Drug target affinity prediction method based on graph neural network

Abstract

The invention discloses a drug target affinity prediction method based on a graph neural network, and relates to a drug target affinity prediction method. The invention aims to solve the problem of low prediction precision of drug target affinity in the existing method. The method comprises the following steps: learning drug global features from a drug atomic graph; extracting amino acid positive example features from the positive example protein hydrophobic diagram, the positive example protein hydrogen bond diagram and the positive example protein contact diagram, and performing fusion pooling to obtain amino acid positive example fusion pooling features as protein global features; extracting sub-graphs from the drug atomic graph to obtain drug sub-graph features as drug local information; subgraphs are extracted from the positive example protein hydrophobic graph, the positive example protein hydrogen bond graph and the positive example protein contact graph, and protein subgraph features are obtained and serve as protein local information. The drug local features and the drug global features are spliced into drug features, the protein local features and the protein global features are spliced into protein features, and the drug local features and the protein global features are spliced into drug-protein combined features used for drug target affinity prediction.

Description

Technical Field

[0001] The invention belongs to the field of bioinformatics, and in particular relates to a method for predicting drug target affinity. Background Art

[0002] Drugs can change the way proteins work to achieve therapeutic effects by interacting with specific regions of proteins such as enzyme active sites or receptor binding sites. Proteins that interact with drugs are called target proteins, and local regions of target proteins that interact with drugs are called targets. The strength of drug-protein interaction is called drug target affinity (DTA). The stronger the affinity, the greater the possibility of interaction. The indicators that quantify affinity include dissociation constant (Kd), inhibition constant (Ki), and half-inhibitory concentration (IC50). The larger the values, the smaller the affinity, which means the smaller the possibility of interaction, which means that such drugs are less likely to be used as a treatment for diseases caused by the target protein.

[0003] In the past few decades, high-throughput screening (HTS) experiments have greatly accelerated the calculation of drug target affinity strength. However, HTS experiments are costly and laborious, and cannot meet the needs of revealing the affinity strength of millions of existing drug compounds and thousands of target proteins. Therefore, it is very necessary to develop efficient computational methods to make full use of the biological data of known drug molecule-target protein affinity to clarify the mechanism of action of drugs in the human body.

[0004] In recent years, researchers have proposed many prediction methods for drug target affinity in the field of machine learning and deep learning. In machine learning, Kron-RLS uses kernel methods to extract drug features and protein features from drug-drug similarity features and protein-protein similarity features and uses random forests to obtain drug target affinity strength; SimBoost has the same representation of proteins and drug compounds as Kron-RLS. It constructs features for drugs, targets, and drug target pairs, and extracts feature vectors of drug target pairs through feature engineering to train gradient boosting machines to predict binding affinity. In the field of deep learning, drug target affinity is mainly divided into two categories, namely sequence-based and graph-based methods. For sequence-based methods, they start more from the primary structure of proteins. Since the amino acid sequence, as the primary structure of proteins, is the basic information of the higher-level structure of proteins and their external functions. For example, DeepDTA uses a convolutional neural network (CNN) to extract features from the simplified molecular input line entry system (SMILES) sequence of compounds and the amino acid sequence of proteins to infer the affinity between drugs and proteins; while graph-based methods focus on representing the spatial structure of drugs and proteins. For example, DGraphDTA constructs drug atom graphs and protein amino acid contact graphs based on the RDKit software package and the Psconsc4 algorithm, and uses a graph neural network (GNN) based on the representation theory that adjacent nodes should have similarities to extract features of drugs and proteins, thereby predicting the affinity between drugs and target proteins.

[0005] Although previous methods have made some progress in DTA prediction methods, there is still room for improvement. Most of the DTA methods proposed previously have failed to fully utilize the interaction force information that supports the spatial structure of drugs and proteins. At the same time, the use of cross-attention based on dot product calculations also ignores the fact that atoms with interaction forces are not all similar, and there are also practical situations such as feature complementarity between anions and cations. As a result, the extracted drug and protein features cannot fully represent the drugs and proteins. Existing DTA prediction methods fail to fully utilize known drug and protein data, and also cause the model to omit structural information extraction and feature representation for drugs and proteins, resulting in low accuracy in drug target affinity prediction. Summary of the invention

[0006] The purpose of the present invention is to address the problem that the existing DTA prediction methods fail to fully utilize known drug and protein data, resulting in low prediction accuracy of drug target affinity.

[0007] A drug target affinity prediction method based on a graph neural network, which realizes the prediction of drug target affinity based on a drug target affinity prediction model, wherein the drug target affinity prediction model includes a data acquisition unit, a drug feature extraction unit, a protein feature extraction unit, a subgraph feature processing unit, and an affinity prediction unit;

[0008] The processing of the data acquisition unit includes:

[0009] Step a1: For each drug, obtain the drug atomic features and chemical bond features, take the atoms contained in the drug as nodes, the atomic features as node features, and the chemical bond features as edges, and then construct the corresponding drug atomic graph G d ;

[0010] Step a2: Obtain the hydrophobic interaction forces and hydrogen bonds of the protein amino acids, and construct the protein hydrophobic graph G with the amino acids contained in the protein as nodes and the hydrophobic interaction forces between the amino acids as edges. h The protein hydrogen bond graph G is constructed with the amino acids contained in the protein as nodes and the hydrogen bonds between the amino acids as edges. hy ; For protein amino acids, the Euclidean space distance data of each amino acid in the protein is obtained, and the amino acids contained in the protein are taken as nodes, and the Euclidean space distance between amino acids less than the distance threshold is taken as edges to construct the protein contact graph G cont ;

[0011] The processing of the drug feature extraction unit includes:

[0012] Step b1: Use the edge aggregation graph attention network algorithm EGAT and the co-embedding graph neural network algorithm CensNet to analyze the drug atomic graph G corresponding to each drug. d Extract atomic features NF d and chemical bond characteristics EF d , using the average pooling algorithm to obtain the atomic pooling feature NP d and chemical bond pooling feature EP d , and then splice to get the drug global feature GP d =NP d ||EP d ;

[0013] Step b2: Processing drug global features GP based on feature fusion algorithm d and atomic features NF d The final drug global feature FuNP is obtained by fusing the node-level atomic features and the graph-level drug global features and pooling them. d ;

[0014] The processing of the protein feature extraction unit includes:

[0015] Step c1: Use three GAT networks to analyze the hydrophobic graph G of the positive protein h 、Positive example protein hydrogen bond diagram G hy 、Positive example protein contact graph G cont Extract the positive amino acid features NF corresponding to the three graphs of the protein h；i NF hy；j NF cont；k ;

[0016] Step c4: Use the feature fusion algorithm to fuse the three amino acid positive feature in c1 to obtain the amino acid fusion positive feature NF fusion；i , NF fusion；i Represents the fusion feature of the i-th amino acid feature in the three positive example graphs of the protein;

[0017] Step c5: Use the average pooling algorithm to process the amino acid fusion positive example features to obtain the amino acid positive example fusion pooling features Among them, PF fusion represents the amino acid positive example fusion pooling feature;

[0018] The processing of the sub-graph feature processing unit includes:

[0019] Step d1: Randomly extract nodes from the drug atom graph with a fixed probability to form a node list, and use the k-hop algorithm to extract nodes from the drug atom graph G d The drug subgraph with the node in the node list as the central node is extracted from the node list, and the node feature set of each drug subgraph is averaged and pooled to obtain the corresponding drug subgraph pooling feature. Then, all drug subgraph pooling features are concatenated into the drug subgraph pooling total feature SubP d ;

