Drug-disease association prediction method based on graph learning and related equipment

Abstract

The application discloses a drug-disease correlation relationship prediction method based on graph learning and related equipment, and the method comprises the following steps: obtaining a drug similarity network and a disease similarity network; obtaining a drug feature vector corresponding to each drug to be judged according to the drug similarity network, obtaining a disease feature vector corresponding to each disease to be judged according to the disease similarity network, combining the drug feature vector and the disease feature vector to obtain a plurality of combined feature vectors, and obtaining a drug-disease correlation relationship corresponding to each combined feature vector, wherein the combined feature vector is obtained by splicing the drug feature vector and the disease feature vector, and the drug-disease correlation relationship corresponding to the combined feature vector is the correlation relationship between the drug to be judged and the disease to be judged corresponding to the combined feature vector. The application is beneficial to improving the accuracy of drug-disease correlation relationship prediction.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to a method and related equipment for predicting drug-disease associations based on graph learning. Background Technology

[0002] With the development of science and technology, data analysis technology has been widely applied in various fields, such as processing and analyzing medical data. For example, data analysis technology can be used to calculate drug repositioning, that is, to predict the relationship between drugs and diseases.

[0003] In existing technologies, predicting the association between drugs and diseases (i.e., calculating drug relocation) is typically based on classical machine learning methods. Known drug-disease association pairs are used as positive sample labels, transforming drug relocation into a binary classification problem. Furthermore, drug and disease information are used as input features to train the machine learning classifier. Specifically, when predicting the association between drugs and diseases using classical machine learning methods, prediction can only be made for one drug-disease association pair, failing to fully consider the associations between drugs and between diseases, thus hindering the improvement of the accuracy of drug-disease association prediction.

[0004] Therefore, existing technologies still need improvement and development. Summary of the Invention

[0005] The main objective of this invention is to provide a method and related equipment for predicting drug-disease associations based on graph learning. This aims to solve the problem that existing technologies based on classical machine learning methods cannot fully consider the associations between drugs and between diseases when predicting the associations between drugs and diseases, which is not conducive to improving the accuracy of drug-disease association prediction.

[0006] To achieve the above objectives, a first aspect of the present invention provides a method for predicting drug-disease associations based on graph learning, wherein the method includes:

[0007] Obtain a drug similarity network, wherein the drug similarity network includes multiple nodes and multiple edges, one node in the drug similarity network represents a drug to be judged, and the drugs to be judged represented by two nodes connected by an edge in the drug similarity network have a similarity relationship.

[0008] Obtain a disease similarity network, wherein the disease similarity network includes multiple nodes and multiple edges, one node in the disease similarity network represents a disease to be judged, and the diseases to be judged represented by two nodes connected by an edge in the disease similarity network have a similarity relationship.

[0009] The drug feature vectors corresponding to each of the drugs to be judged are obtained according to the drug similarity network, and the disease feature vectors corresponding to each of the diseases to be judged are obtained according to the disease similarity network. Multiple combined feature vectors are obtained by combining the drug feature vectors and the disease feature vectors. The drug-disease association relationship corresponding to each of the combined feature vectors is obtained. A combined feature vector is obtained by concatenating a drug feature vector and a disease feature vector. The drug-disease association relationship corresponding to a combined feature vector is the association relationship between the drug to be judged and the disease to be judged corresponding to the combined feature vector.

[0010] Optionally, the above-mentioned drug feature vectors corresponding to each of the drugs to be judged are obtained according to the drug similarity network, and the disease feature vectors corresponding to each of the diseases to be judged are obtained according to the disease similarity network. Multiple combined feature vectors are obtained by combining the drug feature vectors and the disease feature vectors, and the drug-disease association relationships corresponding to each of the combined feature vectors are obtained respectively, including:

[0011] The drug similarity network and the disease similarity network mentioned above are input into the trained hybrid graph embedding association prediction model, wherein the hybrid graph embedding association prediction model includes a drug feature processing unit, a disease feature processing unit, a feature vector concatenation unit, and a classification unit.

[0012] The drug feature processing unit described above obtains the drug feature vectors corresponding to each of the drugs to be judged based on the drug similarity network described above.

[0013] The disease feature processing unit described above obtains the disease feature vectors corresponding to each of the diseases to be judged based on the disease similarity network described above.

[0014] Through the above feature vector concatenation unit, multiple combined feature vectors are obtained by combining the above drug feature vectors and the above disease feature vectors;

[0015] The drug-disease association relationship corresponding to each of the above-mentioned combined feature vectors is obtained through the above classification units.

[0016] Optionally, the aforementioned drug feature processing unit includes a first graph convolutional network layer, a first graph attention subunit, a second graph convolutional network layer, and a first convolutional neural network layer. The process of obtaining drug feature vectors corresponding to each drug to be judged based on the drug similarity network through the aforementioned drug feature processing unit includes:

[0017] Based on the above drug similarity network, the adjacency matrix of the drug similarity network and the initial drug node representation are obtained, wherein the initial drug node representation is used to reflect the features corresponding to the drug to be judged in the above drug similarity network;

[0018] The drug similarity network adjacency matrix and the drug node representation are input into the first graph convolutional network layer to trigger feature processing and obtain the first output drug node representation of the first graph convolutional network layer.

[0019] The adjacency matrix of the drug similarity network and the first output drug node representation are input into the first graph attention subunit to trigger the first graph attention subunit to perform feature processing and obtain the second output drug node representation output by the first graph attention subunit.

[0020] The adjacency matrix of the drug similarity network and the second output drug node representation are input into the second graph convolutional network layer to trigger the second graph convolutional network layer to perform feature processing and obtain the third output drug node representation output by the second graph convolutional network layer.

[0021] The first output drug node representation and the third output drug node representation are input into the first convolutional neural network layer to obtain the drug feature vector output by the first convolutional neural network layer.

[0022] Optionally, the aforementioned disease feature processing unit includes a third graph convolutional network layer, a second graph attention subunit, a fourth graph convolutional network layer, and a second convolutional neural network layer. The process of obtaining disease feature vectors corresponding to each of the diseases to be judged based on the disease similarity network through the aforementioned disease feature processing unit includes:

[0023] Based on the above disease similarity network, obtain the disease similarity network adjacency matrix and the initial disease node representation, wherein the above initial disease node representation is used to reflect the features corresponding to the disease to be judged in the above disease similarity network;

[0024] The above-mentioned disease similarity network adjacency matrix and the above-mentioned disease node representation are input into the above-mentioned third graph convolutional network layer to trigger the above-mentioned third graph convolutional network layer to perform feature processing and obtain the first output disease node representation of the above-mentioned third graph convolutional network layer.

[0025] The above-mentioned disease similarity network adjacency matrix and the above-mentioned first output disease node representation are input into the above-mentioned second graph attention subunit to trigger the above-mentioned second graph attention subunit to perform feature processing and obtain the second output disease node representation output by the above-mentioned second graph attention subunit;

[0026] The above-mentioned disease similarity network adjacency matrix and the above-mentioned second output disease node representation are input into the above-mentioned fourth graph convolutional network layer to trigger the above-mentioned fourth graph convolutional network layer to perform feature processing and obtain the third output disease node representation output by the above-mentioned fourth graph convolutional network layer.

[0027] The first output disease node representation and the third output disease node representation are input into the second convolutional neural network layer to obtain the disease feature vector output by the second convolutional neural network layer.

