Adverse reaction prediction method and system based on drug symmetric element path

Abstract

The present application relates to a kind of adverse reaction prediction method and system based on drug symmetry meta path, the prediction method is by designing the information aggregation mechanism based on meta path, make full use of the complex semantic information in HIN, meta path level split not only helps to learn high-quality long path node representation, can also learn more representative node features, effectively capture semantic information, improve the accuracy and stability of prediction, provide strong guarantee for drug safety use.

Description

Technical Field

[0001] This invention belongs to the field of adverse drug reaction prediction, specifically relating to an adverse reaction prediction method and system based on drug symmetric metapath. Background Technology

[0002] According to the authoritative definition given by the World Health Organization (WHO) Collaborating Centre for International Drug Testing, adverse drug reactions (ADRs) specifically refer to a series of harmful conditions that suddenly arise when a qualified drug is used normally and as prescribed by a doctor. These include side effects that contradict the original treatment purpose, worrying toxic reactions, unavoidable allergic reactions, far-reaching "three-way effects" (causing fetal malformations, gene mutations, and inducing cancer), as well as lingering aftereffects and chain-reaction-like secondary reactions.

[0003] Due to the harmfulness and severity of adverse drug reactions, measures need to be taken to strictly monitor the safety of marketed drugs. Once a highly suspected adverse reaction signal is detected, it must be reported to the relevant authorities for research, analysis and management to reduce potential medication risks.

[0004] In the field of predicting drug side effect associations (DSAs), there are currently three key technological approaches: docking-based methods, machine learning-based methods, and network-based methods.

[0005] (1) Docking-based methods: predicting drug side effect associations (DSAs) by identifying the binding of ligands to target binding sites, performing docking prediction through steps such as molecular shape matching, mechanical optimization and energy minimization, or attempting to dock one or more small molecules to receptor sites to find potential ligands.

[0006] (2) Machine learning-based methods: Treating DSA prediction as a classification problem, using known DSAs as training data to train a classifier. Traditional machine learning methods include using small molecule chemical structure information to train a random forest classifier; deep learning methods include combining outer product operations, multilayer perceptrons and convolutional neural networks to learn the representation of drugs and side effects, using deep graph convolutional neural networks and bidirectional long short-term memory networks to extract features and then using fully connected layers for prediction, and using deep neural network models to extract drug features for prediction.

[0007] (3) Network-based methods: These methods construct a network by treating DSA-related biological entities as nodes and connections as edges, and model DSA prediction as a link prediction problem. One type of method uses an adjacency matrix to represent the network, employs matrix factorization and diffusion methods to learn drug and side effect representations; it then projects the association matrix and introduces graph regularization to construct a graph, using linear neighborhood similarity to measure drug similarity and propagating known side effect information based on the similarity graph for prediction. Another type of method applies graph neural networks to learn features, such as using graph attention networks to integrate data for feature learning, connecting graph embeddings and node embeddings as features, and predicting through matrix completion.

[0008] These existing methods all have certain drawbacks or shortcomings. For example, they fail to make full use of various data that may affect the occurrence of drug side effects, resulting in the inability to learn high-quality features to improve the prediction performance of DSAs; or they have difficulty capturing the complex semantic relationships between biological entities during the occurrence of drug side effects, as different biological entities are heterogeneous, which poses a challenge to capturing the complex semantic relationships between drugs and side effects; or they lack interpretability of the predicted DSAs, only providing the prediction results without clearly explaining the basis for the prediction, while interpretability is very important for doctors' prescription decisions.

[0009] Given the need for further optimization in current drug side effect association (DSAs) prediction techniques, this study innovatively proposes a heterogeneous information network (HIN)-based model (MPGNN-DSA) specifically for DSA prediction. In constructing this HIN, multiple DSA-related datasets are comprehensively integrated, acting as a "data hub" that gathers information from various sources. This allows for the comprehensive mining and full utilization of the complex semantic associations inherent within the data, laying a solid foundation for subsequent accurate predictions. Furthermore, a carefully designed meta-path-based feature learning mechanism is employed. This mechanism acts like a precise "scalpel," skillfully cutting into the complex structure of the HIN and efficiently capturing the complex semantics hidden deep within the network through meta-paths. It's akin to exploring precious treasures along hidden veins on an ancient map, tailoring high-quality representations for drugs and side effects to clearly present their features. Moreover, meta-path segmentation techniques are used to perform refined feature fusion on the displayed features. This cutting-edge method is akin to equipping a predictive "telescope" with a high-powered lens, capturing semantic information along longer pathways for predicted drug side effects. This allows different meta-paths to learn drug and side effect representations independently, enabling them to exchange information and aggregate more valuable data, thus significantly improving the accuracy and reliability of predictions. By ranking the predicted drug side effects according to their probability and selecting the top K most likely side effects as key areas of focus, this method achieves proactive prediction of potential drug side effects, injecting new momentum into the field of pharmaceutical safety and providing solid and reliable decision-making references. Summary of the Invention