[0020] Step d2: randomly extract nodes from the protein hydrophobicity map, the positive protein hydrogen bond map, and the positive protein contact map with a fixed probability, and each of them forms a corresponding node list; use the k-hop algorithm to extract protein subgraphs from the positive protein hydrophobicity map, the positive protein hydrogen bond map, and the positive protein contact map, and use the average pooling algorithm for the node feature set of each protein subgraph to obtain the corresponding protein subgraph pooling feature; then all protein subgraph pooling features are spliced ​​into the protein subgraph pooling total feature SubP p ;

[0021] Step d3: Use additive attention and cross attention algorithms to calculate the interactive attention coefficient Atten for each pair of drug subgraphs and protein subgraphs based on the drug subgraph pooling features and protein subgraph pooling total features. za ;

[0022] Step d4: Use matrix multiplication to calculate the total pooling features of the drug subgraph in step d1 and the interactive attention coefficient in step d3 to obtain the final drug subgraph cross-fusion feature SubFu that fused the protein information. d；z ; Use matrix multiplication to calculate the protein subgraph pooling total features of step d2 and the interactive attention coefficient of step d3 to calculate the protein subgraph cross-fusion feature SubFu of the final fusion compound information P；a ;

[0023] Step d5: Use the average pooling algorithm to process the drug subgraph cross-fusion features SubFu in step d4 d；z , protein subgraph cross fusion feature SubFu P；a Get the drug subgraph cross fusion pooling feature SubPFu d , protein subgraph cross fusion pooling feature SubPFu p ;

[0024] The processing of the affinity prediction unit includes:

[0025] Step e1: Use the splicing method to process the drug global features of step b2 and the drug subgraph cross-fusion pooling features of step d5 to obtain the drug feature F drug =FuNP d ||SubPFu d ;

[0026] Step e2: Use the splicing method to process the amino acid positive example fusion pooling features of c5 and the protein subgraph cross fusion pooling features of d5 to obtain the protein feature F protein =PF fusion ||SubPFu p ;

[0027] Step e3: Use the splicing method to splice the drug features and protein features to obtain the drug-protein joint feature F conbind =F drug ||F protein ;

[0028] Step e4: Target F conbind , machine learning technology is used to predict the affinity prediction value affinity.

[0029] Furthermore, in step b1, the edge aggregation graph attention network algorithm EGAT and the co-embedded graph neural network algorithm CensNet are used to map the drug atomic graph G corresponding to each drug. d Extract atomic features NF d and chemical bond characteristics EF d The specific process includes:

[0030] Step b11: Use EGAT to directly pass the neighbor atom feature N d；j and its associated chemical bond characteristics Get the atomic features U in the updated drug atom graph d；i ;

[0031] e ji =LeakyReLU((W α [W ν N d；i ||W ν N d；j ||W ρ ξ d；ij ]))

[0032]

[0033] Among them, W ν , W ρ , W ε is the weight parameter, || is the concatenation operation, LeakyReLU(·) and σ(·) are nonlinear activation functions, N d；i Represents the drug atom graph G d The i-th initial atomic feature, U d；i Represents the drug atom graph G d The characteristics of the i-th atom updated by the EGAT algorithm, ξ d；ij Represents the drug atomic graph G d The initial characteristics of the chemical bond connecting the i-th atom to the j-th atom;

[0034] Step b12: Update the atomic feature NF using the chemical bond feature through CensNet d , and use the atomic features to update the chemical bond features EF d :

[0035] NF d =[PΛ(W e ξ d ) T +I]W d U d

[0036] EF d =[PΛ(W d U d ) T +I]W e ξ d

[0037] Among them, U d Represents the drug atomic graph G d All atomic features U after EGAT update d；iThe feature set concatenated by row, ξ d Represents the drug atomic graph G d All chemical bond characteristics of ξ d；ij The feature set obtained by concatenating rows; W e , W d is the weight parameter; P represents the correlation matrix between chemical bonds and atoms, where the element P i”j” =1 indicates that there is a correlation between the i-th chemical bond and the j-th atom, I is the unit matrix; T is the transposition operation, Λ is the matrix diagonal operation; NF d Represents the set of atomic features after being updated by the CensNet algorithm, EF d Represents the set of chemical bond features after being updated by the CensNet algorithm.

[0038] Furthermore, the process of obtaining the final drug global features by fusing the node-level atomic features and the graph-level drug global features and performing pooling in step b2 includes:

[0039]

[0040] Among them, NF d；i For NF d Atomic features in FuNP d；i represents the i-th final atomic feature obtained by integrating the global features of the drug, The square of the 2-norm representing the global characteristics of the drug, FuNP d Represents the global characteristics of the drug obtained after pooling.

[0041] Furthermore, in step c1, three graph attention networks GAT are used to analyze the positive protein hydrophobic graph G h 、Positive example protein hydrogen bond diagram G hy 、Positive example protein contact graph G cont The specific process of extracting the positive example features of amino acids corresponding to the three graphs of proteins includes:

[0042] Step c11: Use GAT to h Extract positive amino acid features:

[0043] e ij =LeakyReLU(W 1 [W 2 N h；i ||W 3 N h；j ])

[0044]

[0045] NF h；i =GAT(Nh；i )

[0046] Among them, W 1 , W 2 , W 3 , W 4 is the weight parameter, NF h；i G h The i-th node feature after GAT extraction represents the updated feature of the i-th amino acid under the hydrophobic interaction force;

[0047] Step c12: Use GAT to hy Extract amino acid positive feature NF hy；j =GAT(N hy；j ), the extraction process is the same as step c11; wherein, NF hy；j G hy The characteristics of the jth amino acid after GAT extraction;

[0048] Step c13: Use GAT to cont Extract amino acid positive feature NF cont；k =GAT(N cont；k ), the extraction process is the same as step c11; wherein, NF cont；k G cont The features of the kth amino acid after GAT extraction.

[0049] Furthermore, in step c4, the feature fusion algorithm is used to fuse the three amino acid positive example features in c1 to obtain the amino acid fusion positive example feature:

[0050] NF temp；i =NF h；i T NF h；i +NF hy；i T NF hy；i +NF h；i T NF hy；i cos(NF h；i T ,NF hy；i )

[0051] NF fusion；i =NF temp；i T NF temp；i +NF cont；i T NF cont；i +NF cont；i T NF temp；i cos(NF temp；i ,NFcont；i )

[0052] Among them, NF fusion；i Represents the fusion feature of the i-th amino acid feature in the three positive example graphs of the protein.

[0053] Furthermore, in step d1, the k-hop algorithm is used to extract the drug atom graph G d Extract the drug subgraph with the node in the node list as the central node, and use the average pooling algorithm for each drug subgraph to obtain the corresponding drug subgraph pooling feature, and then splice all drug subgraph pooling features into the drug subgraph pooling total feature. The specific process includes:

[0054] Taking the node in the node list as the central node, the k-hop algorithm is used to extract the drug subgraph from the drug atomic graph, and the average pooling algorithm is used for each drug subgraph to obtain the corresponding drug subgraph pooling feature:

[0055]

[0056] Among them, k-hop(·) is a k-hop subgraph algorithm, which returns r is the feature set of the central node and its e-order neighbor nodes for The pooling feature of the r-th drug subgraph obtained by average pooling, Z0 is the number of nodes in the r-th drug subgraph, SubP d is the collection of drug subgraph pooling features.

