graph convolutional network to extract the protein three-dimensional information P G2 .

[0056] Based on protein structure, potential binding pockets are identified and output, and protein pocket features P are output. G2 =(V,E). V represents the amino acid residues of the protein, and E represents the interaction between the residues. The deep convolutional network and pooling are then used to extract the three-dimensional structural features of the protein to enhance the model's understanding of the complex three-dimensional structure of the protein. The specific implementation is as follows:

[0057] The deep convolutional network and pooling formula are expressed as: G2-GNN =WGNN(σ(GNN(GNN(P G2 ))))+b and P G2′ =σ(W1(meanpool(P G2-GNN )+maxpool(P G2-GNN ))+b2). ​​Where σ represents the activation function, W and W1 represent trainable weights, b and b2 represent bias terms, meanpool() and maxpool() represent average pooling and maximum pooling, respectively. Deep convolutional network methods can improve the accuracy and efficiency of protein three-dimensional structure analysis. The comprehensive pooling strategy aims to enhance the generalization ability of the network and optimize feature expression, thereby improving the accuracy and efficiency of the model in subsequent applications.

[0058] S5: The protein amino acid sequence is first input into an embedding layer to obtain the one-dimensional sequence information of the protein. Subsequently, this one-dimensional sequence information is further input into a convolutional interactive attention module to effectively extract and enhance the key features of the protein sequence;

[0059] The protein amino acid sequence is taken as input and passed to the embedding layer. The role of the embedding layer is to transform the discrete amino acid sequence P∈R b×n Mapped into a high-dimensional continuous vector P∈R b×n×d b represents the number of samples, n represents the length of the protein and d represents the dimension of each amino acid in the protein sequence;

[0060] Specifically, Figure 2As shown in Figure 1, the target feature P is split into two independent sequence features P1 and P2, and a convolutional neural network (CNN) is used for separate training to extract local features of each sequence:

[0061] Use pooling technology to process The maximum pooling and average pooling methods are combined to retain the key information in the sequence. By integrating the advantages of the two pooling methods, the information retention effect is optimized and the model's ability to capture key biological features is improved:

[0062] Using interactive attention networks to locate and emphasize important binding sites in protein sequences: Where d is the normalization factor of the dimension. By scaling the dot product attention mechanism, the model can effectively distinguish the importance of different sequence features and assign weights.

[0063] The two attention vectors CA1 and CA2 are concatenated and the nonlinear relationship in the protein sequence is learned through full connection. This step enables the model to not only learn the influence of a single feature, but also capture and model the complex patterns generated by the interaction of multiple features: P = W([CA1; CA2]) + b.

[0064] S6: Build a multi-feature fusion engine to integrate the sequence information, two-dimensional structure information, three-dimensional structure information and drug molecule structure information of proteins. After inputting these multi-dimensional data into the fusion engine, the system can process and output the corresponding binding affinity value.

[0065] In order to integrate global and local information and more effectively simulate the complex interactions between drugs and targets, a multimodal feature fusion is used to integrate protein sequence information P and protein two-dimensional structure information P G1′ , protein pocket map information P G2′ The drug molecule graph is input into the multimodal fusion engine to predict the affinity value. The specific method is:

[0066] First, the protein two-dimensional structure information P G1′ , protein pocket map information P G2′ :P G =P G1′ +P G2′ , α=Tanh(WP G +b), T G =(α)P G1′ +(1-α)P G2′ ; Where Tanh() represents the activation function.

[0067] Using a weighted hybrid strategy, protein structure representation TG and protein sequence representation P in a weighted manner: w = σ(Tanh(W[T G ; P] + b)); Next, the calculated weight w is used to adjust the synthetic representation W P =Relu(w·[T G ; P]); where Relu() represents the activation function. The fused representation is further processed using a fully connected layer: T = (σ(W1W P +b1);

[0068] After obtaining the refined protein representation T, it is spliced ​​with the drug representation D and the interaction between the drug and the target is captured by the deep predictor: X con =[T;D];y=Relu((W2(Relu(W1X con +b1))+b2)).

[0069] The above is a detailed introduction to a drug target affinity algorithm based on protein trimodality and multi-feature fusion provided by the present invention. The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the various embodiments can be referred to each other.

