A drug-target interaction prediction method based on multi-domain heterogeneous graph aggregation learning

Abstract

This invention discloses a drug-target interaction prediction method based on multivariate domain heterogeneous graph aggregation learning. The method is broadly divided into the following steps: calculating drug or target similarity information to obtain initial features; defining metapaths and metadomain graphs; constructing a heterogeneous graph using heterogeneous bioinformatics network data; using a lightweight feature aggregator to pre-calculate the information aggregation of each neighbor node in the heterogeneous graph; using a feature projection step to convert the dimensions of the feature aggregation vector to be consistent; using a multivariate domain graph semantic feature fusion module to aggregate metadomain graph-level semantic information; using contrastive learning to perform feature optimization to obtain final node features; and inputting the final node features into a multilayer perceptron to predict drug-target interaction scores. This invention utilizes a simple and efficient training process to model complex biological entity relationships and introduces a lightweight metapath-level feature aggregator to address the excessive complexity and computational effort caused by repeated neighbor information aggregation during the training cycle.

Description

Technical Field

[0001] The present invention relates to the field of biological information processing, and in particular to a drug-target interaction prediction method based on multi-domain heterogeneous graph aggregation learning. Background Art

[0002] Drug development for specific diseases is shifting from a "one drug for multiple diseases" model to a "one disease, one drug" model, with an emphasis on optimizing drug selectivity and efficacy. However, failures in drug development due to incorrect target selection are common. High-precision prediction of drug-target interactions (DTIs) can significantly narrow the search scope for candidate drugs for downstream drug discovery experiments and reduce the economic and time losses caused by failed clinical trials. Since identifying drug-target interactions using biological experiments is time-consuming and expensive, it is necessary to use computational methods to determine potential interactions between drugs and targets.

[0003] In recent years, the development of artificial intelligence and machine learning technologies has provided new perspectives for predicting potential interactions between drugs and targets. With the explosive growth of medical data and advances in deep learning technology, heterogeneous biological data offers a multifaceted perspective for modeling relationships between drugs and targets. However, these high-dimensional datasets pose significant challenges to researchers due to their high dimensionality, complex structure, and unique types. Traditional computational methods face challenges in processing these complex data.

[0004] Due to the superiority of Graph Neural Networks (GNNs) in graph analysis, GNN-based methods are used to predict drug-target interactions using heterogeneous biological data. Their characteristic is that they utilize multi-dimensional information of biological entities such as drugs, targets, drug side effects, and diseases as network input features, which are further converted into a set of feature matrices that can be processed by deep learning models to achieve the prediction of drug-target interactions. However, these methods only consider the drug-target related network and ignore the deep semantic relationships that are integrated across networks. At the same time, these methods use a large number of redundant modules to embed semantic information such as node types and edge types into the propagated messages, resulting in excessive complexity and computational effort caused by repeated aggregation of neighbor information in each training cycle, hindering the model from achieving better performance. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention provides a drug-target interaction prediction method based on multi-domain heterogeneous graph aggregation learning with efficient algorithm, high prediction accuracy and strong robustness.

[0006] The technical solution of the present invention to solve the above problems is: a drug-target interaction prediction method based on multi-domain heterogeneous graph aggregation learning, comprising the following steps:

[0007] 1) Calculate the similarity information of drugs or targets to obtain the initial feature matrix;

[0008] 2) Define metapath and metadomain graph;

[0009] 3) constructing a heterogeneous graph using known drug-disease associations, known drug-side effect associations, known target-disease associations, known drug-target associations, drug-drug interactions, target-target interactions, and the initial feature matrix obtained in step 1);

[0010] 4) Using a lightweight meta-path-level feature aggregator to pre-compute the meta-path-level information aggregation of each neighbor node in the biological heterogeneous graph;

[0011] 5) Using the feature projection step to transform the feature aggregation vectors of different dimensions into feature vectors with consistent dimensions;

[0012] 6) Use the Transformer-based multi-domain graph semantic feature fusion module to aggregate the meta-domain graph-level semantic information and obtain the feature matrix before comparison optimization;

[0013] 7) Use contrastive learning to optimize features and obtain the final node features;

[0014] 8) The final node features are input into a multi-layer perceptron to predict the drug-target interaction score matrix. The above-mentioned drug-target interaction prediction method based on multi-domain heterogeneous graph aggregation learning, the step 1) specifically includes the following steps:

[0015] 1-1) For drug nodes, the topological fingerprint of each drug is calculated using the RDKit software tool, and then the structural similarity between drugs is calculated using the Tanimoto method, and finally the drug structure similarity matrix is ​​obtained. Where d represents the drug;

[0016] 1-2) For target nodes, the sequence similarity between the two targets is calculated using the Smith-Waterman score based on their primary sequences, and the target sequence similarity matrix is ​​finally obtained. Where t represents the drug.

[0017] In the above-mentioned drug-target interaction prediction method based on multi-domain heterogeneous graph aggregation learning, the specific process of step 2) is as follows:

[0018] 2-1) Define meta-path. Meta-path is defined as a combination of several edge types. εi ∈ε respectively represent the i-th node type and i-th edge type, i represents the i-th node, and given the meta-path A meta path instance is a sequence of nodes that follows Definition:

[0019]

[0020] in is the starting node, for One of the meta-path based neighbors of Indicates the first node type Start to the lth node type A meta-path that ends with a node, describing the node type and The l-1 order composite neighbor relationship ε between them is:

[0021] ε=ε1∧ε2∧...∧ε l-1 (2)

[0022] Among them, ∧ represents the composite relation operation, ε l-1 Indicates the type of the l-1th edge;

[0023] 2-2) Define the metadomain graph, given a metapath and a with edge type ε l-1 Node in Indicates the node type, n indicates the number of nodes on this meta-path, Indicates the type of the nth node, metadomain graph Indicates that the meta path The domain, intermediate nodes and nodes on the meta-path A directed graph composed of .