[0010] To address the shortcomings of existing technologies, this invention constructs a heterogeneous graph based on the MPGNN-DSA algorithm. By performing explicit information transfer and feature fusion along carefully designed meta-paths in the graph, adverse drug reactions can be accurately predicted.

[0011] The specific technical solution of the present invention is as follows:

[0012] In a first aspect, the present invention provides a method for predicting adverse reactions based on drug symmetric metapaths, comprising the steps of:

[0013] 1) Collect data on the interrelationships between drug Dr, protein P, disease Di, and adverse reaction A.

[0014] 2) Construct a basic model for predicting adverse drug reactions based on symmetric meta-pathways, including the following steps:

[0015] 2.1 Constructing a heterogeneous information network: Based on the collected data, construct a matrix of interrelationships to form a heterogeneous information network;

[0016] 2.2 Model Training; Select a symmetrical long meta-path to segment from the inside out, perform feature fusion on the segmented meta-paths, update the node features based on the feature fusion, reconstruct the relationship matrix between nodes, and obtain the trained basic model for drug pair adverse reaction prediction based on symmetrical meta-paths.

[0017] 3) Use the drug-adverse reaction prediction basic model based on symmetric metapath trained in step 2) to predict the relationship between drugs and adverse reactions.

[0018] The core of this prediction method lies in the explicit information transmission and feature fusion along the meta-path to accurately predict adverse drug reactions. This process relies on constructing a reasonable heterogeneous graph structure, accurately selecting the meta-path, and effectively fusing node features.

[0019] In the prediction method of this invention, the explicit feature vector of a drug is constructed mainly based on the drug's indication characteristics and its biological information characteristics.

[0020] In some preferred embodiments, in step 1), the indication characteristics of the drug are extracted using SIDER (Side Effect Resource); the bioinformatics characteristics of the drug are extracted from the DrugBank database; the STITCH database collects drug-to-drug association information; the STRING database collects protein-protein interaction related information; and the public comparative toxicology genomics database collects drug-to-disease and protein-to-disease relationship information.

[0021] The SIDER database (Drug and Side Effects Resource Database) (see http: / / sideeffects.embl.de / ) contains information on marketed drugs and their recorded adverse reactions and indications. Much of this information was extracted from public documents and materials using text mining techniques, primarily including the frequency of drug side effects, drug and side effect classifications, and related information. The drug indication features extracted from the SIDER database in this invention can be denoted as V. A-Dr .

[0022] The DrugBank database (see https: / / go.drugbank.com / ) integrates bioinformatics and cheminformatics resources, providing comprehensive molecular information on detailed drug data, drug targets, and their mechanisms, including medicinal chemistry, pharmacology, pharmacokinetics, ADME, and their interactions. This invention utilizes drug bioinformatics features extracted from the DrugBank database, denoted as VA-P.

[0023] The STITCH database (see https: / / ngdc.cncb.ac.cn / databasecommons / database / id / 208) integrates data resources from diverse fields, covering not only key knowledge in biology and chemistry but also deeply linking drug-drug interactions in practical application scenarios. From this database, we can accurately capture the synergistic and antagonistic relationships between drugs, as well as the detailed dynamics of their cooperation or mutual restraint in different disease treatment systems. This invention, based on drug-to-drug features collected from the STITCH database, can be labeled V. Dr-Dr .

[0024] The STRING database (see https: / / cn.string-db.org) is built upon public databases and literature information. It collects content from multiple public databases such as UniProt, KEGG, NCBI, and Gene Ontology, integrating these data to generate a comprehensive and wide-ranging protein-protein interaction network database. The protein-protein interaction information collected in this invention based on the STRING database can be denoted as V. P-P .

