Methods, devices, and readable storage media for predicting drug-target interactions

Abstract

This application relates to the field of digital healthcare technology and discloses a method, apparatus, and medium for predicting drug-target interactions. The method includes: acquiring drug graph data and target graph data; mapping each vector in the drug graph data and target graph data to a high-dimensional vector; performing neighborhood information propagation on the high-dimensional vectors of each node based on an intra-graph information propagation layer to obtain a set of intra-graph feature vectors for drug nodes and a set of intra-graph feature vectors for target nodes; performing cross-graph information propagation on the high-dimensional vectors of each node based on an inter-graph information propagation layer to obtain a set of inter-graph feature vectors for drug nodes and a set of inter-graph feature vectors for target nodes; and determining the prediction result based on the similarity between the drug graph representation results and the target graph representation results. This application constructs the prediction of drug-target interactions into a graph matching task by using molecular and protein graphs as inputs, enabling the prediction model of drug-target interactions to have better generalization ability and more accurately predict drug-target interactions.

Description

Technical Field

[0001] This disclosure relates to the field of digital healthcare technology, and more specifically, to a method, apparatus, and readable storage medium for predicting drug-target interactions. Background Technology

[0002] Drug discovery is the process of identifying new candidate compounds with potential therapeutic effects. Predicting drug-target interactions (DTI) is an essential step in the drug discovery process. Experimental methods for determining drug-target interactions, such as X-ray diffraction, nuclear magnetic resonance (NMR), and laser Raman spectroscopy, are both time-consuming and resource-intensive.

[0003] In recent years, machine learning-based models have been widely used to predict drug molecule-target protein interactions, thereby enabling rapid drug screening.

[0004] Deep learning models, with their powerful fitting capabilities, can effectively predict the interaction between drugs and targets to a certain extent. However, since the DTI task requires processing inputs from both drug molecules and protein molecules, fusing the feature representations from these two perspectives is a crucial issue. Current techniques typically involve extracting features from each branch separately and then training by concatenating these features. This approach fails to preserve the original structural information of the molecules, resulting in suboptimal prediction performance. Summary of the Invention

[0005] In view of the above, this application provides a method, apparatus and readable storage medium for predicting drug-target interactions, aiming to solve the problem of fusing and learning two input features, namely drug molecules and target protein molecules, to achieve more accurate prediction of drug-target interactions.

[0006] In a first aspect, embodiments of this application provide a method for predicting drug-target interactions. The method is based on a drug-target interaction prediction model, which includes a coding layer, a propagation layer, and an aggregation layer connected in sequence. The propagation layer includes an intra-graph information propagation layer and an inter-graph information propagation layer.

[0007] The method includes:

[0008] Acquire drug map data and target map data, wherein the drug map data and the target map data each include multiple node vectors and multiple edge vectors;

[0009] Based on the encoding layer, each vector in the drug map data and the target map data is mapped to a high-dimensional vector, which includes node high-dimensional vectors and edge high-dimensional vectors.

[0010] Based on the graph information propagation layer, neighborhood information propagation is performed on the high-dimensional vectors of each node in the drug graph data and the target graph data according to the high-dimensional vectors obtained by mapping, to obtain the drug node graph feature vector set and the target node graph feature vector set.

[0011] Based on the inter-graph information propagation layer, cross-graph information propagation is performed on the high-dimensional vectors of each node in the drug graph data and the target graph data according to the high-dimensional vectors obtained by mapping, so as to obtain the set of feature vectors of nodes between the drug graph and the set of feature vectors of nodes between the target graph.

[0012] Based on the aggregation layer, the drug graph representation result is determined according to the drug node feature vector set, the target graph representation result is determined according to the target node feature vector set, and the prediction result of drug-target interaction is determined according to the similarity between the drug graph representation result and the target graph representation result; wherein, the drug node feature vector set includes the intra-drug feature vector set and the inter-drug graph node feature vector set, and the target node feature vector set includes the intra-target node feature vector set and the inter-target graph node feature vector set.

[0013] Secondly, embodiments of this application also provide a drug target interaction prediction device, the device being equipped with a drug target interaction prediction model, the model comprising a coding layer, a propagation layer, and an aggregation layer connected in sequence, wherein the propagation layer comprises an intra-graph information propagation layer and an inter-graph information propagation layer;

[0014] The device includes:

[0015] An acquisition module is used to acquire drug map data and target map data, wherein the drug map data and the target map data respectively include multiple node vectors and multiple edge vectors;

[0016] The encoding module is used to map each vector in the drug map data and the target map data into a high-dimensional vector based on the encoding layer. The high-dimensional vector includes node high-dimensional vectors and edge high-dimensional vectors.

[0017] The first propagation module is used to propagate neighborhood information to each node's high-dimensional vector in the drug graph data and the target graph data based on the graph information propagation layer and the high-dimensional vector obtained by mapping, so as to obtain the drug node graph feature vector set and the target node graph feature vector set.

[0018] The second propagation module is used to propagate cross-graph information based on the inter-graph information propagation layer and according to the high-dimensional vectors obtained by mapping, respectively, to the high-dimensional vectors of each node in the drug graph data and the target graph data, so as to obtain the set of feature vectors of nodes between the drug graph and the set of feature vectors of nodes between the target graph.

[0019] An aggregation module is used to determine drug graph representation results based on the aggregation layer, according to the drug node feature vector set, to determine target graph representation results based on the target node feature vector set, and to determine the prediction results of drug-target interaction based on the similarity between the drug graph representation results and the target graph representation results; wherein, the drug node feature vector set includes the intra-drug feature vector set and the inter-drug graph node feature vector set, and the target node feature vector set includes the intra-target node feature vector set and the inter-target graph node feature vector set.

[0020] Thirdly, embodiments of this application also provide a computer-readable storage medium storing one or more programs, which, when executed by an electronic device including multiple applications, cause the electronic device to perform the steps of any of the above-described methods for predicting drug-target interactions.

[0021] Fourthly, embodiments of this application also provide a computer-readable storage medium storing one or more programs that, when executed by an electronic device including multiple applications, cause the electronic device to perform the steps of any of the above-described methods for predicting drug-target interactions.

[0022] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects:

[0023] This application constructs a novel drug-target interaction prediction model. This model transforms the prediction task of drug-target interaction into a graph matching task by taking drug graph data and target graph data as input. Unlike existing technologies that extract features from drug molecules and target proteins separately and then roughly train the model by splicing features, the proposed model utilizes intra-graph neighborhood information and cross-graph information to propagate information between the high-dimensional vectors of each node in both the drug and target graph data. This enables information interaction, allowing the feature vector sets of drug nodes and target nodes to converge more node and structural information. This results in a model with better generalization ability, enabling more accurate prediction of drug-target interactions and providing valuable reference information for drug selection. It significantly saves economic, human, and time costs in drug selection, shortens the drug development cycle, and improves drug development efficiency. Attached Figure Description

[0024] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0025] Figure 1 A flowchart illustrating a method for predicting drug-target interactions according to an embodiment of this application is shown.

[0026] Figure 2 A schematic diagram of a drug-target interaction prediction model according to this application is shown;

[0027] Figure 3 A schematic diagram of the structure of an encoding layer according to an embodiment of this application is shown;

[0028] Figure 4 A schematic diagram of the structure of an in-figure information propagation layer according to an embodiment of this application is shown;

[0029] Figure 5 A schematic diagram of the structure of an inter-graph information propagation layer according to an embodiment of this application is shown;

[0030] Figure 6 A schematic diagram of the structure of an aggregate layer according to an embodiment of this application is shown;

[0031] Figure 7 A flowchart illustrating a method for predicting drug-target interactions according to yet another embodiment of this application is shown.

[0032] Figure 8 A schematic diagram of the structure of a drug target interaction prediction device according to an embodiment of this application is shown;

[0033] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0035] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0036] In DTI tasks, it is necessary to comprehensively process two input features from drug molecules and target protein molecules. In existing technologies, the features of the two branches are usually extracted separately and then trained by splicing the features. However, this method cannot retain the original structural information of drug molecules and target molecules, resulting in poor prediction performance and very limited scalability and generalization.

[0037] Based on this, this application proposes a method for predicting drug-target interactions. By taking molecular and protein graphs as inputs, this application innovatively constructs the prediction of drug-target interactions into a graph matching task, which enables the prediction model of drug-target interactions to have better generalization ability and can more accurately predict drug-target interactions.

