A Drug Target Affinity Prediction Method Based on Graph Neural Networks

Abstract

This invention discloses a drug target affinity prediction method based on graph neural networks. To address the low accuracy of existing methods, this invention learns global drug features from the drug atomic map; extracts positive amino acid features from positive protein hydrophobic maps, hydrogen bond maps, and contact maps, and performs fusion pooling to obtain amino acid positive example fusion pooling features as global protein features; extracts sub-maps from the drug atomic map to obtain drug sub-map features as local drug information; and extracts sub-maps from positive protein hydrophobic maps, hydrogen bond maps, and contact maps to obtain protein sub-map features as local protein information. The drug local features and global drug features are concatenated to form drug features, and the protein local features and global protein features are concatenated to form protein features. These two concatenations form a drug-protein joint feature for drug target affinity prediction.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology, specifically relating to a method for predicting drug target affinity. Background Technology

[0002] Drugs can achieve therapeutic effects by interacting with specific regions of proteins, such as enzyme active sites or receptor binding sites, thereby altering the way proteins function. The protein that interacts with the drug is called the target protein, the local region within the target protein that interacts with the drug is called the target site, and the strength of the drug-protein interaction is called drug target affinity (DTA). Higher affinity indicates a greater likelihood of interaction. Indicators for quantifying affinity include the dissociation constant (Kd), inhibition constant (Ki), and half-maximal inhibitory concentration (IC50). Higher values ​​for these three values ​​indicate lower affinity, meaning a lower likelihood of interaction, and consequently, a less likely drug to be used to treat diseases caused by that target protein.

[0003] Over the past few decades, high-throughput screening (HTS) experiments have significantly accelerated the calculation of drug target affinity strengths. However, HTS experiments are costly and labor-intensive, and cannot meet the needs of revealing the affinity strengths of millions of existing drug compounds and thousands of target proteins. Therefore, there is a great need to develop efficient computational methods to fully utilize biological data on known drug molecule-target protein affinities to elucidate the mechanisms of drug action in the human body.

[0004] In recent years, researchers have proposed many methods for predicting drug target affinity in the fields of machine learning and deep learning. In machine learning, Kron-RLS uses a kernel method to extract drug and protein features from drug-drug similarity features and protein-protein similarity features, and uses random forests to obtain drug target affinity strength. SimBoost represents proteins and drug compounds in the same way as Kron-RLS. It constructs features for drugs, targets, and drug-target pairs, and uses feature engineering to extract feature vectors of drug-target pairs to train a gradient booster to predict binding affinity. In the field of deep learning, drug target affinity is mainly divided into two categories: sequence-based and graph-based methods. Sequence-based methods often start with the primary structure of proteins, as the amino acid sequence, as the primary structure of proteins, provides fundamental information about the higher-order structure and the functions they exhibit. 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. Graph-based methods, on the other hand, focus on representing the spatial structure of drugs and proteins. For instance, DGraphDTA constructs drug atomic graphs and protein amino acid contact graphs based on the RDKit software package and the Pconsc4 algorithm. It uses a graph neural network (GNN) based on the theory that adjacent nodes should have similar representations to extract features of drugs and proteins, thereby predicting the affinity between drugs and target proteins.

[0005] While previous methods have made some progress in DTA prediction, there is still room for improvement. Most proposed DTA methods have failed to fully utilize the interaction forces that underpin the spatial structure of drugs and proteins. Furthermore, the use of cross-attention based on dot product calculations ignores the fact that atoms with interacting forces are not all similar, and that complementary features exist, similar to those between cations and anions. This results in the extracted drug and protein features not fully representing the drugs and proteins. The current DTA prediction methods' failure to fully utilize known drug and protein data also leads to omissions in the extraction of structural information and feature representation of drugs and proteins, resulting in low accuracy in drug target affinity prediction. Summary of the Invention

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

[0007] A drug target affinity prediction method based on graph neural networks is disclosed. The method predicts drug target affinity based on a drug target affinity prediction model, which 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 data acquisition unit's processing includes:

[0009] Step a1: For each drug, obtain the drug's atomic features and chemical bond features. Using the atoms contained in the drug as nodes, the atomic features as node features, and the chemical bond features as edges, a corresponding drug atomic graph G is constructed. d ;

[0010] Step a2: For protein amino acids, obtain the hydrophobic interactions and hydrogen bonds of the protein. Construct a protein hydrophobic graph G using the amino acids contained in the protein as nodes and the hydrophobic interactions between amino acids as edges. h A protein hydrogen bond graph G is constructed using amino acids as nodes and hydrogen bonds between amino acids as edges. hy For protein amino acids, the Euclidean spatial distance data of each amino acid in the protein is obtained. Using the amino acids in the protein as nodes and the Euclidean spatial distance between amino acids less than a distance threshold as edges, a protein contact graph G is constructed. 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 process the drug atom graph G corresponding to each drug. d Extracting Atomic Features (NF) d and chemical bond characteristics EF d The atomic pooling feature NP is obtained using the average pooling algorithm. d and chemical bond pooling characteristics EP d The global drug characteristics (GP) are obtained by splicing. d =NP d ||EP d ;

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

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

[0015] Step c1: Use three GAT networks to analyze the hydrophobicity map G of positive protein samples. h Example: Protein hydrogen bond diagram G hy Positive example protein contact diagram G cont Extract the positive amino acid features (NF) corresponding to the three protein graphs. h；i NF hy；j NF cont；k ;

[0016] Step c4: Use a feature fusion algorithm to fuse the three amino acid positive example features in c1 to obtain the amino acid fused positive example feature NF. fusion；i NF fusion；i This represents the fusion feature of the i-th amino acid in the three positive protein images;