[0057] Furthermore, in step d2, the specific process of extracting protein subgraphs from the positive example protein hydrophobicity map, the positive example protein hydrogen bond map, and the positive example protein contact map using the k-hop algorithm, and obtaining the corresponding protein subgraph pooling features using the average pooling algorithm for each protein subgraph includes the following steps:

[0058] The k-hop algorithm is used to extract protein subgraphs from the positive protein hydrophobic graph, the positive protein hydrogen bond graph, and the positive protein contact graph, and the average pooling algorithm is used for each protein subgraph to obtain the corresponding protein subgraph pooling features:

[0059]

[0060] Among them, Z1 is the number of nodes in the rth subgraph of the positive protein hydrogen bond graph, Z2 is the number of nodes in the rth subgraph of the positive protein hydrophobic graph, and Z3 is the number of nodes in the rth subgraph of the positive protein contact graph.

[0061] Furthermore, in step d3, the additive attention and cross attention algorithms are used to calculate the interactive attention coefficient Atten for each pair of drug subgraphs and protein subgraphs based on the drug subgraph pooling features and protein subgraph pooling total features. za The process includes:

[0062] Step d31: Calculate the cross-attention coefficient using the cross-attention algorithm:

[0063]

[0064] Here, z actually refers to the subgraph index of the drug atom graph; SubP d；z SubP d The zth row of ; a actually refers to the subgraph index of the protein; SubP p；a SubP p The ath row of ; is the weight parameter, CrossAtten za represents the cross-attention coefficient between the zth subgraph of the drug atom graph and the ath subgraph of the protein;

[0065] Step d32: Calculate the additive attention coefficient using the additive attention algorithm:

[0066]

[0067] Among them, W O , is the weight parameter, AddAtten za represents the additive attention coefficient between the zth subgraph of the drug atom graph and the ath subgraph of the protein;

[0068] Step d33: Calculate the final interaction attention coefficient:

[0069] Atten za =W a CrossAtten za +(1-W a )AddAtten za

[0070] Among them, W a is the weight parameter; Atten za Represents the interaction attention coefficient between the zth subgraph of the drug atom graph and the ath subgraph of the protein.

[0071] Furthermore, the drug target affinity prediction model is pre-trained. In the process of training the drug target affinity prediction model, the data acquisition unit also uses a negative sampling strategy to train G in the process of performing step a2. h , Ghy , G cont Generate corresponding negative example graph

[0072] The processing of the protein feature extraction unit also includes:

[0073] Step c2: Use the average pooling algorithm to process c1 separately to extract the three positive amino acid features corresponding to the three graphs of the protein, and obtain the corresponding three positive amino acid pooling features PF h PF hy PF cont ;

[0074] Step c3: Use three GAT networks to analyze the negative protein hydrophobicity map Negative protein hydrogen bond diagram Negative protein contact map Extract the negative features of amino acids corresponding to the three graphs of protein

[0075] At the same time, in step c4, the feature fusion algorithm is used to fuse the three amino acid negative example features in c3 to obtain the amino acid fusion negative example feature

[0076]

[0077] in, It represents the fusion feature of the i-th amino acid feature in the three negative example images of the protein;

[0078] Based on the processing of the drug target affinity prediction model, the total loss function Loss is used to realize the training of the drug target affinity prediction model;

[0079]

[0080] in, To adjust the weights, each sub-loss is entered as follows:

[0081]

[0082] in, A and B in the equation represent the computational objects, and W a represents a learnable matrix; represents mathematical expectation; A', B', and C in the equation represent the computational objects, and W b , W c represents a learnable matrix.

[0083] Furthermore, in step a2, a negative sampling strategy is used to h , G hy , Gcont Generate corresponding negative example graph The specific process includes:

[0084] G h , G hy , G cont The node feature matrix N h 、N hy 、N cont The feature N of each node in the graph h；i 、N hy；i 、N cont；i Each row is spliced ​​together; randomly shuffle N h 、N hy 、N cont The rows of the feature matrix Create three negative example graphs without changing the adjacency matrix: Negative example protein hydrophobicity graph Negative protein hydrogen bond diagram Negative protein contact map

[0085] The beneficial effects of the present invention are:

[0086] This application uses GAT to extract the amino acid features of protein hydrophobic graphs, protein hydrogen bond graphs, and protein contact graphs, provides protein structural information from multiple amino acid interaction forces, and uses feature fusion methods to not only fuse information from multiple protein graph structures, but also fuse local and global information. EGAT and CensNet are used to learn atomic features and chemical bond features from drug atom graphs to highlight the structural features of drugs. And fuse global and local information. At the same time, the combination of linear attention and cross attention is used on subgraph features to alleviate the shortcomings of cross attention based on dot product method that cannot maintain good information transmission at feature-dissimilar nodes, and the comparative learning method is used to improve the ability to extract protein features. This allows the present invention to provide more stable and reliable results while improving performance. A large number of experiments were conducted on the Davis and KIBA data sets to verify that this application is superior to several of the most advanced DTA prediction methods under different experimental settings. It has good robustness and generalization. The prediction method of this application makes full use of known drug and protein data to improve the accuracy of drug target affinity prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] Figure 1 This is a prediction flow chart of the present invention.

[0088] Figure 2 A comparison chart of the prediction performance of a five-fold cross-validation experiment on the Davis dataset with other comparison methods.

[0089] Figure 3A comparison chart of the prediction performance of a five-fold cross validation experiment on the KIBA dataset with other comparison methods.

[0090] Figure 4 A comparison chart of the prediction performance of a five-fold cross-validation experiment using different model settings on the Davis dataset.

[0091] Figure 5 A comparison chart of the prediction performance of a five-fold cross-validation experiment using different model settings on the KIBA dataset. DETAILED DESCRIPTION

[0092] Specific implementation method 1: Combination Figure 1 To explain this embodiment,

[0093] This embodiment is a drug target affinity prediction method based on a graph neural network, which realizes the prediction of drug target affinity based on a drug target affinity prediction model. The drug target affinity prediction model includes a data acquisition unit, a drug feature extraction unit, a protein feature extraction unit, a subgraph feature processing unit, and an affinity prediction unit. The embodiment itself also includes a process of training the model; wherein,

[0094] AData acquisition unit:

[0095] Step a1: Obtain the unique identifier of each drug in the PubChem database, i.e., PubChem ID; obtain the simplified molecular linear input specification (SMILES) data of the drug according to the PubChem ID; then input the SMILES data of the drug into the RDKit software package, use RDKit to calculate the physicochemical properties of each atom and chemical bond in the drug, and then obtain the i-th atom V corresponding to the drug. d；i The atomic characteristics N d；i , and the chemical bond E d；ij Chemical bond characteristics ξ d；ij ; Set atomic feature N d；i and chemical bond characteristics ξ d；ij They are recorded as initial atomic features and initial chemical bond features respectively;

[0096] Atomic features include but are not limited to the one-hot encoding of the atomic symbol, the one-hot encoding of the degree of the atom in the drug, the number of neighbors of the atom in the molecule, the one-hot encoding of the number of hydrogen bonds of the atom in the drug, whether the atom is an aromatic structure, etc. Select one or more of them as the atomic feature N d；i ;

[0097] The chemical bond characteristics include, but are not limited to, the bond energy, bond length, bond rotation, whether the bond has a ring, whether it is a single bond, whether it is a double bond, whether it is a triple bond, etc., and one or more of them are selected as the chemical bond characteristics ξ d；ij ;

[0098] The atoms contained in the drug are used as nodes, the initial atomic features are used as node features, the chemical bonds between two atoms in the drug are used as edges, and the initial features of the chemical bonds are used as edge features to construct the corresponding drug atom graph G. d =(V d ,E d ), V d represents a node, E d Represents an edge.