[0028] Optionally, the above-mentioned drug-disease association relationships are obtained through the above-mentioned classification units for each of the above-mentioned combined feature vectors, including:

[0029] Each of the above-mentioned combined feature vectors is input into the above-mentioned classification unit, and the above-mentioned classification unit classifies each of the above-mentioned combined feature vectors to obtain the drug-disease association relationship corresponding to each of the above-mentioned combined feature vectors. Among them, the drug-disease association relationship corresponding to a combination feature vector is that the drug to be judged and the disease to be judged corresponding to the combination feature vector are related or that the drug to be judged and the disease to be judged corresponding to the combination feature vector are not related.

[0030] Optionally, the above classification unit is a random forest classifier.

[0031] Optionally, the above hybrid graph embedding association prediction model is trained according to the following steps:

[0032] The training drug similarity network and the training disease similarity network in the training data are input into the above hybrid graph embedding association prediction model. The predicted drug-disease association is obtained through the above hybrid graph embedding association prediction model. The training data includes multiple training network groups. Each training network group includes a training drug similarity network, a training disease similarity network, and labeled drug-disease association.

[0033] Based on the predicted drug-disease association and the labeled drug-disease association, the model parameters of the hybrid graph embedding association prediction model are adjusted, and the steps of inputting the training drug similarity network and the training disease similarity network from the training data into the hybrid graph embedding association prediction model are continued until the preset training conditions are met, so as to obtain the trained hybrid graph embedding association prediction model.

[0034] A second aspect of the present invention provides a drug-disease association prediction system based on graph learning, wherein the aforementioned drug-disease association prediction system based on graph learning includes:

[0035] The drug network acquisition module is used to acquire a drug similarity network, wherein the drug similarity network includes multiple nodes and multiple edges, one node in the drug similarity network represents a drug to be judged, and the drugs to be judged represented by two nodes connected by an edge in the drug similarity network have a similarity relationship.

[0036] The disease network acquisition module is used to acquire a disease similarity network, wherein the disease similarity network includes multiple nodes and multiple edges, one node in the disease similarity network represents a disease to be judged, and the diseases to be judged represented by two nodes connected by an edge in the disease similarity network have a similarity relationship.

[0037] The similarity network processing module is used to obtain drug feature vectors corresponding to each of the drugs to be judged according to the drug similarity network, obtain disease feature vectors corresponding to each of the diseases to be judged according to the disease similarity network, combine the drug feature vectors and the disease feature vectors to obtain multiple combined feature vectors, and obtain the drug-disease association relationship corresponding to each of the combined feature vectors respectively. Among them, one of the combined feature vectors is obtained by concatenating one of the drug feature vectors and one of the disease feature vectors, and the drug-disease association relationship corresponding to one combined feature vector is the association relationship between the drug to be judged and the disease to be judged corresponding to the combined feature vector.

[0038] A third aspect of the present invention provides a smart terminal, the smart terminal including a memory, a processor, and a graph learning-based drug-disease association prediction program stored in the memory and executable on the processor, wherein the graph learning-based drug-disease association prediction program, when executed by the processor, implements the steps of any of the graph learning-based drug-disease association prediction methods.

[0039] A fourth aspect of the present invention provides a computer-readable storage medium storing a graph learning-based drug-disease association prediction program, wherein when the graph learning-based drug-disease association prediction program is executed by a processor, it implements the steps of any of the graph learning-based drug-disease association prediction methods described above.

[0040] As can be seen from the above, in the present invention, a drug similarity network is obtained, wherein the drug similarity network includes multiple nodes and multiple edges, one node in the drug similarity network represents a drug to be judged, and the drugs to be judged represented by two nodes connected by an edge in the drug similarity network have a similarity relationship; a disease similarity network is obtained, wherein the disease similarity network includes multiple nodes and multiple edges, one node in the disease similarity network represents a disease to be judged, and the diseases to be judged represented by two nodes connected by an edge in the disease similarity network have a similarity relationship; drug feature vectors corresponding to each drug to be judged are obtained according to the drug similarity network, and disease feature vectors corresponding to each disease to be judged are obtained according to the disease similarity network; multiple combined feature vectors are obtained by combining the drug feature vectors and the disease feature vectors; and the drug-disease association relationship corresponding to each combined feature vector is obtained respectively. A combined feature vector is obtained by concatenating a drug feature vector and a disease feature vector, and the drug-disease association relationship corresponding to a combined feature vector is the association relationship between the drug to be judged and the disease to be judged corresponding to that combined feature vector.

[0041] Compared to existing technologies, the present invention combines a drug similarity network (reflecting similarity relationships between multiple drugs) and a disease similarity network (reflecting similarity relationships between multiple diseases) to predict the association between a drug and a disease. This approach fully considers both the drug-drug and disease-drug associations (i.e., similarity relationships), improving the accuracy of drug-disease association prediction. For example, if drug A is known to be associated with disease a, and drug A has a very high similarity to drug B, then drug B is highly likely to also be associated with disease a. The present invention fully considers the impact of this similarity relationship, thereby improving prediction accuracy. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention, 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 of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart illustrating a method for predicting drug-disease association based on graph learning, provided in an embodiment of the present invention.

[0044] Figure 2This is a schematic diagram of the structure of a hybrid graph embedding association prediction model provided in an embodiment of the present invention;

[0045] Figure 3 This is a schematic diagram of a drug-disease association prediction process provided by an embodiment of the present invention;

[0046] Figure 4 This is a schematic diagram of the structure of a graph attention subunit provided in an embodiment of the present invention;

[0047] Figure 5 This is a schematic diagram of a training network group provided in an embodiment of the present invention;

[0048] Figure 6 This is a schematic diagram of the structure of a drug-disease association prediction system based on graph learning provided in an embodiment of the present invention;

[0049] Figure 7 This is a block diagram illustrating the internal structure of a smart terminal provided in an embodiment of the present invention. Detailed Implementation

[0050] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0051] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0052] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0053] With the development of science and technology, data analysis techniques are widely used in various fields, such as processing and analyzing medical data. For example, data analysis techniques can be used to calculate drug repositioning, that is, to predict the association between drugs and diseases. Calculating drug repositioning can help accelerate drug development and reduce experimental costs.

[0054] In existing technologies, predicting the association between drugs and diseases is typically based on classic machine learning methods (i.e., calculating drug relocation). Known drug-disease association pairs are used as positive sample labels, transforming drug relocation into a binary classification problem. Drug and disease information are then used as input features to train the machine learning classifier. Specifically, when predicting drug-disease associations using classic machine learning methods, the learning process involves multiple drug-disease association pairs as input. Therefore, the learning process only learns the relationships between each individual drug-disease association pair. During prediction, only one drug-disease association pair can be used, meaning only the association between a single drug and a single disease is considered. This fails to adequately consider the associations between drugs and between diseases, hindering the accuracy of drug-disease association prediction. Furthermore, the accuracy of the prediction is significantly limited by the quality of the input data during the learning process.

[0055] Drug development is an extremely lengthy and expensive process, and computational methods with cheaper and more labor-saving solutions can accelerate drug discovery. Identifying drug-disease associations (DDAs) can effectively screen candidate related drugs and guide subsequent validation experiments to achieve better experimental results. Initially, most reoriented drugs were discovered incidentally in clinical settings. With the development of high-throughput technologies, the increase in large-scale genomics, pharmacology, and chemistry datasets has made it possible to predict the relationship between drugs and diseases. The development of high-accuracy (ACC) computational methods has profound implications, but still faces significant challenges.