[0017] Step c5: Process the amino acid fusion positive example features using the average pooling algorithm to obtain the amino acid positive example fusion pooling features. Among them, PF fusion This indicates the positive example fusion pooling characteristics of amino acids;

[0018] The subgraph feature processing unit includes the following:

[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 Drug subgraphs are extracted from the node list, with each node as the center node. For each drug subgraph, the average pooling algorithm is used to obtain the corresponding drug subgraph pooling features. Then, all drug subgraph pooling features are concatenated to form the total drug subgraph pooling feature, SubP. d ;

[0020] Step d2: Randomly extract nodes from the protein hydrophobic map, positive protein hydrogen bond map, and positive protein contact map with fixed probabilities, forming corresponding node lists for each. Use the k-hop algorithm to extract protein subgraphs from each of the three maps, and apply average pooling to the node feature set of each subgraph to obtain the corresponding protein subgraph pooling features. Then, concatenate all protein subgraph pooling features to form the total protein subgraph pooling feature SubP. p ;

[0021] Step d3: Calculate the interaction attention coefficient (Atten) for each pair of drug subgraphs and protein subgraphs using additive attention and cross-attention algorithms on the pooling features of the drug subgraph and the total pooling features of the protein subgraph. za ;

[0022] Step d4: Calculate the final drug subgraph cross-fusion feature (SubFu) of fused protein information using matrix multiplication on the total pooling features of the drug subgraph from step d1 and the interaction attention coefficients from step d3. d；z The final fused compound information protein subgraph cross-fusion feature (SubFu) is calculated by using matrix multiplication on the total pooling features of the protein subgraph in step d2 and the interaction attention coefficients in step d3. P；a ;

[0023] Step d5: Process the drug subgraph cross-fusion features (SubFu) from step d4 using the average pooling algorithm. d；z Protein subgraph cross-fusion feature SubFu P；a Obtain 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: Process the global drug features from step b2 and the drug subgraph cross-fusion pooling features from step d5 using a concatenation method to obtain drug features F. drug =FuNP d ||SubPFu d ;

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

[0027] Step e3: Use a splicing method to splice drug features and protein features to obtain drug-protein combined features F. conbind =F drug ||F protein ;

[0028] Step e4: For F conbind The affinity value is predicted using machine learning techniques.

[0029] Furthermore, step b1 involves using the edge aggregation graph attention network algorithm EGAT and the co-embedding graph neural network algorithm CensNet to process the drug atom graph G corresponding to each drug. d Extracting Atomic Features (NF) d and chemical bond characteristics EF d The specific process includes:

[0030] Step b11: Utilize EGAT directly through the neighboring atom feature N d；j and the characteristics of the associated chemical bonds Atomic features U obtained from the updated drug atomic diagram 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 Here, || represents the concatenation operation, LeakyReLU(·) and σ(·) are non-linear activation functions, and N is the weight parameter. d；i Drug atomic diagram G d The i-th initial atomic feature, U d；i Drug atomic diagram G d The feature of the i-th atom updated by the EGAT algorithm, ξ d；ij Drug atomic diagram G d Initial characteristics of the chemical bond connecting the i-th atom and the j-th atom;

[0034] Step b12: Update atomic features NF using chemical bond features via CensNet d Simultaneously, the chemical bond feature EF is updated using atomic features. d :

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

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

[0037] Among them, U d Drug atomic diagram G d All atomic features U after EGAT update d；iThe feature set concatenated by row, ξ d Drug atomic diagram G d All chemical bond characteristics ξ d；ij The feature set obtained by concatenating rows; W e W d The weighting parameter is P; P represents the correlation matrix between chemical bonds and atoms, where 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 identity matrix; T is the transpose operation; Λ is the matrix diagonal operation; NF d EF represents the set of atomic features updated by the CensNet algorithm. d This represents the set of chemical bond features after being updated by the CensNet algorithm.

[0038] Furthermore, step b2, which describes the process of fusing node-level atomic features and graph-level global drug features and then pooling them to obtain the final global drug features, includes:

[0039]

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

[0041] Furthermore, in step c1, three graph attention networks (GATs) are used to analyze the hydrophobic map G of positive protein samples. h Example: Protein hydrogen bond diagram G hy Positive example protein contact diagram G cont The specific process of extracting the positive amino acid features corresponding to the three protein graphs includes:

[0042] Step c11: Use GAT to process G h Extracting positive amino acid characteristics:

[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 For weight parameters, NF h；i For G h The i-th node feature extracted by GAT represents the feature of the i-th amino acid updated under hydrophobic interaction force;

[0047] Step c12: Use GAT to process G hy Extracting positive amino acid features (NF) hy；j =GAT(N hy；j The extraction process is the same as step c11; among which, NF hy；j For G hy The characteristics of the j-th amino acid after GAT extraction;

[0048] Step c13: Use GAT on G cont Extracting positive amino acid features (NF) cont；k =GAT(N cont；k The extraction process is the same as step c11; among which, NF cont；k For G cont The characteristics of the kth amino acid after GAT extraction.

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

[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 This represents the fusion feature of the i-th amino acid feature in the three positive examples of the protein.

[0053] Furthermore, in step d1, the k-hop algorithm is used to extract the drug atomic diagram G. d The specific process of extracting drug subgraphs centered on nodes in the node list, applying average pooling to each drug subgraph to obtain corresponding drug subgraph pooling features, and then concatenating all drug subgraph pooling features to form the total drug subgraph pooling features includes:

[0054] Using nodes in the node list as center nodes, the k-hop algorithm is used to extract drug subgraphs from the drug atomic graph, and the average pooling algorithm is applied to each drug subgraph to obtain the corresponding drug subgraph pooling features:

[0055]

[0056] Here, k-hop(·) is the k-hop subgraph algorithm, which returns a subgraph starting from node n. r The feature set of the central node and its e-order neighbor nodes for The pooling feature of the r-th drug subgraph obtained after average pooling, where Z0 is the number of nodes in the r-th drug subgraph, SubP d It is a set of pooling features for drug subgraphs.