[0099] Step a2: Use Protein Tools prediction software to calculate the hydrophobic interaction force and hydrogen bond data of the protein for the given protein amino acid sequence, and input the protein amino acid sequence data set into the protein sequence comparison toolkit HHbilts to calculate the amino acid sequence comparison data, and input the calculated amino acid sequence comparison data into the Psconc4 toolkit to obtain the Euclidean space distance data of each amino acid in the protein; the initial characteristics of each amino acid are the physicochemical properties of the amino acid.

[0100] The protein hydrophobic graph G is constructed by taking the amino acids contained in the protein as nodes and the hydrophobic interactions between the amino acids as edges. h =(V h ,Ε h ) ; The amino acids contained in the protein are used as nodes and the hydrogen bonds between the amino acids are used as edges to construct the protein hydrogen bond graph G hy =(V hy ,Ε hy ); The amino acids contained in the protein are taken as nodes, and the Euclidean space distance between the amino acids is less than As edges, construct the protein contact graph G cont =(V cont ,Ε cont ).

[0101] The i-th node V in the three graph structures of the protein h；i 、V hy；i 、V cont；i The initial features are denoted as N h；i 、N hy；i 、N cont；iIn fact, the initial features of the three graphs are consistent. Since the relative positions of amino acids in the spatial structure of proteins play an important role in the functional expression of proteins, changing the relative positions of amino acids affects the function of proteins. In order to obtain protein data with a large functional gap with the original protein, a negative sampling strategy is used to h , G hy , G cont Generate the corresponding negative example map Specifically: For G h , G hy , G cont The node feature matrix N h 、N hy 、N cont The feature N of each node in the graph h；i 、N hy；i 、N cont；i Each row is spliced ​​together; randomly shuffle N h 、N hy 、N cont The rows of the feature matrix Create three negative example graphs without changing the adjacency matrix: Negative example protein hydrophobicity graph Negative protein hydrogen bond diagram Negative protein contact map To distinguish the graph structure from the negative samples, we call G h , G hy , G cont These are the positive example protein hydrophobic map, positive example protein hydrogen bond map, and positive example protein contact map.

[0102] The positive example protein hydrophobic graph G in step a2 h 、Positive example protein hydrogen bond diagram G hy 、Positive example protein contact graph G cont , Negative protein hydrophobicity map Negative protein hydrogen bond diagram Negative protein contact map The construction process is as follows:

[0103] Step a21: Input the protein amino acid sequence into the Protein Tools prediction software to obtain protein hydrophobic interaction force data and hydrogen bond data, and use the Psconc4 prediction algorithm to obtain protein amino acid contact distance data;

[0104] Step a22: Under the normal amino acid feature order, amino acids are used as nodes, and hydrophobic interactions, hydrogen bonds, and contact distances between amino acids are used as three different types of edge sets to construct a positive example protein hydrophobic graph, a positive example protein hydrogen bond graph, and a positive example protein contact graph.

[0105] Step a23: With the amino acid feature order shuffled, the negative protein hydrophobic graph, negative protein hydrogen bond graph, and negative protein contact graph are constructed using amino acids as nodes and hydrophobic interactions, hydrogen bonds, and contact distances between amino acids as three different types of edge sets.

[0106] B Drug feature extraction unit:

[0107] Step b1: Use the edge aggregation graph attention network algorithm and the co-embedding graph neural network algorithm to analyze the drug atom graph G corresponding to each drug. d Extract atomic features and chemical bond features, the edge aggregated graph attention network algorithm is abbreviated as EGAT (Edge Aggregated Graph Attention Networks), and the co-embedding graph neural network algorithm is abbreviated as CensNet (Co-embedding of Nodes and Edges with Graph Neural Networks). At the same time, the average pooling algorithm and splicing method are used to process atomic features and chemical bond features to obtain the global features of the drug;

[0108] Step b11: Use EGAT to directly pass the neighbor atom feature N d；j and its associated chemical bond characteristics Get the atomic features U in the updated drug atom graph d；i ;

[0109] e ji =LeakyReLU((W α [W ν N d；i ||W ν N d；j ||W ρ ξ d；ij ]))

[0110]

[0111] Among them, W ν , W ρ , W ε is a learnable weight parameter, || is a concatenation operation, LeakyReLU(·) and σ(·) are nonlinear activation functions, and N d；i Represents the drug atom graph G d The i-th initial atomic feature, U d；i Represents the drug atom graph G d The characteristics of the i-th atom updated by the EGAT algorithm, ξ d；ij Represents the drug atom graph G d The initial characteristics of the chemical bond connecting the i-th atom to the j-th atom.

[0112] Step b12: Update the atomic feature NF using the chemical bond feature through CensNet d , and use the atomic features to update the chemical bond features EF d :

[0113] NF d =[PΛ(W e ξ d ) T +I]W d U d

[0114] EF d =[PΛ(W d U d ) T +I]W e ξ d

[0115] Among them, U d Represents the drug atomic graph G d All atomic features U after EGAT update d；i The feature set concatenated by row, ξ d Represents the drug atomic graph G d All chemical bond characteristics of ξ d；ij The feature set obtained by concatenating rows. e , W d is a learnable weight parameter; P represents the association matrix between chemical bonds and atoms, where the element P i”j” =1 indicates that there is a correlation between the i-th chemical bond and the j-th atom, I is the unit matrix; T is the transposition operation, Λ is the matrix diagonal operation; NF d Represents the set of atomic features after being updated by the CensNet algorithm, EF d Represents the set of chemical bond features after being updated by the CensNet algorithm.

[0116] Step b13: Use the average pooling algorithm and splicing method to process the atomic features and chemical bond features to obtain the drug global feature GP d :

[0117]

[0118] GP d =NP d ||EP d

[0119] Among them, NF d；i For NF d Atomic features in NP d Indicates atomic pooling features; EFd；t For NF d Chemical bond characteristics in EP d Indicates the chemical bond pooling feature, GP d is the global characteristic of the drug, n is the number of atoms in the drug, and m is the number of chemical bonds in the drug.

[0120] Step b2: Processing drug global features GP based on feature fusion algorithm d and atomic features NF d The final drug global feature FuNP is obtained by fusing the node-level atomic features and the graph-level drug global features and pooling them. d , as new atomic features;

[0121]

[0122] Among them, FuNP d；i represents the i-th final atomic feature obtained by integrating the global features of the drug, The square of the 2-norm representing the global characteristics of the drug, FuNP d Represents the global characteristics of the drug obtained after pooling.

[0123] C Protein feature extraction unit:

[0124] Step c1: Use three graph attention networks (GAT) to analyze the positive protein hydrophobic graph G h 、Positive example protein hydrogen bond diagram G hy 、Positive example protein contact graph G cont The positive amino acid features corresponding to the three graphs of the protein are extracted; the specific process includes:

[0125] Step c11: Use GAT to h Extract positive amino acid features:

[0126] e ij =LeakyReLU(W 1 [W 2 N h；i ||W 3 N h；j ])

[0127]

[0128] NF h；i =GAT(N h；i )

[0129] Among them, W 1 , W 2 , W 3 , W4 is the learnable weight parameter, NF h；i G h The i-th node feature after GAT extraction represents the updated feature of the i-th amino acid under the hydrophobic interaction force.

