A method and apparatus for predicting relationships of multi-source information using a prediction model

Abstract

The embodiment of the present disclosure provides a method and device for predicting the relationship of multi-source information using a prediction model, the method comprising: inputting first multi-source information and second multi-source information into the prediction model and outputting the relationship of the first multi-source information and the second multi-source information through the prediction model; wherein the first multi-source information is a drug, and the second multi-source information is a disease or a target; the first multi-source information is a target, and the second multi-source information is a target; the prediction model comprises an encoder and a decoder; the encoder is used to calculate the embedding representation of a first node corresponding to the first multi-source information and the embedding representation of a second node corresponding to the second multi-source information in a predetermined network, and the predetermined network at least comprises one of the following: a drug-disease association relationship network, a drug-target interaction network, and a disease-target association relationship network; and the decoder is used to predict the association relationship or the interaction relationship of the first multi-source information and the second multi-source information according to the embedding representation of the first node and the embedding representation of the second node.

Description

Technical Field

[0001] This disclosure relates to the fields of artificial intelligence and drug repositioning, and in particular to a method and apparatus for predicting multi-source information relationships using a predictive model. Background Technology

[0002] Current drug relocation methods typically use relatively simple data, capturing only partial information; or they may use rich data, but based on various features, they ignore network structure information. The emergence of graph convolutional neural networks (GCNs) has utilized network structure information, but it simply processes various types of information into a large heterogeneous graph, completely ignoring the domain-specific features.

[0003] Existing drug relocation methods can be categorized into two types: those based on similarity calculation and those based on network structure. However, both have some problems: For those based on similarity calculation, the assumption is that similar drugs apply to similar diseases, and network structure information is not obtained; for those based on network structure, domain-specific features are ignored, and all nodes are projected together into the same vector space. Furthermore, many methods rely heavily on predefined features such as drug similarity, while ignoring the data that the disease domain can provide.

[0004] It is evident that existing technologies lack accurate methods for drug repositioning, which hinders the development of medicine or pharmacology. Summary of the Invention

[0005] In view of this, the present disclosure proposes a method, apparatus, storage medium and electronic device for predicting multi-source information relationships using a predictive model, in order to solve the following problems of the prior art: the prior art does not have a relatively accurate method for drug repositioning, which is not conducive to the development of medicine or pharmacology.

[0006] On one hand, this disclosure proposes a method for predicting the relationship between multi-source information using a prediction model, comprising: inputting first multi-source information and second multi-source information into a prediction model, and outputting the relationship between the first multi-source information and the second multi-source information through the prediction model; wherein the first multi-source information and the second multi-source information include at least one of the following: the first multi-source information is a drug and the second multi-source information is a disease; the first multi-source information is a drug and the second multi-source information is a target; or the first multi-source information is a target and the second multi-source information is a target.

[0007] The prediction model includes an encoder and a decoder; the encoder is used to calculate the embedding representation of a first node corresponding to the first multi-source information and the embedding representation of a second node corresponding to the second multi-source information in a predetermined network, wherein the predetermined network includes at least one of the following: a drug-disease association network, a drug-target interaction network, and a disease-target association network; the decoder is used to predict the association or interaction relationship between the first multi-source information and the second multi-source information based on the embedding representation of the first node and the embedding representation of the second node.

[0008] In some embodiments, the training method of the encoder in the prediction model includes: extracting network topology representations of drugs and targets from a drug target association network using a deep walk algorithm, and learning the sequence representation of each node in the drug target association network according to random walk theory to obtain the embedding representation of a first drug node and the embedding representation of a first target node in the drug target association network; extracting network topology representations of diseases and targets from a disease target association network using a deep walk algorithm, and learning the sequence representation of each node in the disease target association network according to random walk theory to obtain the embedding representation of a first disease node and the embedding representation of a second target node in the disease target association network; using the first target node's... The embedding representations of the third drug node, the second target node, the first drug node, and the first disease node are used as inputs. A graph convolutional neural network algorithm and a bipartite graph attention convolutional layer are used to determine the embedding representations of the third drug node, the third disease node, and the fifth target node. Specifically, the embedding representations of the third drug node and the third disease node are used to predict drug-disease associations; the embedding representations of the third drug node are used to predict drug-drug interactions; the embedding representations of the third drug node and the fifth target node are used to predict drug-target interactions; and the embedding representations of the fifth target node are used to predict target-target interactions.

[0009] In some embodiments, the step of using the embedding representations of the first target node, the second target node, the first drug node, and the first disease node as input, and using a graph convolutional neural network algorithm and a bipartite graph attention convolutional layer to determine the embedding representations of the third drug node, the third disease node, and the fifth target node includes:

[0010] The embedding representation of the first target node is used as input, and the target-to-target interaction network is used as the adjacency matrix. The graph convolutional neural network algorithm is used to learn the embedding representation of the third target node.

[0011] Using the embedding representation of the second target node as input, and the target-to-target interaction network as the adjacency matrix, a graph convolutional neural network algorithm is used to learn the embedding representation of the fourth target node. Using the embedding representations of the first drug node and the first disease node as input, and the drug-disease association network as the adjacency matrix, a graph convolutional neural network algorithm is used to learn the embedding representations of the second drug node and the second disease node. Using the embedding representations of the fourth target node and the second drug node as input, and the drug-target interaction network as the adjacency matrix, a graph convolutional neural network algorithm is used to learn the embedding representations of the fourth target node and the second drug node. The bipartite graph attention convolutional layer learns the embedding representation of the third drug node; taking the embedding representations of the third target node and the second disease node as input, and the disease target association network as the adjacency matrix, the bipartite graph attention convolutional layer learns the embedding representation of the third disease node; the embedding representations of the third target node and the fourth target node learn their respective node weight parameters through an attention mechanism; and the embedding representation of the fifth target node is learned by weighted summation based on the embedding representation of the third target node and its weight parameters, and the embedding representation of the fourth target node and its weight parameters.

[0012] In some embodiments, the training method of the decoder in the prediction model includes: taking the embedding representation of the third drug node and the embedding representation of the third disease node as input, and using a predetermined prediction probability score matrix to learn the drug-disease association; taking the embedding representation of the third drug node and the embedding representation of the fifth target node as input, and using the predetermined prediction probability score matrix to learn the drug-target interaction; taking the embedding representation of the fifth target node as input, and using the predetermined prediction probability score matrix to learn the association between targets; wherein the training loss is adopted to learn the association by minimizing the weighted binary cross-entropy loss.

