A method and device for predicting drug-target interaction and a storage medium

Abstract

The application belongs to the technical field of drug target relationship prediction, and relates to a drug and target interaction prediction method and device and a storage medium; based on drugs, targets, diseases, side effects and their interactions, a heterogeneous biological network graph is constructed to obtain initial feature vectors of each node; the drug node and the target node are taken as target nodes to obtain four types of neighbor nodes and four types of meta-path nodes of the target nodes; the initial feature vectors of the four types of neighbor nodes and the four types of meta-path nodes of the target nodes are subjected to feature extraction to obtain neighbor view embedding representation and meta-path view embedding representation; the neighbor view embedding representation and the meta-path view embedding representation are projected into query vectors, key vectors and value vectors, and cross-attention values are calculated; based on the neighbor view embedding representation, the meta-path view embedding representation and the cross-attention values of each target node, a fusion representation is obtained; and based on the fusion representation, interaction prediction of the drug node and the target node is carried out.

Description

Technical Field

[0001] This invention relates to the field of drug target relationship prediction technology, and in particular to a method, apparatus and storage medium for predicting drug-target interaction. Background Technology

[0002] Drug-target interaction (DTI) refers to the specific binding between a drug molecule and a biological target (such as a protein or nucleic acid). DTI prediction is a crucial step in drug discovery and retargeting. By presupposing the potential binding relationships between a drug and its target protein, it not only helps in understanding the drug's mechanism of action but also significantly shortens the new drug development cycle, reduces development costs, and provides a theoretical basis for drug repurposing. DTI prediction is widely used in drug retargeting, mechanism of action analysis, and personalized medicine.

[0003] Currently, many methods are available for predicting drug-target interactions (DTI). While traditional experimental methods (such as surface plasmon resonance and enzyme-linked immunosorbent assay) offer high accuracy, they are time-consuming, expensive, and have low throughput, making them unsuitable for large-scale screening. Therefore, developing efficient and accurate computationally driven DTI prediction methods has become a research hotspot in artificial intelligence and bioinformatics in recent years.

[0004] Existing computationally driven DTI prediction methods are mainly divided into three categories: ligand-based methods, structure-based methods, and network-based methods. Among them, network-based methods have attracted much attention due to their ability to effectively integrate multi-source heterogeneous biomedical data. These methods typically model biological entities such as drugs, targets, diseases, and side effects as nodes in a heterogeneous graph and use graph neural networks to learn node representations, thereby predicting the potential associations between drugs and targets. However, existing heterogeneous graph-based DTI prediction models usually rely only on local neighbor aggregation to learn node representations. Although this can effectively capture the neighborhood information directly connected to drugs or targets in the heterogeneous graph, it ignores the indirect associations implied by higher-order semantic paths. For example, drug pairs associated through co-treatment of diseases are correlated. This ignores the global semantic context in the heterogeneous graph, resulting in incomplete representation information learned by the model during prediction, and consequently, lower prediction accuracy when predicting interactions based on the feature information of drugs and targets.

[0005] In summary, existing methods for predicting drug-target interactions suffer from low accuracy. Summary of the Invention

[0006] Therefore, the technical problem to be solved by the present invention is to overcome the problem of low accuracy of prediction results in existing drug-target interaction prediction methods.

[0007] To address the aforementioned technical problems, this invention provides a method for predicting drug-target interactions, comprising:

[0008] By using drugs, targets, diseases, and side effects as nodes and the interactions between nodes as edges, a heterogeneous biological network graph is constructed, and the initial feature vectors of each node are obtained.

[0009] Drug nodes and target nodes are used as target nodes. The nodes connected to each target node are classified to obtain four types of neighbor nodes for each target node. The four types of nodes with the same connected nodes as each target node are obtained to obtain four types of meta-path nodes for each target node.

[0010] Feature extraction and cross-relation fusion are performed on the initial feature vectors of the four types of neighboring nodes of each target node to obtain the neighbor view embedding representation of each target node; feature extraction and aggregation are performed on the initial feature vectors of the four types of meta-path nodes of each target node to obtain the meta-path view embedding representation of each target node.

[0011] The neighbor view embedding representation of each target node is projected as a first query vector, a first key vector, and a first value vector, and the meta path view embedding representation of each target node is projected as a second query vector, a second key vector, and a second value vector, thereby calculating the first cross attention value and the second cross attention value of each target node.

[0012] The neighbor view embedding representation, meta-path view embedding representation, first cross attention value and second cross attention value of each target node are residually fused to obtain the fused representation of each target node;

[0013] The fused representations of drug nodes and target nodes are concatenated and input into the multilayer perceptron prediction head, outputting the interaction probability of drug nodes and target nodes.

[0014] Preferably, obtaining the initial feature vector of each node includes:

[0015] The chemical structure of each drug node is encoded using SMILES sequences to obtain the encoded sequences; principal component analysis is then used to project the encoded sequences of each drug node to obtain the initial feature vectors of each drug node.

[0016] The context-aware embedding of the amino acid sequence of each target node is extracted using a protein language model. The dimensionality of the average sequence representation output by the last layer of the protein language model is reduced to obtain the initial feature vector of each target node.

[0017] Unsupervised graph embedding is used to randomly traverse the heterogeneous biological network graph to generate node sequences for each disease node and side effect node. The Skip-gram model is used to extract features from the node sequences of each disease node and side effect node to generate vector representations of each disease node and side effect node. The vector representations of each disease node and side effect node are then dimensionality-reduced to obtain the initial feature vectors of each disease node and side effect node.