[0130] Step c12: Use GAT to hy Extract amino acid positive feature NF hy；j =GAT(N hy；j ), the extraction process is the same as step c11; wherein, NF hy；j G hy The features of the jth amino acid in after GAT extraction.

[0131] Step c13: Use GAT to cont Extract amino acid positive feature NF cont；k =GAT(N cont；k ), the extraction process is the same as step c11; wherein, NF cont；k G cont The features of the kth amino acid after GAT extraction.

[0132] Step c2: using the average pooling algorithm to process the three amino acid positive example features respectively to obtain the corresponding three amino acid positive example pooling features; the specific process includes:

[0133] Step c21: Use the average pooling algorithm to h The extracted amino acid positive features are pooled to obtain the corresponding protein hydrophobicity map amino acid pooling features:

[0134]

[0135] Where R is the number of amino acids in the protein hydrophobicity map.

[0136] Step c22: Use the average pooling algorithm to hy The extracted amino acid positive features are pooled to obtain the corresponding protein hydrogen bond graph amino acid pooling features:

[0137]

[0138] Step c23: Use the average pooling algorithm to cont The extracted amino acid positive example features are pooled to obtain the corresponding protein distance graph amino acid pooling features:

[0139]

[0140] Step c3: Use three GAT networks to analyze the negative protein hydrophobicity map Negative protein hydrogen bond diagram Negative protein contact map The negative features of amino acids corresponding to the three graphs of proteins are extracted; the specific process includes:

[0141] Step c31: Use GAT to extract amino acid negative features from the protein hydrophobicity map:

[0142]

[0143] in, is the characteristic of the i-th amino acid in the negative protein hydrophobicity map.

[0144] Step c32: Use GAT to extract amino acid negative features from protein hydrogen bond graph:

[0145]

[0146] in, is the characteristic of the jth amino acid in the negative protein hydrogen bond map.

[0147] Step c33: Use GAT to extract amino acid negative features from the protein contact graph:

[0148]

[0149] in, is the feature of the kth amino acid in the negative protein contact graph.

[0150] Step c4: using a feature fusion algorithm to fuse the three amino acid positive feature in c1 to obtain an amino acid fusion positive feature. Similarly, using a feature fusion algorithm to fuse the three amino acid negative feature in c3 to obtain an amino acid fusion negative feature. The specific process includes:

[0151] Step c41: Use the feature fusion algorithm to fuse the three amino acid positive example features in c1 to obtain the amino acid fusion positive example feature:

[0152] NF temp；i =NF h；i T NF h；i +NF hy；i T NF hy；i +NF h；i T NF hy；i cos(NF h；i T ,NF hy；i )

[0153] NF fusion；i =NF temp；i T NFtemp；i +NF cont；i T NF cont；i +NF cont；i T NF temp；i cos(NF temp；i ,NF cont；i )

[0154] Among them, NF fusion；i Represents the fusion feature of the i-th amino acid feature in the three positive example graphs of the protein.

[0155] Step c42: Use the feature fusion algorithm to fuse the three amino acid negative example features in c3 to obtain the amino acid fusion negative example feature:

[0156]

[0157] in, Represents the fusion feature of the i-th amino acid feature in the three negative example graphs of the protein.

[0158] Step c5: Use the average pooling algorithm to process the amino acid fusion positive example features to obtain the amino acid positive example fusion pooling features

[0159] Among them, PF fusion Represents the amino acid positive example fusion pooling features.

[0160] D sub-graph feature processing unit:

[0161] Step d1: Randomly extract a node list from the drug atom graph with a fixed probability, and use the k-hop algorithm to extract a node list from the drug atom graph G d Extract the drug subgraph with the node in the node list as the central node, and use the average pooling algorithm for the node feature set of each drug subgraph to obtain the corresponding drug subgraph pooling feature, and then all drug subgraph pooling features are spliced ​​into the drug subgraph pooling total feature; the specific process includes:

[0162] Step d11: randomly extract s nodes from the entire set of nodes in the drug atom graph with a fixed probability to form a node list;

[0163] Step d12: Taking the node in the node list as the central node, use the k-hop algorithm to extract the drug subgraph from the drug atomic graph, and use the average pooling algorithm for each drug subgraph to obtain the corresponding drug subgraph pooling feature:

[0164]

[0165] Among them, k-hop(·) is a k-hop subgraph algorithm, which returns ris the feature set of the central node and its e-order neighbor nodes for The pooling feature of the r-th drug subgraph obtained by average pooling, Z0 is the number of nodes in the r-th drug subgraph, SubP d is the set of drug subgraph pooling features, and s is the number of subgraphs in the drug atomic graph.

[0166] Step d2: Randomly extract node lists from the protein hydrophobicity map, the positive protein hydrogen bond map, and the positive protein contact map with a fixed probability, use the k-hop algorithm to extract protein subgraphs from the positive protein hydrophobicity map, the positive protein hydrogen bond map, and the positive protein contact map, and use the average pooling algorithm for each protein subgraph node feature set to obtain the corresponding protein subgraph pooling feature. Then all protein subgraph pooling features are spliced ​​into the total protein subgraph pooling feature; the specific process includes:

[0167] Step d21: randomly extracting nodes from the full set of nodes of the positive example protein hydrophobicity map, the positive example protein hydrogen bond map, and the positive example protein contact map with a fixed probability, and forming corresponding node lists respectively;

[0168] Step d22: Use the k-hop algorithm to extract protein subgraphs from the positive protein hydrophobicity map, the positive protein hydrogen bond map, and the positive protein contact map, and use the average pooling algorithm for each protein subgraph to obtain the corresponding protein subgraph pooling feature:

[0169]

[0170]

[0171] Among them, Z1 is the number of nodes in the rth subgraph of the positive protein hydrogen bond graph, Z2 is the number of nodes in the rth subgraph of the positive protein hydrophobic graph, and Z3 is the number of nodes in the rth subgraph of the positive protein contact graph.

[0172] Step d22: All protein subgraph pooling features are concatenated into protein subgraph pooling total features:

[0173]

[0174] Step d3: Use additive attention and cross attention algorithms to calculate the interactive attention coefficient of each pair of drug subgraphs and protein subgraphs for the drug subgraph pooling features and protein subgraph pooling total features.

[0175] Step d31: Calculate the cross-attention coefficient using the cross-attention algorithm:

[0176]

[0177] Here, z actually refers to the subgraph index of the drug atom graph, and its value range is [1, s]; SubP d；z SubP d The zth row of the protein; a actually refers to the subgraph index of the protein. Since it contains three graph structures, the value range of a is [1,3s]. p；a SubP p The ath row of . is a learnable weight parameter, CrossAtten za Represents the cross-attention coefficient between the z-th subgraph of the drug atom graph and the a-th subgraph of the protein.

[0178] Step d32: Calculate the additive attention coefficient using the additive attention algorithm:

[0179]

[0180] Among them, W O , is a learnable weight parameter, AddAtten za Represents the additive attention coefficient between the zth subgraph of the drug atom graph and the ath subgraph of the protein.

[0181] Step d33: Calculate the final interaction attention coefficient:

[0182] Atten za =W a CrossAtten za +(1-W a )AddAtten za

[0183] Among them, W a is the weight parameter; Atten za Represents the interaction attention coefficient between the zth subgraph of the drug atom graph and the ath subgraph of the protein.