[0013] On the other hand, embodiments of this disclosure propose an apparatus for predicting multi-source information relationships using a prediction model, comprising: an input module for inputting first multi-source information and second multi-source information into a prediction model, and outputting the relationship between the first multi-source information and the second multi-source information through the prediction model; wherein the first multi-source information and the second multi-source information include at least one of the following: the first multi-source information is a drug and the second multi-source information is a disease; the first multi-source information is a drug and the second multi-source information is a target; the first multi-source information is a target and the second multi-source information is a target; the prediction model includes: an encoder and a decoder; the encoder is used to calculate the embedding representation of a first node corresponding to the first multi-source information and the embedding representation of a second node corresponding to the second multi-source information in a predetermined network, wherein the predetermined network includes at least one of the following: a drug-disease association network, a drug-target interaction network, and a disease-target association network; the decoder is used to predict the association or interaction relationship between the first multi-source information and the second multi-source information based on the embedding representation of the first node and the embedding representation of the second node.

[0014] In some embodiments, the model further includes: an encoder training module for training the encoder in the prediction model, wherein the encoder training method includes: extracting the network topology representation of drugs and targets from the drug target association network using a deep walk algorithm, and learning the sequence representation of each node in the drug target association network according to the random walk theory, to obtain the embedding representation of the first drug node and the embedding representation of the first target node in the drug target association network.

[0015] A deep walk algorithm is used to extract the network topology representation of diseases and targets from the disease target association network. Following random walk theory, the sequence representation of each node in the disease target association network is learned to obtain the embedding representations of the first disease node and the second target node. Using the embedding representations of the first target node, the second target node, the first drug node, and the first disease node as input, a graph convolutional neural network algorithm and a bipartite graph attention convolutional layer are used to determine the embedding representations of the third drug node, the third disease node, and the fifth target node. Specifically, the embedding representations of the third drug node and the third disease node are used to predict drug-disease associations; the embedding representations of the third drug node are used to predict drug-drug interactions; the embedding representations of the third drug node and the fifth target node are used to predict drug-target interactions; and the embedding representations of the fifth target node are used to predict target-target interactions.

[0016] In some embodiments, when determining the embedding representations of the third drug node, the third disease node, and the fifth target node, the encoder training module is specifically configured to: use the embedding representation of the first target node as input, the target-to-target interaction network as an adjacency matrix, and use a graph convolutional neural network algorithm to learn the embedding representation of the third target node; use the embedding representation of the second target node as input, the target-to-target interaction network as an adjacency matrix, and use a graph convolutional neural network algorithm to learn the embedding representation of the fourth target node; use the embedding representations of the first drug node and the first disease node as input, the drug-disease association network as an adjacency matrix, and use a graph convolutional neural network algorithm to learn the embedding representations of the second drug node and the second target node. The embedding representation of the disease node is obtained by: taking the embedding representations of the fourth target node and the second drug node as input, using the drug target interaction network as the adjacency matrix, and learning the embedding representation of the third drug node using a bipartite graph attention convolutional layer; taking the embedding representations of the third target node and the second disease node as input, using the disease target association network as the adjacency matrix, and learning the embedding representation of the third disease node using a bipartite graph attention convolutional layer; learning the weight parameters of each node by using the embedding representations of the third target node and the fourth target node through an attention mechanism, and learning the embedding representation of the fifth target node by weighted summation based on the embedding representation of the third target node and its weight parameters, and the embedding representation of the fourth target node and its weight parameters.

[0017] In some embodiments, the model further includes: a decoder training module for training the decoder in the prediction model, wherein the training method of the decoder includes: taking the embedding representation of the third drug node and the embedding representation of the third disease node as input, and using a predetermined prediction probability score matrix to learn the drug-disease association; taking the embedding representation of the third drug node and the embedding representation of the fifth target node as input, and using the predetermined prediction probability score matrix to learn the drug-target interaction; taking the embedding representation of the fifth target node as input, and using the predetermined prediction probability score matrix to learn the association between targets; wherein the training loss is adopted to learn the association by minimizing the weighted binary cross-entropy loss.

[0018] On the other hand, embodiments of this disclosure provide a storage medium storing a computer program that, when executed by a processor, implements the methods provided in any embodiment of this disclosure.

[0019] On the other hand, embodiments of this disclosure provide an electronic device, including at least a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the method provided in any embodiment of this disclosure when executing the computer program in the memory.

[0020] This disclosure uses a trained prediction model to predict the relationship between two known multi-source information sources. Given existing known drug-disease association networks, drug-target interaction networks, and disease-target association networks, it learns the relationships between these networks and then determines the relationships between unknown drugs, diseases, and targets based on the known relationships between them. This not only enables drug relocation but also allows for the study of diseases and targets, with high prediction accuracy. It provides a clearer approach for scientific research, saves more research resources, and benefits the development of medicine or pharmacology. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart of a method for predicting multi-source information relationships using a predictive model, provided in the first embodiment of this disclosure;

[0023] Figure 2 This is a diagram illustrating the heterogeneous information distribution fusion strategy provided in the first embodiment of this disclosure.

[0024] Figure 3 A flowchart illustrating node embedding learning provided in the first embodiment of this disclosure;

[0025] Figure 4 This is a schematic diagram of the apparatus for predicting multi-source information relationships using a prediction model, provided in the second embodiment of this disclosure. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the described embodiments of this disclosure without creative effort are within the scope of protection of this disclosure.

[0027] Unless otherwise defined, the technical or scientific terms used in this disclosure shall have the ordinary meaning understood by one of ordinary skill in the art to which this disclosure pertains. The terms “first,” “second,” and similar terms used in this disclosure do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as “comprising” or “including” mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as “connected” or “linked” are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as “upper,” “lower,” “left,” and “right” are used only to indicate relative positional relationships, and these relative positional relationships may change accordingly when the absolute position of the described objects changes.

[0028] To keep the following description of the embodiments of this disclosure clear and concise, detailed descriptions of known functions and known components are omitted.

[0029] The first embodiment of this disclosure provides a method for predicting multi-source information relationships using a predictive model. The process of this method is as follows: Figure 1 As shown, steps S101 to S102 are included:

[0030] S101, input the first multi-source information and the second multi-source information into the prediction model;

[0031] S102, output the relationship between the first multi-source information and the second multi-source information through the prediction model; wherein, the first multi-source information and the second multi-source information include at least one of the following: the first multi-source information is a drug and the second multi-source information is a disease; the first multi-source information is a drug and the second multi-source information is a target; the first multi-source information is a target and the second multi-source information is a target.

[0032] The aforementioned prediction model includes: an encoder and a decoder;

[0033] The encoder is used to compute the embedding representation of the first node corresponding to the first multi-source information and the embedding representation of the second node corresponding to the second multi-source information in a predetermined network, wherein the predetermined network includes at least one of the following: drug-disease association network, drug-target interaction network, and disease-target association network.

[0034] The decoder is used to predict the association or interaction relationship between the first multi-source information and the second multi-source information based on the embedding representation of the first node and the embedding representation of the second node.