[0024] In the above-mentioned drug-target interaction prediction method based on multi-domain heterogeneous graph aggregation learning, the step 3) includes the following steps:

[0025] 3-1) Collect data sets for drug-target interaction prediction research, including 708 drugs, 4192 side effects, 5603 diseases and 1512 targets, as well as drug-disease known association matrices Drug-side effect known association matrix Target-disease known association matrix Drug-target known association matrix Drug-Drug Interaction Matrix and target-target interaction matrix where d, t, s, and I symbols represent drug, target, side effect, and disease bioentities, respectively, and dI, ds, tI, dt, dd, and tt represent drug-disease, drug-side effect, target-disease, drug-target, drug-drug, and target-target associations, respectively;

[0026] 3-2) Using the Deep Graph Learning Library (DGL Library) to build heterogeneous graphs Final heterogeneous graph The correlation matrix It can be defined as:

[0027]

[0028] in, and Respectively and Transpose of a matrix;

[0029] 3-3) The similarity matrix of drugs and the target similarity matrix As the chemical structure characteristics of drugs and sequence structure characteristics of targets, isomerization graph Initial node feature matrix of drugs and targets Can be initialized as:

[0030]

[0031] in It is a common symbol for expressing the dimension of the matrix, and N(d) and N(t) represent the number of drugs and targets, respectively.

[0032] In the above-mentioned drug-target interaction prediction method based on multi-domain heterogeneous graph aggregation learning, the process of step 4) is as follows:

[0033] 4-1) A lightweight meta-path-level feature aggregator is used to pre-calculate the meta-path-level information aggregation of each neighboring node in the biological heterogeneous graph. Taking the i-th drug node as an example, the aggregation process can be expressed as:

[0034]

[0035] Where d represents the drug, Represents the characteristic matrix of the i-th drug node after the above aggregation process, Indicates the i-th drug node in the meta-domain graph The feature matrix after information aggregation, It is the meta-domain graph of the given i-th drug node Set Φ d The generated list of different semantic feature matrices, corresponds to the metadomain graph The set of all meta-path instances in , It represents The number of meta-path instances, is a meta-path instance containing the target node i and the starting node j, is the neighbor node feature vector x on the meta-path j The attention weights of the target can be obtained by the same logic. Different semantic feature matrices correspond to different meta-domain graphs of the target. t represents the target;

[0036] 4-2) Use adjacency matrix multiplication to further simplify formula (5), and add variable-length meta-path instances to the meta-domain graph to enhance the perception domain of local node feature aggregation. Let is a l-1 length metapath, where d represents the drug, t1t2t3…t l-1 Represents a meta path The intermediate nodes on t l-1 Represents the l-1th intermediate node. Taking drug node d as an example, the specific simplification process is as follows:

[0037]

[0038] in Indicates that drug node d is in the meta-domain graph The feature matrix after the above simplified aggregation is: Represented in the metadomain graph The adjacency matrix between the drug node d and the intermediate node t1 in After row normalization, the matrices d-t1, t1-t2 and t l-1 -d represents drug node d-intermediate node t1, intermediate node t1-intermediate node t2 and intermediate node t l-1 - drug node d, Indicates that in the meta path The feature aggregation weight vector of the drug neighbor nodes on It is a common symbol for expressing the matrix dimension. N(d) and N(t) represent the number of drugs and targets respectively. ⊙ represents the product operation of vector elements. Represents the initial feature matrix set of all drug nodes. Similarly, the target node t in the meta-domain graph can be obtained The feature matrix after simplified aggregation

[0039] In the above-mentioned drug-target interaction prediction method based on multi-domain heterogeneous graph aggregation learning, the process of step 5) is as follows:

[0040] 5-1) When aggregating meta-domain graph-level information in the subsequent process, feature embedding vectors of different dimensions may bring challenges. The feature projection step converts the feature aggregation vectors of different dimensions into features with consistent dimensions. Use a multi-layer perceptron (MLP) module Where v represents the node type in the heterogeneous graph, and the multi-layer perception block contains a normalization layer, a nonlinear layer, and a dropout layer between two consecutive linear layers. As an example of the feature projection step, the specific projection process is as follows:

[0041]

[0042] in Indicates that drug node d is in the meta-domain graph The feature matrix after the simplified aggregation of formula (6) is: Represented in the metadomain graph Perform feature projection on the feature matrix ·, Indicates that drug node d is in the meta-domain graph The feature matrix after feature projection is the same as the drug domain graph The feature projection process is similar to that of the target node t. Perform feature projection to obtain the target's feature projection matrix

[0043] In the above-mentioned drug-target interaction prediction method based on multi-domain heterogeneous graph aggregation learning, the process of step 6) is as follows:

[0044] 6-1) Taking the drug node as an example, use the predefined drug different meta-domain graph list And the semantic vector set of each drug node after the feature projection step Where n represents the number of drug meta-domain graphs. The Transformer-based multi-domain graph semantic feature fusion module learns the degree of mutual influence between each pair of semantic vectors. The process of multi-domain graph semantic feature fusion can be expressed as:

[0045]

[0046] in and Representing the ith, jth and rth meta-domain graphs of drugs (or targets), each semantic vector Each maps a query vector A key vector and a vector of values Mutual attention weight coefficient is the query vector With key vector The dot product result after softmax normalization, and Respectively express and The transpose of, EXP(·) represents the exponential function with the natural constant e as the base, Represents a vector and Perform product operation, current semantics The output vector is a vector of all values Add the weighted sum of the remaining connections, The query vectors are Key vector Sum value vector The trainable shared parameters, θ is also a trainable parameter shared by all drug meta-domain graphs, and finally the feature fusion matrix list of each meta-domain graph will be obtained. Similarly, for the different meta-domain graph lists of target t Where m represents the number of meta-domain graphs of target t. Using the same Transformer-based multi-domain graph semantic feature fusion step, we can obtain the feature fusion matrix list of each meta-domain graph of target t.

[0047] 6-2) Perform weighted summation on the feature fusion matrix list of each meta-domain graph and obtain the drug feature matrix before comparison optimization and target feature matrix The process is as follows:

[0048]

[0049] in, Represents the computational domain graph and meta-domain diagram The feature fusion matrix and The trainable weight coefficients between them.

[0050] In the above-mentioned drug-target interaction prediction method based on multi-domain heterogeneous graph aggregation learning, the process of step 7) is as follows:

[0051] 7-1) The feature matrix of the drug node and the feature matrix of the target node A new feature matrix is ​​formed by concatenating Subsequently, feature optimization and update will be performed on the feature matrix, and a topological map will be constructed based on the original drug-target association map. and semantic graph In the topology diagram In the topology graph, if two drug-target pair nodes have the same drug or target pointing to these two nodes, they will be connected, thus constructing the topological graph adjacency matrix M p , for semantic graph By calculating the cosine similarity between drug-target pair nodes, sorting them, and connecting the drug-target pair nodes with their first k most similar nodes, the semantic graph adjacency matrix M is constructed. s Finally, two graph convolutional neural network (GCN) modules are used to extract the topological graph and semantic graph The node features of the enhanced view are learned and used for graph contrastive learning to supervise the optimization features. The contrastive learning optimization process is as follows:

[0052]

[0053] where I is the identity matrix, and Represented in the topology diagram and semantic graph The feature matrix extracted by the l+1th convolution kernel, D p and D s They are the topological graph adjacency matrix M p And the semantic graph adjacency matrix M s The degree matrix of p and D s The diagonals of are all 1, represents the regularization of the matrix, is the trainable weight matrix in the GCN module, and The feature matrices are The updated feature matrix from the topological map and semantic map, Relu(·) represents the activation function in the neural network;

[0054] 7-2) Use the InfoCNE method to segment negative samples and classify them into different categories. Then calculate the loss by evaluating the similarity between sample features. Then use the contrast loss calculated by InfoNCE to strengthen the consistency of model prediction and expand the distinction between positive and negative samples in the feature space. The contrast loss function for:

[0055]

[0056] Where N is the number of nodes in the original drug-target association graph, Z i It is a topological diagram Node i in the semantic graph The set of all positive samples in , exp(·) represents the exponential function with the natural constant e as the base, S(·,·) represents the function for calculating cosine similarity, and τ represents an adjustable scalar parameter. and Represent nodes i, j and i respectively ′ In the topology diagram and semantic graph The feature matrix extracted by the GCN module.

[0057] In the above-mentioned drug-target interaction prediction method based on multi-domain heterogeneous graph aggregation learning, the process of step 8) is as follows:

[0058] 8-1) will eventually and Input into the multi-layer perceptron (MLP), and use MLP to learn the feature interaction between drugs and target nodes to obtain the interaction prediction score matrix The MLP learning prediction process is as follows:

[0059]

[0060] Where ⊙ represents the Hadamard product, W,W 2 …W L and ε 1 ,ε 2 ,…ε L are all represented as parameter matrices that need to be trained. represents the interaction prediction score between drug d and target t;

[0061] 8-2) Use data labels to optimize model performance and use the cross entropy loss function to calculate classification loss. The cross entropy loss function The definition is as follows:

[0062]

[0063] Where N is the number of nodes in the original drug-target association graph, Y i represents the label of sample i, represents the predicted association probability of sample i, (d, t) represents the drug-target pair sample,

[0064] Loss function for final method training It consists of three parts, defined as follows:

[0065]

[0066] in is the contrastive loss function, is the classification loss calculated using the cross entropy loss function, λ is a hyperparameter between 0 and 1, and λ represents the ratio between the contrast loss and the classification loss. Represents L2 regularization of all trainable model parameters, and the parameter matrix of L2 regularization is Θ to prevent model overfitting.

[0067] The beneficial effects of the present invention are:

[0068] This method can model complex relationships between biological entities using a simple and efficient training process. It introduces a lightweight meta-path-level feature aggregator, resolving the problem of excessive complexity and computational effort caused by repeated aggregation of neighbor information in each training cycle. The Transformer-based multi-domain graph semantic feature fusion module improves the generalization performance of the method by fusing deep semantic relationships across multi-domain graphs and obtaining high-order associations between meta-domain graphs. It uses an attention mechanism to enhance the interpretability of the method, and finally employs contrastive learning to enhance the robustness of the method. This method has high accuracy and robustness, and has considerable market value in fields such as drug design and development and smart healthcare. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is the overall flow chart of the present invention. DETAILED DESCRIPTION

[0070] The present invention will be further described below with reference to the accompanying drawings and examples.