[0018] Preferably, a multi-layered heterogeneous graph convolutional network is used to extract features and fuse them across relationships from the initial feature vectors of the four types of neighboring nodes of each target node, resulting in a neighbor view embedding representation for each target node, including:

[0019] For each type of neighboring nodes of each target node, calculate the sum of the reciprocal of the normalized value of the initial feature vector of all nodes in that type of neighboring nodes and the product of the first node representation of each node in the (k-1)th heterogeneous graph convolutional layer of that type of neighboring nodes, and obtain the aggregation result of that type of neighboring nodes in the kth heterogeneous graph convolutional layer.

[0020] By utilizing the learnable weight matrix of each type of neighboring nodes of the target node in the k-th layer of the heterogeneous graph convolutional layer, a linear transformation is performed on the aggregation result of each type of neighboring nodes of the target node in the k-th layer of the heterogeneous graph convolutional layer, to obtain the node representation of each type of neighboring nodes of the target node in the k-th layer of the heterogeneous graph convolutional layer.

[0021] The node representations of each target node in the k-th heterogeneous graph convolutional layer are fused using an attention mechanism or mean pooling to obtain the first node representation of each target node in the k-th heterogeneous graph convolutional layer.

[0022] Based on the first node representation of each target node in the Kth layer of the heterogeneous graph convolutional layer, the neighbor view embedding representation of each target node is obtained;

[0023] Where 2≤k≤K, K represents the number of heterogeneous graph convolutional layers in the heterogeneous graph convolutional network.

[0024] Preferably, a multi-layered heterogeneous graph convolutional network is used to extract and aggregate the initial feature vectors of the four types of meta-path nodes for each target node, resulting in a meta-path view embedding representation for each target node, including:

[0025] For each class of meta-path nodes of each target node, calculate the sum of the reciprocal of the mean of the initial feature vectors of all nodes in the class of meta-path node, the learnable transformation matrix of the class of meta-path node in the k-th layer of the heterogeneous graph convolutional network, and the product of the second node representation of each node in the (k-1)-th layer of the heterogeneous graph convolutional layer, to obtain the feature representation of the class of meta-path node in the k-th layer of the heterogeneous graph convolutional layer.

[0026] Calculate the product of the self-loop information parameter of the k-th heterogeneous graph convolutional layer and the second node representation of each target node in the (k-1)-th heterogeneous graph convolutional layer. Based on the sum of this product and the feature representations of various meta-path nodes of the target node in the k-th heterogeneous graph convolutional layer, obtain the second node representation of the target node in the k-th heterogeneous graph convolutional layer.

[0027] Based on the second node representation of each target node in the Kth layer of the heterogeneous graph convolutional layer, the neighbor view embedding representation of each target node is obtained;

[0028] Where 2≤k≤K, K represents the number of heterogeneous graph convolutional layers in the heterogeneous graph convolutional network.

[0029] Preferably, calculating the first cross-attention value and the second cross-attention value for each target node includes:

[0030] Calculate the first inner product of the second query vector and the first key vector of each target node. Scale the first inner product using a scaling factor and then use a softmax layer to obtain the first attention weight of the scaled first inner product. Based on the product of the first attention weight and the first value vector of the target node, obtain the first cross-attention value of the target node.

[0031] Calculate the second inner product of the first query vector and the second key vector of each target node. Scale the second inner product using a scaling factor and then use a softmax layer to obtain the second attention weight of the scaled second inner product. Based on the product of the second attention weight and the second value vector of the target node, obtain the second cross-attention value of the target node.

[0032] Preferably, the neighbor view embedding of the target node is represented as follows:

[0033] ,

[0034] in, The neighbor view embedding representation of the i-th target node; Represents a nonlinear activation function; Indicates the type of neighboring nodes; Represents the target node's... Neighbor-like nodes in the th Learnable relational attention weights in heterogeneous graph convolutional layers; Represents the target node's... Neighbor-like nodes in the th Learnable weight matrix of heterogeneous graph convolutional layer; Represents the i-th target node. Neighbor-like nodes; express Nodes in; Represents the i-th target node. The normalized values ​​of the initial feature vectors of all nodes in the neighboring nodes; Represents a node The first node of the (K-1)th heterogeneous graph convolutional layer is represented; K represents the number of heterogeneous graph convolutional layers in the heterogeneous graph convolutional network.

[0035] The metapath view embedding of the target node is represented as follows:

[0036] ,

[0037] ,

[0038] in, The meta-path view embedding representation represents the i-th target node; This represents the parameter that preserves the self-loop information of the k-th heterogeneous graph convolutional layer; This represents the second node representation of the i-th target node in the (k-1)-th heterogeneous graph convolutional layer. The sum of the feature representations of various meta-path nodes of the i-th target node in the k-th heterogeneous graph convolutional layer represents the sum of the feature representations of each type of meta-path node. Indicates the metapath type of the i-th target node; Represents the i-th target node. Feature representation of class path nodes in the Kth layer of heterogeneous graph convolutional layer; Represents the i-th target node. The mean of the initial feature vectors of all nodes in the class path node; Represents the i-th target node. Metapath nodes; express Nodes in; express The second node of the heterogeneous graph convolutional layer at layer K-1 is represented; Indicates the first The learnable transformation matrix of class path nodes in the k-th layer heterogeneous graph convolutional network.