[0035] This disclosure uses a trained prediction model to predict the relationship between two known multi-source information sources. Given existing known drug-disease association networks, drug-target interaction networks, and disease-target association networks, it learns the relationships between these networks and then determines the relationships between unknown drugs, diseases, and targets based on the known relationships between them. This not only enables drug relocation but also allows for the study of diseases and targets, with high prediction accuracy. It provides a clearer approach for scientific research, saves more research resources, and benefits the development of medicine or pharmacology.

[0036] The first and second multi-source information input into the prediction model have an unknown relationship. Therefore, if the first multi-source information is a drug, then the second multi-source information is a disease or a target. If the first multi-source information is a target, then the second multi-source information can also be a target.

[0037] The training methods for the encoder in the above prediction model include:

[0038] I. The deep walk algorithm is used to extract the network topology representation of drugs and targets from the drug target association network, and the sequence representation of each node in the drug target association network is learned according to the random walk theory to obtain the embedding representation of the first drug node and the embedding representation of the first target node in the drug target association network.

[0039] The deep walk algorithm is used to extract the network topology representation of diseases and targets from the disease target association network, and the sequence representation of each node in the disease target association network is learned according to the random walk theory to obtain the embedding representation of the first disease node and the embedding representation of the second target node in the disease target association network.

[0040] 2. Using the embedding representations of the first target node, the second target node, the first drug node, and the first disease node as input, the graph convolutional neural network algorithm and the bipartite graph attention convolutional layer are used to determine the embedding representations of the third drug node, the third disease node, and the fifth target node.

[0041] Specifically, the embedding representations of the third drug node and the third disease node are used to predict drug-disease association; the embedding representations of the third drug node are used to predict drug-drug interactions; the embedding representations of the third drug node and the fifth target node are used to predict drug-target interactions; and the embedding representations of the fifth target node are used to predict target-target interactions.

[0042] Preferably, the embedding representations of the first target node, the second target node, the first drug node, and the first disease node are used as input, and a graph convolutional neural network algorithm and a bipartite graph attention convolutional layer are used to determine the embedding representations of the third drug node, the third disease node, and the fifth target node, including steps 1-4:

[0043] 1. Using the embedding representation of the first target node as input, and the target-to-target interaction network as the adjacency matrix (the adjacency matrix that must be used for the algorithm implementation, and the algorithm uses this adjacency matrix), the graph convolutional neural network algorithm is used to learn the embedding representation of the third target node; using the embedding representation of the second target node as input, and the target-to-target interaction network as the adjacency matrix, the graph convolutional neural network algorithm is used to learn the embedding representation of the fourth target node;

[0044] 2. Using the embedding representations of the first drug node and the first disease node as input, and the drug-disease association network as the adjacency matrix, the graph convolutional neural network algorithm is used to learn the embedding representations of the second drug node and the second disease node.

[0045] 3. Using the embedding representations of the fourth target node and the second drug node as input, and the drug target interaction network as the adjacency matrix, a bipartite graph attention convolutional layer is used to learn the embedding representation of the third drug node; using the embedding representations of the third target node and the second disease node as input, and the disease target association network as the adjacency matrix, a bipartite graph attention convolutional layer is used to learn the embedding representation of the third disease node.

[0046] 4. The embedding representations of the third target node and the fourth target node are learned through an attention mechanism to obtain the weight parameters of their respective nodes. The embedding representation of the fifth target node is learned by weighted summation based on the embedding representation of the third target node and its weight parameters, and the embedding representation of the fourth target node and its weight parameters.

[0047] The training methods for the decoder in the above prediction model include:

[0048] The embedding representations of the third drug node and the third disease node are used as inputs, and a predetermined prediction probability score matrix is ​​used to learn the drug-disease association. The embedding representations of the third drug node and the fifth target node are used as inputs, and a predetermined prediction probability score matrix is ​​used to learn the drug-target interaction. The embedding representation of the fifth target node is used as input, and a predetermined prediction probability score matrix is ​​used to learn the association between targets. The training loss is to minimize the weighted binary cross-entropy loss to learn the association.

[0049] The following is combined with Figure 2 and Figure 3 The training process of the above prediction model will be explained in detail. Figure 2 A diagram illustrating the heterogeneous information distribution fusion strategy. Figure 3 This is a flowchart for node embedding learning. The terms "node embedding," "node embedding representation," and other terms related to "embedding" used in the following process are alternative descriptions of "node embedding representation," and will not be further explained.

[0050] Information networks are ubiquitous in the real world, including social networks, citation networks, e-commerce websites, recommendation sites, biological networks, and medical systems. With the rapid development of information technology, information networks are often massive in scale, involving millions or even hundreds of millions of nodes. Each node contains complex attributes and characteristics; for example, a social network may have tens of thousands of posts and users. Simultaneously, data in massive networks is often sparse. For instance, in social media, a large proportion of words or attributes are typically only useful to a small number of users. The ever-increasing network size and data sparsity pose challenges to network analysis methods, requiring them to be scalable and effectively address the data sparsity problem.

[0051] Network representation learning (also known as representation learning or embedding learning) aims to embed nodes into a low-dimensional vector space using information such as the network's topology and node content, while preserving the network's inherent structural and content features. Currently, much research in network representation learning focuses on homogeneous information networks containing only one type of node and one type of edge. Models such as DeepWalk and LINE (based on the neighborhood similarity assumption) learn based on the network's topology. Homogeneous information networks are a simplification of real-world information networks, often extracting only a portion of the real-world network's information or failing to distinguish the differences in objects and relationships within the real-world network, resulting in incomplete or lost information.

[0052] Heterogeneous information networks (also known as heterogeneous networks) contain various types of nodes and edges. Different types of nodes represent different objects, and different types of edges represent different interaction relationships between objects. For example, a disease control network includes objects such as doctors, patients, diseases, and drugs. Doctors and patients are connected through a "treatment / being treated" relationship, and patients and drugs are connected through a "taking / being taken" relationship. Different types of objects and edges have different semantics, distinguishing the differences between data objects and their relationships in the real world. At the same time, information can be propagated between different objects through different types of relationships in the network, reflecting the similarity or influence between objects. Because heterogeneous information networks contain more comprehensive and richer structural and semantic information, representation learning for heterogeneous information networks can not only effectively alleviate the problems of high dimensionality and sparsity of network data, but also integrate different types of heterogeneous information in the network, making the learned feature representations more meaningful and valuable. However, due to the special characteristics of heterogeneous information networks, representation learning methods for homogeneous information networks cannot be directly applied to heterogeneous information networks. The heterogeneity of nodes and edges in a network requires representation learning to not only comprehensively extract and utilize the network's multi-dimensional information, but also to effectively integrate this information, while capturing the embedding uncertainty caused by various attributes as much as possible.