[0056] In existing technologies, classic machine learning methods are commonly used for drug repositioning. In some applications, network propagation methods, matrix factorization or completion methods, and deep learning methods can also be employed. For example, some researchers have proposed a Laplace-regularized least squares algorithm combined with a similarity kernel fusion method to predict drug-disease associations, known as DDA-SKF. Network propagation methods first construct a heterogeneous network of drugs and diseases, then use network-based algorithms (such as random walks) to predict the probability of drug-disease associations. For instance, some researchers have proposed a method called MBiRW, which uses a similarity measurer to construct a heterogeneous network of drugs and diseases, then employs a double random walk algorithm to predict potential drug-disease associations. Matrix factorization or completion methods treat drug repositioning as a recommendation system, recommending new drugs or indications based on known drug-disease associations and other acquired information. Some researchers have proposed a similarity-constrained matrix factorization method for predicting drug-disease associations called SCMFDD. It maps drug and disease features to a low-rank space to solve for constraint optimization. Some researchers have proposed a multi-similarity bilinear matrix factorization (MSBMF) to extract effective drug and disease representations, which can be used to infer unknown drug-disease associations. With continuous development in recent years, these types of computational drug relocation methods have achieved good performance. However, some key drawbacks still limit their ability to achieve higher accuracy and practical applications, such as their high dependence on the quality of input features in machine learning methods, representational biases in highly heterogeneous networks, and weak representations of drug-disease associations caused by linear multiplication in matrix factorization or completion methods. Unlike the linear multiplication of latent features in traditional matrix factorization, some researchers have developed a network-based deep learning approach. Deep learning-based relocation methods are also increasingly widely used. However, effectively utilizing the relationships between biological entities to capture biological interactions to enhance the prediction of drug-disease associations remains challenging.

[0057] To address at least one of the aforementioned problems, the present invention provides the following solutions: First, a drug similarity network is obtained, comprising multiple nodes and edges. Each node in the drug similarity network represents a drug to be judged, and two nodes connected by an edge in the drug similarity network represent drugs to be judged that have a similarity relationship. Second, a disease similarity network is obtained, comprising multiple nodes and edges. Each node in the disease similarity network represents a disease to be judged, and two nodes connected by an edge in the disease similarity network represent diseases to be judged that have a similarity relationship. Third, drug feature vectors corresponding to each drug to be judged are obtained from the drug similarity network, and disease feature vectors corresponding to each disease to be judged are obtained from the disease similarity network. Multiple combined feature vectors are obtained by combining the drug feature vectors and the disease feature vectors. The drug-disease association relationships corresponding to each combined feature vector are then obtained. Each combined feature vector is obtained by concatenating a drug feature vector and a disease feature vector, and the drug-disease association relationship corresponding to a combined feature vector is the association relationship between the drug to be judged and the disease to be judged corresponding to that combined feature vector.

[0058] Compared to existing technologies, the present invention combines a drug similarity network (reflecting similarity relationships between multiple drugs) and a disease similarity network (reflecting similarity relationships between multiple diseases) to predict the association between a drug and a disease. This approach fully considers both the association (i.e., similarity) between drugs and the association (i.e., similarity) between diseases, thus improving the accuracy of drug-disease association prediction. For example, if drug A is known to be associated with disease a, and drug A has a very high similarity to drug B, then drug B is highly likely to also be associated with disease a. The present invention fully considers the impact of this similarity, thereby improving prediction accuracy and efficiency.

[0059] Currently, Graph Convolutional Networks (GCNs) are receiving increasing attention and are being applied to many network-related prediction tasks to learn network topology and preserve node-level vector embeddings. GCNs extend convolutional neural networks for processing graph data, are easily embedded into end-to-end architectures, perform specific tasks with graph inputs, capture structural information of the graph through message passing between graph nodes, and maintain high interpretability. They have shown compelling performance in biomedical network analysis, such as microRNA-disease association prediction, drug side effect prediction, and miRNA-drug resistance association prediction.

[0060] This embodiment provides a graph learning-based drug-disease association prediction method. It combines drug similarity networks and disease similarity networks, considering the similarity relationships between drugs and between diseases, to predict drug-disease associations, thereby improving prediction accuracy and achieving better drug relocation. Specifically, this embodiment proposes a hybrid graph representation learning method called GTDDA, constructing drug-drug similarity networks and disease-disease similarity networks to represent the relationship between drugs and diseases, respectively. Secondly, a hybrid graph embedding model combining graph convolutional networks and transformers is introduced to learn feature representations of drugs and diseases. Finally, the learned representations are concatenated to establish a prediction model for identifying drug-disease associations, and drug-disease association prediction is performed based on the feature representations obtained from the concatenation. Through testing experiments, it is determined that GTDDA outperforms several advanced drug-disease association prediction methods on several public databases, providing effective predictions for drug relocation.

[0061] In one application scenario, a hybrid graph representation learning framework called GTDDA is used. First, the transformer encoder is combined with the attention aggregator in a Graph Attention Network (GAT), creating a module called Graph Attention Module Utilizing Transformer Encoder Structure (GATT). This module primarily connects a residual attention layer and a normalization layer from the transformer framework after the GAT aggregator, thereby strengthening the signal from node aggregation and achieving better graph feature representation. Then, a hybrid graph embedding model integrating GCN and GATT is used to predict the association between drugs and diseases.

[0062] The hybrid graph embedding association prediction model proposed in this embodiment performs well in prediction. Specifically, it utilizes two similarity networks to represent the relationship between drugs and diseases, including a drug-drug similarity network (i.e., drug similarity network) and a disease-disease similarity network (i.e., disease similarity network). A hybrid graph embedding model integrating GCN and GATT (i.e., hybrid graph embedding association prediction model) is used to simultaneously learn feature representations of drugs and diseases. The GATT modules (e.g., drug feature processing unit and disease feature processing unit) leverage the advantages of GAT and transformer structures, thereby improving the quality of the learned features. After concatenating and merging the learned representations, the drug-disease association prediction is performed using the classification unit (e.g., classifier) ​​in the hybrid graph embedding association prediction model.

[0063] like Figure 1As shown, this embodiment of the invention provides a method for predicting drug-disease associations based on graph learning. Specifically, the method includes the following steps:

[0064] Step S100: Obtain a drug similarity network, wherein the drug similarity network includes multiple nodes and multiple edges, one node in the drug similarity network represents a drug to be judged, and the drugs to be judged represented by two nodes connected by an edge in the drug similarity network have a similarity relationship.

[0065] In this embodiment, the drugs to be judged are those for which drug-disease association prediction is required. Multiple drugs can be judged simultaneously. These drugs may or may not have similar relationships. When a similar relationship exists, the nodes in the corresponding drug similarity network are connected by an edge; otherwise, no edge exists between the two nodes. It should be noted that in this embodiment, the similarity relationship between the drugs to be judged can be determined based on a pre-set database, or the similarity between the two drugs can be calculated using a preset similarity calculation method. If the similarity is greater than a preset drug similarity threshold, the two drugs are considered to have a similar relationship. Other methods can also be used for judgment, which are not specifically limited here.

[0066] Specifically, drugs typically possess different biological or chemical characteristics. A drug can be encoded as a binary feature vector. Therefore, drugs can be transformed into feature vectors, and different drug-drug similarities can be calculated based on these characteristics using different similarity measures. For example, the Jaccard index or cosine similarity can be used as indicators to measure drug-drug similarity.