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

[0058] The k-hop algorithm is used to extract protein sub-maps from the hydrophobic maps, hydrogen bond maps, and contact maps of positive protein samples. For each protein sub-map, the average pooling algorithm is applied to obtain the corresponding protein sub-map pooling features.

[0059]

[0060] Where Z1 is the number of nodes in the r-th subgraph of the hydrogen bond graph of the positive example protein, Z2 is the number of nodes in the r-th subgraph of the hydrophobic graph of the positive example protein, and Z3 is the number of nodes in the r-th subgraph of the contact graph of the positive example protein.

[0061] Further, step d3 involves using additive attention and cross-attention algorithms to calculate the interaction attention coefficient (Atten) for each pair of drug subgraphs and protein subgraphs based on the pooling features of the drug subgraph and the total pooling features of the protein subgraph. za The process includes:

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

[0063]

[0064] Here, 'z' actually refers to the subgraph index of the drug atom diagram; SubP d；z It is SubP d The z-th line; 'a' actually refers to the protein subgraph index; SubP p；a For SubP p The a-th row; CrossAtten is the weighting parameter. za The cross-attention coefficient represents the relationship between the z-th subgraph of the drug atomic map and the a-th subgraph of the protein.

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

[0066]

[0067] Among them, W O , For weight parameters, AddAtten za The additive attention coefficient represents the relationship between the z-th subgraph of the drug atomic map and the a-th 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 For weight parameters; Atten za This represents the interaction attention coefficient between the z-th subgraph of the drug atomic diagram and the a-th subgraph of the protein.

[0071] Furthermore, the drug target affinity prediction model is pre-trained. During the training process of the drug target affinity prediction model, the data acquisition unit also employs a negative sampling strategy for G during step a2 processing. h Ghy G cont Generate corresponding negative example graphs

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

[0073] Step c2: Use the average pooling algorithm to process the three amino acid positive feature samples corresponding to the three protein images extracted in c1, and obtain the corresponding three amino acid positive feature pooling feature PF. h PF hy PF cont ;

[0074] Step c3: Use three GAT networks to plot the hydrophobicity of negative example proteins. Negative example protein hydrogen bond diagram Negative example protein contact diagram Extract the amino acid negative features corresponding to the three protein graphs.

[0075] Simultaneously, in step c4, a feature fusion algorithm is used to fuse the three amino acid negative example features from c3 to obtain the amino acid fusion negative example features.

[0076]

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

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

[0079]

[0080] in, To adjust the weights, the sub-losses are as follows:

[0081]

[0082] in, In this context, A and B represent the objects being calculated, and W... a Represents a learnable matrix; Represents the mathematical expectation; In this context, A', B', and C represent the computational objects, and W... b W c This represents a learnable matrix.

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

[0084] G h G hy G cont The node feature matrix N h N hy N cont Each node in the graph has a feature N. h；i N hy；i N cont；i Each row is concatenated; N is randomly shuffled. h N hy N cont The rows yield the feature matrix Create three negative example graphs without changing the adjacency matrix: a negative example protein hydrophobic graph. Negative example protein hydrogen bond diagram Negative example protein contact diagram

[0085] The beneficial effects of this invention are:

[0086] This application employs GAT to extract amino acid features from protein hydrophobic maps, protein hydrogen bond maps, and protein contact maps, providing protein structural information from multiple amino acid interactions. It utilizes feature fusion to integrate information from multiple protein map structures, as well as local and global information. EGAT and CensNet are used to learn atomic and chemical bond features from drug atomic maps, highlighting the drug's structural characteristics. Global and local information are also fused. Simultaneously, a combination of linear and cross-attention is used on subgraph features to mitigate the shortcomings of dot-product-based cross-attention in maintaining good information transfer at nodes with dissimilar features. Contrastive learning is employed to improve protein feature extraction capabilities. This allows the invention to provide more stable and reliable results while improving performance. Extensive experiments on the Davis and KIBA datasets verify that this application outperforms several state-of-the-art DTA prediction methods under different experimental settings. It exhibits good robustness and generalization. The prediction method in this application fully utilizes known drug and protein data to improve the accuracy of drug target affinity prediction. Attached Figure Description

[0087] Figure 1 This is a flowchart of the prediction process of the present invention.

[0088] Figure 2 This graph shows a comparison of prediction performance with other comparative methods in a five-fold cross-validation experiment on the Davis dataset.

[0089] Figure 3This is a graph comparing the prediction performance of five-fold cross-validation experiments with other methods on the KIBA dataset.

[0090] Figure 4 This graph compares the prediction performance of different model settings using five-fold cross-validation experiments on the Davis dataset.

[0091] Figure 5 This graph compares the prediction performance of different model settings using five-fold cross-validation experiments on the KIBA dataset. Detailed Implementation

[0092] Specific implementation method one: Combining Figure 1 This implementation method is described below.

[0093] This embodiment presents a drug target affinity prediction method based on graph neural networks. It uses a drug target affinity prediction model to predict drug target affinity. This 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 also includes a model training process.