[0070] The contents not described in detail in the specification of the present invention belong to the prior art known to the professional and technical personnel in this field.

Claims

1. A drug target affinity algorithm based on protein trimodality and multi-feature fusion, comprising the following steps: S1: RDKit is responsible for parsing the drug SMILES (Simplified Molecular Input Linear Expression) string, building an accurate molecular structure model, and inputting it into the graph residual agent attention module to extract its potential features D; S2: Use AlphaFold3 to obtain the high-precision three-dimensional structure of the target protein; S3: The three-dimensional structure data of proteins are converted into residue-level graphical representations using BioPython tools and then input into the deep graph isomorphism network to extract and analyze the two-dimensional structure information of proteins. G1 ; S4: Use the DoGSiteScorer tool to identify the binding pocket position on the protein surface and input it into the deep graph convolutional network to extract the protein three-dimensional information P G2 ; S5: The protein amino acid sequence is first input into an embedding layer, and the one-dimensional sequence information of the protein is obtained through this process; Subsequently, these one-dimensional sequence information are further input into a convolutional interactive attention module to effectively extract and enhance the key features of protein sequences; S6: Construct a multi-feature fuser to integrate protein sequence information, two-dimensional structure information, three-dimensional structure information and drug molecule structure information; after inputting these multi-dimensional data into the fuser, the system can process and output the corresponding binding affinity value.

2. The drug target affinity algorithm based on protein trimodality and multi-feature fusion according to claim 1, characterized in that: In S1, through the design of the composite network structure, not only can the basic information of the drug molecular structure be captured, but also the deep biological activity characteristics of the molecule can be learned through the attention mechanism and multi-layer neural network. The specific steps are as follows: The SMILES format of the drug is converted into the corresponding molecular graph representation G(X,E) through the chemical information tool library RDKit, where X is the atomic attribute matrix, reflecting the chemical properties of each atom in the molecular graph; E is the edge attribute matrix, representing the type of chemical bonds between atoms and their characteristics; In order to fully extract the structural information of the drug molecular graph, the converted molecular graph G is input into the convolutional neural network (GCN) to learn the deep feature representation of the molecular graph: M 3 =graph 3 (G) M 4 =M 3 +graph(M 3 ) In the above formula, M i Represents the features of the i-th layer, graph 3 (·) represents a three-layer graph neural network, and graph(·) represents a one-layer graph neural network; In each layer M i , the node feature update can be expressed as: In the above formula, represents the adjacency matrix of the drug molecule graph, express The degree matrix, W (i) represents the weight parameter of the i-th layer, σ represents the activation function; The output feature M after processing by the graph convolutional neural network 4 , is input into the agent attention network, which strengthens the ability to capture key information by learning more critical atomic nodes in the drug molecule graph; calculates the attention score A, and updates the node representation through the attention mechanism: A=Wpooling(Q)+b D proxy =Attn(Q,A,Attn(A,K,v)) In the above formula, Q, K, and V are determined by M 4 The mappings represent query vector, key vector and value vector respectively, W represents the weight of the aggregation layer, and b represents the bias term; By aggregating features and the output of the proxy attention network, this information is fed into a feed-forward neural network (FNN) to further learn high-level representations of nonlinear relationships and features: D=pooling(W2(W1(σ(G+D proxy ))+b1)+b2) W1, W2 represent weight matrices, b1, b2 represent bias vectors, and σ represents an activation function, which is used to introduce nonlinearity and enhance the expressiveness of the model.

3. The drug target affinity algorithm based on protein trimodality and multi-feature fusion according to claim 1, characterized in that: In S2, the protein three-dimensional structure P = {A1, A2, ..., A n }, where A i =(x i ,y i ,z i ) represents the three-dimensional coordinates of the ith atom in the protein, and n represents the total number of atoms in the protein.

4. The drug target affinity algorithm based on protein trimodality and multi-feature fusion according to claim 1, characterized in that: In S3, protein 3D structure data are accurately converted into residue-level graphical representations using the bioinformatics tool BioPython. G1 =(V,E), where v i Indicates that each node corresponds to a residue, and E represents the interaction relationship between residues; finally, it is input into the deep graph isomorphism network to extract the two-dimensional structural features; the node information update rule in the graph isomorphism neural network is: in represents the feature of node v after the i-th iteration, N (V) Represents the set of neighbor nodes of node v; ∈ (i) represents a trainable parameter, MLP i Represents the multilayer perceptron of the i-th layer.