[0053] Given a graph G = (V, E), where V represents the set of nodes and E represents the set of edges, each node v in the node set belongs to an object type in the object set O, where O = {type1, type2, ..., typei, ...}; each edge in the edge set belongs to a relation type in the relation type set R, where R = {..., typei-typej, ...} (i and j are types in the object set, and i can be equal to j).

[0054] The object sets involved in this disclosure are of three types: drug node set, disease node set, and target node set; the relationship type sets are {drug-disease, drug-target, disease-target, target-target}. Each relationship type set is a small network, consisting of three bipartite networks and one homogeneous network, namely, the drug-disease association relationship bipartite network, the drug-target interaction relationship bipartite network, the disease-target association relationship bipartite network, and the target-target interaction relationship homogeneous network.

[0055] Based on this, the present disclosure proposes a node representation method for this type of data; or any other heterogeneous network that can be abstracted into this structure can use the present disclosure to realize the representation learning and downstream analysis application of network nodes, all of which are within the equivalent protection scope of the present disclosure.

[0056] The learning process for node embedding representations can be divided into five parts (i.e., all the contents included in steps one and two above), specifically including the following steps 1-5:

[0057] 1) The embeddings of the first drug node and the first target node are learned through a drug-target interaction network, and the embeddings of the first disease node and the second target node are learned through a disease-target association network. The first drug node embedding and the first target node embedding learn the structural information of the drug-target interaction relationship; the first disease node embedding and the second target node embedding learn the structural information of the disease-target association relationship.

[0058] 2) Using the first and second target node embeddings as inputs, the third and fourth target node embeddings are learned through the target-to-target interaction network. The third and fourth target node embeddings update the structural information of target-to-target interactions based on the original structural information of drug-to-target interactions and disease-to-target associations.

[0059] 3) Using the first drug node embedding and the first disease node embedding as input, the second drug node embedding and the second disease node embedding are learned through the drug-disease association network. The second drug node embedding and the second disease node embedding update the structural information of the drug-disease interaction relationship based on the original structural information of the drug-target interaction relationship and the structural information of the disease-target association relationship.

[0060] 4) Using the fourth target node embedding and the second drug node embedding as input, the third drug node embedding is learned through the drug target interaction network. Based on the previously included features, the structural information of the drug-target interaction relationship is updated. Using the third target node embedding and the second disease node embedding as input, the third disease node embedding is learned through the disease target association network. Based on the previously included features, the structural information of the disease-target association relationship is updated.

[0061] 5) The third and fourth target node embeddings are learned by the attention mechanism to obtain the weight parameters of their respective nodes, and then the weighted sum is used to learn the fifth target node embedding.

[0062] Through the above steps, we learned the embedding of the third drug node, the embedding of the third disease node, and the embedding of the fifth target node.

[0063] Third drug node embedding and third disease node embedding can be used to predict new drug-disease associations; third drug node embeddings can predict new drug-drug interactions; third drug node embedding and fifth target node embedding can predict new drug-target interactions; and fifth target node embeddings can predict new target-target interactions. The term "new" in the above context refers to relationships unknown prior to the application of the embodiments disclosed herein.

[0064] The specific implementation steps of the solution are as follows (the labels appearing in the following formulas are only for explanation of the content in their corresponding parts):

[0065] I. Encoding process (i.e., the embedding representation learning process of drug, disease, and target nodes)

[0066] 1) The embedding representation learning process of the first drug node, the first disease node, and the first and second target nodes:

[0067] a) First, the embedding representation of the first drug node and the first target node:

[0068] The practical problem is processing the network of relationships between the drug node set and the target node set, denoted as G. rt Let the number of drug nodes be u1 and the number of target nodes be v. Here, we can adopt various methods to learn their low-dimensional node embeddings, such as matrix factorization, random walk, graph neural network and other algorithms.

[0069] A rt Let A be the adjacency matrix of the network connecting drug nodes and target nodes. If a node in the drug node set is associated with a node in the target node set, then A... rt (i, j) = 1, otherwise, A rt (i, j) = 0.

[0070] The network topology information observed in a binary network is more complex because it represents the relationship between pairs of nodes. This scheme uses DeepWalk to extract the network representations of drug nodes and target nodes from a given binary network, which is an effective graph representation learning algorithm.

[0071] DeepWalk takes pairs of nodes as input and learns the sequence representation of each node according to the random walk theory. The output of DeepWalk is the node representation vector obtained from the skip-gram model. In a graph G = (V, E), suppose a path from v0 to v... i-1 (1≤i≤|V|), where |V| is the number of nodes, and the random traversal sequence is denoted as {v0,…,v...}. i-1The random walk follows the rule A, meaning it only visits nodes connected by edges, and defines the next node to reach v. i The probability of a node is:

[0072] Pr(v i |(v0,…,v i-1 ))

[0073] Our goal is to obtain the vector representation of each node in V, and a mapping function Φ: v∈V→R |V|×d Specifically, it is a |V|×d matrix;

[0074] Pr(v i |(Φ(v0),…,Φ(v i-1 )))

[0075]

[0076] In the above equation, w determines v. i The range of neighboring nodes. By solving the minimization problem of the above equation, we can obtain Φ(V)∈R. |V|×d |V|=u1+v represents the network of all nodes, using matrix X. rt To represent the vector representation of drugs and targets, X rt =Φ(V), therefore we have:

[0077]

[0078]

[0079] Among them, X rt1 For the node embedding of the first drug node, X rt2 Node embedding for the first target node.

[0080] b) Embedded representation of the first disease node and the second target node:

[0081] The practical problem is to process the association network between the disease node set and the target node set, denoted as G. dt Let the number of disease nodes be u2, the number of target nodes be v, and A dt Let A be the adjacency matrix of the association network between disease nodes and target nodes. If a node in the disease node set is associated with a node in the target node set, then A... dt (i,j) = 1, otherwise, A dt (i,j)=0;

[0082] Similar to the embedding representation learning process of the first drug node and the first target node mentioned above, we can obtain the following:

[0083]

[0084] Among them, X dt1 For the node embedding of the first disease node, X dt2 The node embedding for the second target node, u2 is the number of disease nodes, and v is the number of target nodes.

[0085] The above steps 2-4 adopt a graph neural network approach, so it needs to be divided into multiple learning layers to continuously update the learned node embeddings.

[0086] 2) The embedding representation learning process of the third and fourth target nodes:

[0087] The practical problem involves handling the interaction relationships between the target node set nodes, denoted as G. tt The number of target nodes is v, A tt Let A be the adjacency matrix of interactions between target nodes. If target node i in the set of target nodes has an interaction relationship with target node j, then A... tt (i,j) = 1, otherwise, A tt (i,j)=0;

[0088] a) Third target node embedding learning process:

[0089] For the third target node embedding learning, network G tt The initial node embedding is the node embedding matrix of the first target node in the output of process 1, i.e., X. rt2 ∈R v×d , means as follows:

[0090]

[0091] v is G tt The number of nodes in the array, where d is the initial embedding feature dimension of each node.