[0094] Data Acquisition Unit A:

[0095] Step a1: Obtain the unique identifier for each drug in the PubChem database, i.e., PubChem ID; obtain the Simplified Molecular Input Line Entry System (SMILES) data for the drug based on the PubChem ID; then input the SMILES data of the drug into the RDKit software package, and use RDKit to calculate the physicochemical properties of each atom and chemical bond in the drug, thereby obtaining the i-th atom V of the drug. d；i Atomic characteristics N d；i and chemical bond E d；ij Chemical bond characteristics ξ d；ij ;to use atomic features N d；i and chemical bond characteristics ξ d；ij These are respectively denoted as initial atomic features and initial chemical bond features;

[0096] Atomic features include, but are not limited to, the one-heat code of the atomic symbol, the one-heat code of the atom's degree in the drug, the number of the atom's neighbors in the molecule, the one-heat code of the number of hydrogen bonds in the drug, and whether the atom has an aromatic structure. One or more of these are selected as atomic feature N. d；i ;

[0097] Chemical bond characteristics include, but are not limited to, selecting one or more of the following as chemical bond characteristics: bond energy, bond length, bond rotation, presence of a ring, whether it is a single bond, double bond, or triple bond. d；ij ;

[0098] Using the atoms contained in the drug as nodes, the initial atomic features as node features, and the chemical bonds between two atoms in the drug as edges, the initial features of the chemical bonds as edge features, a corresponding drug atom graph G is constructed. d =(V d E d V d Represents a node, E d Represents an edge.

[0099] Step a2: Use the Protein Tools prediction software to calculate the hydrophobic interaction forces and hydrogen bond data of the given protein amino acid sequence, and input the protein amino acid sequence dataset into the HHbilts protein sequence comparison toolkit to calculate the amino acid sequence comparison data. Input the calculated amino acid sequence comparison data into the Pconsc4 toolkit to obtain the Euclidean spatial 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] A protein hydrophobic graph G is constructed using amino acids as nodes and the hydrophobic interactions between amino acids as edges. h =(V h ,E h A protein hydrogen bond graph G is constructed using amino acids as nodes and hydrogen bonds between amino acids as edges. hy =(V hy ,E hy ); Taking the amino acids contained in a protein as nodes, the Euclidean distance between amino acids is less than 1 / 3. As edges, construct the protein contact graph G. cont =(V cont ,E cont ).

[0101] In the three-graph structure of a protein, the i-th node V 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 a protein's spatial structure play a crucial role in its functional expression, altering these positions affects protein function. To obtain protein data that significantly deviates from the original protein's function, a negative sampling strategy is employed for G... h G hy G cont Generate corresponding negative example graphs Specifically, for G h G hy G cont The node feature matrix N h N hy N cont Each node in the graph has a feature N. h；i N hy；i N cont；i Each row is concatenated; N is randomly shuffled. h N hy N cont The rows yield the feature matrix Create three negative example graphs without changing the adjacency matrix: a negative example protein hydrophobic graph. Negative example protein hydrogen bond diagram Negative example protein contact diagram To distinguish it from the graph structure of negative samples, it is called G h G hy G cont These are examples of hydrophobic diagrams, hydrogen bond diagrams, and contact diagrams for positive protein samples.

[0102] Hydrophobic diagram G of the positive example protein in step a2 h Example: Protein hydrogen bond diagram G hy Positive example protein contact diagram G cont Negative example protein hydrophobic diagram Negative example protein hydrogen bond diagram Negative example protein contact diagram 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 Pconsc4 prediction algorithm to obtain protein amino acid contact distance data.

[0104] Step a22: Under the normal amino acid characteristic sequence, construct positive protein hydrophobic map, positive protein hydrogen bond map, and positive protein contact map using amino acids as nodes and hydrophobic interactions, hydrogen bonds, and contact distances between amino acids as three different types of edge sets.

[0105] Step a23: With the amino acid characteristic sequence shuffled, construct negative protein hydrophobic map, negative protein hydrogen bond map, and negative protein contact map 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 process the drug atom graph G corresponding to each drug. d Atomic and chemical bond features are extracted. 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 the concatenation method are used to process atomic and chemical bond features to obtain global drug features.

[0108] Step b11: Utilize EGAT directly through the neighboring atom feature N d；j and the characteristics of the associated chemical bonds Atomic features U obtained from the updated drug atomic diagram 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 Here are the learnable weight parameters, || represents the concatenation operation, LeakyReLU(·) and σ(·) are non-linear activation functions, and N... d；i Drug atomic diagram G d The i-th initial atomic feature, U d；i Drug atomic diagram G d The feature of the i-th atom updated by the EGAT algorithm, ξ d；ij Drug atomic diagram G d The initial characteristics of the chemical bond connecting the i-th atom and the j-th atom.

[0112] Step b12: Update atomic features NF using chemical bond features via CensNet d Simultaneously, the chemical bond feature EF is updated using atomic features. d :

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

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

[0115] Among them, U d Drug atomic diagram G d All atomic features U after EGAT update d；i The feature set concatenated by row, ξ d Drug atomic diagram G d All chemical bond characteristics ξ d；ij The feature set obtained by concatenating rows. W e W d These are learnable weight parameters; P represents the correlation matrix between chemical bonds and atoms, where elements P i”j” =1 indicates that there is a correlation between the i”th chemical bond and the j”th atom; I is the identity matrix; T is the transpose operation; Λ is the matrix diagonal operation; NF d EF represents the set of atomic features updated by the CensNet algorithm. d This represents the set of chemical bond features after being updated by the CensNet algorithm.

[0116] Step b13: Process atomic features and chemical bond features using average pooling algorithm and concatenation method to obtain the global drug feature (GP). d :

[0117]

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

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

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

[0121]

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

[0123] C protein feature extraction unit:

[0124] Step c1: Use three Graph Attention Networks (GATs) to process the hydrophobic map G of the positive protein example. h Example: Protein hydrogen bond diagram G hy Positive example protein contact diagram G cont Extracting positive amino acid features corresponding to the three protein images; the specific process includes:

[0125] Step c11: Use GAT to process G h Extracting positive amino acid characteristics:

[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 For learnable weight parameters, NF h；i For G h The i-th node feature extracted by GAT represents the feature of the i-th amino acid updated under hydrophobic interaction.