[0039] Preferably, the first query vector of the target node is represented as:

[0040] ,

[0041] in, This represents the first query vector for the i-th target node; This represents the first learnable projection matrix; The neighbor view embedding representation of the i-th target node;

[0042] The first key vector of the target node is represented as:

[0043] ,

[0044] in, This represents the first key vector of the i-th target node; This represents the second learnable projection matrix;

[0045] The first value vector of the target node is represented as:

[0046] ,

[0047] in, This represents the first value vector of the i-th target node; This represents the third learnable projection matrix;

[0048] The second query vector of the target node is represented as:

[0049] ,

[0050] in, This represents the second query vector for the i-th target node; This represents the fourth learnable projection matrix; The meta-path view embedding representation represents the i-th target node;

[0051] The second key vector of the target node is represented as:

[0052] ,

[0053] in, This represents the second key vector of the i-th target node; This represents the fifth learnable projection matrix;

[0054] The second value vector of the target node is represented as:

[0055] ,

[0056] in, This represents the second value vector of the i-th target node; This represents the sixth learnable projection matrix;

[0057] The formula for calculating the first cross-attention value of the target node is:

[0058] ,

[0059] in, This represents the first cross-attention value of the i-th target node; Indicates transpose; Indicates the scaling factor;

[0060] The formula for calculating the second cross-attention value of the target node is:

[0061] ,

[0062] in, This represents the second cross-attention value of the i-th target node.

[0063] Preferably, the fusion of target nodes is represented as follows:

[0064] ,

[0065] in, This represents the fused representation of the i-th target node; The neighbor view embedding representation of the i-th target node; This represents the first cross-attention value of the i-th target node; The meta-path view embedding representation represents the i-th target node; This represents the second cross-attention value of the i-th target node.

[0066] The present invention also provides a device for predicting drug-target interactions, comprising:

[0067] The heterogeneous graph construction and feature acquisition module is used to construct a heterogeneous biological network graph by using drugs, targets, diseases and side effects as nodes and the interactions between nodes as edges, and to obtain the initial feature vectors of each node.

[0068] The neighbor node and metapath node acquisition module is used to take drug nodes and target nodes as target nodes, classify the nodes connected to each target node to obtain four types of neighbor nodes for each target node; and acquire four types of nodes that have the same connected nodes as each target node to obtain four types of metapath nodes for each target node.

[0069] The dual-view feature extraction module is used to extract features and fuse them across relationships from the initial feature vectors of the four types of neighboring nodes of each target node to obtain the neighboring view embedding representation of each target node; and to extract and aggregate features from the initial feature vectors of the four types of meta-path nodes of each target node to obtain the meta-path view embedding representation of each target node.

[0070] The bidirectional cross-attention calculation module is used to project the neighbor view embedding representation of each target node into a first query vector, a first key vector, and a first value vector, and to project the meta-path view embedding representation of each target node into a second query vector, a second key vector, and a second value vector, thereby calculating the first cross-attention value and the second cross-attention value of each target node.

[0071] The dual-view feature fusion module is used to perform residual fusion of the neighbor view embedding representation, meta-path view embedding representation, first cross-attention value and second cross-attention value of each target node to obtain the fused representation of each target node;

[0072] The prediction module is used to concatenate the fused representations of drug nodes and target nodes and input them into the multilayer perceptron prediction head, outputting the interaction probability of drug nodes and target nodes.

[0073] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method for predicting drug-target interactions.

[0074] The drug-target interaction prediction method provided in this application constructs a heterogeneous biological network graph and introduces a meta-path view based on the local neighbor view. The local neighbor view aggregates the neighborhood information of each drug node and target node, while the meta-path view extracts the global semantic aggregation information of each drug node and target node, obtaining the local direct association information and global indirect semantic information of each drug node and target node. Furthermore, to achieve deep interaction and complementary information extraction between the neighbor view and the meta-path view, a bidirectional cross-attention fusion mechanism is designed. By enabling mutual queries and key-value pairs between the two views, their cross-view dependencies are dynamically modeled, generating a more discriminative fusion representation. This not only preserves the original semantics of each view but also explicitly models the dynamic interaction between local structure and higher-order semantics through bidirectional cross-attention, significantly improving the richness and task relevance of node representations. Finally, a lightweight multilayer perceptron is used to complete the final prediction of drug-target interactions based on the feature-rich fusion representations of drug nodes and target nodes. This fully extracts and deeply fuses multi-view information, effectively improving the prediction accuracy of drug-target interactions. Attached Figure Description

[0075] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0076] Figure 1 A schematic diagram of the drug-target interaction prediction method provided in this application;

[0077] Figure 2 The heterogeneous biological network provided in this application; wherein, Figure 2 (a) in the diagram is a schematic of node types. Figure 2(b) in the diagram is a heterogeneous biological network diagram. Figure 2 (c) in the diagram is an example of a metapath;

[0078] Figure 3 This is a schematic diagram illustrating the principle framework of the drug-target interaction prediction method provided in this application. Detailed Implementation

[0079] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0080] Please see Figure 1 , Figure 1 The diagram shows a flowchart of the drug-target interaction prediction method provided in this application, which specifically includes steps S10 to S60:

[0081] S10: Using drugs, targets, diseases, and side effects as nodes, and the interactions between nodes as edges, construct a heterogeneous biological network graph and obtain the initial feature vectors of each node.

[0082] S20: Using drug nodes and target nodes as target nodes, classify the nodes connected to each target node to obtain four types of neighbor nodes for each target node; obtain four types of nodes with the same connected nodes as each target node to obtain four types of metapath nodes for each target node.

[0083] For example, such as Figure 2 The image shows a heterogeneous biological network provided in this application, wherein, Figure 2 (a) in the diagram is a schematic of node types. Figure 2 (b) in the diagram is a heterogeneous biological network diagram. Figure 2 (c) in the diagram is an example of a metapath.