[0025] The Public Comparative Toxicology Genomics Database (CTD, http: / / ctdbase.org / ) tightly links toxicological information related to chemicals, genes, phenotypes, diseases, and exposures. By integrating literature sources and carefully curated interactive content, it creates a rich knowledge base capable of effectively coordinating key information such as chemical exposures and their resulting biological effects from heterogeneous cross-species data. This invention is based on drug-to-disease and protein-to-disease relationship information collected from the Public Comparative Toxicology Genomics Database, which can be tagged as V. Dr-Di V P-Di .

[0026] In some preferred embodiments, the heterogeneous graph network used is composed of the relationships between drugs, targets, diseases, and adverse reactions. Therefore, based on the collected data, the key matrices are first constructed as follows: drug-drug (Dr_Dr), drug-target (Dr_P), protein-protein (P_P), drug-disease (Dr_Di), drug-side effect (A_Dr), and protein-disease (P_Di).

[0027] In some preferred embodiments, each bit of these relation matrices consists of {0, 1}. In some specific embodiments, they are shown below:

[0028]

[0029] Matrix[x][y] represents the relationship between nodes. Matrix[i][j] = 1 means that there is a relationship between node i and node j, and Matrix[i][j] = 0 means that there is no relationship between node i and node j.

[0030] In this invention, it is believed that the semantic information features of long meta-paths should be aggregated. However, previous aggregation methods cannot capture the features of long paths. Therefore, a method for segmenting based on symmetrical long meta-paths is proposed, and the segmented meta-paths are fused for features. Furthermore, the feature fusion method is improved, and the symmetrical long meta-path is denoted as MP.

[0031] In some preferred embodiments, the symmetric long-member path is selected as MP = [[A-Dr,Dr-A],[[Di-Dr,Dr-Di],[P-Dr,Dr-P]]]; where A-Dr represents the relationship between adverse reactions and drugs, Dr-A represents the relationship between drugs and adverse reactions, Di-Dr represents the relationship between diseases and drugs, Dr-Di represents the relationship between drugs and diseases, P-Dr represents the relationship between targets and drugs, and Dr-P represents the relationship between drugs and targets.

[0032] According to MP, long-term paths are divided into short paths Dr-Di-Dr, Dr-P-Dr, and A-Dr-A, where A-Dr-A is denoted as... Let Dr be the intermediate node of the shortest path in the aggregation, denoted as Am, and A be the starting node, denoted as A.

[0033] The edge type uses a mapping function.

[0034] The shortest path after partitioning is denoted as If the path is divided into K short paths, the set of short paths is denoted as .

[0035] Long path segmentation is performed from the inside out, and node feature fusion is also performed from the inside out. The target node is defined as an adverse reaction, and its type is denoted as As.

[0036] The short path feature fusion after segmentation involves taking an intermediate node j of type Am and fusing it along the short path set. Perform feature fusion.

[0037] In some preferred embodiments, the short path feature fusion formula is as follows:

[0038]

[0039] The relationships between each neighbor of the target node i (i.e., As) are as follows: One of the intermediate nodes is Used to receive along The characteristics of all neighbor nodes of type As on the relation path, this process is denoted as The accumulated features are fused with their original feature embeddings, and finally a fully connected layer is formed along the r1 path. At this point, the intermediate node type is A. m It has already incorporated the features of the target node and updated its own features.

[0040] The short path features have been fused and updated through the above process, and the intermediate node feature representations are updated. Finally, the features of all nodes along the r1 path are fused and the features of the final target node are updated.

[0041] In some preferred embodiments, the formula for feature fusion along the nodes of path r1 is as follows:

[0042]

[0043] Among them, intermediate node type A m The mean is accumulated along all r1 paths, which explicitly merges multiple updated features passed along the r2 path, denoted as . Given an initial target node of type As with K short meta-paths, set the path along these K short meta-paths. The features on the data are accumulated to the intermediate nodes, denoted as... After all intermediate nodes receive the returned information, they fuse the features of the original target node of type A, and finally perform a full connection at one layer. At this point, the starting node of type As completes the feature fusion of the long path.

[0044] In some preferred embodiments, after fusing the long and short path features described above, the final updated node features are obtained using the following feature fusion formula:

[0045]

[0046] in, Let h represent the set of neighboring nodes of node i. i This indicates the node i that needs to be updated. This indicates the current node i being updated.

[0047] After updating the features of Di, Dr, A, and P type nodes, it is necessary to reconstruct the relationship matrix between nodes.