[0038] Graph matching refers to establishing correspondences between nodes in two or more graph structures. In computer vision, graph matching algorithms are commonly used to solve key-to-key matching relationships between multiple images, such as in image retrieval. This application innovatively applies graph matching technology to the prediction of drug-target interactions, proposing a method for predicting drug-target interactions.

[0039] Figure 1 This diagram illustrates a flowchart of a method for predicting drug-target interactions according to an embodiment of this application. Figure 1 It can be seen that this application includes at least steps S101 to S105:

[0040] Step S101: Obtain drug map data and target map data, wherein the drug map data and the target map data respectively include multiple node vectors and multiple edge vectors.

[0041] The method for predicting drug-target interactions in this application is based on a drug-target interaction prediction model. Figure 2 A schematic diagram of a drug target interaction prediction model according to this application is shown. Figure 2 It can be seen that the drug-target interaction prediction model 200 includes a coding layer 201, a propagation layer 202 and an aggregation layer 203 connected in sequence. The propagation layer 202 includes an intra-graph information propagation layer 2021 and an inter-graph information propagation layer 2022.

[0042] First, drug graph data and target graph data are acquired, wherein the drug graph data and target graph data each include multiple node vectors and multiple edge vectors. In some embodiments of this application, before the step of acquiring drug graph data and target graph data, the method further includes: converting the drug molecule to be tested and the target protein molecule to be tested into drug graph structures and target graph structures, respectively; processing the drug graph structures and target graph structures based on graph neural networks to obtain drug graph data and target graph data, wherein the drug graph data includes multiple node vectors with atoms as nodes and multiple edge vectors with inter-atomic chemical bonds as edges; the target graph data includes multiple node vectors with amino acid residues as nodes and multiple edge vectors with spatial relationships between amino acid residues as edges.

[0043] Traditional machine learning can assist drug development to some extent, but it requires molecular descriptors as feature inputs. The choice of different molecular descriptors has a significant impact on the performance of machine learning models. Therefore, most traditional machine learning methods require complex and time-consuming feature engineering. In contrast, the emerging deep learning technology can extract features directly from the "original" structure of drugs, thereby bypassing feature engineering and shortening the development cycle.

[0044] In this application, the drug molecule to be tested and the target protein molecule to be tested can be first converted into drug map structures and target map structures, respectively. In some embodiments, open-source chemical information toolkits, such as RDKit, Chemistry Development Kit (CDK), Open Babel, etc., can be used to convert the chemical formula of the drug molecule to be tested into a drug map structure and the amino acid sequence of the target protein molecule to be tested into a target map structure.

[0045] Then, based on graph neural networks, the drug graph structure and target graph structure are processed respectively to obtain drug graph data and target graph data. The drug graph data includes multiple node vectors with atoms as nodes and multiple edge vectors with chemical bonds between atoms as edges; the target graph data includes multiple node vectors with amino acid residues as nodes and multiple edge vectors with spatial relationships between amino acid residues as edges.

[0046] Here, for drug map data, node vectors are used to characterize the properties of atoms in drug molecules, such as atom type and atomic orientation, while edge vectors are used to characterize the properties of chemical bonds between atoms in drug molecules, such as bond type, chemical bond orientation, and saturation. For target map data, node vectors are used to characterize the properties of amino acid residues in the corresponding target protein molecule, such as structure, while edge vectors are used to characterize the spatial relationships between amino acid residues in the target protein molecule, such as distance and angle.

[0047] In some embodiments of this application, the drug graph structure can be processed based on graph neural networks, such as Graph Neural Networks (GNN) models, Graph Isomorphic Net (GIN) models, Graph Convolutional Network (GCN) models, and Graph Attention Networks (GAT) models, to obtain drug graph data. The drug graph data includes multiple node vectors with atoms as nodes and multiple edge vectors with inter-atomic chemical bonds as edges. For example, if the drug molecule to be tested includes atoms 1 and 2, and atoms 1 and 2 are connected by chemical bond 1, then the drug graph data includes node vector 1 with atom 1 as a node, node vector 2 with atom 2 as a node, and edge vector 1 with chemical bond 1 as an edge.

[0048] Similarly, based on a graph neural network, the target graph structure is processed to obtain target graph data. This data includes multiple node vectors with amino acid residues as nodes and multiple edge vectors with the spatial relationships between amino acid residues as edges. For example, consider amino acid residues 1, 2, and 3 in the target protein molecule. If the nearest amino acid residue is selected based on the average distance between atoms, and amino acid residue 2 is chosen as the neighbor node of amino acid residue 1, then the spatial relationship 1 between amino acid residue 1 and amino acid residue 2 is: the distance between amino acid residue 1 and amino acid residue 2 is distance 1, and the angle between them is angle 1. Therefore, the target graph data includes node vector 3 with amino acid residue 1 as a node, node vector 4 with amino acid residue 2 as a node, and edge vector 2 with the spatial relationship 1 between amino acid residue 1 and amino acid residue 2 as edges.

[0049] Step S102: Based on the encoding layer, each vector in the drug map data and the target map data is mapped to a high-dimensional vector, wherein the high-dimensional vector includes node high-dimensional vectors and edge high-dimensional vectors.

[0050] To utilize more details of the vectors in the drug map data and the target map data, the vectors in the drug map data and the target map data can be mapped to high-dimensional vectors based on the encoding layer. The high-dimensional vectors include node high-dimensional vectors and edge high-dimensional vectors.

[0051] In some embodiments of this application, node vector 1 (x1, x2, x3), node vector 2 (y1, y2, y3), and edge vector 1 (e1, e2, e3) from the drug graph data can be input into the above-mentioned encoding layer to obtain node high-dimensional vector 1 (x1, x2, x3, x4, x5), node high-dimensional vector 2 (y1, y2, y3, y4, y5), and edge high-dimensional vector 1 (e1, e2, e3, e4, e5).

[0052] In other embodiments of this application, in the above method, the encoding layer includes a multilayer perceptron node encoding layer and a multilayer perceptron edge encoding layer; the step of mapping each vector in the drug map data and the target map data into multiple node high-dimensional vectors and multiple edge high-dimensional vectors based on the encoding layer includes: for any node vector in the drug map data and the target map data, inputting the node vector to the multilayer perceptron node encoding layer to obtain the corresponding node high-dimensional vector; for any edge vector in the drug map data and the target map data, inputting the edge vector to the multilayer perceptron edge encoding layer to obtain the corresponding edge high-dimensional vector.

[0053] Figure 3 A schematic diagram of the structure of the coding layer according to an embodiment of this application is shown. Figure 3 It can be seen that the coding layer 201 includes the multilayer perceptron node coding layer 2011 and the multilayer perceptron edge coding layer 2012.

[0054] In some embodiments of this application, after obtaining drug map data and target map data, for any node vector in the drug map data and target map data, the node vector can be input to the node encoding layer 2011 of the multilayer perceptron to obtain the corresponding high-dimensional node vector; for any edge vector in the drug map data and target map data, the edge vector is input to the edge encoding layer 2012 of the multilayer perceptron to obtain the corresponding high-dimensional edge vector.

[0055] Step S103: Based on the graph information propagation layer, according to the high-dimensional vector obtained by mapping, neighborhood information propagation is performed on the high-dimensional vectors of each node in the drug graph data and the target graph data respectively to obtain the set of feature vectors in the drug node graph and the set of feature vectors in the target node graph.

[0056] The high-dimensional vectors of the obtained drug graph data and target graph data can be input into the graph information propagation layer to perform neighborhood information propagation on the high-dimensional vectors of each node in the drug graph data and target graph data, respectively, to obtain the graph feature vector set of drug node and graph feature vector set of target node.

[0057] For example, in some embodiments of this application, for node A in the drug graph data, node B is connected to node A by edge BA, and node C is connected to node A by edge CA. When propagating neighborhood information for the high-dimensional vector A corresponding to node A, the high-dimensional vectors B, BA, C, and CA can be fused with node A to obtain the propagated high-dimensional vector A. This process is repeated to obtain multiple propagated high-dimensional vectors, which together form the feature vector set within the drug node graph.

[0058] Step S104: Based on the inter-graph information propagation layer, according to the high-dimensional vectors obtained by mapping, cross-graph information propagation is performed on the high-dimensional vectors of each node in the drug graph data and the target graph data respectively, to obtain the drug graph inter-node feature vector set and the target graph inter-node feature vector set.