[0092] This process employs a Graph Convolutional Neural Network (GCN), and the specific process is explained below:

[0093] Generative Networks (GCNs) are multi-layered connected neural network architectures used to learn low-dimensional representations of nodes within a graph structure. Each layer of a GCN directly aggregates information about connected neighbors in the graph, using the reconstructed embeddings as input to the next layer. The spectral graph convolution theorem defines convolution in the Fourier domain based on the normalized graph Laplacian operator:

[0094]

[0095] I is the identity matrix. The degree matrix is ​​represented by an asymmetric matrix, as the degree distribution of nodes in a HIN node can vary significantly. Instead of the symmetric L, let L represent the transition probability matrix.

[0096] In this embodiment, each resulting network is convolved using the transition probability matrix P as a Fourier basis. Specifically, let P = ΦΛΦ -1 Λ and Φ are the eigenvector matrix and the diagonal matrix of eigenvalues ​​of P, respectively. The convolution on each network is defined as follows:

[0097] G θ ★H 0 =G θ (P)H 0

[0098] =G θ (ΦΛΦ -1 )H 0

[0099] =ΦG θ (Λ)Φ -1 H 0

[0100] G θ ★H 0 Give signal H 0 In the Fourier domain of the figure, with filter G θ The product of Φ and denoted by Φ represents the output of the graph convolution. - 1 H 0 It is signal H 0 The Fourier transform of the target node. To perform convolution on the local neighbors of the target node, we define G... θ (Λ) is a K-order polynomial filter:

[0101]

[0102] θ∈R K It is a vector of polynomial coefficients. Therefore,

[0103]

[0104] Specifically, given a network G with a corresponding adjacency matrix... Convolutions on G only depend on nodes at most K steps away from the target node. In other words, the output signal after the convolution operation is approximately defined by a K-order local spectral filter on the network. The filter parameter θ... k It can be shared across the entire network G. Generalizing, the convolution operation on network G is defined as:

[0105]

[0106] in, and Let represent the trainable weight matrix of layer 0 and the embedding matrix of the first layer, respectively. d is the output node embedding dimension, and σ(·) is the activation function, using the ReLU(·) function.

[0107]

[0108] in, and These are the node embedding matrix and trainable weight parameter matrix of layer l, respectively, and the node embedding matrix of layer l+1.

[0109] b) The fourth target node embedding learning process:

[0110] Similar to the embedding representation learning process of the third target node mentioned above, we can obtain the following:

[0111] For the fourth target node embedding learning, network G tt The initial node embedding is the node embedding matrix of the second target node in the output of process 1, that is, X. dt2 ∈R v×d , means as follows:

[0112]

[0113] v is G tt The number of nodes in the array, where d is the initial embedding feature dimension for each node.

[0114]

[0115] in, and Let represent the input matrix of layer 0, the trainable weight matrix, and the first-layer node embedding matrix, respectively. d is the output node embedding dimension, and σ(·) is the activation function, using the ReLU(·) function.

[0116]

[0117] in, and These are the node embedding matrix and trainable weight parameter matrix of layer l, respectively, and the node embedding matrix of layer l+1.

[0118] 3) The process of learning representations of the second drug node embedding and the second disease node embedding:

[0119] This refers to the processing of drug-disease association networks, denoted as G. rdThe number of drug nodes is u1, the number of disease nodes is u2, and A rd Let A be the adjacency matrix of a drug-disease association network. If a drug node is associated with a disease node, then A... rd (i,j) = 1, otherwise, A rd (i,j)=0;

[0120] set up This is the adjacency matrix between all drug nodes and disease nodes. Since the similarity between diseases is not considered, and the relationship between non-disease nodes is not considered, it is filled with a matrix of 0s.

[0121] Network G rd The initialization node embedding is It is expressed as follows:

[0122]

[0123] in,

[0124] For the initialized drug and disease nodes, since we need to use the previous first drug node embedding and first disease node embedding as input, this solution still uses GCN to implement the following process:

[0125]

[0126] in, and These represent the first layer, where d is the output node embedding dimension, and σ(·) is the activation function, using the ReLU(·) function:

[0127]

[0128]

[0129] in, and These are the node embeddings of layer l and the trainable weight matrix and node embeddings of layer l+1, respectively.

[0130] 4) The learning process of third drug node embedding and third disease node embedding representation:

[0131] A bipartite graph bg(u,v,ε) is equivalent to g(u∪v,ε), where u and v represent two sets of node domains, u i and v j Let and represent the i-th and j-th nodes of u and v, respectively. All edges in the bipartite graph lie strictly between u and v, and e ij Indicate u i and v jThe features of the two sets of nodes between the edges can be represented as X. u ∈R M×P and X v ∈R N×Q M and N are the number of nodes in the two groups, respectively; Represents node u i Similarly, the eigenvectors of [the vectors] Represents node v j eigenvectors.

[0132] Since there are no intra-domain edges in a bipartite graph, bipartite graph convolution only performs message passing and node feature aggregation through inter-domain edges. For message MP v→u From domain v to domain u, we define a general bipartite graph convolution:

[0133]

[0134] Our bipartite graph convolutional layer, based on the GAT framework, is called a bipartite graph attention convolutional layer (bga). Because the attention mechanism considers the features of two sets of nodes, it is specifically designed for X. u and X v The learnable matrix W is defined. u ∈R P×S and W v ∈R Q×S

[0135] The BGA formula is:

[0136]

[0137] The weighting coefficients are calculated using the following formula:

[0138]

[0139] This step involves learning the embeddings of the third drug node and the third disease node.

[0140] a) The learning process of the third drug node embedding:

[0141] A drug target bipartite graph is needed, where the features of the drug nodes are the output of the drug node embedding in the third step, i.e., the second drug node embedding. The feature of the target node is the output of the target node embedding in step 2b, i.e., the fourth target node embedding.

[0142] The i-th representation of the third drug node embedding is: Where ε rt Let the set of edges be the bipartite graph of the drug target. Let i be the i-th drug node.

[0143] Therefore, the third drug node is embedded as follows:

[0144]

[0145] Where M represents the number of drug nodes.

[0146] b) The learning process of embedding the third disease node:

[0147] We need a bipartite graph of disease targets, where the features of disease nodes are the output of the disease node embedding in the third step, i.e., the second disease node embedding. The characteristic of the target node is the output of the target node embedding in step 2a, i.e., the third target node embedding.

[0148] Therefore, the j-th representation of the third disease node embedding is: Where ε dt Let the set of edges be the bipartite graph of disease targets. Let j be the j-th disease node.

[0149] Therefore, the third disease node is embedded as follows:

[0150]

[0151] Where N is the number of disease nodes.