[0067] Step S200: Obtain a disease similarity network, wherein the disease similarity network includes multiple nodes and multiple edges, one node in the disease similarity network represents a disease to be judged, and the diseases to be judged represented by two nodes connected by an edge in the disease similarity network have a similarity relationship.

[0068] In this embodiment, the diseases to be determined are those requiring drug-disease association prediction. Multiple diseases can be determined simultaneously. These diseases may or may not have similar relationships. When a similarity exists, nodes in the corresponding disease similarity network are connected by an edge; otherwise, no edge exists between the two nodes. It should be noted that in this embodiment, the similarity relationship between the diseases to be determined can be determined based on a pre-set database, or the similarity between the two diseases can be calculated using a preset similarity calculation method. If the similarity is greater than a preset disease similarity threshold, the two diseases are considered to have a similar relationship. Other methods can also be used for determination, which are not specifically limited here.

[0069] Specifically, Medical Subject Headings (MeSH) identifiers can be used to calculate semantic similarity as the probability of interaction. A disease's MeSH identifier can be represented as a hierarchical directed acyclic graph (DAG). The similarity between diseases can then be calculated based on the DAG. Alternatively, MimMiner can be used to obtain feature vectors from medical descriptions in the OMIM database through text mining analysis, thereby measuring the similarity between diseases. For drug similarity networks, let Gr denote the drug similarity network (i.e., drug-drug similarity network) of N drugs to be judged, and its adjacency matrix is ​​derived from the drug similarity matrix S. r Composition. Specifically, if drug j is based on the top k nearest neighbors of drug i in the drug similarity matrix, then the (i,j) term of the adjacency matrix is ​​S. r (i,j); otherwise, the value of this term is 0. It should be noted that the adjacency matrix corresponding to the disease similarity network can be determined according to the similarity principle, which will not be elaborated here.

[0070] In one application scenario, similar drugs in a drug similarity network are connected by edges, and the corresponding values ​​in the adjacency matrix are their similarity scores. Dissimilar drugs are not connected by edges, and their corresponding values ​​in the adjacency matrix are 0. In this embodiment, drug j is considered similar if it is based on the top k nearest neighbors of drug i in the drug similarity matrix. Similarly, if disease j is based on the top k nearest neighbors of disease i in the disease similarity matrix, it is considered similar in this embodiment. It should be noted that in the dataset shown in this embodiment, each drug has a similarity greater than 0 with at least a few other drugs, so there are no isolated points in the similarity network graph.

[0071] Step S300: Obtain drug feature vectors corresponding to each of the drugs to be judged according to the drug similarity network, obtain disease feature vectors corresponding to each of the diseases to be judged according to the disease similarity network, combine the drug feature vectors and the disease feature vectors to obtain multiple combined feature vectors, and obtain the drug-disease association relationship corresponding to each of the combined feature vectors respectively. Among them, a combined feature vector is obtained by concatenating a drug feature vector and a disease feature vector, and the drug-disease association relationship corresponding to a combined feature vector is the association relationship between the drug to be judged and the disease to be judged corresponding to the combined feature vector.

[0072] Among them, the drug feature vector is a vector used to reflect the feature information of the drug to be judged, which can be obtained through encoding; the disease feature vector is a vector used to reflect the feature information of the disease to be judged, which can also be obtained through encoding.

[0073] Specifically, the process involves obtaining drug feature vectors corresponding to each drug to be judged based on the drug similarity network, obtaining disease feature vectors corresponding to each disease to be judged based on the disease similarity network, combining the drug feature vectors and disease feature vectors to obtain multiple combined feature vectors, and obtaining the drug-disease association relationship corresponding to each of the combined feature vectors. This includes: inputting the drug similarity network and the disease similarity network into a trained hybrid graph embedding association prediction model, wherein the hybrid graph embedding association prediction model includes a drug feature processing unit, a disease feature processing unit, a feature vector concatenation unit, and a classification unit; obtaining drug feature vectors corresponding to each drug to be judged based on the drug similarity network through the drug feature processing unit; obtaining disease feature vectors corresponding to each disease to be judged based on the disease similarity network through the disease feature processing unit; obtaining multiple combined feature vectors by combining the drug feature vectors and disease feature vectors through the feature vector concatenation unit; and obtaining the drug-disease association relationship corresponding to each of the combined feature vectors through the classification unit.

[0074] In this embodiment, a trained hybrid graph embedding association prediction model is used to predict the association between drugs and diseases. Figure 2 This is a schematic diagram of the structure of a hybrid graph embedding association prediction model provided in an embodiment of the present invention, such as... Figure 2 As shown, the hybrid graph embedding association prediction model includes a drug feature processing unit, a disease feature processing unit, a feature vector concatenation unit, and a classification unit.

[0075] The aforementioned drug feature processing unit includes a first graph convolutional network layer, a first graph attention subunit, a second graph convolutional network layer, and a first convolutional neural network layer. The process of obtaining drug feature vectors corresponding to each drug to be judged based on the drug similarity network through the aforementioned drug feature processing unit includes: obtaining a drug similarity network adjacency matrix and an initial drug node representation based on the drug similarity network, wherein the initial drug node representation is used to reflect the features corresponding to the drugs to be judged in the drug similarity network; inputting the drug similarity network adjacency matrix and the drug node representation into the first graph convolutional network layer to trigger feature processing and obtain the first feature vector output by the first graph convolutional network layer. The process involves: first, inputting the drug similarity network adjacency matrix and the first output drug node representation into the first graph attention subunit to trigger feature processing and obtain a second output drug node representation; second, inputting the drug similarity network adjacency matrix and the second output drug node representation into the second graph convolutional network layer to trigger feature processing and obtain a third output drug node representation; and third, inputting the first and third output drug node representations into the first convolutional neural network layer to obtain a drug feature vector output by the first convolutional neural network layer.

[0076] The aforementioned disease feature processing unit includes a third graph convolutional network layer, a second graph attention subunit, a fourth graph convolutional network layer, and a second convolutional neural network layer. The process of obtaining disease feature vectors corresponding to each of the diseases to be judged based on the disease similarity network through the aforementioned disease feature processing unit includes: obtaining a disease similarity network adjacency matrix and an initial disease node representation based on the aforementioned disease similarity network, wherein the initial disease node representation is used to reflect the features corresponding to the diseases to be judged in the aforementioned disease similarity network; inputting the disease similarity network adjacency matrix and the disease node representation into the aforementioned third graph convolutional network layer to trigger feature processing by the aforementioned third graph convolutional network layer and obtaining the first output of the aforementioned third graph convolutional network layer. The disease node representation is generated; the disease similarity network adjacency matrix and the first output disease node representation are input into the second graph attention subunit to trigger feature processing and obtain the second output disease node representation; the disease similarity network adjacency matrix and the second output disease node representation are input into the fourth graph convolutional network layer to trigger feature processing and obtain the third output disease node representation; the first output disease node representation and the third output disease node representation are input into the second convolutional neural network layer to obtain the disease feature vector output by the second convolutional neural network layer.

[0077] In this embodiment, a GTDDA framework for drug-disease association prediction is proposed (i.e., the model framework corresponding to the hybrid graph embedding association prediction model mentioned above). The drug feature processing unit and the disease feature processing unit are units used for graph learning, which are used to learn feature representations from the drug similarity network and the disease similarity network, respectively.