[0059] Based on the inter-graph information propagation layer, cross-graph information propagation is performed on the high-dimensional vectors of each node in the drug graph data and target graph data according to the high-dimensional vectors obtained by mapping, so as to obtain the set of feature vectors of nodes between the drug graph and the set of feature vectors of nodes between the target graph.

[0060] For example, in some embodiments of this application, for node high-dimensional vectors 11, 12, and 13 in drug graph data, and node high-dimensional vectors 21, 22, and 23 in target graph data, node high-dimensional vector 11 can be propagated across the graph using node high-dimensional vectors 12, 13, 21, 22, and 23 to obtain propagated node high-dimensional vector 11; node high-dimensional vector 12 can be propagated across the graph using node high-dimensional vectors 11, 13, 21, 22, and 23 to obtain propagated node high-dimensional vector 12. Similarly, propagated node high-dimensional vectors 13, 21, 22, and 23 can be obtained.

[0061] The high-dimensional vectors 11, 12, and 13 of the propagated nodes form the set of node feature vectors between the drug map nodes; the high-dimensional vectors 21, 22, and 23 of the propagated nodes form the set of node feature vectors between the target map nodes.

[0062] Step S105: Based on the aggregation layer, determine the drug graph representation result according to the drug node feature vector set, determine the target graph representation result according to the target node feature vector set, and determine the prediction result of drug-target interaction according to the similarity between the drug graph representation result and the target graph representation result; wherein, the drug node feature vector set includes the intra-drug feature vector set and the inter-drug graph node feature vector set, and the target node feature vector set includes the intra-target node feature vector set and the inter-target graph node feature vector set.

[0063] Here, the drug node feature vector set includes the feature vector set within the drug node graph and the feature vector set between drug nodes, and the target node feature vector set includes the feature vector set within the target node graph and the feature vector set between target nodes.

[0064] In some embodiments of this application, the set of feature vectors within the drug node graph or the set of feature vectors between nodes in the drug graph can be input to the aggregation layer to obtain the drug graph representation result; the set of feature vectors within the target node graph or the set of feature vectors between nodes in the target graph can be input to the aggregation layer to obtain the target graph representation result.

[0065] In order to gather more information within the graph data and between the graph data, and to make full use of the features of the two graph data, in some other embodiments of this application, the drug node feature vector set can be input into the aggregation layer to obtain the drug graph representation result; and the target node feature vector set can be input into the aggregation layer to obtain the target graph representation result.

[0066] After obtaining the drug image representation results and the target image representation results, similarity calculation formulas, such as cosine similarity calculation formulas and distance calculation formulas, can be used to calculate the similarity value between the two vectors of the drug image representation results and the target image representation results. If the similarity value is greater than or equal to the preset similarity value, the predicted result of drug-target interaction is determined to be that the drug target has an interaction; if the similarity value is less than the preset similarity value, the predicted result of drug-target interaction is determined to be that the drug target does not have an interaction.

[0067] from Figure 1As shown in the method, this application constructs a novel drug-target interaction prediction model. This model transforms the prediction task of drug-target interaction into a graph matching task by taking drug graph data and target graph data as input. Unlike existing technologies that extract features from drug molecules and target proteins separately and then roughly train the model by splicing features, the model proposed in this application can utilize intra-graph neighborhood information and cross-graph information to propagate information between the high-dimensional vectors of each node in the drug graph data and target graph data, achieving information interaction. This allows the feature vector sets of drug nodes and target nodes to converge more node and structural information, giving the model better generalization ability and enabling accurate prediction of drug-target interactions. It provides valuable reference information for drug selection, greatly saving economic, human, and time costs in drug selection, shortening the drug development cycle, and improving drug development efficiency.

[0068] In some embodiments of this application, in the above method, the graph information propagation layer includes a graph information cascade layer and a first node network layer; the step of propagating neighborhood information on each node high-dimensional vector in the drug graph data and the target graph data based on the graph information propagation layer, according to the multiple node high-dimensional vectors and multiple edge high-dimensional vectors obtained by mapping, to obtain a set of drug node graph feature vectors and a set of target node graph feature vectors, includes: based on the graph information cascade layer, for any first target node high-dimensional vector in the drug graph data and the target graph data, respectively, the graph neighborhood node high-dimensional vectors of the first target node high-dimensional vector and the matching graph neighborhood edge high-dimensional vectors are respectively... The quantity is concatenated with the high-dimensional vector of the first target node to obtain several intra-graph concatenated vectors; based on the first node network layer, information accumulation processing is performed on the several intra-graph concatenated vectors to obtain the intra-graph node neighborhood information of the high-dimensional vector of the first target node; and using the intra-graph node neighborhood information, neighborhood information propagation is performed on the high-dimensional vector of the first target node to obtain the first node intra-graph feature vector corresponding to the high-dimensional vector of the first target node; the intra-graph feature vectors corresponding to the high-dimensional vectors of each node in the drug graph data form the drug node intra-graph feature vector set; the intra-graph feature vectors corresponding to the high-dimensional vectors of each node in the target graph data form the target node intra-graph feature vector set.

[0069] Figure 4 This diagram illustrates the structure of an in-figure information propagation layer according to an embodiment of this application. Figure 4 As can be seen, the intra-graph information propagation layer 2021 includes the intra-graph information cascading layer 20211 and the first node network layer 20212.

[0070] Based on the graph information cascade layer, for any high-dimensional vector of the first target node in the drug graph data and target graph data, the high-dimensional vectors of each of the graph's intra-graph neighboring nodes and the matching high-dimensional vectors of the graph's intra-graph neighboring edges are concatenated with the high-dimensional vector of the first target node to obtain several graph cascaded vectors.

[0071] In some embodiments of this application, the in-graph cascade vector can be determined according to the following formula (1).

[0072] m j→i =f cascade1 (h i ,h j ,e ij Equation (1);

[0073] Among them, f cascade1 h represents the cascading layer of information within the diagram. i h represents the high-dimensional vector of the i-th first target node; j Represents the high-dimensional vector of the j-th neighborhood node in the graph; e ij Let represent the high-dimensional vector of the graph neighborhood edge from node j to node i.

[0074] Specifically, for any high-dimensional vector A of the first target node in the drug graph data, the high-dimensional vectors of the intra-graph neighboring nodes of the first target node A are: node high-dimensional vector A1 and node high-dimensional vector A2. The high-dimensional vector of the intra-graph neighboring edge that matches node high-dimensional vector A1 is the intra-graph neighboring edge high-dimensional vector A1, and the high-dimensional vector of the intra-graph neighboring edge that matches node high-dimensional vector A2 is the intra-graph neighboring edge high-dimensional vector A2. Substituting the high-dimensional vector A of the first target node, node high-dimensional vector A1, and intra-graph neighboring edge high-dimensional vector A1 into the above formula (1), we obtain the intra-graph cascaded vector A1. Similarly, substituting the high-dimensional vector A of the first target node, node high-dimensional vector A2, and intra-graph neighboring edge high-dimensional vector A2 into the above formula (1), we can obtain the intra-graph cascaded vector A2.

[0075] Then, based on the first node network layer, information accumulation processing is performed on several intra-graph concatenated vectors to obtain the intra-graph node neighborhood information of the high-dimensional vector of the first target node; and using the intra-graph node neighborhood information, neighborhood information propagation is performed on the high-dimensional vector of the first target node to obtain the intra-graph feature vector of the first node corresponding to the high-dimensional vector of the first target node.

[0076] Specifically, in some embodiments of this application, the feature vector within the first node graph can be determined according to the following formula (2):

[0077] h' i =f node (h i ,Σ (j,i)∈E m j→iEquation (2);

[0078] Among them, f node This represents the first node in the network layer.

[0079] Specifically, information accumulation processing can be performed on several intra-graph concatenated vectors. For example, the intra-graph concatenated vectors A1 and A2 can be accumulated to obtain the intra-graph node neighborhood information A1 of the high-dimensional vector A of the first target node. Then, the intra-graph node neighborhood information A1 and the high-dimensional vector A of the first target node are input into the first node network layer. The neighborhood information A1 is used to propagate neighborhood information of the high-dimensional vector A of the first target node to obtain the intra-graph feature vector A of the first node corresponding to the high-dimensional vector A of the first target node.