[0071] like Figure 1 As shown, a drug-target interaction prediction method based on multi-domain heterogeneous graph aggregation learning includes the following steps:

[0072] 1) Calculate the similarity information of drugs or targets to obtain the initial feature matrix.

[0073] Step 1) specifically includes the following steps:

[0074] 1-1) For drug nodes, the topological fingerprint of each drug is calculated using the RDKit software tool, and then the structural similarity between drugs is calculated using the Tanimoto method, and finally the drug structure similarity matrix is ​​obtained. Where d represents the drug;

[0075] 1-2) For target nodes, the sequence similarity between the two targets is calculated using the Smith-Waterman score based on their primary sequences, and the target sequence similarity matrix is ​​finally obtained. Where t represents the drug.

[0076] 2) Define the meta-path and meta-domain graph.

[0077] The specific process of step 2) is as follows:

[0078] 2-1) Define meta-path. Meta-path is defined as a combination of several edge types. ε i ∈ε respectively represent the i-th node type and i-th edge type, i represents the i-th node, and given the meta-path A meta path instance is a sequence of nodes that follows Definition:

[0079]

[0080] in is the starting node, for One of the meta-path based neighbors of Indicates the first node type Start to the lth node type A meta-path that ends with a node, describing the node type and The l-1 order composite neighbor relationship ε between them is:

[0081] ε=ε1∧ε2∧...∧ε l-1 (2)

[0082] Among them, ∧ represents the composite relation operation, ε l-1 Indicates the type of the l-1th edge;

[0083] 2-2) Define the metadomain graph, given a metapath and a with edge type ε l-1 Node in Indicates the node type, n indicates the number of nodes on this meta-path, Indicates the type of the nth node, metadomain graph Indicates that the meta path The domain, intermediate nodes and nodes on the meta-path A directed graph composed of .

[0084] 3) Constructing a heterogeneous graph using known drug-disease associations, known drug-side effect associations, known target-disease associations, known drug-target associations, drug-drug interactions, target-target interactions, and the initial feature matrix obtained in step 1).

[0085] Step 3) includes the following steps:

[0086] 3-1) Collect data sets for drug-target interaction prediction research, including 708 drugs, 4192 side effects, 5603 diseases and 1512 targets, as well as drug-disease known association matrices Drug-side effect known association matrix Target-disease known association matrix Drug-target known association matrix Drug-Drug Interaction Matrix and target-target interaction matrix where d, t, s, and I symbols represent drug, target, side effect, and disease bioentities, respectively, and dI, ds, tI, dt, dd, and tt represent drug-disease, drug-side effect, target-disease, drug-target, drug-drug, and target-target associations, respectively;

[0087] 3-2) Using the Deep Graph Learning Library (DGL Library) to build heterogeneous graphs Final heterogeneous graph The correlation matrix It can be defined as:

[0088]

[0089] in, and Respectively and Transpose of a matrix;

[0090] 3-3) The similarity matrix of drugs and the target similarity matrix As the chemical structure characteristics of drugs and sequence structure characteristics of targets, isomerization graph Initial node feature matrix of drugs and targets Can be initialized as:

[0091]

[0092] in It is a common symbol for expressing the dimension of the matrix, and N(d) and N(t) represent the number of drugs and targets, respectively.

[0093] 4) A lightweight meta-path-level feature aggregator is used to pre-compute the meta-path-level information aggregation of each neighbor node in the biological heterogeneous graph.

[0094] The process of step 4) is as follows:

[0095] 4-1) A lightweight meta-path-level feature aggregator is used to pre-calculate the meta-path-level information aggregation of each neighboring node in the biological heterogeneous graph. Taking the i-th drug node as an example, the aggregation process can be expressed as:

[0096]

[0097] Where d represents the drug, Represents the characteristic matrix of the i-th drug node after the above aggregation process, Indicates the i-th drug node in the meta-domain graph The feature matrix after information aggregation, It is the meta-domain graph of the given i-th drug node Set Φ d The generated list of different semantic feature matrices, corresponds to the metadomain graph The set of all meta-path instances in , It represents The number of meta-path instances, is a meta-path instance containing the target node i and the starting node j, is the neighbor node feature vector x on the meta-path j The attention weights of the target can be obtained by the same logic. Different semantic feature matrices correspond to different meta-domain graphs of the target. t represents target;

[0098] 4-2) Use adjacency matrix multiplication to further simplify formula (5), and add variable-length meta-path instances to the meta-domain graph to enhance the perception domain of local node feature aggregation. Let is a l-1 length metapath, where d represents the drug, t1t2t3…t l-1 Represents a meta path The intermediate nodes on t l-1 Represents the l-1th intermediate node. Taking drug node d as an example, the specific simplification process is as follows:

[0099]

[0100] in Indicates that drug node d is in the meta-domain graph The feature matrix after the above simplified aggregation is: Represented in the metadomain graph The adjacency matrix between the drug node d and the intermediate node t1 in After row normalization, the matrices d-t1, t1-t2 and t l-1-d represents drug node d-intermediate node t1, intermediate node t1-intermediate node t2 and intermediate node t l-1 - drug node d, Indicates that in the meta path The feature aggregation weight vector of the drug neighbor nodes on It is a common symbol for expressing the matrix dimension. N(d) and N(t) represent the number of drugs and targets respectively. ⊙ represents the product operation of vector elements. Represents the initial feature matrix set of all drug nodes. Similarly, the target node t in the meta-domain graph can be obtained The feature matrix after simplified aggregation

[0101] 5) A feature projection step is used to transform the feature aggregation vectors of different dimensions into feature vectors with consistent dimensions.