[0130] Step c12: Use GAT to process G hy Extracting positive amino acid features (NF) hy；j =GAT(N hy；j The extraction process is the same as step c11; among which, NF hy；j For G hy The characteristics of the j-th amino acid after GAT extraction.

[0131] Step c13: Use GAT on G cont Extracting positive amino acid features (NF) cont；k =GAT(N cont；k The extraction process is the same as step c11; among which, NF cont；k For G cont The characteristics of the kth amino acid after GAT extraction.

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

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

[0134]

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

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

[0137]

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

[0139]

[0140] Step c3: Use three GAT networks to plot the hydrophobicity of negative example proteins. Negative example protein hydrogen bond diagram Negative example protein contact diagram Extracting the amino acid negative features corresponding to the three protein maps; the specific process includes:

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

[0142]

[0143] in, The feature of the i-th amino acid in the hydrophobic map of a negative example protein.

[0144] Step c32: Extract negative amino acid features from the protein hydrogen bond map using GAT:

[0145]

[0146] in, This is a feature of the j-th amino acid in the hydrogen bond diagram of a negative example protein.

[0147] Step c33: Extract amino acid negative features from protein contact maps using GAT:

[0148]

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

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

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

[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 This represents the fusion feature of the i-th amino acid feature in the three positive examples of the protein.

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

[0156]

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

[0158] Step c5: Process the amino acid fusion positive example features using the average pooling algorithm to obtain the amino acid positive example fusion pooling features.

[0159] Among them, PF fusion This indicates the positive example fusion pooling characteristics of amino acids.

[0160] D-subgraph feature processing unit:

[0161] Step d1: Randomly extract a list of nodes from the drug atom graph with a fixed probability, and use the k-hop algorithm to extract nodes from the drug atom graph G. d The process involves extracting drug subgraphs centered on nodes from a node list, applying average pooling to the node feature set of each drug subgraph to obtain corresponding drug subgraph pooling features, and then concatenating all drug subgraph pooling features into a total drug subgraph pooling feature. The specific process includes:

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

[0163] Step d12: Using a node in the node list as the center node, extract drug subgraphs from the drug atomic graph using the k-hop algorithm, and apply average pooling to each drug subgraph to obtain the corresponding drug subgraph pooling features:

[0164]

[0165] Here, k-hop(·) is the k-hop subgraph algorithm, which returns a subgraph starting from node n. rThe feature set of the central node and its e-order neighbor nodes for The pooling feature of the r-th drug subgraph obtained after average pooling, where Z0 is the number of nodes in the r-th drug subgraph, SubP d is the set of pooling features of drug subgraphs, and s is the number of subgraphs of the drug atom graph.

[0166] Step d2: Randomly extract node lists from the protein hydrophobic map, positive protein hydrogen bond map, and positive protein contact map with fixed probabilities. Use the k-hop algorithm to extract protein sub-graphs from these maps, and apply average pooling to the node feature set of each sub-graph to obtain the corresponding protein sub-graph pooling features. Then, concatenate all protein sub-graph pooling features to form the total protein sub-graph pooling features. The specific process includes:

[0167] Step d21: Randomly select nodes from the complete set of nodes of positive protein hydrophobic graph, positive protein hydrogen bond graph, and positive protein contact graph with a fixed probability, and form corresponding node lists for each.

[0168] Step d22: Use the k-hop algorithm to extract protein sub-maps from the positive protein hydrophobic map, positive protein hydrogen bond map, and positive protein contact map, and use the average pooling algorithm for each protein sub-map to obtain the corresponding protein sub-map pooling features:

[0169]

[0170]

[0171] Where Z1 is the number of nodes in the r-th subgraph of the hydrogen bond graph of the positive example protein, Z2 is the number of nodes in the r-th subgraph of the hydrophobic graph of the positive example protein, and Z3 is the number of nodes in the r-th subgraph of the contact graph of the positive example protein.

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

[0173]

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

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

[0176]

[0177] Here, z actually refers to the subgraph index of the drug atom diagram, with a value range of [1, s]; SubP d；z It is SubP d The z-th row; 'a' actually refers to the subgraph index of the protein. Since it contains three graph structures, the value of 'a' ranges from [1, 3s]. SubP p；a For SubP p The a-th row. CrossAtten is a learnable weight parameter. za This represents the cross-attention coefficient between the z-th subgraph of the drug atomic map 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 , AddAtten is a learnable weight parameter. za This represents the additive attention coefficient between the z-th subgraph of the drug atomic map and the a-th 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 For weight parameters; Atten za This represents the interaction attention coefficient between the z-th subgraph of the drug atomic diagram and the a-th subgraph of the protein.

[0184] Step d4: Calculate the final drug subgraph cross-fusion features by combining the total pooling features of the drug subgraph from step d1 and the interaction attention coefficients from step d3 using matrix multiplication. Similarly, calculate the final protein subgraph cross-fusion features by combining the total pooling features of the protein subgraph and the interaction attention coefficients using matrix multiplication; the specific process includes:

[0185] Step d41: Calculate the drug subgraph cross-fusion features that fuse protein information using matrix multiplication on the total pooling features of the drug subgraph from step d1 and the interaction attention coefficients from step d3.

[0186] SubFu d；z =Attenza SubP d；z

[0187] Step d42: Calculate the cross-fusion features of the protein subgraph based on the pooling total features and interaction attention coefficients of the protein subgraph using matrix multiplication:

[0188]

[0189] Step d5: Process the drug subgraph cross-fusion features and protein subgraph cross-fusion features from step d4 using the average pooling algorithm to obtain the drug subgraph cross-fusion pooling features and protein subgraph cross-fusion pooling features respectively.