[0080] In the drug graph data, the node graph feature vectors corresponding to the high-dimensional vectors of each node form the drug node graph feature vector set; in the target graph data, the node graph feature vectors corresponding to the high-dimensional vectors of each node form the target node graph feature vector set. For example, in some embodiments of this application, for node high-dimensional vectors A, B, and C in the drug graph data, using the above formulas (1) and (2), the corresponding node graph feature vectors A, B, and C can be obtained respectively, and the node graph feature vectors A, B, and C form the drug node graph feature vector set.

[0081] In some embodiments of this application, in the above method, the inter-graph information propagation layer includes an inter-graph information concatenation layer, a cross-graph information matching layer, and a second node network layer; the step of performing cross-graph information propagation on the high-dimensional vectors of each node in the drug graph data and the target graph data based on the inter-graph information propagation layer and the high-dimensional vectors of each node in the drug graph data and the target graph data, respectively, to obtain a set of node feature vectors in the drug graph and a set of node feature vectors in the target graph, includes: based on the inter-graph information concatenation layer, for any high-dimensional vector of a second target node in the drug graph data and the target graph data, concatenating the high-dimensional vectors of each cross-graph neighbor node and the matched high-dimensional vectors of cross-graph neighbor edges of the high-dimensional vector of the second target node with the high-dimensional vector of the second target node, to obtain several cross-graph concatenation vectors; based on the cross-graph information matching layer, concatenating the high-dimensional vectors of the second target node with the high-dimensional vector of each node in the drug graph data and the target graph data, respectively, to obtain several cross-graph concatenation vectors; based on the cross-graph information matching layer, concatenating the high-dimensional vectors ... The high-dimensional vectors of each cross-graph neighborhood node of the vector are matched with the high-dimensional vector of the second target node to obtain several cross-graph matching vectors. Based on the second node network layer, the information accumulation processing is performed on the several cross-graph concatenated vectors and the several cross-graph matching vectors to obtain the cross-graph node neighborhood information and cross-graph node matching information of the high-dimensional vector of the second target node. And according to the cross-graph node neighborhood information and the cross-graph node matching information, cross-graph information propagation is performed on the high-dimensional vector of the second target node to obtain the second node inter-graph feature vector corresponding to the high-dimensional vector of the second target node. The node inter-graph feature vectors corresponding to the high-dimensional vectors of each node in the drug graph data form the drug graph inter-graph feature vector set. The node inter-graph feature vectors corresponding to the high-dimensional vectors of each node in the target graph data form the target graph inter-graph feature vector set.

[0082] Figure 5 This illustration shows a schematic diagram of the structure of an inter-graph information propagation layer according to an embodiment of this application. Figure 5 It can be seen that the inter-graph information propagation layer 2022 includes the inter-graph information cascading layer 20221, the cross-graph information matching layer 20222, and the second node network layer 20223.

[0083] Here, cross-graph neighborhood is used to define the nodes that are relatively close in the drug map data and target map data. For example, for nodes 1, 2 and 3 in the drug map data, and nodes 4, 5 and 6 in the target map data, if nodes 1 and 4 are the closest, then the midpoint between nodes 1 and 4 can be used as the center. If the distance from node 2 to the center is 4, the distance from node 3 to the center is 3, the distance from node 5 to the center is 6, and the distance from node 6 to the center is 10, and the radius is a preset distance value of 5, then the cross-graph neighborhood of the drug map data and target map data includes nodes 1, 2, 3 and 4.

[0084] Based on the inter-graph information cascade layer, for any high-dimensional vector of a second target node in the drug graph data and target graph data, the high-dimensional vectors of each cross-graph neighbor node and the high-dimensional vector of the matched cross-graph neighbor edge are concatenated with the high-dimensional vector of the second target node to obtain several cross-graph cascaded vectors. Specifically, in some embodiments of this application, the cross-graph cascaded vectors can be determined according to the following formula (3):

[0085]

[0086] Among them, f cascade2 This represents a cascading layer of information between graphs.

[0087] For example, if the cross-graph neighborhood of drug graph data and target graph data includes nodes 1, 2, 3 and 4, where nodes 1, 2 and 3 belong to drug graph data and are connected in pairs, and node 4 belongs to target graph data. For the high-dimensional vector 1 of the second target node corresponding to node 1 in drug graph data, the high-dimensional vector 1 of the second target node, the high-dimensional vector 2 of the cross-graph neighborhood node corresponding to node 2, and the matching high-dimensional vector 12 of the cross-graph neighborhood edge corresponding to nodes 1 and 2 can be substituted into the above formula (3) to obtain the cross-graph concatenation vector 11; the high-dimensional vector 1 of the second target node, the high-dimensional vector 3 of the cross-graph neighborhood node corresponding to node 3, and the matching high-dimensional vector 13 of the cross-graph neighborhood edge corresponding to nodes 1 and 3 can be substituted into the above formula (3) to obtain the cross-graph concatenation vector 12.

[0088] There is a matching relationship between drug molecules and target proteins. In computers, based on the idea of ​​graph matching, cross-graph matching vectors can be used to describe the degree of matching between two nodes from different graph data.

[0089] Based on a cross-graph information matching layer, node matching can be performed between the high-dimensional vectors of each cross-graph neighboring node of the high-dimensional vector of the second target node and the high-dimensional vector of the second target node to obtain one or more cross-graph matching vectors. In some embodiments of this application, in the above method, the cross-graph matching vector is determined according to the following method: obtaining distance information between the high-dimensional vector of the second target node and the high-dimensional vectors of the cross-graph neighboring nodes to be matched based on a distance metric function; obtaining comprehensive distance information between the high-dimensional vector of the second target node and each of the high-dimensional vectors of the cross-graph neighboring nodes based on the distance metric function; using the ratio between the distance information and the comprehensive distance information as weight data between the high-dimensional vector of the second target node and the high-dimensional vectors of the cross-graph neighboring nodes to be matched; determining the cross-graph matching vector between the high-dimensional vector of the second target node and the high-dimensional vectors of the cross-graph neighboring nodes to be matched based on the weight data and the difference between the high-dimensional vector of the second target node and the high-dimensional vectors of the cross-graph neighboring nodes to be matched.

[0090] Based on the distance metric function, the distance information between the high-dimensional vector of the second target node and the high-dimensional vector of the cross-graph neighboring nodes to be matched is obtained; based on the distance metric function, the comprehensive distance information between the high-dimensional vector of the second target node and the high-dimensional vectors of each cross-graph neighboring node is obtained; the ratio between the distance information and the comprehensive distance information is used as the weight data between the high-dimensional vector of the second target node and the high-dimensional vector of the cross-graph neighboring nodes to be matched. Specifically, in some embodiments of this application, the weight data between the high-dimensional vector of the second target node and the high-dimensional vector of the cross-graph neighboring nodes to be matched can be determined according to the following formula (4):

[0091]

[0092] Among them, s h Represents the distance metric function; h k This represents the high-dimensional vector of the k-th cross-graph neighborhood node to be matched.

[0093] Specifically, if the cross-graph neighborhood of drug graph data and target graph data includes nodes 1, 2, 3, and 4, where nodes 1, 2, and 3 belong to the drug graph data and are paired, and node 4 belongs to the target graph data, then for node 4, a distance metric function, such as Euclidean distance, cosine similarity, Manhattan distance, Hamming distance, or Chebyshev distance, can be used to obtain the distance 1 between the high-dimensional vector 4 of the second target node corresponding to node 4 and the high-dimensional vector 1 of the cross-graph neighborhood node to be matched corresponding to node 1. This distance is then processed using an exponential function with the natural constant e as its base to obtain the distance information 1. Similarly, based on the distance metric function, we can obtain the distance information 2 between the high-dimensional vector 4 of the second target node and the high-dimensional vector 2 of the cross-graph neighboring nodes corresponding to node 2, and the distance information 3 between the high-dimensional vector 4 of the second target node and the high-dimensional vector 3 of the cross-graph neighboring nodes corresponding to node 3. We add distance information 2 and distance information 3 to obtain the comprehensive distance information. Then, we use the ratio between distance information 1 and the comprehensive distance information as the weight data a between the high-dimensional vector 4 of the second target node and the high-dimensional vector 1 of the cross-graph neighboring nodes to be matched. 14 .