[0084] In one embodiment of this application, a heterogeneous biological network graph containing four types of nodes—drug (d), target (drug, p), disease (drug, s), and side effects (drug, e)—was constructed based on the publicly available Luo dataset. , where the set of nodes Representing drug, target, disease, and side effect nodes respectively, the set It includes various types of associations, such as drug-drug interactions and drug-target interactions. and These represent the set of node types and the set of edge types, respectively (edge ​​types include dd, dp, ds, de, pp, pd, and ps). When the graph is heterogeneous, this application further introduces the concept of meta-path to characterize higher-order semantic relations. Meta-path is defined as a composite relational path between node types in a heterogeneous graph, formally represented as... For example, the metapath " "Describes drug pairs that are associated through their shared treatment of diseases, while..." This reflects the functional similarity between target proteins involved in the same disease. Specifically, taking node a as an example, if node a is connected to node c, and nodes b, c, and d are all connected to node c, then nodes b, c, and d are all meta-path nodes of node a. Furthermore, based on the type of nodes b, c, and d, they can be classified into various types of meta-path nodes of node a. These meta-paths provide a structural basis for subsequent high-order semantic feature extraction.

[0085] S30: Perform feature extraction and cross-relation fusion on the initial feature vectors of the four types of neighboring nodes of each target node to obtain the neighbor view embedding representation of each target node; perform feature extraction and aggregation on the initial feature vectors of the four types of meta-path nodes of each target node to obtain the meta-path view embedding representation of each target node.

[0086] S40: Project the neighbor view embedding representation of each target node into a first query vector, a first key vector, and a first value vector, and project the meta-path view embedding representation of each target node into a second query vector, a second key vector, and a second value vector, thereby calculating the first cross-attention value and the second cross-attention value of each target node.

[0087] S50: Perform residual fusion on the neighbor view embedding representation, meta-path view embedding representation, first cross-attention value and second cross-attention value of each target node to obtain the fused representation of each target node.

[0088] S60: The fused representations of the drug node and target node are concatenated and input into the Multilayer Perceptron (MLP) prediction head, and the interaction probability of the drug node and target node is output.

[0089] Furthermore, based on the construction of the heterogeneous biological graph, it is necessary to extract features from each node to obtain the initial feature representation of each node, which will then serve as the input for subsequent graph neural network learning. Since drug nodes and target nodes have direct sequence or structural information, while disease nodes and side effect nodes lack direct sequence or structural information, different initial feature extraction methods are required for different types of nodes.

[0090] Specifically, obtaining the initial feature vectors of each node includes steps 1-1 to 1-3:

[0091] Step 1-1: Encode the chemical structure of each drug node using SMILES sequences to obtain the encoded sequences; project the encoded sequences of each drug node using principal component analysis to obtain the initial feature vectors of each drug node.

[0092] For example, the encoded sequence of the drug node can be projected onto a 128-dimensional real space to obtain the initial feature vector of the drug node. .

[0093] Step 1-2: Use the protein language model to extract the context-aware embedding of the amino acid sequence of each target node, reduce the dimensionality of the average sequence representation output by the last layer of the protein language model, and obtain the initial feature vector of each target node.

[0094] For example, the dimensionality of the sequence average representation can be reduced to 128 dimensions to obtain the initial feature vector of the target node. .

[0095] Steps 1-3: Randomly traverse the heterogeneous biological network graph using unsupervised graph embedding to generate node sequences for each disease node and side effect node; extract features from the node sequences of each disease node and side effect node using the Skip-gram model to generate vector representations for each disease node and side effect node; reduce the dimensionality of the vector representations of each disease node and side effect node to obtain the initial feature vectors for each disease node and side effect node.

[0096] For example, the vector representations of each disease node and side effect node are reduced to 128 dimensions to obtain the initial feature vector of the disease node. and the initial feature vector of the side effect node Ultimately, the initial feature vectors of all nodes are unified to 128 dimensions, thus yielding the node feature matrix. .

[0097] Furthermore, such as Figure 3As shown, after obtaining the initial feature vectors of each node, this application first extracts the local structural features of drug nodes and target nodes from the neighbor view, aiming to capture the contextual information of drug nodes and target nodes that are directly connected and indirectly adjacent in the heterogeneous graph. Specifically, a heterogeneous graph convolutional network (HGCN) is used as the neighbor view encoder. For any drug node or target node, its aggregation process in the k-th heterogeneous graph convolutional layer is grouped according to edge type. For example, for drug node a, which is directly connected to target node b, disease node c, and side effect node d, the edge types of its direct connection include dp, ds, and de. The initial feature representations of all neighbor nodes connected by each edge type can be aggregated.

[0098] Specifically, this application utilizes a multi-layered heterogeneous graph convolutional network to extract features and fuse cross-relationships of the initial feature vectors of the four types of neighboring nodes of each target node, obtaining the neighbor view embedding representation of each target node, including steps 2-1 to 2-4:

[0099] Step 2-1: For each type of neighboring nodes of each target node, calculate the sum of the reciprocal of the normalized value of the initial feature vector of all nodes in that type of neighboring nodes and the product of the first node representation of each node in the (k-1)th heterogeneous graph convolutional layer of that type of neighboring nodes, to obtain the aggregation result of that type of neighboring nodes in the kth heterogeneous graph convolutional layer.

[0100] Step 2-2: Using the learnable weight matrix of each type of neighboring nodes of the target node in the k-th layer of the heterogeneous graph convolutional layer, perform a linear transformation on the aggregation result of each type of neighboring nodes of the target node in the k-th layer of the heterogeneous graph convolutional layer to obtain the node representation of each type of neighboring nodes of the target node in the k-th layer of the heterogeneous graph convolutional layer.