[0190]

[0191] E-affinity prediction unit:

[0192] Step e1: Process the global drug features from step b2 and the drug subgraph cross-fusion pooling features from step d5 using a concatenation method to obtain the drug features:

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

[0194] Step e2: Process the amino acid positive example fusion pooling feature of c5 and the protein subgraph cross-fusion pooling feature of d5 using a splicing method to obtain the protein features:

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

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

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

[0198] Step e4: For F conbind The affinity value is predicted using machine learning techniques.

[0199] In this embodiment, machine learning technology is implemented using neural networks. In fact, any neural network can be used for prediction; this embodiment uses a fully connected neural network for specific predictions. Specifically, a three-layer fully connected neural network (FC) is used to process the drug-protein joint features to obtain the affinity prediction value.

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

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

[0202] Error calculation is performed, and the entire model is trained. The process of calculating errors to guide training includes:

[0203] (1) The mutual information of amino acid positive features, amino acid negative features, and amino acid positive pooling features for each graph structure of the protein is calculated using a bilinear function, and the training objective for calculating the mutual information is calculated using cross-entropy maximization, including:

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

[0205]

[0206] in, Let W represent a bilinear function, where A and B are replaceable parameters. a Represents a learnable matrix; It represents the 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 former to have the maximum mutual information and the latter to have the minimum mutual information.

[0208]

[0209] (13) The mutual information of amino acid positive example features, amino acid positive example pooling features and amino acid negative example features after GAT update, and the mutual information of amino acid negative example features, amino acid positive example pooling features and input amino acid positive example features after GAT update are used as a set of cross-entropy calculation targets, requiring the former to have the maximum mutual information and the latter to have the minimum mutual information.

[0210]

[0211] Where A', B', and C are replaceable parameters, W b W c Represents a learnable matrix;

[0212] (2) The mutual information of amino acid positive example features, amino acid positive example fusion features, amino acid negative example features, amino acid negative example fusion features, and amino acid positive example fusion pooling features is calculated using a bilinear function, including:

[0213] (21) The mutual information between amino acid positive example fusion features and amino acid positive example pooling features, and the mutual information between amino acid negative example fusion features and amino acid positive example pooling features are used as a set of cross-entropy calculation targets, requiring the former to have the maximum mutual information and the latter to have the minimum mutual information.

[0214]

[0215] (22) Mutual information between positive amino acid fusion features and positive amino acid features, and mutual information between positive amino acid fusion features and negative amino acid features are used as a set of cross-entropy calculation targets, requiring the former to have the maximum mutual information and the latter to have the minimum mutual information.

[0216]

[0217] (23) Mutual information of amino acid positive example features, amino acid positive example fusion pooling features and amino acid fusion negative example features. The mutual information of amino acid positive example features, amino acid positive example fusion pooling features and amino acid fusion positive example features is used as a set of cross-entropy calculation targets for discriminator loss. The former is required to have the maximum mutual information and the latter to have the minimum mutual information.

[0218]

[0219] (3) Calculate the distance between the predicted affinity value and the label value using the mean square error function, and require the distance to be minimized (MSE loss to be minimized);

[0220]

[0221] Based on the losses in the above three aspects, parameters are used to coordinate the weight of each computational objective in the final training objective;

[0222]

[0223] in, To adjust the weights.

[0224] The overall network model is obtained after training with the loss function described above. This network model is then used for actual affinity prediction.

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

[0226] The affinity prediction effect of the present invention will be verified through the following examples. In the evaluation, the root mean squared error (RMSE), mean squared error (MSE), concordance index (CI), Pearson correlation coefficient, and adjusted R-squared (RM2) index were used as evaluation indicators.

[0227]

[0228]

[0229] Example:

[0230] To evaluate the impact of chemical bonds on prediction performance, this example establishes network models with and without chemical bonds, calculates the corresponding MSE, RMSE, CI, Person, and RM2 values, and compares them. See Tables 1-2 for details. Figures 2-5 As shown, the model performance is optimal when chemical bonds are present, indicating that chemical bonds improve the extraction of hidden features of drugs.

[0231] To evaluate the superiority of this application, comparative experiments were conducted on the Davis and KIBA datasets to compare it with other state-of-the-art drug target affinity prediction methods, including DeepDTA, GraphDTA, WideDTA, DeepDTAF, AttentionDTA, FusionDTA, DeepCDA, DGraphDTA, and GraphCL-DTA. The results are shown in Table 1. Furthermore, the application performs well on both the Davis and KIBA datasets, demonstrating its robustness and generalization ability.

[0232] Table 1. Performance comparison results of independent tests on Davis and KIBA datasets.

[0233]

[0234]

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

[0236]

[0237]