[0078] It should be noted that the drug feature processing unit and the disease feature processing unit used in this embodiment have the same structure, both consisting of two graph convolutional network layers (i.e., GCN layers), one graph attention subunit (i.e., one GATT subunit), and one convolutional neural network layer (i.e., CNN layer). In one application scenario, the hybrid graph embedding association prediction model may only have one feature processing unit, which processes the drug similarity network and the disease similarity network. In this embodiment, two feature processing units (i.e., drug feature processing unit and disease feature processing unit) are set up respectively, and both are trained separately on drug data and disease data during the training process, thus achieving better accuracy in feature extraction and processing.

[0079] Figure 3 This is a schematic diagram of a drug-disease association prediction process provided by an embodiment of the present invention. Figure 3This specifically demonstrates the process of predicting drug-disease associations using a hybrid graph embedding association prediction model. For example... Figure 3 As shown, in this embodiment, the feature processing unit (drug feature processing unit or disease feature processing unit) encloses two graph convolutional layers (i.e., GCN layers), and there is a GATT sub-unit between the two GCN layers. Finally, a CNN fusion layer is used to integrate the outputs of different GCN layers.

[0080] Figure 3 In this model, Gr and Gd represent the input graph networks, namely the drug similarity network and the disease similarity network, respectively. The adjacency matrix can be obtained from the corresponding graph network. In this embodiment, the adjacency matrix consists of disease similarity, drug similarity, and drug-disease associations.

[0081] It should be noted that the first, second, third, and fourth convolutional network layers in this embodiment are essentially GCN layers, and they process data in a similar way. Therefore, we will use one GCN layer as an example for specific explanation.

[0082] GCN is a framework that uses deep learning to directly learn graph-structured data. By mapping nodes in the graph, it integrates the feature information of the nodes and their neighbors. This gives each node in the graph a regular expression, which is then fed into a convolutional network to generate a new representation of the node. It can effectively utilize the structural information of the graph. In this embodiment, GCN is used to obtain feature representations of drugs and diseases from drug similarity networks and disease similarity networks, respectively. Given a network G, its adjacency matrix and input node representations (i.e., initial drug node representations or initial disease representations) can be represented as follows: and Where N is the number of nodes, and F is the dimension of the node features. The output node representation H... 1 It can be obtained through a GCN layer, as shown in formulas (1) and (2) below:

[0083] H 1 =GCN(S,H 0 (1)

[0084]

[0085] in, It is a degree matrix; σ is the trainable weight matrix (i.e., the pre-trained model parameters); σ(·) is the ReLU activation function. It should be noted that the first, second, third, and fourth convolutional network layers in this embodiment are all GCN layers, therefore their data processing can refer to the above formulas (1) and (2). When formula (1) represents the processing of the first convolutional network layer, the corresponding S represents the drug similarity network adjacency matrix.

[0086] In this embodiment, the input drug similarity network (and / or disease similarity network) can be viewed as a graph network, including nodes and edges. For example, in the disease similarity network, each disease to be judged is a node, and the similarity between diseases is the edge weight. The input node representation (i.e., the initial disease node representation) is then used. This represents the node representations of all nodes, with each row representing the representation of one node. In the initial input layer (i.e., the first or third graph convolutional network layer), the feature attribute H... 0 It is the feature matrix of the input network (i.e., drug similarity network or disease similarity network). If there are N nodes in the graph and each node has F-dimensional features, then the feature matrix is ​​an N×F matrix.

[0087] Figure 4 This is a schematic diagram of a graph attention subunit provided in an embodiment of the present invention. It should be noted that the structures of the first graph attention subunit and the second graph attention subunit in this embodiment are completely identical. Figure 4 The structure in the diagram can represent either the first graph attention subunit or the second graph attention subunit. Figure 3 The GATT subunit (i.e., the GATT layer) in the middle is Figure 4 In the graph attention subunit, such as Figure 3 As shown, this embodiment introduces a GATT layer between two GCN layers. The output data of the first GCN layer serves as the input data for the GATT layer, and the output data of the GATT layer serves as the input data for the second GCN layer. This helps the second GCN layer extract high-level features of the drug or disease. GATT is very similar to the originally proposed transformer structure. The key component of GATT is the proposed multi-head attention part, which is an attention block tailored to the structural input data. It stacks several dot-product attention layers in parallel. Each dot-product attention layer takes a set of queries Q, keys K, and values ​​V as input, then calculates the dot product of all keys and applies a softmax function to obtain the weights of the values. In addition, GATT can be seen as a variant of GAT on a fully connected graph. GAT is a neural network that uses multi-head attention to assign different weights to neighboring nodes based on their importance.

[0088] For a network G, after passing through a GCN layer, its output serves as the input to a GATT layer. The output node representation of the GATT layer (i.e., the second output drug node representation or the second output disease node representation) is H. 2 As shown in formula (3) below:

[0089] H 2 =GATT(S,H 1 (3)

[0090] Here, S remains the adjacency matrix of the original input network G (i.e., the adjacency matrix of the drug similarity network or the disease similarity network). In the GATT calculation process, to better utilize the edge features between nodes, this embodiment associates the known edge features with the calculated hidden edge information. Assuming attention scores... Representing an edge<i,j> Implicit information, attempting to define the edges<i,j> Inject available edge information and improve the score This step essentially involves simply combining two values. and e ij The process is accomplished through multiplication, and its calculation is shown in formula (7). Assuming we are currently at level l, the update formulas for subsequent different levels are as follows:

[0091]

[0092]

[0093]

[0094]

[0095] in, and These represent the attention score matrices before and after the improvement, respectively. and These represent the node and edge features after the multi-head attention layer, respectively. E kl This represents the edge feature information matrix obtained from layer l, which is used in the multi-head attention part of the calculation. Q represents the feature output of the node and edge obtained after passing through the graph attention module, respectively. kl ,K kl V kl These three represent the query, key, and value matrices used in multi-head attention, respectively. Essentially, they are all linear transformations of the original matrix. `k` represents the number of attention heads, ranging from 1 to L. `n` represents the feature dimension of the hidden layer. `||` represents the concatenation operation. `h`... i and e ij It represents the feature information of nodes and edges. The softmax value represents the feature of node j in layer l. j These are commonly used activation functions. The characteristics representing node i at layer l. This represents the edge features between node i and node j in layer l. Then, the attention output is concatenated using a multi-head concatenation method. and The data is then passed to a separate feedforward network, followed by residual connections and normalization layers, as shown in the following formula:

[0096]

[0097]

[0098]

[0099] The above are the update formulas for node features. Edge feature updates use the same calculation method and will not be repeated here. Among them, Both represent trainable weight matrices (i.e., model parameters that can be pre-trained), and n represents the feature dimension of the hidden layer. and This represents the node characteristics of the intermediate layer. represents the feature of node i in layer l+1, Norm represents the normalization function, and ReLU is a commonly used activation function.

[0100] It should be noted that the attention score is continuously updated during training, and its update formula is shown in formula (7); the available edge information is the known edge weights in the input graph network. It also incorporates the edge information between two nodes into the attention calculation, thereby improving the attention score. Initially, the attention score can be calculated from Q, K, and V, referring to the multi-head attention mechanism.

[0101] Based on the aforementioned GCN and GATT layers, drug feature processing units and disease feature processing units can be constructed, respectively used to construct feature representations of nodes from the corresponding similarity networks. In this embodiment, for the input drug similarity network G... r Its drug similarity network adjacency matrix is ​​represented by γ, γ ij ∈γ represents the similarity between drugs. For the input node features of the network (i.e., the initial drug node representation), N r F is the number of nodes. r The dimensions of drug features are defined. GCN and GATT layers are introduced alternately to obtain graph feature representations of drug nodes at different levels. The specific steps are shown in formula (11) below:

[0102]

[0103] In order to combine the output features of different GCN layers (i.e., the first output drug node representation) and (i.e., the third output drug node representation), the final drug representation X is obtained using a one-dimensional convolutional neural network (CNN). r (i.e., drug feature vector). Among them, This represents the second output drug node from the GATT layer.