[0102] The process of step 5) is as follows:

[0103] 5-1) When aggregating meta-domain graph-level information in the subsequent process, feature embedding vectors of different dimensions may bring challenges. The feature projection step converts the feature aggregation vectors of different dimensions into features with consistent dimensions. Use a multi-layer perceptron (MLP) module Where v represents the node type in the heterogeneous graph, and the multi-layer perception block contains a normalization layer, a nonlinear layer, and a dropout layer between two consecutive linear layers. As an example of the feature projection step, the specific projection process is as follows:

[0104]

[0105] in Indicates that drug node d is in the meta-domain graph The feature matrix after the simplified aggregation of formula (6) is: Represented in the metadomain graph Perform feature projection on the feature matrix ·, Indicates that drug node d is in the meta-domain graph The feature matrix after feature projection is the same as the drug domain graph The feature projection process is similar to that of the target node t. Perform feature projection to obtain the target's feature projection matrix

[0106] 6) Use the Transformer-based multi-domain graph semantic feature fusion module to aggregate meta-domain graph-level semantic information and obtain the feature matrix before comparison optimization.

[0107] The process of step 6) is as follows:

[0108] 6-1) Taking the drug node as an example, use the predefined drug different meta-domain graph list And the semantic vector set of each drug node after the feature projection step Where n represents the number of drug meta-domain graphs. The Transformer-based multi-domain graph semantic feature fusion module learns the degree of mutual influence between each pair of semantic vectors. The process of multi-domain graph semantic feature fusion can be expressed as:

[0109]

[0110] in and Representing the ith, jth and rth meta-domain graphs of drugs (or targets), each semantic vector Each maps a query vector A key vector and a vector of values Mutual attention weight coefficient is the query vector With key vector The dot product result after softmax normalization, and Respectively express and The transpose of, EXP(·) represents the exponential function with the natural constant e as the base, Represents a vector and Perform product operation, current semantics The output vector is a vector of all values Add the weighted sum of the remaining connections, The query vectors are Key vector Sum value vector The trainable shared parameters, θ is also a trainable parameter shared by all drug meta-domain graphs, and finally the feature fusion matrix list of each meta-domain graph will be obtained. Similarly, for the different meta-domain graph lists of target t Where m represents the number of meta-domain graphs of target t. Using the same Transformer-based multi-domain graph semantic feature fusion step, we can obtain the feature fusion matrix list of each meta-domain graph of target t.

[0111] 6-2) Perform weighted summation on the feature fusion matrix list of each meta-domain graph and obtain the drug feature matrix before comparison optimization and target feature matrix The process is as follows:

[0112]

[0113] in, Represents the computational domain graph and meta-domain diagram The feature fusion matrix and The trainable weight coefficients between them.

[0114] 7) Use contrastive learning to optimize features and obtain the final node features.

[0115] The process of step 7) is as follows:

[0116] 7-1) The feature matrix of the drug node and the feature matrix of the target node A new feature matrix is ​​formed by concatenating Subsequently, feature optimization and update will be performed on the feature matrix, and a topological map will be constructed based on the original drug-target association map. and semantic graph In the topology diagram In the topology graph, if two drug-target pair nodes have the same drug or target pointing to these two nodes, they will be connected, thus constructing the topological graph adjacency matrix M p , for semantic graph By calculating the cosine similarity between drug-target pair nodes, sorting them, and connecting the drug-target pair nodes with their first k most similar nodes, the semantic graph adjacency matrix M is constructed. s Finally, two graph convolutional neural network (GCN) modules are used to extract the topological graph and semantic graph The node features of the enhanced view are learned and used for graph contrastive learning to supervise the optimization features. The contrastive learning optimization process is as follows:

[0117]

[0118] where I is the identity matrix, and Represented in the topology diagram and semantic graph The feature matrix extracted by the l+1th convolution kernel, D p and D s They are the topological graph adjacency matrix M pAnd the semantic graph adjacency matrix M s The degree matrix of p and D s The diagonals of are all 1, represents the regularization of the matrix, is the trainable weight matrix in the GCN module, and The feature matrices are The updated feature matrix from the topological map and semantic map, Relu(·) represents the activation function in the neural network;

[0119] 7-2) Use the InfoCNE method to segment negative samples and classify them into different categories. Then calculate the loss by evaluating the similarity between sample features. Then use the contrast loss calculated by InfoNCE to strengthen the consistency of model prediction and expand the distinction between positive and negative samples in the feature space. The contrast loss function for:

[0120]

[0121] Where N is the number of nodes in the original drug-target association graph, Z i It is a topological diagram Node i in the semantic graph The set of all positive samples in , exp(·) represents the exponential function with the natural constant e as the base, S(·,·) represents the function for calculating cosine similarity, and τ represents an adjustable scalar parameter. and Represent nodes i, j and i respectively ′ In the topology diagram and semantic graph The feature matrix extracted by the GCN module.