[0048] Each type of node is obtained by multiplying the updated node feature representation by a learnable diagonal matrix, and then by the transpose of the updated node feature representation matrix, as shown in the following formula:

[0049] Matrix = XΛX T ;

[0050] In this process, a diagonal matrix, denoted as Λ, is generated based on the feature dimensions of all nodes. The value of Λ is randomly generated and conforms to a normal distribution with a mean of 0 and a standard deviation of 0.1. The node feature representation of each type of node after feature fusion through long-dimensional paths is denoted as X.

[0051] The reconstructed relationship matrix between drugs, diseases, targets, and adverse reactions is shown below:

[0052]

[0053] The loss value after training is used to calculate the relationship matrix between drugs, diseases, targets, and adverse reactions. The loss function is shown below:

[0054] Matrix_loss=SUM[(Matrix-re_Matrix) 2 ],

[0055] Where Matrix is ​​the original feature matrix of this type of node, and re_Matrix is ​​the feature matrix of this type of node after training.

[0056] In some preferred embodiments of the present invention, the relationship between a drug and an adverse reaction is determined based on the value of the trained feature matrix re_A_Dr[x][y]. For example, in some embodiments, the value of the i-th adverse reaction and the j-th drug in the matrix is ​​re_A_Dr[i][j]. If the value is greater than a set threshold of 0.5, it indicates that there is a relationship between the i-th adverse reaction and the j-th drug; otherwise, there is no relationship. Furthermore, the larger the value, the stronger the corresponding relationship. The values ​​are then sorted, and the top K adverse reactions with the strongest correlation are selected as potential adverse reactions of the drug.

[0057] This invention also provides an adverse reaction prediction system based on drug symmetric metapaths. This system is constructed based on the aforementioned prediction method and includes:

[0058] The model training module is used to collect data on the interrelationships between drug Dr, protein P, disease Di, and adverse reaction A, and to train a basic model for predicting adverse drug reactions based on symmetric metapaths; and

[0059] The prediction module is used to predict the relationship between drugs and adverse reactions using a trained basic model for drug adverse reaction prediction based on symmetric metapaths.

[0060] The beneficial effects of this invention are:

[0061] The adverse drug reaction prediction method of this invention is based on the assumption that the interrelationships between drugs and other elements (such as targets, diseases, etc.) in a complex network can reveal potential adverse reaction patterns. It constructs a heterogeneous graph using multi-dimensional information of the drug, and optimizes the consideration of adverse drug reaction-related factors by performing explicit information transmission and feature fusion along carefully designed meta-paths in the graph. It identifies nodes and their relationship paths that are semantically closely related to the target drug in the heterogeneous graph structure. It uses these associations to predict the potential association degree of the target drug with the target adverse reaction. Finally, it ranks the adverse reactions according to the predicted association degree of the target drug with each adverse reaction, and selects the top K adverse reactions with the highest association degree as potential adverse reactions of the target drug.

[0062] This method, through the design of a meta-path-based information aggregation mechanism, fully utilizes the complex semantic information in HIN. Meta-path-level segmentation not only helps learn high-quality long-path node representations but also learns more representative node features, effectively capturing semantic information and improving prediction accuracy and stability. The prediction method of this invention achieves an AUC of 0.977 and an AUPR of 0.841, outperforming all baseline models. Moreover, compared to the better-performing baseline model simpleHGN, this model has a lower standard deviation in prediction results, indicating stronger prediction stability across different data samples. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of the meta-path selection method of the present invention;

[0064] Figure 2 This is a schematic diagram of the prediction method of the present invention. Detailed Implementation

[0065] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0066] Example 1

[0067] like Figure 2 As shown, this method for predicting adverse drug reactions based on symmetric metapaths includes the following steps:

[0068] S1 collects data on the interrelationships between drug Dr, protein P, disease Di, and adverse reaction A.

[0069] Data collection can be performed by extracting and collecting data from existing databases.

[0070] The data in this embodiment comes from the SIDER (Side Effect Resource) database to extract the indication characteristics of each drug; the DrugBank database to extract the biological information characteristics of each drug, including targets, enzymes, transport proteins, carriers, etc.; the STITCH database to collect drug-to-drug association information; the STRING database to collect protein-protein interaction related information; and the Comparative Toxicology Genomics Database (CTD) to collect drug-to-disease and protein-to-disease relationship information.