[0104] Correspondingly, GCN and GATT are used to analyze the disease similarity network G... d Learning multi-level node features (i.e., the first output disease node representation) (i.e., the second output disease node representation) and (i.e., the third output disease node representation), as shown in the following formula (12):

[0105]

[0106] in, Let γ' be the initial disease node representation, and γ' be the adjacency matrix of the disease similarity network. ij ∈γ' represents the similarity between diseases, and the initial features of this disease network are: N d F is the number of nodes. d This represents a dimension of disease characteristics. Furthermore, a one-dimensional convolutional neural network is used to process the output of the GCN layer. and By integrating these components, disease symptom X can be obtained. d (i.e., disease feature vector).

[0107] The initial input node features (i.e., the initial drug node representation and initial disease node representation during training) are randomly generated by the `torch.randn` function, representing the feature information of that node. These node features change continuously during training, with the ultimate goal of obtaining high-quality encoding vectors for drugs and diseases through this model. In calculating X... r At that time, first of all and The new matrix is ​​obtained by concatenating the elements using the torch.cat function. Then, this new matrix is ​​input into a one-dimensional convolution to obtain the output, which is the final representation of the drug.

[0108] Specifically, the process of obtaining the drug-disease association relationship corresponding to each of the above-mentioned combination feature vectors through the above-mentioned classification unit includes: inputting each of the above-mentioned combination feature vectors into the above-mentioned classification unit, classifying each of the above-mentioned combination feature vectors through the above-mentioned classification unit, and obtaining the drug-disease association relationship corresponding to each of the above-mentioned combination feature vectors. In this case, the drug-disease association relationship corresponding to a combination feature vector is either that the drug to be judged and the disease to be judged corresponding to the combination feature vector are related, or that the drug to be judged and the disease to be judged corresponding to the combination feature vector are not related.

[0109] In one application scenario, the similarity between the drug and the disease to be identified can be calculated based on the above classification unit. If the similarity is greater than a preset association threshold, they are considered related; otherwise, they are not related. In this embodiment, the above classification unit is a random forest classifier.

[0110] Furthermore, according to X r and X d The representation of M can be used to obtain the drug-disease prediction preference matrix M as shown in formula (13), where M is the drug-disease prediction preference matrix M. ij The higher the value of drug i, the greater the likelihood that drug i is associated with disease j.

[0111] M = X r T X d (13)

[0112] In this embodiment, the binary cross-entropy function (BCE) is used to measure the difference between the preference matrix M and the known adjacency matrix R (i.e., the drug-disease association matrix, determined based on the known database during training, or labeled with drug-disease associations), and this difference is used as the loss function for training the graph representation learning model. By minimizing the aforementioned loss function on the drug-disease association training database, the final graph representation matrices X corresponding to drugs and diseases can be obtained after training. r and X d This is used to predict novel drug-disease associations. During training, model parameters are adjusted, such as the weight parameters mentioned above.

[0113] To build a predictive model for identifying drug-disease associations, a training set was constructed from a selected database, where known association pairs were used as positive samples, and an equal number of randomly selected drug-disease association pairs were used as negative samples. The feature representations of drugs and diseases obtained using GTDDA were concatenated to form a combined description, thus representing the drug-disease pair. A combined description of a drug-disease pair is as follows: X r (i) represents the drug feature matrix X r The i-th row, X d (j) represents the disease feature matrix X d Line j:

[0114] z ij =[X r (i),X d (j)] (14)

[0115] z ij X represents the feature vector of the i-th drug. r (i) and the j-th disease feature vector X d (j) The combined feature vector obtained by splicing.

[0116] Random Forest (RF) is a powerful ensemble classifier that uses multiple decision trees to mitigate overfitting to training data. In this embodiment, RF was trained based on the aforementioned combined description to predict the association between drugs and diseases. RF can also be pre-trained directly using the RandomForestClassifier library; this is not specifically limited here.

[0117] Specifically, the hybrid graph embedding association prediction model is trained according to the following steps: The training drug similarity network and training disease similarity network from the training data are input into the hybrid graph embedding association prediction model. The predicted drug-disease association corresponding to the training drug similarity network (and the training disease similarity network) is obtained through the hybrid graph embedding association prediction model. The training data includes multiple training network groups, each including a training drug similarity network, a training disease similarity network, and the labeled drug-disease association corresponding to the training drug similarity network (and the training disease similarity network). Based on the predicted drug-disease association and the labeled drug-disease association, the model parameters of the hybrid graph embedding association prediction model are adjusted, and the steps of inputting the training drug similarity network and training disease similarity network from the training data into the hybrid graph embedding association prediction model are continued until the preset training conditions are met, thus obtaining the trained hybrid graph embedding association prediction model.

[0118] The aforementioned drug similarity network is the drug similarity network used during model training, and the aforementioned disease similarity network is the disease similarity network used during model training. It should be noted that the data used during model training is similar to the data and processing procedures used when making predictions using the trained model, and will not be repeated here. In this embodiment, the loss value between the predicted drug-disease association and the labeled drug-disease association is calculated to adjust the model parameters. The specific method for calculating the loss value can be set according to actual needs and is not specifically limited here. The aforementioned preset training conditions can be that the number of iterations reaches a preset iteration threshold or the obtained loss value is less than a preset loss threshold, or other preset training conditions for stopping training, which are not specifically limited here.

[0119] Figure 5 This is a schematic diagram of a training network group provided in an embodiment of the present invention, such as... Figure 5 As shown, in this embodiment, a training network group includes a drug similarity network, a disease similarity network, and a network for labeling drug-disease associations. In this embodiment, training and testing data can be obtained from existing datasets, for example, from the three benchmark datasets Fdataset, Cdataset, and Rdataset. Table 1 summarizes the statistical data for the three datasets, including the number of drugs, the number of diseases, and the number of known drug-disease associations.

[0120] Table 1

[0121] Dataset Drug quantity Number of diseases Number of associations Fdataset 593 313 1933 Cdataset 663 409 2352 Rdataset 894 454 2704

[0122] Table 1 shows the first dataset, which includes 593 drugs, 313 diseases, and 1933 confirmed drug-disease associations. Drugs were extracted from the Drug Library (DB) database, a comprehensive database containing extensive information on drugs and their targets. Diseases were collected from the Online Mendelian Inheritance in Humans (OMIM) database, a public resource for information on human genes and diseases. The second dataset contains 2352 associations between 663 drugs from the DB database and 409 diseases from the OMIM database. The third dataset includes 894 drugs, 454 diseases, and 2704 known associations between them.

[0123] By utilizing the above datasets, the corresponding drug-drug similarity matrices were obtained. Disease-Disease Similarity Matrix Drug-disease association matrix R∈{0,1} N×M Let M and N represent the number of diseases and drugs, respectively, with the superscript r indicating drugs and the superscript d indicating diseases. Let R be the drug-disease association matrix. If the database confirms that drug r...i With disease d j If it is related, then it is R. ij =1; otherwise R ij =0. Where, take matrix S r Chinese medicine r i and r j The similarity of the drug similarity network G r drug node γ i and γ j The edge weights are taken from matrix S. d Disease d i and d j The similarity between them is used as the similarity network G d Disease Node D i and D j The edge weights.