[0094] Then, based on the weight data and the difference between the high-dimensional vector of the second target node and the high-dimensional vector of the neighboring nodes to be matched in the cross-graph, the cross-graph matching vector between the high-dimensional vector of the second target node and the high-dimensional vector of the neighboring nodes to be matched in the cross-graph is determined. Specifically, the cross-graph matching vector can be determined according to the following formula (5):

[0095] μ ji =a ji (h i -h jEquation (5);

[0096] Specifically, we can first subtract the high-dimensional vector 1 of the neighboring nodes to be matched from the high-dimensional vector 4 of the second target node to obtain the difference, and then combine the above difference with the weight data a. 14 Multiplying them together yields the cross-graph matching vector μ between the high-dimensional vector 4 of the second target node and the high-dimensional vector 1 of the cross-graph neighboring nodes to be matched. 14 .

[0097] Finally, information accumulation processing is performed on several cross-graph concatenated vectors to obtain the cross-graph node neighborhood information of the high-dimensional vector of the second target node; information accumulation processing is performed on several cross-graph matching vectors to obtain the cross-graph node matching information. Then, the high-dimensional vector of the second target node, the cross-graph node neighborhood information, and the cross-graph node matching information are input into the second node network layer, and cross-graph information propagation is performed on the high-dimensional vector of the second target node to obtain the second node inter-graph feature vector corresponding to the high-dimensional vector of the second target node. Specifically, in some embodiments of this application, the high-dimensional vector of the second target node can be determined according to the following formula (6):

[0098]

[0099] Here, the inter-graph feature vectors corresponding to the high-dimensional vectors of each node in the drug graph data form the inter-graph feature vector set of the drug graph; the inter-graph feature vectors corresponding to the high-dimensional vectors of each node in the target graph data form the inter-graph feature vector set of the target graph. In some embodiments of this application, the inter-graph feature vector set of the drug graph and the inter-graph feature vector set of the target graph respectively include: inter-graph feature vectors corresponding to the high-dimensional vectors of nodes belonging to the cross-graph neighborhood, and high-dimensional vectors of nodes not belonging to the cross-graph neighborhood.

[0100] In some embodiments of this application, in the above method, the aggregation layer includes a first processing layer, a second processing layer, and a similarity layer, wherein the first processing layer and the second processing layer are respectively connected to the similarity layer; the step of determining the drug graph representation result based on the aggregation layer according to the drug node feature vector set, determining the target graph representation result based on the target node feature vector set, and determining the prediction result of drug-target interaction based on the similarity between the drug graph representation result and the target graph representation result includes: processing the drug node feature vector set based on the first processing layer to obtain the drug graph representation result; processing the target node feature vector set based on the second processing layer to obtain the target graph representation result; determining the similarity value between the drug graph representation result and the target graph representation result based on the similarity layer; and determining the prediction result of drug-target interaction based on the similarity value and a preset threshold.

[0101] Figure 6A schematic diagram of the structure of a polymer layer according to an embodiment of this application is shown. Figure 6 As can be seen, the aggregation layer 203 includes a first processing layer 2031, a second processing layer 2032 and a similarity layer 2033, wherein the first processing layer 2031 and the second processing layer 2032 are respectively connected to the similarity layer 2033.

[0102] Based on the first processing layer, the drug node feature vector set is processed to obtain the drug graph representation result; based on the second processing layer, the target node feature vector set is processed to obtain the target graph representation result. Both the first and second processing layers are multi-perceptron network structures. Specifically, in some embodiments of this application, the drug node feature vector set can be substituted into formula (7) based on the following formulas (7) and (8) to obtain the drug graph representation result; the target node feature vector set can be substituted into formula (8) to obtain the target graph representation result.

[0103] h G1 =f G1 {h m}m∈V1 equation (7);

[0104] h G2 =f G2 {h n}n∈V2 equation (8);

[0105] Among them, f G1 Indicates the first processing layer; f G2 V1 represents the second processing layer; f represents the set of feature vectors for drug nodes; G2 G1 represents the set of feature vectors of target nodes; G2 represents the drug map data; G3 represents the target map data.

[0106] Then, based on the similarity layer, the similarity value between the drug map representation results and the target map representation results is determined. In some embodiments of this application, the drug map representation results and the target map representation results can be input to the similarity layer f based on the following formula (9). s The similarity value s is obtained:

[0107] s = f s (h G1 h G2 Equation (9);

[0108] In other embodiments of this application, other similarity calculation methods, such as cosine similarity, Euclidean distance, Manhattan distance, etc., can be used to calculate the similarity value between the drug image representation result and the target image representation result.

[0109] After obtaining the similarity value between the drug map characterization results and the target map characterization results, the predicted result of drug-target interaction is determined based on a preset threshold. For example, in some embodiments of this application, if the similarity value between the drug map characterization results and the target map characterization results is 0.9, and the preset threshold is 0.5, then 0.9 is greater than 0.5, and the predicted result of drug-target interaction is determined to be that there is an interaction.

[0110] In some embodiments of this application, in the above method, the drug-target interaction prediction model is trained according to the following method: obtaining a training sample set; wherein the training sample set contains multiple pairs of drug map data and target map data, and the sample label of the training sample set is whether there is an interaction; obtaining an initial drug-target interaction prediction model, the model including a coding layer, a propagation layer, and an aggregation layer connected in sequence, wherein the propagation layer includes an intra-graph information propagation layer and an inter-graph information propagation layer; inputting the training sample set into the initial drug-target interaction prediction model to obtain multiple predicted values; based on a preset loss function, updating the parameters in the initial drug-target interaction prediction model according to the multiple predicted values ​​and sample labels to obtain the drug-target interaction prediction model, wherein the preset loss function is set according to the Euclidean distance between drug targets.

[0111] A training sample set is obtained; wherein the training sample set contains multiple pairs of drug map data and target map data, and the sample labels of the training sample set are whether they interact. For example, in some embodiments of this application, the training sample set may contain drug map data 1 and target map data 1, drug map data 2 and target map data 2, and drug map data 3 and target map data 3. The sample labels corresponding to drug map data 1 and target map data 1 are interactive; the sample labels corresponding to drug map data 2 and target map data 2 are interactive; and the sample labels corresponding to drug map data 3 and target map data 3 are not interactive. In practice, the label value of interactive samples can be set to 1, and the label value of non-interactive samples can be set to 0.

[0112] Then, the initial prediction model for drug-target interaction is obtained, by Figure 2 It can be seen that the model includes an encoding layer, a propagation layer, and an aggregation layer connected in sequence. The propagation layer includes an intra-graph information propagation layer and an inter-graph information propagation layer.

[0113] The training sample set is input into the initial prediction model of drug-target interaction to obtain multiple predicted values. Based on the preset loss function, the parameters in the initial prediction model of drug-target interaction are updated according to the multiple predicted values ​​and sample labels to obtain the drug-target interaction prediction model. The preset loss function is set according to the Euclidean distance between the drug targets. In some embodiments of this application, the preset loss function can be determined according to the following formula (10):

[0114]

[0115] Where t = 1 when the sample label has an interaction, and -1 when the sample label has no interaction; γ represents a threshold; d(G1, G2) is the Euclidean distance between G1 and G2.

[0116] The loss function above indicates that when the drug molecule and the target protein have an interaction, it is desirable for d(G1, G2) < 1 - γ, and when the molecule and protein pair do not have an interaction, it is desirable for d(G1, G2) > 1 + γ.

[0117] In some other embodiments of this application, the drug map data 1 and the target map data 1 can be input into the initial prediction model of drug-target interaction to obtain the prediction value 1. Based on the above formula, the weights in the initial prediction model of drug-target interaction are updated according to the prediction value 1 and the sample label value 1 to obtain the drug-target interaction prediction model 1.

[0118] Then, drug map data 2 and target map data 2 are input into the initial drug-target interaction prediction model to obtain prediction value 2. Based on the above formula, the weights in drug-target interaction prediction model 1 are updated according to prediction value 2 and sample label value 1 to obtain drug-target interaction prediction model 2. Similarly, drug map data 3 and target map data 3 are input into drug-target interaction prediction model 2 to obtain the final drug-target interaction prediction model.

[0119] As can be seen from the above embodiments, when the drug-target interaction prediction model is used for drug prediction tasks of various diseases or viruses with new target proteins, only a small amount of labeled data is needed to obtain good accuracy results, transforming the original manual parameter tuning method that relies on machine learning engineers and experts into a method that can be applied to large-scale, replicable industrial expansion.

[0120] Figure 7 This illustration shows a flowchart of a method for predicting drug-target interactions according to yet another embodiment of the present application. Figure 7 As can be seen, the method for predicting drug-target interactions in this embodiment includes the following steps S701 to S721:

[0121] Step S701: Convert the drug molecule to be tested and the target protein molecule to be tested into drug map structure and target map structure, respectively.