[0184] Step d4: Use matrix multiplication to calculate the cross-fusion features of the drug subgraph pooling of step d1 and the interactive attention coefficient of step d3 to finally fuse the drug subgraph cross-fusion features of the protein information. Similarly, use matrix multiplication to calculate the cross-fusion features of the protein subgraph pooling of the total features and the interactive attention coefficient to finally fuse the compound information; the specific process includes:

[0185] Step d41: Use matrix multiplication to calculate the cross-fusion features of the drug subgraphs of step d1 and the interaction attention coefficients of step d3 to fused protein information:

[0186] SubFu d；z =Attenza SubP d；z

[0187] Step d42: Use matrix multiplication to calculate the protein subgraph cross-fusion features of the fused compound information by pooling the total features and the interaction attention coefficient of the protein subgraph:

[0188]

[0189] Step d5: Use the average pooling algorithm to process the drug subgraph cross-fusion features and protein subgraph cross-fusion features of step d4 to obtain drug subgraph cross-fusion pooling features and protein subgraph cross-fusion pooling features:

[0190]

[0191] E affinity prediction unit:

[0192] Step e1: Use the splicing method to process the drug global features of step b2 and the cross-fusion pooling features of the drug subgraphs of step d5 to obtain the drug features:

[0193] F drug =FuNP d ||SubPFu d

[0194] Step e2: Use the splicing method to process the amino acid positive example fusion pooling features of c5 and the protein subgraph cross fusion pooling features of d5 to obtain the protein features:

[0195] F protein =PF fusion ||SubPFu p

[0196] Step e3: Use the splicing method to splice the drug features and protein features to obtain the drug-protein joint features:

[0197] F conbind =F drug ||F protein

[0198] Step e4: Target F conbind , machine learning technology is used to predict the affinity prediction value affinity.

[0199] In this embodiment, the machine learning technology is implemented by a neural network. In fact, any neural network can be used to implement the prediction. In this embodiment, a fully connected neural network is used for specific prediction, mainly using a three-layer fully connected neural network FC (Full Connection Neural Network, FC) to process the drug-protein joint features to obtain the affinity prediction value:

[0200] affinity=FC1(ReLU(FC2(ReLU(FC3(F conbind ))))

[0201] Among them, FC2(·), FC3(·), FC3(·) represent fully connected layers, and ReLU(·) is the activation function.

[0202] Perform error calculation and train the previous model as a whole. The process of calculating error to guide training includes:

[0203] (1) Using a bilinear function to calculate the mutual information of the amino acid positive feature, amino acid negative feature, and amino acid positive pooling feature of each protein graph structure, and using the cross entropy maximization to calculate the mutual information training objective, including:

[0204] (11) The mutual information between the amino acid positive feature and the amino acid positive pooled feature after GAT update, and the mutual information between the amino acid negative feature and the amino acid positive pooled feature after GAT update are used as a set of cross entropy calculation targets, requiring the mutual information of the former to be the largest (the first expectation maximum), while the mutual information of the latter is the smallest (the second expectation minimum).

[0205]

[0206] in, represents a bilinear function, A and B are replaceable parameters, W a represents a learnable matrix; represents mathematical expectation.

[0207] (12) The mutual information between the positive amino acid features after GAT update and the input positive amino acid features, and the mutual information between the positive amino acid features after GAT update and the negative amino acid features are used as a set of cross entropy calculation targets, requiring the mutual information of the former to be maximized and the mutual information of the latter to be minimized.

[0208]

[0209] (13) The mutual information of the amino acid positive feature, the amino acid positive pooled feature and the amino acid negative feature after GAT update, the mutual information of the amino acid negative feature, the amino acid positive pooled feature and the input amino acid positive feature after GAT update are used as a set of cross entropy calculation targets, requiring the mutual information of the former to be maximized and the mutual information of the latter to be minimized.

[0210]

[0211] Among them, A', B', C are replaceable parameters, W b , W c represents a learnable matrix;

[0212] (2) Using a bilinear function to calculate the mutual information of the amino acid positive example feature, the amino acid positive example fusion feature, the amino acid negative example feature, the amino acid negative example fusion feature, and the amino acid positive example fusion pooling feature, including:

[0213] (21) The mutual information between the amino acid positive example fusion feature and the amino acid positive example pooling feature, and the mutual information between the amino acid negative example fusion feature and the amino acid positive example pooling feature are used as a set of cross entropy calculation targets, requiring the mutual information of the former to be maximized and the mutual information of the latter to be minimized;

[0214]

[0215] (22) The mutual information between the amino acid positive example fusion feature and the amino acid positive example feature, and the mutual information between the amino acid positive example fusion feature and the amino acid negative example feature are used as a set of cross entropy calculation targets, requiring the mutual information of the former to be the largest and the mutual information of the latter to be the smallest;

[0216]

[0217] (23) The mutual information of amino acid positive features, amino acid positive fusion pooling features and amino acid fusion negative features, the mutual information of amino acid positive features, amino acid positive fusion pooling features and amino acid fusion positive features are used as the discriminator loss as a set of cross entropy calculation targets, requiring the mutual information of the former to be maximized and the mutual information of the latter to be minimized.

[0218]

[0219] (3) Use the mean square error function to calculate the distance between the affinity prediction value and the label value, requiring the distance to be minimum (minimum MSE loss);

[0220]

[0221] Based on the above three losses, use parameters to coordinate the weight of each calculation target in the final training target;

[0222]

[0223] in, To adjust the weight.

[0224] After training with the above loss function, the overall network model is obtained, and the network model is used for actual affinity prediction.

[0225] The constructed model refers to all the contents involved in step a1 to step e4. (1) to (3) are actually the training process of the model. The process of using the model is actually the processing process of step a1 to step e4.

[0226] The affinity prediction effect of the present invention is verified by examples. When evaluating, the root mean square error (RMSE), mean square error (MSE), concordance index (CI), Pearson correlation coefficient Person, and adjusted R-Square (RM2) index are used as evaluation indicators.

[0227]

[0228]

[0229] Example:

[0230] In order to evaluate the impact of chemical bonds on prediction performance, this example establishes network models with and without chemical bonds, and calculates the corresponding MSE, RMSE, CI, Person, and RM2 values, and compares them. Figure 2-Figure 5 As shown, the model performance is optimal when there are chemical bonds, which indicates that chemical bonds improve the hidden feature extraction of drugs.

[0231] In order to evaluate the superiority of the present application, the present application is compared with other advanced drug target affinity prediction methods DeepDTA, GraphDTA, WideDTA, DeepDTAF, AttentionDTA, FusionDTA, DeepCDA, DGraphDTA, and GraphCL-DTA on the Davis and KIBA datasets, and the results are shown in Table 1. Secondly, the performance is good on both the Davis and KIBA datasets, indicating that the present application has good robustness and generalization ability.

[0232] Table 1. Comparison of independent test performance on Davis and KIBA datasets

[0233]

[0234]

[0235] Table 2 Comparison results under different settings on Davis and KIBA datasets

[0236]

[0237]

[0238] Among them, s1 means no hydrogen bond, s2 means no hydrophobic bond, s3 means no amino acid distance, s4 means hydrophobic-hydrogen bond-contact fusion order, s5 means hydrophobic-contact-hydrogen bond, s6 means contact-hydrogen bond-hydrophobic, s7 means contact-hydrophobic-hydrogen bond, s8 means hydrogen bond-contact-hydrophobic, s9 means hydrogen bond-hydrophobic-contact, s10 means no chemical bond, s11 means no local subgraph information, s12 means no global information fusion, s13 means no linear attention, s14 means no cross attention, and My means the best combination.