[0101] Steps 2-3: Use attention mechanism or mean pooling to fuse the node representations of each target node's various neighbor nodes in the k-th layer of the heterogeneous graph convolutional layer, to obtain the first node representation of each target node in the k-th layer of the heterogeneous graph convolutional layer.

[0102] Steps 2-4: Based on the first node representation of each target node in the Kth heterogeneous graph convolutional layer, obtain the neighbor view embedding representation of each target node; where 2≤k≤K, and K represents the number of heterogeneous graph convolutional layers in the heterogeneous graph convolutional network.

[0103] Specifically, the neighbor view embedding of the target node is represented as follows:

[0104] ,

[0105] in, The neighbor view embedding representation of the i-th target node; This represents a non-linear activation function, such as ReLU; Indicates the type of neighboring nodes; Represents the target node's... Neighbor-like nodes in the th Learnable relational attention weights in heterogeneous graph convolutional layers; Represents the target node's... Neighbor-like nodes in the th Learnable weight matrix of heterogeneous graph convolutional layer; Represents the i-th target node. Neighbor-like nodes; express Nodes in; Represents the i-th target node. The normalized values ​​of the initial feature vectors of all nodes in the neighboring nodes; Represents a node The first node of the (K-1)th heterogeneous graph convolutional layer is represented; K represents the number of heterogeneous graph convolutional layers in the heterogeneous graph convolutional network. .

[0106] In one specific embodiment of this application, two layers of HGCN are stacked to capture second-order neighborhood information, ultimately obtaining each drug node. and target nodes The neighbor view embedding representation is denoted as follows: and ,in With a hidden layer dimension (e.g., 128), this representation effectively encodes the structural roles and semantic context of drug nodes and target nodes in a local heterogeneous neighborhood.

[0107] like Figure 3 As shown, to further capture the higher-order semantic relationships contained in heterogeneous graphs, this application introduces a meta-path view to extract features from drug nodes and target nodes. This view explicitly models the relationships between cross-type nodes through predefined semantic meta-paths, thereby revealing the underlying biological logic. Specifically, meta-paths such as dd, dp, ds, de, pp, pd, and ps constitute the basic semantic units for information propagation in heterogeneous graphs. Based on this, this application employs a meta-path-based message passing mechanism for multi-level iteration to gradually expand the receptive field and integrate higher-order contextual information.

[0108] Specifically, this application utilizes a multi-layered heterogeneous graph convolutional network to extract and aggregate the initial feature vectors of the four types of meta-path nodes for each target node, obtaining the meta-path view embedding representation of each target node, including steps 3-1 to 3-3:

[0109] Step 3-1: For each class of meta-path nodes of each target node, calculate the sum of the inverse of the mean of the initial feature vectors of all nodes in the class of meta-path node, the learnable transformation matrix of the class of meta-path node in the k-th layer of the heterogeneous graph convolutional network, and the product of the second node representation of each node in the (k-1)-th layer of the heterogeneous graph convolutional layer, to obtain the feature representation of the class of meta-path node in the k-th layer of the heterogeneous graph convolutional layer.

[0110] Step 3-2: Calculate the product of the self-loop information parameter of the k-th heterogeneous graph convolutional layer and the second node representation of each target node in the (k-1)-th heterogeneous graph convolutional layer. Based on the sum of this product and the feature representations of various meta-path nodes of the target node in the k-th heterogeneous graph convolutional layer, obtain the second node representation of the target node in the k-th heterogeneous graph convolutional layer.

[0111] Step 3-3: Based on the second node representation of each target node in the Kth heterogeneous graph convolutional layer, obtain the neighbor view embedding representation of each target node; where 2≤k≤K, and K represents the number of heterogeneous graph convolutional layers in the heterogeneous graph convolutional network.

[0112] Specifically, the metapath view embedding of the target node is represented as follows:

[0113] ,

[0114] ,

[0115] in, The meta-path view embedding representation represents the i-th target node; This represents the parameter that preserves the self-loop information of the k-th heterogeneous graph convolutional layer; This represents the second node representation of the i-th target node in the (k-1)-th heterogeneous graph convolutional layer. The sum of the feature representations of various meta-path nodes of the i-th target node in the k-th heterogeneous graph convolutional layer represents the sum of the feature representations of each type of meta-path node. Indicates the metapath type of the i-th target node; Represents the i-th target node. Feature representation of class path nodes in the Kth layer of heterogeneous graph convolutional layer; Represents the i-th target node. The mean of the initial feature vectors of all nodes in the class path node; Represents the i-th target node. Metapath nodes; express Nodes in; express The second node of the heterogeneous graph convolutional layer at layer K-1 is represented; Indicates the first The learnable transformation matrix of class path nodes in the k-th layer heterogeneous graph convolutional network.

[0116] Specifically, in one embodiment of this application, stacking is used. Layers (e.g.) This type of meta-path-aware message passing model can progressively extend local interactions to higher-order neighborhoods containing multi-hop semantics, ultimately taking the first path. Layer output as drug node and target nodes The embedded representations under the metapath view are denoted as follows: and This representation effectively integrates global contextual information propagated through multiple semantic paths, significantly enhancing the semantic richness and discriminative ability of node features.

[0117] Furthermore, such as Figure 3 As shown, after extracting the embedded representations of drug nodes and target nodes in the neighbor view and meta-path view, in order to realize the deep interaction and complementary information extraction between the neighbor view and the meta-path view, this application designs a bidirectional cross-attention fusion mechanism. This mechanism dynamically models the cross-view dependency relationship by enabling the two views to query each other and use key-value pairs, thereby generating a more discriminative fusion representation.