[0122] Step S702: Based on the graph neural network, process the drug graph structure and the target graph structure respectively to obtain drug graph data and target graph data. The drug graph data includes multiple node vectors with atoms as nodes and multiple edge vectors with chemical bonds between atoms as edges. The target graph data includes multiple node vectors with amino acid residues as nodes and multiple edge vectors with spatial relationships between amino acid residues as edges.

[0123] Step S703: Obtain the training sample set; wherein, the training sample set contains multiple sets of paired drug map data and target map data, and the sample label of the training sample set is whether there is an interaction.

[0124] Step S704: Obtain the initial prediction model of drug-target interaction. The model includes a coding layer, a propagation layer, and an aggregation layer connected in sequence. The propagation layer includes an intra-graph information propagation layer and an inter-graph information propagation layer.

[0125] Step S705: Input the training sample set into the initial prediction model of drug-target interaction to obtain multiple prediction values.

[0126] Step S706: Based on the preset loss function, update the parameters in the initial prediction model of drug-target interaction according to multiple predicted values ​​and sample labels to obtain the drug-target interaction prediction model. The preset loss function is set according to the Euclidean distance between drug targets.

[0127] Step S707: The coding layer of the drug-target interaction prediction model includes a multilayer perceptron node coding layer and a multilayer perceptron edge coding layer; for any node vector in the drug graph data and target graph data, the node vector is input to the multilayer perceptron node coding layer to obtain the corresponding high-dimensional node vector; for any edge vector in the drug graph data and target graph data, the edge vector is input to the multilayer perceptron edge coding layer to obtain the corresponding high-dimensional edge vector.

[0128] Step S708: The intra-graph information propagation layer of the drug-target interaction prediction model includes an intra-graph information cascade layer and a first node network layer; based on the intra-graph information cascade layer, for any high-dimensional vector of the first target node in the drug graph data and target graph data, the high-dimensional vectors of each intra-graph neighbor node and the matching intra-graph neighbor edge high-dimensional vector of the high-dimensional vector of the first target node are concatenated with the high-dimensional vector of the first target node to obtain several intra-graph cascaded vectors;

[0129] Step S709: Based on the first node network layer, perform information accumulation processing on several intra-graph concatenated vectors to obtain the intra-graph node neighborhood information of the high-dimensional vector of the first target node.

[0130] Step S710: Based on the first node network layer, using the neighborhood information of the nodes in the graph, the neighborhood information is propagated to the high-dimensional vector of the first target node to obtain the first node graph feature vector corresponding to the high-dimensional vector of the first target node; the node graph feature vectors corresponding to the high-dimensional vectors of each node in the drug graph data form the drug node graph feature vector set; the node graph feature vectors corresponding to the high-dimensional vectors of each node in the target graph data form the target node graph feature vector set.

[0131] Step S711: The graph information propagation layer of the drug-target interaction prediction model includes a graph information cascade layer, a cross-graph information matching layer, and a second node network layer. Based on the graph information cascade layer, for any high-dimensional vector of the second target node in the drug graph data and target graph data, the high-dimensional vectors of each cross-graph neighbor node and the high-dimensional vector of the matched cross-graph neighbor edge are concatenated with the high-dimensional vector of the second target node to obtain several cross-graph cascaded vectors.

[0132] Step S712: Based on the distance metric function, obtain the distance information between the high-dimensional vector of the second target node and the high-dimensional vector of the cross-graph neighboring node to be matched.

[0133] Step S713: Based on the distance metric function, obtain the comprehensive distance information between the high-dimensional vector of the second target node and the high-dimensional vectors of each cross-graph neighbor node.

[0134] Step S714: Use the ratio between the distance information and the comprehensive distance information as the weight data between the high-dimensional vector of the second target node and the high-dimensional vector of the cross-graph neighboring node to be matched.

[0135] Step S715: Based on the weight data and the difference between the high-dimensional vector of the second target node and the high-dimensional vector of the cross-graph neighboring nodes to be matched, determine the cross-graph matching vector between the high-dimensional vector of the second target node and the high-dimensional vector of the cross-graph neighboring nodes to be matched. Return to step S712 and continue to execute steps S713, S714, and S715, repeating the process to perform node matching between the high-dimensional vector of each cross-graph neighboring node of the high-dimensional vector of the second target node and the high-dimensional vector of the second target node, thereby obtaining several cross-graph matching vectors.

[0136] Step S716: Based on the second node network layer, perform information accumulation processing on several cross-graph concatenation vectors and several cross-graph matching vectors respectively to obtain the cross-graph node neighborhood information and cross-graph node matching information of the high-dimensional vector of the second target node.

[0137] Step S717: Based on the second node network layer, cross-graph information propagation is performed on the high-dimensional vector of the second target node by utilizing cross-graph node neighborhood information and cross-graph node matching information to obtain the second node inter-graph feature vector corresponding to the high-dimensional vector of the second target node; the inter-graph feature vectors corresponding to the high-dimensional vectors of each node in the drug graph data form the drug graph inter-node feature vector set; the inter-graph feature vectors corresponding to the high-dimensional vectors of each node in the target graph data form the target graph inter-node feature vector set.

[0138] Step S718: The aggregation layer of the drug-target interaction prediction model includes a first processing layer, a second processing layer, and a similarity layer. The first and second processing layers are connected to the similarity layer, respectively. Based on the first processing layer, the feature vector set of drug nodes is processed to obtain the drug graph representation result.

[0139] Step S719: Based on the second processing layer, process the target node feature vector set to obtain the target map representation result.

[0140] Step S720: Based on the similarity layer, determine the similarity value between the drug map representation results and the target map representation results.

[0141] Step S721: Based on the similarity value and the preset threshold, determine the prediction result of drug-target interaction.

[0142] Figure 8 A schematic diagram of a drug target interaction prediction device according to an embodiment of this application is shown. The drug target interaction prediction device is equipped with a drug target interaction prediction model, which includes a coding layer, a propagation layer, and an aggregation layer connected in sequence. The propagation layer includes an intra-graph information propagation layer and an inter-graph information propagation layer.

[0143] The device 800 includes an acquisition module 801, an encoding module 802, a first propagation module 803, a second propagation module 804, and an aggregation module 805, specifically:

[0144] The acquisition module 801 is used to acquire drug map data and target map data, wherein the drug map data and the target map data respectively include multiple node vectors and multiple edge vectors.

[0145] The encoding module 802 is used to map each vector in the drug map data and the target map data into a high-dimensional vector based on the encoding layer. The high-dimensional vector includes node high-dimensional vectors and edge high-dimensional vectors.

[0146] The first propagation module 803 is used to propagate neighborhood information to each node's high-dimensional vector in the drug graph data and the target graph data based on the graph information propagation layer and the high-dimensional vector obtained by mapping, so as to obtain the drug node graph feature vector set and the target node graph feature vector set.

[0147] The second propagation module 804 is used to propagate cross-graph information based on the inter-graph information propagation layer and according to the high-dimensional vectors obtained by mapping, respectively, to the high-dimensional vectors of each node in the drug graph data and the target graph data, so as to obtain the set of feature vectors of nodes between the drug graph and the set of feature vectors of nodes between the target graph.

[0148] The aggregation module 805 is used to determine the drug graph representation result based on the drug node feature vector set, determine the target graph representation result based on the target node feature vector set, and determine the prediction result of drug-target interaction based on the similarity between the drug graph representation result and the target graph representation result, wherein the drug node feature vector set includes the intra-drug feature vector set and the inter-drug graph node feature vector set, and the target node feature vector set includes the intra-target node feature vector set and the inter-target graph node feature vector set.

[0149] In some embodiments of this application, the above-described apparatus further includes a graph processing module, which is used to convert the drug molecule to be tested and the target protein molecule to be tested into a drug graph structure and a target graph structure, respectively; and to process the drug graph structure and the target graph structure based on a graph neural network to obtain drug graph data and target graph data, wherein the drug graph data includes multiple node vectors with atoms as nodes and multiple edge vectors with inter-atomic chemical bonds as edges; the target graph data includes multiple node vectors with amino acid residues as nodes and multiple edge vectors with spatial relationships between amino acid residues as edges.