[0239] The above calculation examples of the present invention are only used to explain the calculation model and calculation process of the present invention in detail, and are not intended to limit the implementation methods of the present invention. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made based on the above description. It is impossible to list all the implementation methods here. All obvious changes or modifications derived from the technical solution of the present invention are still within the scope of protection of the present invention.

Claims

1. A drug target affinity prediction method based on graph neural network, characterized in that: A drug target affinity prediction model is used to predict drug target affinity, wherein the drug target affinity prediction model includes a data acquisition unit, a drug feature extraction unit, a protein feature extraction unit, a subgraph feature processing unit, and an affinity prediction unit; The processing of the data acquisition unit includes: Step a1: For each drug, obtain the drug atomic features and chemical bond features, take the atoms contained in the drug as nodes, the atomic features as node features, and the chemical bond features as edges, and then construct the corresponding drug atomic graph G d ; Step a2: Obtain the hydrophobic interaction forces and hydrogen bonds of the protein amino acids, and construct the protein hydrophobic graph G with the amino acids contained in the protein as nodes and the hydrophobic interaction forces between the amino acids as edges. h The protein hydrogen bond graph G is constructed with the amino acids contained in the protein as nodes and the hydrogen bonds between the amino acids as edges. hy ; For protein amino acids, the Euclidean space distance data of each amino acid in the protein is obtained, and the amino acids contained in the protein are taken as nodes, and the Euclidean space distance between amino acids less than the distance threshold is taken as edges to construct the protein contact graph G cont ; The processing of the drug feature extraction unit includes: Step b1: Use the edge aggregation graph attention network algorithm EGAT and the co-embedding graph neural network algorithm CensNet to analyze the drug atomic graph G corresponding to each drug. d Extract atomic features NF d and chemical bond characteristics EF d , using the average pooling algorithm to obtain the atomic pooling feature NP d and chemical bond pooling feature EP d , and then splice to get the drug global feature GP d =NP d ||EP d ; Step b2: Processing drug global features GP based on feature fusion algorithm d and atomic features NF d The final drug global feature FuNP is obtained by fusing the node-level atomic features and the graph-level drug global features and pooling them. d ; The processing of the protein feature extraction unit includes: Step c1: Use three GAT networks to analyze the hydrophobic graph G of the positive protein h 、Positive example protein hydrogen bond diagram G hy 、Positive example protein contact graph G cont Extract the positive amino acid features NF corresponding to the three graphs of the protein h；i NF hy；j NF cont；k ; Step c4: Use the feature fusion algorithm to fuse the three amino acid positive feature in c1 to obtain the amino acid fusion positive feature NF fusion；i , NF fusion；i Represents the fusion feature of the i-th amino acid feature in the three positive example graphs of the protein; Step c5: Use the average pooling algorithm to process the amino acid fusion positive example features to obtain the amino acid positive example fusion pooling features Among them, PF fusion represents the amino acid positive example fusion pooling feature; The processing of the sub-graph feature processing unit includes: Step d1: Randomly extract nodes from the drug atom graph with a fixed probability to form a node list, and use the k-hop algorithm to extract nodes from the drug atom graph G d The drug subgraph with the node in the node list as the central node is extracted from the node list, and the node feature set of each drug subgraph is averaged and pooled to obtain the corresponding drug subgraph pooling feature. Then, all drug subgraph pooling features are concatenated into the drug subgraph pooling total feature SubP d ; Step d2: randomly extract nodes from the protein hydrophobicity map, the positive protein hydrogen bond map, and the positive protein contact map with a fixed probability, and each of them forms a corresponding node list; use the k-hop algorithm to extract protein subgraphs from the positive protein hydrophobicity map, the positive protein hydrogen bond map, and the positive protein contact map, and use the average pooling algorithm for the node feature set of each protein subgraph to obtain the corresponding protein subgraph pooling feature; then all protein subgraph pooling features are spliced ​​into the protein subgraph pooling total feature SubP p ; Step d3: Use additive attention and cross attention algorithms to calculate the interactive attention coefficient Atten for each pair of drug subgraphs and protein subgraphs based on the drug subgraph pooling features and protein subgraph pooling total features. za ; Step d4: Use matrix multiplication to calculate the total pooling features of the drug subgraph in step d1 and the interactive attention coefficient in step d3 to obtain the final drug subgraph cross-fusion feature SubFu that fused the protein information. d；z ; Use matrix multiplication to calculate the protein subgraph pooling total features of step d2 and the interactive attention coefficient of step d3 to calculate the protein subgraph cross-fusion feature SubFu of the final fusion compound information P；a ; Step d5: Use the average pooling algorithm to process the drug subgraph cross-fusion features SubFu in step d4 d；z , protein subgraph cross fusion feature SubFu P；a Get the drug subgraph cross fusion pooling feature SubPFu d , protein subgraph cross fusion pooling feature SubPFu p ; The processing of the affinity prediction unit includes: Step e1: Use the splicing method to process the drug global features of step b2 and the drug subgraph cross-fusion pooling features of step d5 to obtain the drug feature F drug =FuNP d ||SubPFu d ; Step e2: Use the splicing method to process the amino acid positive example fusion pooling features of c5 and the protein subgraph cross fusion pooling features of d5 to obtain the protein feature F protein =PF fusion ||SubPFu p ; Step e3: Use the splicing method to splice the drug features and protein features to obtain the drug-protein joint feature F conbind =F drug ||F protein ; Step e4: Target F conbind , machine learning technology is used to predict the affinity prediction value affinity.

2. A drug target affinity prediction method based on graph neural network according to claim 1, characterized in that: Step b1 uses the edge aggregation graph attention network algorithm EGAT and the co-embedded graph neural network algorithm CensNet to analyze the drug atomic graph G corresponding to each drug. d Extract atomic features NF d and chemical bond characteristics EF d The specific process includes: Step b11: Use EGAT to directly pass the neighbor atom feature N d；j and its associated chemical bond characteristics Get the atomic features U in the updated drug atom graph d；i ; and ji =LeakyReLU((W α [IN ν N d；i ||In ν N d；j ||In ρ ξ d；ij ])) Among them, W ν , W ρ , W ε is the weight parameter, || is the concatenation operation, LeakyReLU(·) and σ(·) are nonlinear activation functions, N d；i Represents the drug atom graph G d The i-th initial atomic feature, U d；i Represents the drug atom graph G d The characteristics of the i-th atom updated by the EGAT algorithm, ξ d；ij Represents the drug atom graph G d The initial characteristics of the chemical bond connecting the i-th atom to the j-th atom; Step b12: Update the atomic feature NF using the chemical bond feature through CensNet d , and use the atomic features to update the chemical bond features EF d : NF d =[PΛ(W e x d )P T +I]W d U d EF d =[PΛ(W d U d )P T +I]W e x d Among them, U d Represents the drug atom graph G d All atomic features U after EGAT update d；i The feature set concatenated by row, ξ d Represents the drug atom graph G d All chemical bond characteristics of ξ d；ij The feature set obtained by concatenating rows; W e , W d is the weight parameter; P represents the correlation matrix between chemical bonds and atoms, where the element P i”j” =1 indicates that there is a correlation between the i-th chemical bond and the j-th atom, I is the unit matrix; T is the transposition operation, Λ is the matrix diagonal operation; NF d Represents the set of atomic features after being updated by the CensNet algorithm, EF d Represents the set of chemical bond features after being updated by the CensNet algorithm.