[0118] Specifically, the first cross-attention value and the second cross-attention value of each target node are calculated, including steps 4-1 to 4-2:

[0119] Step 4-1: Calculate the first inner product of the second query vector and the first key vector of each target node. Scale the first inner product using a scaling factor and then use a softmax layer to obtain the first attention weight of the scaled first inner product. Based on the product of the first attention weight and the first value vector of the target node, obtain the first cross-attention value of the target node.

[0120] Step 4-2: Calculate the second inner product of the first query vector and the second key vector of each target node. Scale the second inner product using a scaling factor and then use a softmax layer to obtain the second attention weight of the scaled second inner product. Based on the product of the second attention weight and the second value vector of the target node, obtain the second cross-attention value of the target node.

[0121] Specifically, the first query vector of the target node is represented as:

[0122] ,

[0123] in, This represents the first query vector for the i-th target node; This represents the first learnable projection matrix; This represents the neighbor view embedding representation of the i-th target node.

[0124] The first key vector of the target node is represented as:

[0125] ,

[0126] in, This represents the first key vector of the i-th target node; This represents the second learnable projection matrix.

[0127] The first value vector of the target node is represented as:

[0128] ,

[0129] in, This represents the first value vector of the i-th target node; This represents the third learnable projection matrix.

[0130] The second query vector of the target node is represented as:

[0131] ,

[0132] in, This represents the second query vector for the i-th target node; This represents the fourth learnable projection matrix; This represents the meta-path view embedding representation of the i-th target node.

[0133] The second key vector of the target node is represented as:

[0134] ,

[0135] in, This represents the second key vector of the i-th target node; This represents the fifth learnable projection matrix.

[0136] The second value vector of the target node is represented as:

[0137] ,

[0138] in, This represents the second value vector of the i-th target node; This represents the sixth learnable projection matrix.

[0139] Furthermore, the formula for calculating the first cross-attention value of the target node is as follows:

[0140] ,

[0141] in, This represents the first cross-attention value of the i-th target node; Indicates transpose; This represents the scaling factor.

[0142] The formula for calculating the second cross-attention value of the target node is:

[0143] ,

[0144] in, This represents the second cross-attention value of the i-th target node.

[0145] Furthermore, the original representation is residually fused with the information enhanced by cross-attention to obtain the fused representation of the target node:

[0146] ,

[0147] in, This represents the fused representation of the i-th target node; The neighbor view embedding representation of the i-th target node; This represents the first cross-attention value of the i-th target node; The meta-path view embedding representation represents the i-th target node; This represents the second cross-attention value of the i-th target node.

[0148] Specifically, the fusion mechanism designed in this application not only preserves the original semantics of each view, but also explicitly models the dynamic interaction between local structure and higher-order semantics through bidirectional cross-attention, which significantly improves the richness and task relevance of node representations and provides a stronger feature foundation for subsequent DTI prediction.

[0149] like Figure 3 As shown, after obtaining the fused representation of the drug node and the target node, this application uses a lightweight multilayer perceptron (MLP) to complete the final prediction of drug-target interaction.

[0150] Specifically, for any drug-target pair Embed its corresponding fusion and By concatenating along the feature dimensions, a joint representation is formed. :

[0151] ,

[0152] in, This indicates a vector concatenation operation.

[0153] After that The input is fed into an MLP prediction head consisting of two fully connected layers, each followed by a ReLU activation function, and the output layer uses a Sigmoid function to map the result to... The interval represents the probability that the drug-target pair interacts. :

[0154] ,

[0155] in , Represents the learnable weight matrix. , Indicates the bias term. Indicates the hidden layer dimension of the MLP. This represents the Sigmoid function.

[0156] Based on the drug-target interaction prediction method provided in the above embodiments, this application also provides a drug-target interaction prediction device, which specifically includes:

[0157] The heterogeneous graph construction and feature acquisition module is used to construct a heterogeneous biological network graph by using drugs, targets, diseases, and side effects as nodes and the interactions between nodes as edges, and to obtain the initial feature vectors of each node.

[0158] The neighbor node and metapath node acquisition module is used to take drug nodes and target nodes as target nodes, classify the nodes connected to each target node to obtain four types of neighbor nodes for each target node; and acquire four types of nodes with the same connected nodes as each target node to obtain four types of metapath nodes for each target node.

[0159] The dual-view feature extraction module is used to extract features and fuse them across relationships from the initial feature vectors of the four types of neighboring nodes of each target node to obtain the neighboring view embedding representation of each target node; and to extract and aggregate features from the initial feature vectors of the four types of meta-path nodes of each target node to obtain the meta-path view embedding representation of each target node.

[0160] The bidirectional cross-attention calculation module is used to project the neighbor view embedding representation of each target node into a first query vector, a first key vector, and a first value vector, and to project the meta-path view embedding representation of each target node into a second query vector, a second key vector, and a second value vector, thereby calculating the first cross-attention value and the second cross-attention value of each target node.

[0161] The dual-view feature fusion module is used to perform residual fusion of the neighbor view embedding representation, meta-path view embedding representation, first cross-attention value and second cross-attention value of each target node to obtain the fused representation of each target node.

[0162] The prediction module is used to concatenate the fused representations of drug nodes and target nodes and input them into the multilayer perceptron prediction head, outputting the interaction probability of drug nodes and target nodes.

[0163] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method for predicting drug-target interactions.