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

[0239] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for predicting drug target affinity based on graph neural networks, characterized in that, The drug target affinity prediction model is used to predict the affinity of drug targets. 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 data acquisition unit's processing includes: Step a1: For each drug, obtain the drug's atomic features and chemical bond features. Using the atoms contained in the drug as nodes, the atomic features as node features, and the chemical bond features as edges, a corresponding drug atomic graph G is constructed. d ; Step a2: For protein amino acids, obtain the hydrophobic interactions and hydrogen bonds of the protein. Construct a protein hydrophobic graph G using the amino acids contained in the protein as nodes and the hydrophobic interactions between amino acids as edges. h A protein hydrogen bond graph G is constructed using amino acids as nodes and hydrogen bonds between amino acids as edges. hy For protein amino acids, the Euclidean spatial distance data of each amino acid in the protein is obtained. Using the amino acids in the protein as nodes and the Euclidean spatial distance between amino acids less than a distance threshold as edges, a protein contact graph G is constructed. 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 process the drug atom graph G corresponding to each drug. d Extracting Atomic Features (NF) d and chemical bond characteristics EF d The atomic pooling feature NP is obtained using the average pooling algorithm. d and chemical bond pooling characteristics EP d The global drug characteristics (GP) are obtained by splicing. d =NP d ||EP d ; Step b2: Process global drug features (GP) based on feature fusion algorithm d and atomic features NF d The final drug global feature FuNP is obtained by fusing node-level atomic features and graph-level drug global features and then pooling them. d ; The processing of the protein feature extraction unit includes: Step c1: Use three GAT networks to analyze the hydrophobicity map G of positive protein samples. h Example: Protein hydrogen bond diagram G hy Positive example protein contact diagram G cont Extract the positive amino acid features (NF) corresponding to the three protein graphs. h；i NF hy；j NF cont；k ; Step c4: Use a feature fusion algorithm to fuse the three amino acid positive example features in c1 to obtain the amino acid fused positive example feature NF. fusion；i NF fusion；i This represents the fusion feature of the i-th amino acid in the three positive protein images; Step c5: Process the amino acid fusion positive example features using the average pooling algorithm to obtain the amino acid positive example fusion pooling features. Among them, PF fusion This indicates the positive example fusion pooling characteristics of amino acids; The subgraph feature processing unit includes the following: 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 Drug subgraphs are extracted from the node list, with each node as the center node. For each drug subgraph, the average pooling algorithm is used to obtain the corresponding drug subgraph pooling features. Then, all drug subgraph pooling features are concatenated to form the total drug subgraph pooling feature, SubP. d ; Step d2: Randomly extract nodes from the protein hydrophobic map, positive protein hydrogen bond map, and positive protein contact map with fixed probabilities, forming corresponding node lists for each. Use the k-hop algorithm to extract protein subgraphs from each of the three maps, and apply average pooling to the node feature set of each subgraph to obtain the corresponding protein subgraph pooling features. Then, concatenate all protein subgraph pooling features to form the total protein subgraph pooling feature SubP. p ; Step d3: Calculate the interaction attention coefficient (Atten) for each pair of drug subgraphs and protein subgraphs using additive attention and cross-attention algorithms on the pooling features of the drug subgraph and the total pooling features of the protein subgraph. za ; Step d4: Calculate the final drug subgraph cross-fusion feature (SubFu) of fused protein information using matrix multiplication on the total pooling features of the drug subgraph from step d1 and the interaction attention coefficients from step d3. d；z The final fused compound information protein subgraph cross-fusion feature (SubFu) is calculated by using matrix multiplication on the total pooling features of the protein subgraph in step d2 and the interaction attention coefficients in step d3. P；a ; Step d5: Process the drug subgraph cross-fusion features (SubFu) from step d4 using the average pooling algorithm. d；z Protein subgraph cross-fusion feature SubFu P；a Obtain 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: Process the global drug features from step b2 and the drug subgraph cross-fusion pooling features from step d5 using a concatenation method to obtain drug features F. drug =FuNP d ||SubPFu d ; Step e2: Process the amino acid positive example fusion pooling feature of c5 and the protein subgraph cross-fusion pooling feature of d5 using a splicing method to obtain protein feature F. protein =PF fusion ||SubPFu p ; Step e3: Use a splicing method to splice drug features and protein features to obtain drug-protein combined features F. conbind =F drug ||F protein ; Step e4: For F conbind The affinity value is predicted using machine learning techniques.

2. The method for predicting drug target affinity based on graph neural networks according to claim 1, characterized in that, Step b1 describes using the edge aggregation graph attention network algorithm EGAT and the co-embedding graph neural network algorithm CensNet to process the drug atom graph G corresponding to each drug. d Extracting Atomic Features (NF) d and chemical bond characteristics EF d The specific process includes: Step b11: Utilize EGAT directly through the neighboring atom feature N d；j and the characteristics of the associated chemical bonds Atomic features U obtained from the updated drug atomic diagram d；i ; e 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 Here, || represents the concatenation operation, LeakyReLU(·) and σ(·) are non-linear activation functions, and N is the weight parameter. d；i Drug atomic diagram G d The i-th initial atomic feature, U d；i Drug atomic diagram G d The feature of the i-th atom updated by the EGAT algorithm, ξ d；ij Drug atomic diagram G d Initial characteristics of the chemical bond connecting the i-th atom and the j-th atom; Step b12: Update atomic features NF using chemical bond features via CensNet d Simultaneously, the chemical bond feature EF is updated using atomic features. 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 Drug atomic diagram G d All atomic features U after EGAT update d；i The feature set concatenated by row, ξ d Drug atomic diagram G d All chemical bond characteristics ξ d；ij The feature set obtained by concatenating rows; W e W d The weighting parameter is P; P represents the correlation matrix between chemical bonds and atoms, where 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 identity matrix; T is the transpose operation; Λ is the matrix diagonal operation; NF d EF represents the set of atomic features updated by the CensNet algorithm. d This represents the set of chemical bond features after being updated by the CensNet algorithm.

3. The method for predicting drug target affinity based on graph neural networks according to claim 2, characterized in that, Step b2 describes the process of fusing node-level atomic features and graph-level global drug features and then pooling them to obtain the final global drug features. Among them, NF d；i For NF d Atomic features in FuNP d；i This represents the i-th final atomic feature obtained by fusing the global features of the drug. FuNP represents the square of the 2-norm of the global characteristics of a drug. d This represents the global characteristics of the drug obtained after pooling.

4. The method for predicting drug target affinity based on graph neural networks according to claim 1, characterized in that, In step c1, three graph attention networks (GATs) are used to analyze the hydrophobicity graphs of positive protein samples. h Example: Protein hydrogen bond diagram G hy Positive example protein contact diagram G cont The specific process of extracting the positive amino acid features corresponding to the three protein graphs includes: Step c11: Use GAT to process G h Extracting positive amino acid characteristics: e 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 For weight parameters, NF h；i For G h The i-th node feature extracted by GAT represents the feature of the i-th amino acid updated under hydrophobic interaction force; Step c12: Use GAT to process G hy Extracting positive amino acid features (NF) hy；j =GAT(N hy；j The extraction process is the same as step c11; among which, NF hy；j For G hy The characteristics of the j-th amino acid after GAT extraction; Step c13: Use GAT to process G cont Extracting positive amino acid features (NF) cont；k =GAT(N cont；k The extraction process is the same as step c11; among which, NF cont；k For G cont The characteristics of the kth amino acid after GAT extraction.