[0124] As can be seen from the above, in this embodiment, a drug similarity network reflecting the similarity relationships between multiple drugs and a disease similarity network reflecting the similarity relationships between multiple diseases are combined to predict the association between the drug and the disease to be identified. This approach fully considers the association relationships (i.e., similarity relationships) between drugs and between diseases, which helps improve the accuracy of drug-disease association prediction. For example, if it is known that drug A is associated with disease a, and drug A has a very high similarity with drug B, then drug B is very likely also associated with disease a. This embodiment fully considers the impact of such similarity relationships, thereby improving the accuracy of prediction.

[0125] like Figure 6 As shown, corresponding to the above-mentioned graph learning-based drug-disease association prediction method, this embodiment of the invention also provides a graph learning-based drug-disease association prediction system, which includes:

[0126] The drug network acquisition module 410 is used to acquire a drug similarity network, wherein the drug similarity network includes multiple nodes and multiple edges, one node in the drug similarity network represents a drug to be judged, and the drugs to be judged represented by two nodes connected by an edge in the drug similarity network have a similarity relationship.

[0127] The disease network acquisition module 420 is used to acquire a disease similarity network, wherein the disease similarity network includes multiple nodes and multiple edges, one node in the disease similarity network represents a disease to be judged, and the diseases to be judged represented by two nodes connected by an edge in the disease similarity network have a similarity relationship.

[0128] The similarity network processing module 430 is used to obtain drug feature vectors corresponding to each of the drugs to be judged according to the drug similarity network, obtain disease feature vectors corresponding to each of the diseases to be judged according to the disease similarity network, combine the drug feature vectors and the disease feature vectors to obtain multiple combined feature vectors, and obtain the drug-disease association relationship corresponding to each of the combined feature vectors respectively. Among them, one of the combined feature vectors is obtained by concatenating one of the drug feature vectors and one of the disease feature vectors, and the drug-disease association relationship corresponding to one combined feature vector is the association relationship between the drug to be judged and the disease to be judged corresponding to the combined feature vector.

[0129] Specifically, in this embodiment, the specific functions of the above-mentioned graph learning-based drug-disease association prediction system and its modules can be referred to the corresponding descriptions in the above-mentioned graph learning-based drug-disease association prediction method, and will not be repeated here.

[0130] It should be noted that the division of the modules in the above graph learning-based drug-disease association prediction system is not unique and is not intended as a specific limitation.

[0131] Based on the above embodiments, the present invention also provides a smart terminal, the principle block diagram of which can be as follows: Figure 7 As shown. The aforementioned intelligent terminal includes a processor and a memory. The memory of the intelligent terminal includes a graph learning-based drug-disease association prediction program, and the memory provides an environment for the operation of the graph learning-based drug-disease association prediction program. When the graph learning-based drug-disease association prediction program is executed by the processor, it implements the steps of any of the above-mentioned graph learning-based drug-disease association prediction methods. It should be noted that the aforementioned intelligent terminal may also include other functional modules or units, which are not specifically limited here.

[0132] Those skilled in the art will understand that Figure 7 The block diagram shown is merely a partial structural diagram related to the present invention and does not constitute a limitation on the smart terminal to which the present invention is applied. Specifically, the smart terminal may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0133] This invention also provides a computer-readable storage medium storing a graph learning-based drug-disease association prediction program. When executed by a processor, the graph learning-based drug-disease association prediction program implements the steps of any graph learning-based drug-disease association prediction method provided in this invention.

[0134] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0135] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the above system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this invention. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0136] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0137] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0138] In the embodiments provided by this invention, it should be understood that the disclosed systems / smart terminals and methods can be implemented in other ways. For example, the system / smart terminal embodiments described above are merely illustrative. For instance, the division of modules or units described above is merely a logical functional division, and in actual implementation, it can be divided in other ways. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0139] If the integrated modules / units described above are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable storage medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction.

[0140] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not mean that the essence of the corresponding technical solutions deviates from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for predicting drug-disease associations based on graph learning, characterized in that, The method includes: Obtain a drug similarity network, wherein the drug similarity network includes multiple nodes and multiple edges, one node in the drug similarity network represents a drug to be judged, and the drugs to be judged represented by two nodes connected by an edge in the drug similarity network have a similarity relationship; Obtain a disease similarity network, wherein the disease similarity network includes multiple nodes and multiple edges, one node in the disease similarity network represents a disease to be judged, and the diseases to be judged represented by two nodes connected by an edge in the disease similarity network have a similarity relationship; The drug feature vectors corresponding to each drug to be judged are obtained according to the drug similarity network, and the disease feature vectors corresponding to each disease to be judged are obtained according to the disease similarity network. Multiple combined feature vectors are obtained by combining the drug feature vectors and the disease feature vectors. The drug-disease association relationship corresponding to each combined feature vector is obtained respectively. Among them, a combined feature vector is obtained by concatenating a drug feature vector and a disease feature vector. The drug-disease association relationship corresponding to a combined feature vector is the association relationship between the drug to be judged and the disease to be judged corresponding to the combined feature vector. The process of obtaining drug feature vectors corresponding to each drug to be judged based on the drug similarity network, obtaining disease feature vectors corresponding to each disease to be judged based on the disease similarity network, combining the drug feature vectors and disease feature vectors to obtain multiple combined feature vectors, and obtaining the drug-disease association relationship corresponding to each of the combined feature vectors, includes: inputting the drug similarity network and the disease similarity network into a trained hybrid graph embedding association relationship prediction model, wherein the hybrid graph embedding association relationship prediction model includes a drug feature processing unit, a disease feature processing unit, a feature vector concatenation unit, and a classification unit; obtaining drug feature vectors corresponding to each drug to be judged based on the drug similarity network through the drug feature processing unit; obtaining disease feature vectors corresponding to each disease to be judged based on the disease similarity network through the disease feature processing unit; obtaining multiple combined feature vectors by combining the drug feature vectors and disease feature vectors through the feature vector concatenation unit; and obtaining the drug-disease association relationship corresponding to each of the combined feature vectors through the classification unit. The drug feature processing unit includes a first graph convolutional network layer, a first graph attention subunit, a second graph convolutional network layer, and a first convolutional neural network layer; The disease feature processing unit includes a third graph convolutional network layer, a second graph attention subunit, a fourth graph convolutional network layer, and a second convolutional neural network layer. The first graph attention subunit and the second graph attention subunit have the same structure; the output data of the first graph convolutional network layer is used as the input data of the first graph attention subunit, and the output data of the first graph attention subunit is used as the input data of the second graph convolutional network layer; the output data of the third graph convolutional network layer is used as the input data of the second graph attention subunit, and the output data of the second graph attention subunit is used as the input data of the fourth graph convolutional network layer. The first graph attention subunit and the second graph attention subunit are respectively used to: calculate and update the attention score based on the edge information between nodes in the input data; perform weighted aggregation on the features of adjacent nodes based on the updated attention score; and output a new node feature representation after multi-head splicing, residual connection and normalization processing. The updated attention score is represented as follows: ； This represents the improved attention score matrix. This represents the query matrix used in multi-head attention. This represents the key matrix used in multi-head attention. represent Layer nodes Features represent Layer nodes Features This indicates that attention should be paid to the number of heads. This represents the feature dimension of the hidden layer. express The edge feature information matrix obtained from the layer, represent Layer nodes and nodes Edge characteristics between them.