5. The drug target affinity algorithm based on protein trimodality and multi-feature fusion according to claim 1, characterized in that: In S4, the specific steps are as follows: DoGSiteScorer is used to scan protein structures to identify and output potential binding pockets and output protein pocket features P G2 =(V,E); Where V represents the amino acid residues of the protein, and E represents the interaction between the residues. Then, a deep convolutional network and pooling are used to extract the three-dimensional structural features of the protein to enhance the model's understanding of the complex three-dimensional structure of the protein. The specific implementation is as follows: The deep convolutional network and pooling formulas are expressed as: P G2-GNN =WGNN(σ(GNN(GNN(P G2 ))))+b P G2′ =σ(W1(meanpool(P G2-GNN )+maxpool(P G2-GNN ))+b2) Where σ represents the activation function, W and W1 represent trainable weights, b and b2 represent bias terms, meanpool() and maxpool() represent average pooling and maximum pooling respectively; The deep convolutional network method can improve the accuracy and efficiency of protein three-dimensional structure analysis; the comprehensive pooling strategy aims to enhance the generalization ability of the network and optimize feature expression, thereby improving the accuracy and efficiency of the model in subsequent applications.

6. The drug target affinity algorithm based on protein trimodality and multi-feature fusion according to claim 1, characterized in that: In S5, the specific method is: The protein amino acid sequence is taken as input and passed to the embedding layer; the role of the embedding layer is to transform the discrete amino acid sequence P∈R b×n Mapped into a high-dimensional continuous vector P∈R b×n×d ; b represents the number of samples, n represents the protein length and d represents the dimension of each amino acid in the protein sequence; The target feature P is split into two independent sequence features P1 and P2, and a convolutional neural network (CNN) is trained separately to extract local features of each sequence: Use pooling technology to process The maximum pooling and average pooling methods are combined to retain the key information in the sequence. By integrating the advantages of the two pooling methods, the information retention effect is optimized and the model's ability to capture key biological features is improved: Using interactive attention networks to locate and emphasize important binding sites in protein sequences: d is the normalization factor of the dimension; by scaling the dot product attention mechanism, the model can effectively distinguish the importance of different sequence features and assign weights; The two attention vectors CA1 and CA2 are concatenated, and the nonlinear relationship in the protein sequence is learned through full connection. This step enables the model to not only learn the influence of a single feature, but also capture and model the complex patterns generated by the interaction of multiple features: P=W([CA1;CA2])+b.

7. The drug target affinity algorithm based on protein trimodality and multi-feature fusion according to claim 1, characterized in that: In S6, in order to integrate global and local information and more effectively simulate the complex interactions between drugs and targets, a multimodal feature fusion is used to fuse protein sequence information P and protein two-dimensional structure information P G1′ , protein pocket map information P G2′ The drug molecule graph is input into the multimodal fusion device to predict the affinity value; the specific method is: First, the protein two-dimensional structure information P G1 , protein pocket map information P G2 : P G =P G1′ +P G2′ α=Tanh(WP G +b) T G =(a)P G1′ +(1-α)P G2′ Where Tanh() represents the activation function; Then a weighted hybrid strategy is used to characterize protein structure G A weighted combination is performed with the protein sequence representation P; w=α(Tanh(W[T G ; P]+b)) Next, the calculated weight w is used to adjust the synthetic representation: IN P =Relu(w [T G ;P]) Relu() represents the activation function, and the fully connected layer is used to further process the fused representation: T=(α(W1W P +b1)) After obtaining the refined protein representation T, it is spliced ​​with the drug representation D and the interaction between the drug and the target is captured by the deep predictor: X con =[T;D] <h2 style=";text-align:left;direction:ltr">y=Relu((W2Relu(W1X<h2 style=";text-align:left;direction:ltr"> con <h2 style=";text-align:left;direction:ltr"> +b1))+b2))。

Images