[0152] 5) Fifth target node embedding representation learning process:

[0153] The fifth target node embedding representation requires the third and fourth target node embeddings to learn the weight parameters of their respective nodes through an attention mechanism, and then the weighted summation is used to learn the fifth target node embedding. Therefore, the fifth target node embedding also needs to be layered and continuously learned and updated.

[0154] Attention mechanisms, widely used in deep learning tasks such as Natural Language Processing (NLP), image recognition, and speech recognition, have gradually become one of the core technologies attracting attention in the deep learning field. This mechanism is inspired by how the human brain processes information. When processing information received from the outside world, the human brain often focuses its attention on high-value and interesting key information. It can be viewed as a combinatorial function that, by calculating the probability distribution of attention, highlights the influence of a key input on the output. Attention mechanisms were first used in the encoder-decoder framework for NLP tasks, and are now increasingly being used in some graph-based machine learning tasks as well.

[0155] Embedded with the third target node and embedding of the fourth target node As input, the learning weights for each input. It can be represented as follows:

[0156]

[0157] Here att sem This represents the method for performing semantic-level attention. To understand the importance of the third and fourth target node embeddings for the final fifth target node embedding, we first use a non-linear transformation:

[0158]

[0159]

[0160] Where W is the weight matrix, b is the bias vector, and q is the weight matrix. T A semantically trainable weight vector is used to measure the similarity between the embeddings of two target nodes. For the embedding representation of the i-th third target node, Let V be the embedding representation of the j-th fourth target node. V represents the number of target nodes. After obtaining the weight parameters of the two types of target node embeddings, we normalize them using the softmax function. By normalizing the weight parameters using the softmax function, we can obtain the weights of the third target node embedding. Weights embedded with the fourth target node

[0161]

[0162]

[0163] The fifth target node embedding representation is shown below:

[0164]

[0165] Because in the encoding process, apart from the first step of learning the embedding representations of the first drug node, the first disease node, and the first and second target nodes, which are learned directly through a random walk strategy based on the network topology, the remaining node embedding learning process needs to be treated as a layer in the overall training process. This layer contains the four steps 2-5.

[0166] Therefore, it is necessary to consider that the contributions of different levels in the overall process to the node embedding representation will be inconsistent. Thus, an attention mechanism is incorporated into these embeddings to obtain the final third drug node embedding:

[0167]

[0168] Similarly, the final third disease node is embedded:

[0169]

[0170] Final fifth target node embedding:

[0171]

[0172]

[0173] Where, α l β l and γ l The information is automatically learned and initialized to 1 / (l+1), where l = 1, 2, ..., L.

[0174] II. Decoding process (i.e., downstream task, the process of predicting new relationships between nodes)

[0175] 1. Drug-Disease Association Prediction Module

[0176] To reconstruct the association between drugs and diseases, the decoder f(H) in this scheme R H D The formula is as follows:

[0177]

[0178] in, This is the prediction probability score matrix. Drug r i and disease d j The predicted scores associated with each other are determined by the corresponding A′. rd The (i,j) terms are given.

[0179] 2. Drug-Target Interaction Prediction Module

[0180] To reconstruct the association between the drug and the target, the decoder f(H) in this scheme R H T The formula is as follows:

[0181]

[0182] in, This is the prediction probability score matrix. Drug r i and target t j The predicted scores associated with each other are determined by the corresponding A′. rt The (i,j) terms are given.

[0183] 3. Target-to-target interaction prediction module

[0184] To reconstruct the correlation between targets, the decoder f(H) in this scheme T H T The formula is as follows:

[0185]

[0186] Among them, A′ tt ∈R v×v is the prediction probability score matrix. Target point t i and target t j The predicted scores associated with each other are determined by the corresponding A′. tt The (i,j) terms are given.

[0187] III. Optimization Process

[0188] Since the known associations between drugs and diseases, drugs and targets, and targets themselves have been manually verified, they are highly reliable and crucial for improving predictive performance. Therefore, this approach learns parameters by minimizing the weighted binary cross-entropy loss. For the three downstream tasks mentioned earlier, this can be abstracted into the following formula:

[0189]

[0190] In the formula, (i, j) represents the pair of nodes i and j, and S + Let S represent the set of associations of all known node pairs. - This represents the set of all unknown or unobserved drug-disease association pairs. Balance factor. Used to reduce the impact of data imbalance, where |S + |and|S - ||S respectively + and S - The logarithm of the matrix. u×v is the predicted probability score matrix A′. ij Multiply the rows and columns.

[0191] The loss of the drug-disease association prediction module is loss. rd The loss of the drug target interaction prediction module is loss. rt The loss of the target-to-target interaction prediction module is loss. tt Final loss:

[0192] loss = k rd ×loss rd +k rt ×loss rt +k tt ×loss tt

[0193]

[0194] Finally, the model is continuously optimized using the Adam optimizer until the loss no longer decreases significantly, thus obtaining the optimal model parameters.

[0195] This disclosure uses multi-source information, considers network structure features, and takes into account the issues of domain-specific features and the methods of feature fusion between domains. By learning the feature vector embedding representations of various types of nodes in multi-source information, it can not only predict new drug-disease associations in the field of drug relocation, but also predict new interaction relationships between targets or new drug-target interaction relationships based on the learned target embedding information representations, etc., for downstream tasks.

[0196] The second embodiment of this disclosure also provides an apparatus for predicting multi-source information relationships using a predictive model, the structure of which is shown in the figure below. Figure 4 As shown, it includes:

[0197] Input module 10 is used to input the first multi-source information and the second multi-source information into the prediction model;

[0198] The output module 20, coupled to the input module 10, is used to output the relationship between the first multi-source information and the second multi-source information through the prediction model; wherein the first multi-source information and the second multi-source information include at least one of the following: the first multi-source information is a drug and the second multi-source information is a disease; the first multi-source information is a drug and the second multi-source information is a target; or the first multi-source information is a target and the second multi-source information is a target.

[0199] The predictive model includes: an encoder and a decoder;

[0200] An encoder is used to compute the embedding representation of a first node corresponding to a first multi-source information and the embedding representation of a second node corresponding to a second multi-source information in a predetermined network, wherein the predetermined network includes at least one of the following: a drug-disease association network, a drug-target interaction network, and a disease-target association network.

[0201] The decoder is used to predict the association or interaction relationship between the first multi-source information and the second multi-source information based on the embedding representation of the first node and the embedding representation of the second node.

[0202] This disclosure uses a trained prediction model to predict the relationship between two known multi-source information sources. Given existing known drug-disease association networks, drug-target interaction networks, and disease-target association networks, it learns the relationships between these networks and then determines the relationships between unknown drugs, diseases, and targets based on the known relationships between them. This not only enables drug relocation but also allows for the study of diseases and targets, with high prediction accuracy. It provides a clearer approach for scientific research, saves more research resources, and benefits the development of medicine or pharmacology.