[0071] Based on this data, a drug-adverse reaction association matrix Dr_A is constructed, where rows represent drugs and columns represent adverse reactions. Each element value Dr_Aij represents the association strength between drug i and the adverse reaction. The number of drugs is denoted as N, and the number of adverse reactions as M. It can be understood that in the drug-adverse reaction association matrix Dr_A, a subset of elements with a value of 1 indicates a relationship between the drug and the corresponding adverse reaction; conversely, an element with a value of 0 indicates no relationship between the drug and the corresponding adverse reaction. The goal of this invention is to predict the association strength of unknown drug-adverse reactions based on known association strengths, and then sort them in descending order of association strength. The top K adverse reactions with the highest association strengths are identified as potential adverse reactions of the drug, thereby achieving the prediction of potential adverse reactions of the drug.

[0072] Preferably, in this embodiment, 680 drug nodes, 732 target nodes, 5797 disease nodes, and 2600 side effect nodes were extracted based on the DrugBank database, SIDER database, STITCH database, STRING database, and CTD public comparative toxicology genomics database.

[0073] Preferably, in this embodiment, 2954 drug-target relationships are collected based on the DrugBank database. A relationship matrix is ​​constructed based on these relationships, with each element of the matrix consisting of {0, 1}, indicating whether the drug has the carrier protein / enzyme / target protein / transporter protein represented by that dimension.

[0074] Preferably, in this embodiment, 10,821 drug-drug relationships are collected based on the STITCH database. A relationship matrix is ​​constructed based on these relationships, with each element of the matrix consisting of {0, 1}, indicating whether the drug has a corresponding drug in that dimension.

[0075] Preferably, in this embodiment, 69,544 protein-protein relationships were collected based on the STRING database. A relationship matrix was constructed based on these relationships, with each element of the matrix consisting of {0, 1}, indicating whether the carrier protein / enzyme / target protein / transporter has the carrier protein / enzyme / target protein / transporter represented by that dimension.

[0076] Preferably, in this embodiment, 87,291 drug-disease relationships are collected based on the CTD database. A relationship matrix is ​​constructed based on these relationships, with each element of the matrix consisting of {0, 1}, indicating whether the drug has the disease represented by that dimension.

[0077] Preferably, in this embodiment, 432,662 protein-disease relationships were collected based on the CTD database. A relationship matrix was constructed based on these relationships, with each element of the matrix consisting of {0, 1}, indicating whether the carrier protein / enzyme / target protein / transporter protein has the disease represented by that dimension.

[0078] Preferably, in this embodiment, the drug indication features extracted based on the SIDER database are a one-hot vector containing 17,393 high-frequency indications. Each bit of the vector consists of {0, 1}, indicating whether the drug has the indication represented by that dimension.

[0079] S2 constructs a basic model for predicting adverse drug reactions based on symmetric metapaths.

[0080] 2.1 Constructing a heterogeneous information network: Based on the collected data, construct a matrix of interrelationships to form a heterogeneous information network.

[0081] Specifically, based on the data collected above, this embodiment constructs the following key matrices: drug-drug (Dr_Dr), drug-target (Dr_P), protein-protein (P_P), drug-disease (Dr_Di), drug-side effect (A_Dr), and protein-disease (P_Di), as shown below:

[0082]

[0083] Matrix[x][y] represents the relationship between nodes. Matrix[i][j] = 1 means that there is a relationship between node i and node j, and Matrix[i][j] = 0 means that there is no relationship between node i and node j.

[0084] Based on these interrelationship matrices, a heterogeneous information network is formed.

[0085] 2.2 Model Training: A symmetrical long meta-path is selected to be segmented from the inside out. The segmented meta-paths are then fused with features. Based on the updated node features after feature fusion, the relationship matrix between nodes is reconstructed to obtain the trained basic model for predicting adverse drug reactions based on symmetrical meta-paths.

[0086] In this embodiment, the constructed symmetric long-member path is MP = [[A-Dr,Dr-A],[[Di-Dr,Dr-Di],[P-Dr,Dr-P]]], that is, (A-Dr-Di-Dr-A, A-Dr-P-Dr-A), where A-Dr represents the relationship between adverse reactions and drugs, Dr-A represents the relationship between drugs and adverse reactions, Di-Dr represents the relationship between diseases and drugs, Dr-Di represents the relationship between drugs and diseases, P-Dr represents the relationship between targets and drugs, and Dr-P represents the relationship between drugs and targets.