[0122] 8) The final node features are input into a multilayer perceptron to predict the drug-target interaction score matrix.

[0123] The process of step 8) is as follows:

[0124] 8-1) will eventually and Input into the multi-layer perceptron (MLP), and use MLP to learn the feature interaction between drugs and target nodes to obtain the interaction prediction score matrix The MLP learning prediction process is as follows:

[0125]

[0126] Where ⊙ represents the Hadamard product, W,W 2 …WL and ε 1 ,ε 2 ,…ε L are all represented as parameter matrices that need to be trained. represents the interaction prediction score between drug d and target t;

[0127] 8-2) Use data labels to optimize model performance and use the cross entropy loss function to calculate classification loss. The cross entropy loss function The definition is as follows:

[0128]

[0129] Where N is the number of nodes in the original drug-target association graph, Y i represents the label of sample i, represents the predicted association probability of sample i, (d, t) represents the drug-target pair sample,

[0130] Loss function for final method training It consists of three parts, defined as follows:

[0131]

[0132] in is the contrastive loss function, is the classification loss calculated using the cross entropy loss function, λ is a hyperparameter between 0 and 1, and λ represents the ratio between the contrast loss and the classification loss. Represents L2 regularization of all trainable model parameters, and the parameter matrix of L2 regularization is Θ to prevent model overfitting.

[0133] Examples

[0134] [1]Luo Y, Zhao X, Zhou J, et al. A network integration approach for drug-target interaction prediction and computational drug repositioning from heterogeneous information[J]. Nature communications, 2017, 8(1):573.

[0135] Reference [1] discloses a method for predicting drug-target interactions using randomized restart walks (RWR) and inductive matrix completion (IMC). This method extracts drug and target representations from a drug-target similarity network by using randomized restart walks with a dimensionality reduction scheme, and then uses inductive matrix completion to predict new drug-target interactions.

[0136] [2]Wan F, Hong L, Xiao A, et al. NeoDTI: neural integration of neighbor information from a heterogeneous network for discovering new drug–target interactions [J]. Bioinformatics, 2019, 35(1): 104-111.

[0137] Reference [2] discloses a method for predicting drug-target interactions based on graph convolutional neural networks (GCN) and inductive matrix completion (IMC). This method integrates drug or target representations learned from different relational networks using graph convolutional neural networks, and then uses inductive matrix completion to predict new drug-target interactions.

[0138] [3]Li J, Wang J, Lv H, et al. ICHGAN: inductive matrix completion with heterogeneous graph attention networks for drug-target interactions prediction [J]. IEEE / ACM Transactions on Computational Biology and Bioinformatics, 2021, 19(2): 655-665.

[0139] Reference [3] discloses a method for predicting drug-target interactions based on an attention mechanism and inductive matrix completion (IMC). This method uses a two-level attention mechanism to learn drug and target feature representations from a drug-target interaction heterogeneous network, and then inputs the learned features into inductive matrix completion to obtain the prediction results of drug-target interactions.

[0140] In order to verify the prediction accuracy of the method of the present invention, the performance of the method of the present invention was evaluated based on five indicators: the area under the receiver operating characteristic curve (AUC), the area under the precision-recall curve (AUPR), the F1 score (F1_score), the recall rate (Recall) and the precision (Precision). The method was compared with the above three documents. The comparison results are shown in Table 1:

[0141] Table 1 Comparison of performance indicators with other methods

[0142]

[0143] Conclusion: Compared with other methods for predicting drug-target interactions, the method of the present invention achieves the best performance in all performance indicators and has the advantages of high accuracy and strong reliability.

Claims

1. A drug-target interaction prediction method based on multi-domain heterogeneous graph aggregation learning, characterized by: The following steps are involved: 1) Calculate the similarity information of drugs or targets to obtain the initial feature matrix; 2) Define the meta-path and meta-domain graph. Given a meta-path and a with edge type ε l-1 Node in Indicates the node type, n indicates the number of nodes on this meta-path, Indicates the type of the nth node, and the meta-domain graph is represented by the meta-path based on The domain, intermediate nodes and nodes on the meta-path A directed graph composed of 3) constructing a heterogeneous graph using known drug-disease associations, known drug-side effect associations, known target-disease associations, known drug-target associations, drug-drug interactions, target-target interactions, and the initial feature matrix obtained in step 1); 4) A lightweight meta-path-level feature aggregator is used to pre-compute the meta-path-level information aggregation of each neighbor node in the biological heterogeneous graph. The lightweight meta-path-level feature aggregator is a simplified version of the heterogeneous graph neural network aggregator. This aggregator is a meta-path-level neighbor information aggregation module executed once in the preprocessing stage. Its function is to decouple multiple semantic information and generate node feature embeddings; 5) Using the feature projection step to transform the feature aggregation vectors of different dimensions into feature vectors with consistent dimensions; 6) Use the Transformer-based multi-domain graph semantic feature fusion module to aggregate the meta-domain graph-level semantic information and obtain the feature matrix before comparison optimization; 7) Use contrastive learning to optimize features and obtain the final node features; 8) The final node features are input into a multilayer perceptron to predict the drug-target interaction score matrix.