[0203] The aforementioned apparatus further includes an encoder training module for training the encoder in the prediction model, wherein the encoder training method includes:

[0204] I. The deep walk algorithm is used to extract the network topology representation of drugs and targets from the drug target association network, and the sequence representation of each node in the drug target association network is learned according to the random walk theory to obtain the embedding representation of the first drug node and the embedding representation of the first target node in the drug target association network.

[0205] The deep walk algorithm is used to extract the network topology representation of diseases and targets from the disease target association network, and the sequence representation of each node in the disease target association network is learned according to the random walk theory to obtain the embedding representation of the first disease node and the embedding representation of the second target node in the disease target association network.

[0206] 2. Using the embedding representations of the first target node, the second target node, the first drug node, and the first disease node as input, the graph convolutional neural network algorithm and the bipartite graph attention convolutional layer are used to determine the embedding representations of the third drug node, the third disease node, and the fifth target node.

[0207] Specifically, the embedding representations of the third drug node and the third disease node are used to predict drug-disease association; the embedding representations of the third drug node are used to predict drug-drug interactions; the embedding representations of the third drug node and the fifth target node are used to predict drug-target interactions; and the embedding representations of the fifth target node are used to predict target-target interactions.

[0208] Preferably, when determining the embedding representation of the third drug node, the third disease node, and the fifth target node, the encoder training module specifically implements the following steps 1-4:

[0209] 1. Using the embedding representation of the first target node as input and the target-to-target interaction network as the adjacency matrix, the graph convolutional neural network algorithm is used to learn the embedding representation of the third target node; using the embedding representation of the second target node as input and the target-to-target interaction network as the adjacency matrix, the graph convolutional neural network algorithm is used to learn the embedding representation of the fourth target node.

[0210] 2. Using the embedding representations of the first drug node and the first disease node as input, and the drug-disease association network as the adjacency matrix, the graph convolutional neural network algorithm is used to learn the embedding representations of the second drug node and the second disease node.

[0211] 3. Using the embedding representations of the fourth target node and the second drug node as input, and the drug target interaction network as the adjacency matrix, a bipartite graph attention convolutional layer is used to learn the embedding representation of the third drug node; using the embedding representations of the third target node and the second disease node as input, and the disease target association network as the adjacency matrix, a bipartite graph attention convolutional layer is used to learn the embedding representation of the third disease node.

[0212] 4. The embedding representations of the third target node and the fourth target node are learned through an attention mechanism to obtain the weight parameters of their respective nodes. The embedding representation of the fifth target node is learned by weighted summation based on the embedding representation of the third target node and its weight parameters, and the embedding representation of the fourth target node and its weight parameters.

[0213] The aforementioned device further includes a decoder training module for training the decoder in the prediction model. The decoder training method includes: taking the embedding representations of the third drug node and the third disease node as input, and using a predetermined prediction probability score matrix to learn the drug-disease association; taking the embedding representations of the third drug node and the fifth target node as input, and using a predetermined prediction probability score matrix to learn the drug-target interaction; taking the embedding representation of the fifth target node as input, and using a predetermined prediction probability score matrix to learn the association between targets; wherein the training loss is to minimize the weighted binary cross-entropy loss to learn the association.

[0214] The specific implementation of the above training process can be found in the detailed description of the training process of the prediction model in the first embodiment, which will not be repeated here.

[0215] The third embodiment of this disclosure provides a storage medium, which is a computer-readable medium storing a computer program. When the computer program is executed by a processor, it implements the method provided in the first embodiment of this disclosure, which will not be described in detail here.

[0216] The fourth embodiment of this disclosure provides an electronic device, which includes at least a memory 901 and a processor 902. The memory 901 stores a computer program, and the processor 902 implements the method provided in the first embodiment of this disclosure when executing the computer program in the memory 901. The details will not be repeated here.

[0217] Furthermore, although exemplary embodiments have been described herein, their scope includes any and all embodiments based on this disclosure that have equivalent elements, modifications, omissions, combinations (e.g., schemes involving intersections of various embodiments), adaptations, or alterations. Elements in the claims will be interpreted broadly based on the language used in the claims and are not limited to the examples described in this specification or during the implementation of this application, and such examples will be interpreted as non-exclusive. Therefore, this specification and examples are intended to be considered illustrative only, and the true scope and spirit are indicated by the full scope of the following claims and their equivalents.

[0218] The above description is intended to be illustrative and not restrictive. For example, the above examples (or one or more thereof) can be used in combination with each other. Other embodiments may be used by those skilled in the art upon reading the above description. Furthermore, in the above detailed description, various features may be grouped together to simplify the disclosure. This should not be construed as an intention that a feature of the disclosure that is not claimed is necessary for any claim. Rather, the subject matter of this disclosure may be less than all the features of a particular disclosed embodiment. Thus, the following claims are incorporated herein by reference as examples or embodiments, wherein each claim is an independent, separate embodiment, and these embodiments are contemplated to be combined with each other in various combinations or arrangements. The scope of this disclosure should be determined by reference to the appended claims and the full scope of their equivalents.

[0219] The foregoing has provided a detailed description of several embodiments of this disclosure. However, this disclosure is not limited to these specific embodiments. Those skilled in the art can make various variations and modifications based on the concept of this disclosure, and all such variations and modifications should fall within the scope of protection claimed by this disclosure.

Claims

1. A method for predicting multi-source information relationships using a predictive model, characterized in that, include: The first multi-source information and the second multi-source information are input into the prediction model, and the relationship between the first multi-source information and the second multi-source information is output through the prediction model. When the first multi-source information is a drug, the second multi-source information is a disease or a target; when the first multi-source information is a target, the second multi-source information is a target. The prediction model includes: an encoder and a decoder; The encoder is used to calculate the embedding representation of the first node corresponding to the first multi-source information and the embedding representation of the second node corresponding to the second multi-source information in a predetermined network, wherein the predetermined network includes at least one of the following: a drug-disease association network, a drug-target interaction network, and a disease-target association network. The decoder is used to predict the association or interaction relationship between the first multi-source information and the second multi-source information based on the embedding representation of the first node and the embedding representation of the second node. The training method for the encoder in the prediction model includes: The deep walk algorithm is used to extract the network topology representation of drugs and targets from the drug target association network, and the sequence representation of each node in the drug target association network is learned according to the random walk theory to obtain the embedding representation of the first drug node and the embedding representation of the first target node in the drug target association network. The deep walk algorithm is used to extract the network topology representation of diseases and targets from the disease target association network, and the sequence representation of each node in the disease target association network is learned according to the random walk theory to obtain the embedding representation of the first disease node and the embedding representation of the second target node in the disease target association network. Using the embedding representations of the first target node, the second target node, the first drug node, and the first disease node as input, the embedding representations of the third drug node, the third disease node, and the fifth target node are determined using a graph convolutional neural network algorithm and a bipartite graph attention convolutional layer. Specifically, the embedding representations of the third drug node and the third disease node are used to predict drug-disease association; the embedding representations of the third drug node are used to predict drug-drug interactions; the embedding representations of the third drug node and the fifth target node are used to predict drug-target interactions; and the embedding representations of the fifth target node are used to predict target-target interactions.