3. A drug target affinity prediction method based on graph neural network according to claim 2, characterized in that: The process of obtaining the final drug global features by fusing the node-level atomic features and the graph-level drug global features and pooling them in step b2 includes: Among them, NF d；i For NF d Atomic features in FuNP d；i represents the i-th final atomic feature obtained by integrating the global features of the drug, The square of the 2-norm representing the global characteristics of the drug, FuNP d Represents the global characteristics of the drug obtained after pooling.

4. The drug target affinity prediction method based on graph neural network according to claim 1, characterized in that: In step c1, three graph attention networks GAT are used to analyze the positive protein hydrophobic graph G h 、Positive example protein hydrogen bond diagram G hy 、Positive example protein contact graph G cont The specific process of extracting the positive example features of amino acids corresponding to the three graphs of proteins includes: Step c11: Use GAT to h Extract positive amino acid features: and ij =LeakyReLU(W 1 [IN 2 N h；i ||In 3 N h；j ]) NF h；i =GAT(N h；i ) Among them, W 1 , W 2 , W 3 , W 4 is the weight parameter, NF h；i G h The i-th node feature after GAT extraction represents the updated feature of the i-th amino acid under the hydrophobic interaction force; Step c12: Use GAT to hy Extract amino acid positive feature NF hy；j =GAT(N hy；j ), the extraction process is the same as step c11; wherein, NF hy；j G hy The characteristics of the jth amino acid after GAT extraction; Step c13: Use GAT to cont Extract amino acid positive feature NF cont；k =GAT(N cont；k ), the extraction process is the same as step c11; wherein, NF cont；k G cont The features of the kth amino acid after GAT extraction.

5. The drug target affinity prediction method based on graph neural network according to claim 1, characterized in that: In step c4, the feature fusion algorithm is used to fuse the three amino acid positive example features in c1 to obtain the amino acid fusion positive example feature: NF temp；i =NF h；i T NF h；i +NF hy；i T NF hy；i +NF h；i T NF hy；i cos(NF h；i T ,NF hy；i ) NF fusion；i =NF temp；i T NF temp；i +NF cont；i T NF cont；i +NF cont；i T NF temp；i cos(NF temp；i ,NF cont；i ) Among them, NF fusion；i Represents the fusion feature of the i-th amino acid feature in the three positive example graphs of the protein.

6. The method for drug target affinity prediction based on graph neural network according to claim 1, characterized in that: In step d1, the k-hop algorithm is used to extract the drug atom graph G d Extract the drug subgraph with the node in the node list as the central node, and use the average pooling algorithm for each drug subgraph to obtain the corresponding drug subgraph pooling feature, and then splice all drug subgraph pooling features into the drug subgraph pooling total feature. The specific process includes: Taking the node in the node list as the central node, the k-hop algorithm is used to extract the drug subgraph from the drug atomic graph, and the average pooling algorithm is used for each drug subgraph to obtain the corresponding drug subgraph pooling feature: Among them, k-hop(·) is a k-hop subgraph algorithm, which returns r is the feature set of the central node and its e-order neighbor nodes for The pooling feature of the r-th drug subgraph obtained by average pooling, Z0 is the number of nodes in the r-th drug subgraph, SubP d is the collection of drug subgraph pooling features.

7. A drug target affinity prediction method based on graph neural network according to claim 6, characterized in that: In step d2, the specific process of extracting protein subgraphs from the positive example protein hydrophobicity map, the positive example protein hydrogen bond map, and the positive example protein contact map using the k-hop algorithm, and obtaining the corresponding protein subgraph pooling features using the average pooling algorithm for each protein subgraph includes the following steps: The k-hop algorithm is used to extract protein subgraphs from the positive protein hydrophobic graph, the positive protein hydrogen bond graph, and the positive protein contact graph, and the average pooling algorithm is used for each protein subgraph to obtain the corresponding protein subgraph pooling features: Among them, Z1 is the number of nodes in the rth subgraph of the positive protein hydrogen bond graph, Z2 is the number of nodes in the rth subgraph of the positive protein hydrophobic graph, and Z3 is the number of nodes in the rth subgraph of the positive protein contact graph.

8. The method for drug target affinity prediction based on graph neural network according to claim 1, characterized in that: Step d3 uses additive attention and cross attention algorithms to calculate the interactive attention coefficient Atten for each pair of drug subgraphs and protein subgraphs based on the drug subgraph pooling features and protein subgraph pooling total features. za The process includes: Step d31: Calculate the cross-attention coefficient using the cross-attention algorithm: Here, z actually refers to the subgraph index of the drug atom graph; SubP d；z SubP d The zth row of ; a actually refers to the subgraph index of the protein; SubP p；a SubP p The ath row of ; is the weight parameter, CrossAtten za represents the cross-attention coefficient between the zth subgraph of the drug atom graph and the ath subgraph of the protein; Step d32: Calculate the additive attention coefficient using the additive attention algorithm: Among them, W O , is the weight parameter, AddAtten za represents the additive attention coefficient between the zth subgraph of the drug atom graph and the ath subgraph of the protein; Step d33: Calculate the final interaction attention coefficient: Attention za =W a CrossAtten za +(1-W a )AddAttention za Among them, W a is the weight parameter; Atten za Represents the interaction attention coefficient between the zth subgraph of the drug atom graph and the ath subgraph of the protein.

9. A drug target affinity prediction method based on graph neural network according to any one of claims 1 to 8, characterized in that: The drug target affinity prediction model is pre-trained. During the training of the drug target affinity prediction model, the data acquisition unit also uses a negative sampling strategy to train G in the process of performing step a2. h , G hy , G cont Generate corresponding negative example graph The processing of the protein feature extraction unit also includes: Step c2: Use the average pooling algorithm to process c1 separately to extract the three positive amino acid features corresponding to the three graphs of the protein, and obtain the corresponding three positive amino acid pooling features PF h PF hy PF cont ; Step c3: Use three GAT networks to analyze the negative protein hydrophobicity map Negative protein hydrogen bond diagram Negative protein contact map Extract the negative features of amino acids corresponding to the three graphs of protein At the same time, in step c4, the feature fusion algorithm is used to fuse the three amino acid negative example features in c3 to obtain the amino acid fusion negative example feature in, It represents the fusion feature of the i-th amino acid feature in the three negative example images of the protein; Based on the processing of the drug target affinity prediction model, the total loss function Loss is used to realize the training of the drug target affinity prediction model; in, To adjust the weights, each sub-loss is entered as follows: in, A and B in the equation represent the computational objects, and W a represents a learnable matrix; represents mathematical expectation; A', B', and C in the equation represent the computational objects, and W b , W c represents a learnable matrix.

10. A drug target affinity prediction method based on graph neural network according to claim 9, characterized in that: In step a2, negative sampling strategy is used to h , G hy , G cont Generate the corresponding negative example map The specific process includes: G h , G hy , G cont The node feature matrix N h 、N hy 、N cont The feature N of each node in the graph h；i 、N hy；i 、N cont；i Each row is spliced ​​together; randomly shuffle N h 、N hy 、N cont The rows of the feature matrix Create three negative example graphs without changing the adjacency matrix: Negative example protein hydrophobicity graph Negative protein hydrogen bond diagram Negative protein contact map

Images