[0164] The proposed method for predicting drug-target interactions extracts local structural features and high-order semantic features of drugs and targets from neighbor views and meta-path views on a heterogeneous biological network graph. It also innovatively introduces a bidirectional cross-attention mechanism to achieve deep interaction and adaptive fusion of multiple views. Finally, the interaction probability is predicted by splicing and fusing the representation and passing it through a multilayer perceptron. By fully integrating the information from multiple views, it generates semantically rich and discriminative feature representations, thereby improving the predictive performance of drug-target interactions.

[0165] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0166] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0167] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0168] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0169] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for predicting drug-target interactions, characterized in that, The application comprises the following steps: drug, target, disease and side effect are taken as nodes, and the interaction between each node is taken as edge to construct a heterogeneous biological network graph, and the initial feature vector of each node is obtained; drug nodes and target nodes are taken as target nodes, and each target node connected node is classified to obtain four types of neighbor nodes of each target node; four types of nodes with the same connection nodes as each target node are obtained to obtain four types of meta-path nodes of each target node; the initial feature vector of each target node is extracted and cross-relation fusion is performed on the initial feature vector of each target node to obtain the neighbor view embedding representation of each target node; the initial feature vector of each target node is extracted and aggregated to obtain the meta-path view embedding representation of each target node; the neighbor view embedding representation of each target node is projected into a first query vector, a first key vector and a first value vector, and the meta-path view embedding representation of each target node is projected into a second query vector, a second key vector and a second value vector, so as to calculate the first cross-attention value and the second cross-attention value of each target node; the neighbor view embedding representation, the meta-path view embedding representation, the first cross-attention value and the second cross-attention value of each target node are residual fused to obtain the fusion representation of each target node; the fusion representation of the drug nodes and the target nodes is spliced and input into a multi-layer perception prediction head to output the interaction probability of the drug nodes and the target nodes.

2. The method of predicting drug-target interactions according to claim 1, wherein, The initial feature vector of each node comprises the following steps: the chemical structure of each drug node is encoded by using SMILES sequence to obtain an encoded sequence; the principal component analysis method is used to project the encoded sequence of each drug node to obtain the initial feature vector of each drug node; the context-aware embedding of the amino acid sequence of each target node is extracted by using a protein language model, and the sequence average representation output by the last layer of the protein language model is dimensionally reduced to obtain the initial feature vector of each target node; unsupervised graph embedding method is used to randomly walk on the heterogeneous biological network graph to generate node sequences of each disease node and side effect node; the vector representation of each disease node and side effect node is generated by using Skip-gram model to extract the features of the node sequence of each disease node and side effect node; the initial feature vector of each disease node and side effect node is obtained by dimensionally reducing the vector representation of each disease node and side effect node.

3. The method of predicting drug-target interactions according to claim 1, wherein, The initial feature vector of each target node is extracted and cross-relation fusion is performed on the initial feature vector of each target node by using a multi-layer architecture heterogeneous graph convolution network to obtain the neighbor view embedding representation of each target node, comprising: for each type of neighbor node of each target node, the sum of the product of the inverse of the normalized value of the initial feature vector of all nodes in the neighbor node and the first node representation of each node in the neighbor node at the k-1 layer heterogeneous graph convolution layer is calculated to obtain the aggregation result of the neighbor node at the k layer heterogeneous graph convolution layer; The aggregation result of each type of neighbor node of each target node in the kth heterogeneous graph convolution layer is linearly transformed by using the learnable weight matrix of each type of neighbor node of the target node in the kth heterogeneous graph convolution layer to obtain the node representation of each type of neighbor node of each target node in the kth heterogeneous graph convolution layer. The node representations of each type of neighbor node of each target node in the kth heterogeneous graph convolution layer are fused by using an attention mechanism or mean pooling to obtain the first node representation of each target node in the kth heterogeneous graph convolution layer. The neighbor view embedding representation of each target node is obtained based on the first node representation of each target node in the Kth heterogeneous graph convolution layer. 2≤k≤K, and K represents the number of layers of the heterogeneous graph convolution layer in the heterogeneous graph convolution network.

4. The method of predicting drug-target interactions according to claim 1, wherein, The initial feature vectors of the four types of meta-path nodes of each target node are extracted and aggregated by using the multi-layer architecture of the heterogeneous graph convolution network to obtain the meta-path view embedding representation of each target node, including: For each type of meta-path node of each target node, the sum of the inverse of the mean of the initial feature vectors of all nodes in the type of meta-path node, the learnable transformation matrix of the type of meta-path node in the kth heterogeneous graph convolution network, and the product of the second node representation of each node in the type of meta-path node in the k-1th heterogeneous graph convolution layer is calculated to obtain the feature representation of the type of meta-path node in the kth heterogeneous graph convolution layer. The product of the reserved self-loop information parameter of the kth heterogeneous graph convolution layer and the second node representation of each target node in the k-1th heterogeneous graph convolution layer is calculated, and based on the sum of the product and the sum of the feature representations of each type of meta-path node of the target node in the kth heterogeneous graph convolution layer, the second node representation of the target node in the kth heterogeneous graph convolution layer is obtained. The neighbor view embedding representation of each target node is obtained based on the second node representation of each target node in the Kth heterogeneous graph convolution layer. 2≤k≤K, and K represents the number of layers of the heterogeneous graph convolution layer in the heterogeneous graph convolution network.