2. The drug-target interaction prediction method based on multi-domain heterogeneous graph aggregation learning according to claim 1 is characterized in that: The step 1) specifically includes the following steps: 1-1) For drug nodes, the topological fingerprint of each drug is calculated using the RDKit software tool, and then the structural similarity between drugs is calculated using the Tanimoto method, and finally the drug structure similarity matrix is ​​obtained. Where d represents the drug; 1-2) For target nodes, the sequence similarity between the two targets is calculated using the Smith-Waterman score based on their primary sequences, and the target sequence similarity matrix is ​​finally obtained. Where t represents the drug.

3. The drug-target interaction prediction method based on multi-domain heterogeneous graph aggregation learning according to claim 2 is characterized in that: The specific process of step 2) is as follows: 2-1) Define meta-path. Meta-path is defined as a combination of several edge types. ε i ∈ε respectively represent the i-th node type and i-th edge type, i represents the i-th node, and given the meta-path A meta path instance is a sequence of nodes that follows Definition: in is the starting node, for One of the meta-path based neighbors of Indicates the first node type Start to the lth node type A meta-path that ends with a node, describing the node type and The l-1 order composite neighbor relationship ε between them is: ε=ε1∧ε2∧...∧ε l-1 (2) Among them, ∧ represents the composite relation operation, ε l-1 Indicates the type of the l-1th edge; 2-2) Define the metadomain graph, given a metapath and a with edge type ε l-1 Node in Indicates the node type, n indicates the number of nodes on this meta-path, Indicates the type of the nth node, metadomain graph Indicates that the meta path The domain, intermediate nodes and nodes on the meta-path A directed graph composed of .

4. The drug-target interaction prediction method based on multi-domain heterogeneous graph aggregation learning according to claim 3 is characterized in that: The specific process of step 3) is as follows: 3-1) Collect data sets for drug-target interaction prediction research, including 708 drugs, 4192 side effects, 5603 diseases and 1512 targets, as well as drug-disease known association matrices Drug-side effect known association matrix Target-disease known association matrix Drug-target known association matrix Drug-Drug Interaction Matrix and target-target interaction matrix where d, t, s, and I symbols represent drug, target, side effect, and disease bioentities, respectively, and dI, ds, tI, dt, dd, and tt represent drug-disease, drug-side effect, target-disease, drug-target, drug-drug, and target-target associations, respectively; 3-2) Using the Deep Graph Learning Library (DGL Library) to build heterogeneous graphs Final heterogeneous graph The correlation matrix It can be defined as: in, and Respectively and Transpose of a matrix; 3-3) The similarity matrix of drugs and the target similarity matrix As the chemical structure characteristics of drugs and sequence structure characteristics of targets, isomerization graph Initial node feature matrix of drugs and targets Can be initialized as: in It is a common symbol for expressing the dimension of the matrix, and N(d) and N(t) represent the number of drugs and targets, respectively.

5. The drug-target interaction prediction method based on multi-domain heterogeneous graph aggregation learning according to claim 4 is characterized in that: The step 4) comprises the following steps: 4-1) A lightweight meta-path-level feature aggregator is used to pre-calculate the meta-path-level information aggregation of each neighboring node in the biological heterogeneous graph. Taking the i-th drug node as an example, the aggregation process can be expressed as: Where d represents the drug, Represents the characteristic matrix of the i-th drug node after the above aggregation process, Indicates the i-th drug node in the meta-domain graph The feature matrix after information aggregation, It is the meta-domain graph of the given i-th drug node Set Φ d The list of different semantic feature matrices generated, corresponds to the metadomain graph The set of all meta-path instances in , It represents The number of meta-path instances, is a meta-path instance containing the target node i and the starting node j, is the neighbor node feature vector x on the meta-path j The attention weights of the target can be obtained by the same logic. Different semantic feature matrices correspond to different meta-domain graphs of the target. t represents target; 4-2) Use adjacency matrix multiplication to further simplify formula (5), and add variable-length meta-path instances to the meta-domain graph to enhance the perception domain of local node feature aggregation. Let is a l-1 length metapath, where d represents the drug, t1t2t3…t l-1 Represents a meta path The intermediate nodes on t l-1 Represents the l-1th intermediate node. Taking drug node d as an example, the specific simplification process is as follows: in Indicates that drug node d is in the meta-domain graph The feature matrix after the above simplified aggregation is, Represented in the metadomain graph The adjacency matrix between the drug node d and the intermediate node t1 in After row normalization, the matrices d-t1, t1-t2 and t l-1 -d represents drug node d-intermediate node t1, intermediate node t1-intermediate node t2 and intermediate node t l-1 - drug node d, Indicates that in the meta path The feature aggregation weight vector of the drug neighbor nodes on It is a common symbol for expressing the matrix dimension. N(d) and N(t) represent the number of drugs and targets respectively. ⊙ represents the product operation of vector elements. Represents the initial feature matrix set of all drug nodes. Similarly, the target node t in the meta-domain graph can be obtained The feature matrix after simplified aggregation 6. The drug-target interaction prediction method based on multi-domain heterogeneous graph aggregation learning according to claim 5 is characterized in that: The step 5) comprises the following steps: 5-1) When aggregating meta-domain graph-level information in the subsequent process, feature embedding vectors of different dimensions may bring challenges. The feature projection step converts the feature aggregation vectors of different dimensions into features with consistent dimensions. Use a multi-layer perceptron (MLP) module Where v represents the node type in the heterogeneous graph, and the multi-layer perception block contains a normalization layer, a nonlinear layer, and a dropout layer between two consecutive linear layers. As an example of the feature projection step, the specific projection process is as follows: in Indicates that drug node d is in the meta-domain graph The feature matrix after the simplified aggregation of formula (6) is: Represented in the metadomain graph Perform feature projection on the feature matrix ·, Indicates that drug node d is in the meta-domain graph The feature matrix after feature projection is the same as the drug domain graph The feature projection process is similar to that of the target node t. Perform feature projection to obtain the target's feature projection matrix 7. The drug-target interaction prediction method based on multi-domain heterogeneous graph aggregation learning according to claim 6, characterized in that: The step 6) comprises the following steps: 6-1) Taking the drug node as an example, use the predefined drug different meta-domain graph list And the semantic vector set of each drug node after the feature projection step Where n represents the number of drug meta-domain graphs. The Transformer-based multi-domain graph semantic feature fusion module learns the degree of mutual influence between each pair of semantic vectors. The process of multi-domain graph semantic feature fusion can be expressed as: in and Representing the ith, jth and rth meta-domain graphs of drugs (or targets), each semantic vector Each maps a query vector A key vector and a vector of values Mutual attention weight coefficient is the query vector With key vector The dot product result after softmax normalization, and Respectively express and The transpose of, EXP(·) represents the exponential function with the natural constant e as the base, Represents a vector and Perform product operation, current semantics The output vector is a vector of all values Add the weighted sum of the remaining connections, The query vectors are Key vector Sum value vector The trainable shared parameters, θ is also a trainable parameter shared by all drug meta-domain graphs, and finally the feature fusion matrix list of each meta-domain graph will be obtained. Similarly, for the different meta-domain graph lists of target t Where m represents the number of meta-domain graphs of target t. Using the same Transformer-based multi-domain graph semantic feature fusion step, we can obtain the feature fusion matrix list of each meta-domain graph of target t. 6-2) Perform weighted summation on the feature fusion matrix list of each meta-domain graph and obtain the drug feature matrix before comparison optimization and target feature matrix The process is as follows: in, Represents the computational domain graph and metadomain diagram The feature fusion matrix and The trainable weight coefficients between them.