5. The method for predicting drug target affinity based on graph neural networks according to claim 1, characterized in that, In step c4, a feature fusion algorithm is used to fuse the three amino acid positive example features from c1 to obtain the amino acid fused positive example features: 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 This represents the fusion feature of the i-th amino acid feature in the three positive examples of the protein.

6. The method for predicting drug target affinity based on graph neural networks according to claim 1, characterized in that, In step d1, the k-hop algorithm is used to extract the drug atomic diagram G. d The specific process of extracting drug subgraphs centered on nodes in the node list, applying average pooling to each drug subgraph to obtain corresponding drug subgraph pooling features, and then concatenating all drug subgraph pooling features to form the total drug subgraph pooling features includes: Using nodes in the node list as center nodes, the k-hop algorithm is used to extract drug subgraphs from the drug atomic graph, and the average pooling algorithm is applied to each drug subgraph to obtain the corresponding drug subgraph pooling features: Here, k-hop(·) is the k-hop subgraph algorithm, which returns a subgraph starting from node n. r The feature set of the central node and its e-order neighbor nodes for The pooling feature of the r-th drug subgraph obtained after average pooling, where Z0 is the number of nodes in the r-th drug subgraph, SubP d It is a set of pooling features for drug subgraphs.

7. The method for predicting drug target affinity based on graph neural networks according to claim 6, characterized in that, Step d2 involves extracting protein sub-maps from the positive protein hydrophobic map, positive protein hydrogen bond map, and positive protein contact map using the k-hop algorithm, and then applying the average pooling algorithm to each protein sub-map to obtain the corresponding protein sub-map pooling features. The specific process includes the following steps: The k-hop algorithm is used to extract protein sub-maps from the hydrophobic maps, hydrogen bond maps, and contact maps of positive protein samples. For each protein sub-map, the average pooling algorithm is applied to obtain the corresponding protein sub-map pooling features. Where Z1 is the number of nodes in the r-th subgraph of the hydrogen bond graph of the positive example protein, Z2 is the number of nodes in the r-th subgraph of the hydrophobic graph of the positive example protein, and Z3 is the number of nodes in the r-th subgraph of the contact graph of the positive example protein.

8. The method for predicting drug target affinity based on graph neural networks according to claim 1, characterized in that, Step d3 describes using additive attention and cross-attention algorithms to calculate the interaction attention coefficient (Atten) for each pair of drug subgraphs and protein subgraphs based on the pooling features of the drug subgraph and the total pooling features of the protein subgraph. za The process includes: Step d31: Calculate the cross-attention coefficients using the cross-attention algorithm: Here, 'z' actually refers to the subgraph index of the drug atom diagram; SubP d；z It is SubP d The z-th line; 'a' actually refers to the protein subgraph index; SubP p；a For SubP p The a-th row; CrossAtten is a weighting parameter. za The cross-attention coefficient represents the relationship between the z-th subgraph of the drug atomic map and the a-th subgraph of the protein. Step d32: Calculate the additive attention coefficient using the additive attention algorithm: Among them, W O , For weight parameters, AddAtten za The additive attention coefficient represents the relationship between the z-th subgraph of the drug atomic map and the a-th 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 For weight parameters; Atten za This represents the interaction attention coefficient between the z-th subgraph of the drug atomic map and the a-th subgraph of the protein.

9. A method for predicting drug target affinity based on graph neural networks according to any one of claims 1 to 8, characterized in that, The drug target affinity prediction model is pre-trained. During the training process of the drug target affinity prediction model, the data acquisition unit also employs a negative sampling strategy for G during step a2. h G hy G cont Generate corresponding negative example graphs The processing steps of the protein feature extraction unit also include: Step c2: Use the average pooling algorithm to process the three amino acid positive feature samples corresponding to the three protein images extracted in c1, and obtain the corresponding three amino acid positive feature pooling feature PF. h PF hy PF cont ; Step c3: Use three GAT networks to plot the hydrophobicity of negative example proteins. Negative example protein hydrogen bond diagram Negative example protein contact diagram Extract the amino acid negative features corresponding to the three protein graphs. Simultaneously, in step c4, a feature fusion algorithm is used to fuse the three amino acid negative example features from c3 to obtain the amino acid fusion negative example features. in, This represents the fusion feature of the i-th amino acid feature in the three negative example graphs of a protein; Based on the processing of the drug target affinity prediction model, the total loss function Loss is used to train the drug target affinity prediction model. in, To adjust the weights, the sub-losses are as follows: in, In this context, A and B represent the objects being calculated, and W... a Represents a learnable matrix; Represents the mathematical expectation; In this context, A', B', and C represent the computational objects, and W... b W c This represents a learnable matrix.

10. The method for predicting drug target affinity based on graph neural networks according to claim 9, characterized in that, In step a2, a negative sampling strategy is used for G. h G hy G cont Generate corresponding negative example graphs The specific process includes: G h G hy G cont The node feature matrix N h N hy N cont Each node in the graph has a feature N. h；i N hy；i N cont；i Each row is concatenated; N is randomly shuffled. h N hy N cont The rows yield the feature matrix Create three negative example graphs without changing the adjacency matrix: a negative example protein hydrophobic graph. Negative example protein hydrogen bond diagram Negative example protein contact diagram

Images