5. The method of predicting drug-target interactions according to claim 1, wherein, The first cross-attention value and the second cross-attention value of each target node are calculated, including: The first inner product of the second query vector of each target node and its first key vector is calculated, the first inner product is scaled by using a scaling factor, and then the first attention weight of the scaled first inner product is obtained by using a softmax layer; based on the product of the first attention weight and the first value vector of the target node, the first cross-attention value of the target node is obtained. The second inner product of the first query vector of each target node and its second key vector is calculated, the second inner product is scaled by using a scaling factor, and then the second attention weight of the scaled second inner product is obtained by using a softmax layer; based on the product of the second attention weight and the second value vector of the target node, the second cross-attention value of the target node is obtained.

6. The method of predicting drug-target interactions according to claim 1, wherein, The neighbor view embedding representation of the target node is: ， wherein, represents a neighbor view embedding representation of the i-th target node; represents a non-linear activation function; represents a type of neighbor node; represents the i-th target node; class neighbor node at the i-th layer of the heterogeneous graph convolutional layer; represents a learnable relation attention weight of the i-th layer of the heterogeneous graph convolutional layer; represents the i-th target node; class neighbor node at the i-th layer of the heterogeneous graph convolutional layer; represents a learnable weight matrix of the i-th layer of the heterogeneous graph convolutional layer; represents the i-th target node; class neighbor node; represents a node in ; represents the i-th target node; represents an initial feature vector normalization value of all nodes in the i-th target node; represents a node at the first node representation of the K-1-th layer of the heterogeneous graph convolutional layer; K represents a number of layers of the heterogeneous graph convolutional layers in the heterogeneous graph convolutional network; The meta-path view embedding representation of the target node is: ， ， in, The meta-path view embedding representation represents the i-th target node; This represents the parameter that preserves the self-loop information of the k-th heterogeneous graph convolutional layer; This represents the second node representation of the i-th target node in the (k-1)-th heterogeneous graph convolutional layer. The sum of the feature representations of various meta-path nodes of the i-th target node in the k-th heterogeneous graph convolutional layer represents the sum of the feature representations of each type of meta-path node. Indicates the metapath type of the i-th target node; Represents the i-th target node. Feature representation of class path nodes in the Kth layer of heterogeneous graph convolutional layer; Represents the i-th target node. The mean of the initial feature vectors of all nodes in the class path node; Represents the i-th target node. Metapath nodes; express Nodes in; express The second node of the heterogeneous graph convolutional layer at layer K-1 is represented; Indicates the first The learnable transformation matrix of class path nodes in the k-th layer heterogeneous graph convolutional network.

7. The method of predicting drug-target interactions according to claim 1, wherein, The first query vector of the target node is: ， wherein, denotes a first query vector of the i-th target node; denotes a first learnable projection matrix; denotes a neighbor view embedding representation of the i-th target node; The first key vector of the target node is: ， wherein, denotes a first key vector of the i-th target node; denotes a second learnable projection matrix; The first value vector of the target node is: ， wherein, denotes a first value vector of the i-th target node; denotes a third learnable projection matrix; The second query vector of the target node is: ， wherein, denotes a second query vector of the i-th target node; denotes a fourth learnable projection matrix; denotes a meta-path view embedding representation of the i-th target node; The second key vector of the target node is represented as: ， wherein, represents a second key vector of the i-th target node; represents a fifth learnable projection matrix; The second value vector of the target node is represented as: ， wherein, represents a second value vector of the i-th target node; represents a sixth learnable projection matrix; The calculation formula of the first cross-attention value of the target node is: ， wherein, denotes a first cross-attention value for the i-th target node; denotes a transpose; denotes a scaling factor; The calculation formula of the second cross-attention value of the target node is: ， wherein, denotes the second cross-attention value of the i-th target node.

8. The method of predicting drug-target interactions according to claim 1, wherein, The fusion representation of the target node is: ， wherein, denotes a fused representation of the i-th target node; denotes a neighbor view embedding representation of the i-th target node; denotes a first cross-attention value for the i-th target node; denotes a meta-path view embedding representation of the i-th target node; denotes a second cross-attention value for the i-th target node.

9. A device for predicting drug-target interactions, characterized in that, comprises: a heterogeneous graph construction and feature acquisition module, configured to construct a heterogeneous biological network graph by taking drugs, targets, diseases and side effects as nodes and interactions between the nodes as edges, and acquire initial feature vectors of the nodes; a neighbor node and meta-path node acquisition module, configured to take drug nodes and target nodes as target nodes, classify nodes connected to each target node, and obtain four types of neighbor nodes of each target node; four types of nodes having the same connected nodes as each target node are acquired, and four types of meta-path nodes of each target node are obtained; a dual-view feature extraction module, configured to perform feature extraction and cross-relation fusion on initial feature vectors of the four types of neighbor nodes of each target node, and obtain neighbor view embedding representations of each target node; feature extraction and aggregation are performed on initial feature vectors of the four types of meta-path nodes of each target node, and meta-path view embedding representations of each target node are obtained; a bidirectional cross-attention calculation module, configured to project the neighbor view embedding representations of each target node into a first query vector, a first key vector and a first value vector, project the meta-path view embedding representations of each target node into a second query vector, a second key vector and a second value vector, and calculate first cross-attention values and second cross-attention values of each target node; a dual-view feature fusion module, configured to perform residual fusion on the neighbor view embedding representations, the meta-path view embedding representations, the first cross-attention values and the second cross-attention values of each target node, and obtain fusion representations of each target node; a prediction module, configured to concatenate the fusion representations of the drug nodes and the target nodes and input the concatenated fusion representations into a multi-layer perception prediction head, and output interaction probabilities of the drug nodes and the target nodes.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the prediction method of the drug-target interaction according to any one of claims 1 to 8. The computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the prediction method of the drug-target interaction according to any one of claims 1 to 8.

Images