[0150] In some embodiments of this application, in the above-described apparatus, the encoding layer includes a multilayer perceptron node encoding layer and a multilayer perceptron edge encoding layer; the encoding module 802 is used to input any node vector in the drug map data and the target map data into the multilayer perceptron node encoding layer to obtain the corresponding high-dimensional node vector; and to input any edge vector in the drug map data and the target map data into the multilayer perceptron edge encoding layer to obtain the corresponding high-dimensional edge vector.

[0151] In some embodiments of this application, in the above-described apparatus, the intra-graph information propagation layer includes an intra-graph information concatenation layer and a first node network layer; the first propagation module 803 is configured to, based on the intra-graph information concatenation layer, concatenate the intra-graph neighboring node high-dimensional vectors and the matching intra-graph neighboring edge high-dimensional vectors of any first target node high-dimensional vector in the drug graph data and the target graph data with the first target node high-dimensional vector to obtain a plurality of intra-graph concatenated vectors; based on the first node network layer, perform information accumulation processing on the plurality of intra-graph concatenated vectors to obtain the intra-graph node neighborhood information of the first target node high-dimensional vector; and use the intra-graph node neighborhood information to propagate neighborhood information on the first target node high-dimensional vector to obtain the first node intra-graph feature vector corresponding to the first target node high-dimensional vector; the node intra-graph feature vectors corresponding to each node high-dimensional vector in the drug graph data form a drug node intra-graph feature vector set; the node intra-graph feature vectors corresponding to each node high-dimensional vector in the target graph data form a target node intra-graph feature vector set.

[0152] In some embodiments of this application, in the above-described apparatus, the inter-graph information propagation layer includes an inter-graph information concatenation layer, a cross-graph information matching layer, and a second node network layer; the second propagation module 804 is used, based on the inter-graph information concatenation layer, for any high-dimensional vector of a second target node in the drug graph data and the target graph data, to concatenate the high-dimensional vectors of each cross-graph neighboring node and the matched high-dimensional vectors of cross-graph neighboring edges of the high-dimensional vector of the second target node with the high-dimensional vector of the second target node, thereby obtaining a plurality of cross-graph concatenation vectors; based on the cross-graph information matching layer, to perform node matching on the high-dimensional vector of each cross-graph neighboring node of the high-dimensional vector of the second target node with the high-dimensional vector of the second target node, thereby obtaining a plurality of Cross-graph matching vectors; based on the second node network layer, information accumulation processing is performed on the plurality of cross-graph concatenated vectors and the plurality of cross-graph matching vectors respectively to obtain the cross-graph node neighborhood information and cross-graph node matching information of the high-dimensional vector of the second target node; and according to the cross-graph node neighborhood information and the cross-graph node matching information, cross-graph information propagation is performed on the high-dimensional vector of the second target node to obtain the second node inter-graph feature vector corresponding to the high-dimensional vector of the second target node; the inter-graph feature vectors corresponding to the high-dimensional vectors of each node in the drug graph data form the drug graph inter-graph feature vector set; the inter-graph feature vectors corresponding to the high-dimensional vectors of each node in the target graph data form the target graph inter-graph feature vector set.

[0153] In some embodiments of this application, in the above-described apparatus, the second propagation module 804 is configured to determine the cross-graph matching vector according to the following method: obtaining distance information between the high-dimensional vector of the second target node and the high-dimensional vector of the cross-graph neighboring nodes to be matched based on a distance metric function; obtaining comprehensive distance information between the high-dimensional vector of the second target node and each of the high-dimensional vectors of the cross-graph neighboring nodes based on the distance metric function; using the ratio between the distance information and the comprehensive distance information as weight data between the high-dimensional vector of the second target node and the high-dimensional vector of the cross-graph neighboring nodes to be matched; and determining the cross-graph matching vector between the high-dimensional vector of the second target node and the high-dimensional vector of the cross-graph neighboring nodes to be matched based on the weight data and the difference between the high-dimensional vector of the second target node and the high-dimensional vector of the cross-graph neighboring nodes to be matched.

[0154] In some embodiments of this application, in the above-described apparatus, the aggregation layer includes a first processing layer, a second processing layer, and a similarity layer, wherein the first processing layer and the second processing layer are respectively connected to the similarity layer; the aggregation module 805 is used to process the drug node feature vector set based on the first processing layer to obtain a drug graph representation result; to process the target node feature vector set based on the second processing layer to obtain a target graph representation result; to determine the similarity value between the drug graph representation result and the target graph representation result based on the similarity layer; and to determine the prediction result of drug-target interaction based on the similarity value and a preset threshold.

[0155] In some embodiments of this application, in the above-described apparatus, the drug-target interaction prediction model is trained using the following method: obtaining a training sample set; wherein the training sample set contains multiple pairs of drug map data and target map data, and the sample labels of the training sample set are whether there is an interaction; obtaining an initial drug-target interaction prediction model, the model including a coding layer, a propagation layer, and an aggregation layer connected in sequence, wherein the propagation layer includes an intra-graph information propagation layer and an inter-graph information propagation layer; inputting the training sample set into the initial drug-target interaction prediction model to obtain multiple predicted values; updating the parameters in the initial drug-target interaction prediction model based on a preset loss function, according to the multiple predicted values ​​and sample labels, to obtain the drug-target interaction prediction model, wherein the preset loss function is set based on the Euclidean distance between drug targets.

[0156] It should be noted that any of the above-mentioned drug-target interaction prediction devices can be used to implement the aforementioned drug-target interaction prediction methods, which will not be elaborated here.

[0157] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 9 As shown, at the hardware level, this electronic device includes a processor, and optionally also includes an internal bus, a network interface, and memory. The memory may include main memory, such as high-speed random-access memory (RAM), or it may include non-volatile memory, such as at least one disk drive. Of course, this electronic device may also include other hardware required for other business operations.

[0158] The processor, network interface, and memory can be interconnected via an internal bus, which can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 9 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0159] Memory is used to store programs. Specifically, programs may include program code, which includes computer operation instructions. Memory may include main memory and non-volatile memory, and provides instructions and data to the processor.

[0160] The processor reads the corresponding computer program from non-volatile memory into memory and then runs it, forming a predictive device for drug-target interactions at the logical level. The processor executes the program stored in memory and specifically performs the aforementioned method.

[0161] The processor may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in the memory, and the processor reads the information in the memory and, in conjunction with its hardware, completes the steps of the above method.

[0162] This electronic device can execute the drug-target interaction prediction method provided in several embodiments of this application, and realize the drug-target interaction prediction device in Figure 8 The functions of the embodiments shown are not described in detail here.

[0163] This application also proposes a computer-readable storage medium that stores one or more programs, the programs including instructions that, when executed by an electronic device including multiple applications, enable the electronic device to perform the drug target interaction prediction method provided in several embodiments of this application.

[0164] 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.

[0165] 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.

[0166] 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.

[0167] 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.

[0168] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0169] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0170] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0171] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0172] 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.

[0173] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for predicting drug-target interactions, characterized in that, The method is based on a drug-target interaction prediction model, which includes a coding layer, a propagation layer, and an aggregation layer connected in sequence. The propagation layer includes an intra-graph information propagation layer and an inter-graph information propagation layer. The method includes: Acquire drug map data and target map data, wherein the drug map data and the target map data each include multiple node vectors and multiple edge vectors; Based on the encoding layer, each vector in the drug map data and the target map data is mapped to a high-dimensional vector, which includes node high-dimensional vectors and edge high-dimensional vectors. Based on the graph information propagation layer, neighborhood information propagation is performed on the high-dimensional vectors of each node in the drug graph data and the target graph data according to the high-dimensional vectors obtained by mapping, to obtain the drug node graph feature vector set and the target node graph feature vector set. Based on the inter-graph information propagation layer, cross-graph information propagation is performed on the high-dimensional vectors of each node in the drug graph data and the target graph data according to the high-dimensional vectors obtained by mapping, so as to obtain the set of feature vectors of nodes between the drug graph and the set of feature vectors of nodes between the target graph. Based on the aggregation layer, the drug graph representation result is determined according to the drug node feature vector set, the target graph representation result is determined according to the target node feature vector set, and the prediction result of drug-target interaction is determined according to the similarity between the drug graph representation result and the target graph representation result; wherein, the drug node feature vector set includes the intra-drug feature vector set and the inter-drug graph node feature vector set, and the target node feature vector set includes the intra-target node feature vector set and the inter-target graph node feature vector set.