2. The method as described in claim 1, characterized in that, The step of using the embedding representations of the first target node, the second target node, the first drug node, and the first disease node as input, and using a graph convolutional neural network algorithm and a bipartite graph attention convolutional layer to determine the embedding representations of the third drug node, the third disease node, and the fifth target node includes: The embedding representation of the first target node is used as input, and the target-to-target interaction network is used as the adjacency matrix. The graph convolutional neural network algorithm is used to learn the embedding representation of the third target node. The embedding representation of the second target node is used as input, and the target-to-target interaction network is used as the adjacency matrix. The graph convolutional neural network algorithm is used to learn the embedding representation of the fourth target node. Using the embedding representations of the first drug node and the first disease node as input, and the drug-disease association network as the adjacency matrix, the embedding representations of the second drug node and the second disease node are learned using a graph convolutional neural network algorithm. The embedding representations of the fourth target node and the second drug node are used as inputs, and the drug target interaction network is used as the adjacency matrix. The embedding representation of the third drug node is learned by a bipartite graph attention convolutional layer. The embedding representations of the third target node and the second disease node are used as input, and the disease target association network is used as the adjacency matrix. The embedding representation of the third disease node is learned by a bipartite graph attention convolutional layer. The embedding representations of the third target node and the fourth target node are learned through an attention mechanism to obtain the weight parameters of their respective nodes. The embedding representation of the fifth target node is then learned by weighted summation based on the embedding representation of the third target node and its weight parameters, and the embedding representation of the fourth target node and its weight parameters.

3. The method according to any one of claims 1 to 2, characterized in that, The training method for the decoder in the prediction model includes: The embedding representations of the third drug node and the third disease node are used as inputs, and a predetermined prediction probability score matrix is ​​used to learn the drug-disease association relationship. The drug-target interaction relationship is learned by using the embedded representation of the third drug node and the embedded representation of the fifth target node as inputs and the predetermined prediction probability score matrix. Using the embedded representation of the fifth target node as input, the association between targets is learned using the predetermined prediction probability score matrix; The training loss uses the method of minimizing the weighted binary cross-entropy loss to learn the association relationship.

4. An apparatus for predicting multi-source information relationships using a predictive model, characterized in that, include: The input module is used to input the first multi-source information and the second multi-source information into the prediction model, and output the relationship between the first multi-source information and the second multi-source information through the prediction model. When the first multi-source information is a drug, the second multi-source information is a disease or a target; when the first multi-source information is a target, the second multi-source information is a target. The prediction model includes: an encoder and a decoder; The encoder is used to calculate the embedding representation of the first node corresponding to the first multi-source information and the embedding representation of the second node corresponding to the second multi-source information in a predetermined network, wherein the predetermined network includes at least one of the following: a drug-disease association network, a drug-target interaction network, and a disease-target association network. The decoder is used to predict the association or interaction relationship between the first multi-source information and the second multi-source information based on the embedding representation of the first node and the embedding representation of the second node. An encoder training module is used to train the encoder in the prediction model, wherein the training method of the encoder includes: The deep walk algorithm is used to extract the network topology representation of drugs and targets from the drug target association network, and the sequence representation of each node in the drug target association network is learned according to the random walk theory to obtain the embedding representation of the first drug node and the embedding representation of the first target node in the drug target association network. The deep walk algorithm is used to extract the network topology representation of diseases and targets from the disease target association network, and the sequence representation of each node in the disease target association network is learned according to the random walk theory to obtain the embedding representation of the first disease node and the embedding representation of the second target node in the disease target association network. Using the embedding representations of the first target node, the second target node, the first drug node, and the first disease node as input, the embedding representations of the third drug node, the third disease node, and the fifth target node are determined using a graph convolutional neural network algorithm and a bipartite graph attention convolutional layer. Specifically, the embedding representations of the third drug node and the third disease node are used to predict drug-disease association; the embedding representations of the third drug node are used to predict drug-drug interactions; the embedding representations of the third drug node and the fifth target node are used to predict drug-target interactions; and the embedding representations of the fifth target node are used to predict target-target interactions.

5. The apparatus as described in claim 4, characterized in that, The encoder training module, when determining the embedding representations of the third drug node, the third disease node, and the fifth target node, is specifically used for: The embedding representation of the first target node is used as input, and the target-to-target interaction network is used as the adjacency matrix. The graph convolutional neural network algorithm is used to learn the embedding representation of the third target node. The embedding representation of the second target node is used as input, and the target-to-target interaction network is used as the adjacency matrix. The graph convolutional neural network algorithm is used to learn the embedding representation of the fourth target node. Using the embedding representations of the first drug node and the first disease node as input, and the drug-disease association network as the adjacency matrix, the embedding representations of the second drug node and the second disease node are learned using a graph convolutional neural network algorithm. The embedding representations of the fourth target node and the second drug node are used as inputs, and the drug target interaction network is used as the adjacency matrix. The embedding representation of the third drug node is learned by a bipartite graph attention convolutional layer. The embedding representations of the third target node and the second disease node are used as input, and the disease target association network is used as the adjacency matrix. The embedding representation of the third disease node is learned by a bipartite graph attention convolutional layer. The embedding representations of the third target node and the fourth target node are learned through an attention mechanism to obtain the weight parameters of their respective nodes. The embedding representation of the fifth target node is then learned by weighted summation based on the embedding representation of the third target node and its weight parameters, and the embedding representation of the fourth target node and its weight parameters.

6. The apparatus as described in any one of claims 4 to 5, characterized in that, Also includes: A decoder training module is used to train the decoder in the prediction model, wherein the training method of the decoder includes: The embedding representations of the third drug node and the third disease node are used as inputs, and a predetermined prediction probability score matrix is ​​used to learn the drug-disease association relationship. The drug-target interaction relationship is learned by using the embedded representation of the third drug node and the embedded representation of the fifth target node as inputs and the predetermined prediction probability score matrix. Using the embedded representation of the fifth target node as input, the association between targets is learned using the predetermined prediction probability score matrix; The training loss uses the method of minimizing the weighted binary cross-entropy loss to learn the association relationship.

7. A storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 3.

8. An electronic device, comprising at least a memory and a processor, wherein the memory stores a computer program, characterized in that, The processor implements the steps of the method according to any one of claims 1 to 3 when executing a computer program on the memory.

Images