8. The drug-target interaction prediction method based on multi-domain heterogeneous graph aggregation learning according to claim 7, characterized in that: The step 7) comprises the following steps: 7-1) The feature matrix of the drug node and the feature matrix of the target node A new feature matrix is ​​formed by concatenating Subsequently, feature optimization and update will be performed on the feature matrix, and a topological map will be constructed based on the original drug-target association map. and semantic graph In the topology diagram In the topology graph, if two drug-target pair nodes have the same drug or target pointing to these two nodes, they will be connected, thus constructing the topological graph adjacency matrix M p , for semantic graph By calculating the cosine similarity between drug-target pair nodes, sorting them, and connecting the drug-target pair nodes with their first k most similar nodes, the semantic graph adjacency matrix M is constructed. s Finally, two graph convolutional neural network (GCN) modules are used to extract the topological graph and semantic graph The node features of the enhanced view are learned and used for graph contrastive learning to supervise the optimization features. The contrastive learning optimization process is as follows: where I is the identity matrix, and Represented in the topology diagram and semantic graph The feature matrix extracted by the l+1th convolution kernel, D p and D s They are the topological graph adjacency matrix M p And the semantic graph adjacency matrix M s The degree matrix of p and D s The diagonals of are all 1, represents the regularization of the matrix, is the trainable weight matrix in the GCN module, and The feature matrices are The updated feature matrix from the topological map and semantic map, Relu(·) represents the activation function in the neural network; 7-2) Use the InfoCNE method to segment negative samples and classify them into different categories. Then calculate the loss by evaluating the similarity between sample features. Then use the contrast loss calculated by InfoNCE to strengthen the consistency of model prediction and expand the distinction between positive and negative samples in the feature space. The contrast loss function for: Where N is the number of nodes in the original drug-target association graph, Z i It is a topological diagram Node i in the semantic graph The set of all positive samples in , exp(·) represents the exponential function with the natural constant e as the base, S(·,·) represents the function for calculating cosine similarity, and τ represents an adjustable scalar parameter. and Represent nodes i, j and i respectively ′ In the topology diagram and semantic graph The feature matrix extracted by the GCN module.

9. The drug-target interaction prediction method based on multi-domain heterogeneous graph aggregation learning according to claim 8, characterized in that: The step 8) comprises the following steps: 8-1) will eventually and Input into the multi-layer perceptron (MLP), and use MLP to learn the feature interaction between drugs and target nodes to obtain the interaction prediction score matrix The MLP learning prediction process is as follows: Where ⊙ represents the Hadamard product, W,W 2 …W L and ε 1 ,ε 2 ,…ε L are all represented as parameter matrices that need to be trained. represents the interaction prediction score between drug d and target t; 8-2) Use data labels to optimize model performance and use the cross entropy loss function to calculate classification loss. The cross entropy loss function The definition is as follows: Where N is the number of nodes in the original drug-target association graph, Y i represents the label of sample i, represents the predicted association probability of sample i, (d, t) represents the drug-target pair sample, The loss function of the final method training It consists of three parts, defined as follows: in is the contrastive loss function, is the classification loss calculated using the cross entropy loss function, λ is a hyperparameter between 0 and 1, and λ represents the ratio between the contrast loss and the classification loss. Represents L2 regularization of all trainable model parameters, and the parameter matrix of L2 regularization is Θ to prevent model overfitting.

Images