[0087] Metapath segmentation diagram as shown below Figure 1 As shown, according to MP, the long-element path is divided into short paths Dr-Di-Dr, Dr-P-Dr, and A-Dr-A, where A-Dr-A is denoted as... Let Dr be the intermediate node of the shortest path in the aggregation, denoted as Am, and A be the starting node, denoted as As; where, The shortest path is denoted as If the path is divided into K short paths, the set of short paths is denoted as .

[0088] In this embodiment, the long symmetric metapath is divided into two short metapaths Dr-Di-Dr and Dr-P-Dr.

[0089] Long path segmentation is performed from the inside out, and node feature fusion is also performed from the inside out, where the target node is defined as an adverse reaction and its type is As.

[0090] The short path feature fusion after segmentation involves taking an intermediate node j of type Am and fusing it along the short path set. Feature fusion is performed. In this embodiment, the features of an intermediate node Dr are fused along the short path set {Dr-Di-Dr, Dr-P-Dr}.

[0091] The formula for short path feature fusion is as follows:

[0092]

[0093] The relationships between each neighbor of the target node i (i.e., As) are as follows: One of the intermediate nodes is Used to receive along The characteristics of all neighbor nodes of type As on the relation path, this process is denoted as The accumulated features are fused with their original feature embeddings, and finally a fully connected function is performed along the r1 path to complete the short path feature fusion, i.e., the feature representation update of intermediate nodes. Finally, the features of all nodes along the r1 path are fused and the features of the final target node are updated.

[0094] The formula for feature fusion along the nodes of path r1 is as follows:

[0095]

[0096] Among them, the intermediate node type Dr is averaged along all r1 paths, which explicitly integrates multiple updated features passed along the r2 path, denoted as... Given that there are K short meta-paths for the initial target node type A, we set the path along these K short meta-paths. The features on the data are accumulated to the intermediate nodes, denoted as... After all intermediate nodes receive the returned information, they fuse the features of the original target node of type A, and finally perform a full connection at one layer. At this point, the starting node of type As completes the feature fusion of the long path.

[0097] After fusing the long and short path features mentioned above, the final updated node features are obtained using the following feature fusion formula:

[0098]

[0099] in, Let h represent the set of neighboring nodes of node i. i This indicates the node i that needs to be updated. This indicates the current node i being updated.

[0100] After updating the features of Di, Dr, A, and P type nodes, the relationship matrix between nodes is reconstructed.

[0101] Each type of node is obtained by multiplying the updated node feature representation by a learnable diagonal matrix, and then by the transpose of the updated node feature representation matrix, as shown in the following formula:

[0102] Matrix = XΛX T ;

[0103] In this process, a diagonal matrix, denoted as Λ, is generated based on the feature dimensions of all nodes. The value of Λ is randomly generated and conforms to a normal distribution with a mean of 0 and a standard deviation of 0.1. The node feature representation of each type of node after feature fusion through long-dimensional paths is denoted as X.

[0104] The reconstructed relationship matrix between drugs, diseases, targets, and adverse reactions is shown below:

[0105]

[0106] This leads to the trained basic model for predicting adverse drug reactions based on symmetric metapaths.

[0107] The loss value after training is used to calculate the relationship matrix between drugs, diseases, targets, and adverse reactions. The loss function is shown below:

[0108] Matrix_loss=SUM[(Matrix-re_Matrix) 2 ],

[0109] Where Matrix is ​​the original feature matrix of this type of node, and re_Matrix is ​​the feature matrix of this type of node after training.

[0110] S3 uses a trained symmetric metapath-based drug-adverse reaction prediction model to predict the relationship between drugs and adverse reactions.

[0111] The relationship between drugs and adverse reactions is determined based on the value of the feature matrix re_A_Dr[x][y] after training. In some embodiments, the value of the i-th adverse reaction and the j-th drug in the matrix is ​​re_A_Dr[i][j]. If the value is greater than the set threshold of 0.5, it indicates that there is a relationship between the i-th adverse reaction and the j-th drug; otherwise, there is no relationship. The larger the value, the stronger the relationship. The adverse reactions are then sorted according to their values, and the top K adverse reactions with the strongest correlation are selected as potential adverse reactions of the drug.

[0112] The method execution flow provided in the embodiments of this application can be run on devices such as personal computers, servers, and cloud computing platforms.