2. The method for predicting drug-disease association based on graph learning according to claim 1, characterized in that, The step of obtaining the drug feature vector corresponding to each drug to be judged through the drug feature processing unit and the drug similarity network includes: The drug similarity network adjacency matrix and initial drug node representation are obtained based on the drug similarity network, wherein the initial drug node representation is used to reflect the features corresponding to the drug to be judged in the drug similarity network; The drug similarity network adjacency matrix and the drug node representation are input into the first graph convolutional network layer to trigger the first graph convolutional network layer to perform feature processing and obtain the first output drug node representation output by the first graph convolutional network layer; The adjacency matrix of the drug similarity network and the first output drug node representation are input into the first graph attention subunit to trigger the first graph attention subunit to perform feature processing and obtain the second output drug node representation output by the first graph attention subunit; The adjacency matrix of the drug similarity network and the second output drug node representation are input into the second graph convolutional network layer to trigger the second graph convolutional network layer to perform feature processing and obtain the third output drug node representation output by the second graph convolutional network layer. The first output drug node representation and the third output drug node representation are input into the first convolutional neural network layer to obtain the drug feature vector output by the first convolutional neural network layer.

3. The method for predicting drug-disease association based on graph learning according to claim 1, characterized in that, The step of obtaining the disease feature vector corresponding to each disease to be judged through the disease feature processing unit and the disease similarity network includes: The disease similarity network adjacency matrix and initial disease node representation are obtained based on the disease similarity network, wherein the initial disease node representation is used to reflect the features corresponding to the disease to be judged in the disease similarity network; The disease similarity network adjacency matrix and the disease node representation are input into the third graph convolutional network layer to trigger the third graph convolutional network layer to perform feature processing and obtain the first output disease node representation of the third graph convolutional network layer. The disease similarity network adjacency matrix and the first output disease node representation are input into the second graph attention subunit to trigger the second graph attention subunit to perform feature processing and obtain the second output disease node representation output by the second graph attention subunit; The disease similarity network adjacency matrix and the second output disease node representation are input into the fourth graph convolutional network layer to trigger the fourth graph convolutional network layer to perform feature processing and obtain the third output disease node representation output by the fourth graph convolutional network layer. The first output disease node representation and the third output disease node representation are input into the second convolutional neural network layer to obtain the disease feature vector output by the second convolutional neural network layer.

4. The method for predicting drug-disease association based on graph learning according to claim 1, characterized in that, The step of obtaining the drug-disease association relationship corresponding to each of the combined feature vectors through the classification unit includes: Each of the combined feature vectors is input into the classification unit, and the classification unit classifies each of the combined feature vectors to obtain the drug-disease association relationship corresponding to each of the combined feature vectors. The drug-disease association relationship corresponding to a combined feature vector is either that the drug to be judged and the disease to be judged corresponding to the combined feature vector are related, or that the drug to be judged and the disease to be judged corresponding to the combined feature vector are not related.

5. The drug-disease association prediction method based on graph learning according to claim 4, characterized in that, The classification unit is a random forest classifier.

6. The method for predicting drug-disease association based on graph learning according to any one of claims 1-5, characterized in that, The hybrid graph embedding association prediction model is trained according to the following steps: The training drug similarity network and the training disease similarity network in the training data are input into the hybrid graph embedding association prediction model. The predicted drug-disease association is obtained through the hybrid graph embedding association prediction model. The training data includes multiple training network groups. Each training network group includes a training drug similarity network, a training disease similarity network, and labeled drug-disease associations. Based on the predicted drug-disease association and the labeled drug-disease association, the model parameters of the hybrid graph embedding association prediction model are adjusted, and the step of inputting the training drug similarity network and the training disease similarity network from the training data into the hybrid graph embedding association prediction model is continued until the preset training conditions are met, so as to obtain the trained hybrid graph embedding association prediction model.

7. A drug-disease association prediction system based on graph learning, characterized in that, The system includes: A drug network acquisition module is used to acquire a drug similarity network, wherein the drug similarity network includes multiple nodes and multiple edges, one node in the drug similarity network represents a drug to be judged, and the drugs to be judged represented by two nodes connected by an edge in the drug similarity network have a similarity relationship. The disease network acquisition module is used to acquire a disease similarity network, wherein the disease similarity network includes multiple nodes and multiple edges, one node in the disease similarity network represents a disease to be judged, and the diseases to be judged represented by two nodes connected by an edge in the disease similarity network have a similarity relationship. The similarity network processing module is used to obtain drug feature vectors corresponding to each drug to be judged according to the drug similarity network, obtain disease feature vectors corresponding to each disease to be judged according to the disease similarity network, combine the drug feature vectors and the disease feature vectors to obtain multiple combined feature vectors, and obtain the drug-disease association relationship corresponding to each combined feature vector. A combined feature vector is obtained by concatenating a drug feature vector and a disease feature vector, and the drug-disease association relationship corresponding to a combined feature vector is the association relationship between the drug to be judged and the disease to be judged corresponding to that combined feature vector. The similarity network processing module is specifically used for: inputting the drug similarity network and the disease similarity network into a trained hybrid graph embedding association prediction model, wherein the hybrid graph embedding association prediction model includes a drug feature processing unit, a disease feature processing unit, a feature vector concatenation unit, and a classification unit; obtaining drug feature vectors corresponding to each drug to be judged based on the drug similarity network through the drug feature processing unit; obtaining disease feature vectors corresponding to each disease to be judged based on the disease similarity network through the disease feature processing unit; obtaining multiple combined feature vectors by combining the drug feature vectors and the disease feature vectors through the feature vector concatenation unit; and obtaining the drug-disease association relationship corresponding to each of the combined feature vectors through the classification unit. The drug feature processing unit includes a first graph convolutional network layer, a first graph attention subunit, a second graph convolutional network layer, and a first convolutional neural network layer; The disease feature processing unit includes a third graph convolutional network layer, a second graph attention subunit, a fourth graph convolutional network layer, and a second convolutional neural network layer. The first graph attention subunit and the second graph attention subunit have the same structure; the output data of the first graph convolutional network layer is used as the input data of the first graph attention subunit, and the output data of the first graph attention subunit is used as the input data of the second graph convolutional network layer; the output data of the third graph convolutional network layer is used as the input data of the second graph attention subunit, and the output data of the second graph attention subunit is used as the input data of the fourth graph convolutional network layer. The first graph attention subunit and the second graph attention subunit are respectively used to: calculate and update the attention score based on the edge information between nodes in the input data; perform weighted aggregation on the features of adjacent nodes based on the updated attention score; and output a new node feature representation after multi-head splicing, residual connection and normalization processing. The updated attention score is represented as follows: ； This represents the improved attention score matrix. This represents the query matrix used in multi-head attention. This represents the key matrix used in multi-head attention. represent Layer nodes Features represent Layer nodes Features This indicates that attention should be paid to the number of heads. This represents the feature dimension of the hidden layer. express The edge feature information matrix obtained from the layer, represent Layer nodes and nodes Edge characteristics between them.

8. A smart terminal, characterized in that, The smart terminal includes a memory, a processor, and a graph learning-based drug-disease association prediction program stored in the memory and executable on the processor. When the graph learning-based drug-disease association prediction program is executed by the processor, it implements the steps of the graph learning-based drug-disease association prediction method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a graph learning-based drug-disease association prediction program, which, when executed by a processor, implements the steps of the graph learning-based drug-disease association prediction method as described in any one of claims 1-6.

Images