2. The method according to claim 1, characterized in that, Prior to the step of acquiring drug mapping data and target mapping data, the method further includes: The drug molecule to be tested and the target protein molecule to be tested are converted into drug map structure and target map structure, respectively; Based on graph neural networks, the drug graph structure and target graph structure are processed respectively to obtain drug graph data and target graph data. The drug graph data includes multiple node vectors with atoms as nodes and multiple edge vectors with inter-atomic chemical bonds as edges. The target graph data includes multiple node vectors with amino acid residues as nodes and multiple edge vectors with spatial relationships between amino acid residues as edges.

3. The method according to claim 1, characterized in that, The encoding layer includes a multilayer perceptron node encoding layer and a multilayer perceptron edge encoding layer; based on the encoding layer, each vector in the drug map data and the target map data is mapped into multiple node high-dimensional vectors and multiple edge high-dimensional vectors, including: For any node vector in the drug map data and the target map data, the node vector is input into the node encoding layer of the multilayer perceptron to obtain the corresponding high-dimensional node vector; For any edge vector in the drug map data and the target map data, the edge vector is input into the edge encoding layer of the multilayer perceptron to obtain the corresponding high-dimensional edge vector.

4. The method according to claim 1, characterized in that, The intra-graph information propagation layer includes an intra-graph information cascading layer and a first node network layer; Based on the graph-based information propagation layer, neighborhood information propagation is performed on the high-dimensional vectors of each node in the drug graph data and the target graph data according to the multiple high-dimensional vectors of nodes and edges obtained by mapping, respectively, to obtain the set of intra-graph feature vectors of drug nodes and the set of intra-graph feature vectors of target nodes, including: Based on the graph information cascade layer, for any high-dimensional vector of a first target node in the drug graph data and the target graph data, the high-dimensional vectors of each graph neighbor node and the matching graph neighbor edge high-dimensional vector of the first target node high-dimensional vector are concatenated with the first target node high-dimensional vector to obtain several graph cascade vectors. Based on the first node network layer, information accumulation processing is performed on the plurality of intra-graph concatenated vectors to obtain the intra-graph node neighborhood information of the high-dimensional vector of the first target node; and using the intra-graph node neighborhood information, neighborhood information propagation is performed on the high-dimensional vector of the first target node to obtain the first node intra-graph feature vector corresponding to the high-dimensional vector of the first target node. The intra-node feature vectors corresponding to the high-dimensional vectors of each node in the drug graph data form the intra-node feature vector set of the drug node graph; the intra-node feature vectors corresponding to the high-dimensional vectors of each node in the target graph data form the intra-node feature vector set of the target node graph.

5. The method according to claim 1, characterized in that, The inter-graph information propagation layer includes an inter-graph information cascading layer, a cross-graph information matching layer, and a second node network layer. Based on the inter-graph information propagation layer, according to the high-dimensional vectors obtained from the mapping, cross-graph information propagation is performed on the high-dimensional vectors of each node in the drug graph data and the target graph data, respectively, to obtain the drug graph inter-node feature vector set and the target graph inter-node feature vector set, including: Based on the graph information concatenation layer, for any high-dimensional vector of a second target node in the drug graph data and the target graph data, the high-dimensional vectors of each cross-graph neighbor node and the matching high-dimensional vectors of the cross-graph neighbor edges of the high-dimensional vector of the second target node are concatenated with the high-dimensional vector of the second target node to obtain several cross-graph concatenation vectors. Based on the cross-graph information matching layer, the high-dimensional vectors of each cross-graph neighboring node of the high-dimensional vector of the second target node are matched with the high-dimensional vector of the second target node to obtain several cross-graph matching vectors. Based on the second node network layer, information accumulation processing is performed on the plurality of cross-graph concatenated vectors and the plurality of cross-graph matching vectors respectively to obtain the cross-graph node neighborhood information and cross-graph node matching information of the high-dimensional vector of the second target node; and cross-graph information propagation is performed on the high-dimensional vector of the second target node using the cross-graph node neighborhood information and the cross-graph node matching information to obtain the second node inter-graph feature vector corresponding to the high-dimensional vector of the second target node. The node-to-node feature vectors corresponding to the high-dimensional vectors of each node in the drug graph data form the drug graph node feature vector set; the node-to-node feature vectors corresponding to the high-dimensional vectors of each node in the target graph data form the target graph node feature vector set.

6. The method according to claim 5, characterized in that, The cross-graph matching vector is determined according to the following method: Based on the distance metric function, the distance information between the high-dimensional vector of the second target node and the high-dimensional vector of the cross-graph neighboring node to be matched is obtained; Based on the distance metric function, the comprehensive distance information between the high-dimensional vector of the second target node and the high-dimensional vectors of each of the cross-graph neighboring nodes is obtained; The ratio between the distance information and the comprehensive distance information is used as the weight data between the high-dimensional vector of the second target node and the high-dimensional vector of the cross-graph neighboring node to be matched. Based on the weight data and the difference between the high-dimensional vector of the second target node and the high-dimensional vector of the cross-graph neighboring node to be matched, a cross-graph matching vector is determined between the high-dimensional vector of the second target node and the high-dimensional vector of the cross-graph neighboring node to be matched.

7. The method according to claim 1, characterized in that, The aggregation layer includes a first processing layer, a second processing layer, and a similarity layer, wherein the first processing layer and the second processing layer are respectively connected to the similarity layer; Based on the aggregation layer, the drug graph representation result is determined according to the drug node feature vector set, the target graph representation result is determined according to the target node feature vector set, and the prediction result of drug-target interaction is determined based on the similarity between the drug graph representation result and the target graph representation result, including: Based on the first processing layer, the feature vector set of the drug nodes is processed to obtain the drug graph representation result; Based on the second processing layer, the target node feature vector set is processed to obtain the target map representation result; Based on the similarity layer, the similarity value between the drug map representation result and the target map representation result is determined; Based on the similarity value and the preset threshold, the prediction result of drug-target interaction is determined.

8. The method according to any one of claims 1 to 7, characterized in that, The drug-target interaction prediction model was trained using the following method: Obtain a training sample set; wherein the training sample set contains multiple sets of paired drug map data and target map data, and the sample label of the training sample set is whether there is an interaction; An initial prediction model for drug-target interaction is obtained, the model comprising a coding layer, a propagation layer, and an aggregation layer connected in sequence, wherein the propagation layer includes an intra-graph information propagation layer and an inter-graph information propagation layer; The training sample set is input into the initial prediction model for drug-target interaction to obtain multiple predicted values; Based on a preset loss function, the parameters in the initial prediction model of drug-target interaction are updated according to the multiple predicted values ​​and sample labels to obtain the drug-target interaction prediction model. The preset loss function is set according to the Euclidean distance between drug targets.

9. A device for predicting drug-target interactions, characterized in that, The device is equipped with a drug-target interaction prediction model, which includes a coding layer, a propagation layer, and an aggregation layer connected in sequence. The propagation layer includes an intra-graph information propagation layer and an inter-graph information propagation layer. The device includes: An acquisition module is used to acquire drug map data and target map data, wherein the drug map data and the target map data respectively include multiple node vectors and multiple edge vectors; The encoding module is used to map each vector in the drug map data and the target map data into a high-dimensional vector based on the encoding layer. The high-dimensional vector includes node high-dimensional vectors and edge high-dimensional vectors. The first propagation module is used to propagate neighborhood information to each node's high-dimensional vector in the drug graph data and the target graph data based on the graph information propagation layer and the high-dimensional vector obtained by mapping, so as to obtain the drug node graph feature vector set and the target node graph feature vector set. The second propagation module is used to propagate cross-graph information based on the inter-graph information propagation layer and according to the high-dimensional vectors obtained by mapping, respectively, to the high-dimensional vectors of each node in the drug graph data and the target graph data, so as to obtain the set of feature vectors of nodes between the drug graph and the set of feature vectors of nodes between the target graph. An aggregation module is used to determine drug graph representation results based on the aggregation layer, according to the drug node feature vector set, to determine target graph representation results based on the target node feature vector set, and to determine the prediction results of drug-target interaction based on the similarity between the drug graph representation results and the target graph representation results; wherein, the drug node feature vector set includes the intra-drug feature vector set and the inter-drug graph node feature vector set, and the target node feature vector set includes the intra-target node feature vector set and the inter-target graph node feature vector set.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the method for predicting drug-target interactions as described in any one of claims 1 to 8.

Images