[0113] Example 2

[0114] This embodiment provides an adverse reaction prediction system based on the drug symmetric metapath. The system is constructed based on the prediction method of Embodiment 1 and includes:

[0115] The model training module is used to collect data on the interrelationships between drug Dr, protein P, disease Di, and adverse reaction A, and to train a basic model for predicting adverse drug reactions based on symmetric metapaths; and

[0116] The prediction module is used to predict the relationship between drugs and adverse reactions using a trained basic model for drug adverse reaction prediction based on symmetric metapaths.

[0117] The specific prediction system in this embodiment achieves an AUC of 0.977 and an AUPR of 0.841, outperforming all baseline models. Furthermore, compared to the better-performing baseline model simpleHGN, this model exhibits a lower standard deviation in its predictions, indicating stronger predictive stability across different data samples.

[0118] The technical features of the embodiments described above can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. A method for predicting adverse reactions based on drug symmetric meta-pathways, characterized in that, Including the following steps: 1) Collect data on the interrelationships between drug Dr, protein P, disease Di, and adverse reaction A. 2) Construct a basic model for predicting adverse drug reactions based on symmetric meta-pathways, including the following steps: 2.1 Constructing a heterogeneous information network: Based on the collected data, construct a matrix of interrelationships to form a heterogeneous information network; 2.2 Model Training: A symmetrical long meta-path is selected to be segmented from the inside out. The segmented meta-paths are then fused with features. The node features are updated based on the fused features, and the relationship matrix between nodes is reconstructed to obtain the trained basic model for drug pair adverse reaction prediction based on symmetrical meta-paths. 3) Use the drug-adverse reaction prediction basic model based on symmetric metapath trained in step 2) to predict the relationship between drugs and adverse reactions; In step 2.2: the symmetrical long-member path is selected as MP=[[A-Dr,Dr-A],[[Di-Dr,Dr-Di],[P-Dr,Dr-P]]]; where A-Dr represents the relationship between adverse reactions and drugs, Dr-A represents the relationship between drugs and adverse reactions, Di-Dr represents the relationship between diseases and drugs, Dr-Di represents the relationship between drugs and diseases, P-Dr represents the relationship between targets and drugs, and Dr-P represents the relationship between drugs and targets. The symmetric long-element path MP is divided into short paths Dr-Di-Dr, Dr-P-Dr, and A-Dr-A, where A-Dr-A is denoted as... Let Dr be the intermediate node of the aggregated short path, denoted as Am, and A be the starting node, denoted as As. The shortest path is denoted as If the path is divided into K short paths, the set of short paths is denoted as . , =1,2...k; Feature fusion of the segmented short paths involves taking an intermediate node j of type Am and fusing it along the set of short paths. Perform feature fusion; The short path is obtained through feature fusion as follows: ， ， The relationship between each neighbor of target node i is as follows: One intermediate node is Used to receive along The characteristics of all neighbor nodes of type As on the relation path, this process is denoted as The accumulated features are fused with the original feature embedding, and a fully connected function is performed along the r1 path to complete the short path feature fusion, i.e., the intermediate node feature representation update. Finally, the features of all nodes along the r1 path are fused and the features of the final target node are updated. The formula for feature fusion of nodes along path r1 is as follows: ； In this process, the intermediate nodes are averaged along all r1 paths, which explicitly integrates multiple updated features passed along the r2 path, denoted as... There are K short paths starting from the node. Set the path along the K short paths... The features on the data are accumulated to the intermediate nodes, denoted as... After all intermediate nodes receive the returned information, they fuse the features of the original target node (As type) and finally perform a full connection at one layer. The starting node of type As completes the feature fusion of the path.

2. The prediction method according to claim 1, characterized in that, In step 1), the indication characteristics of the drug are extracted from the SIDER database; the bioinformatics characteristics of the drug are extracted from the DrugBank database. The STITCH database collects drug-to-drug association information; the STRING database collects information related to protein-protein interactions. Public comparative toxicology genomics databases collect information on drug-to-disease and protein-to-disease relationships.

3. The prediction method according to claim 1, characterized in that, In step 2.1, the interrelationship matrix includes: drug-drug interrelationship matrix Dr_Dr, drug-target interrelationship matrix Dr_P, protein-protein interrelationship matrix P_P, drug-disease interrelationship matrix Dr_Di, drug-adverse reaction interrelationship matrix A_Dr, and protein-disease interrelationship matrix P_Di.

4. The prediction method according to claim 1, characterized in that, Each bit of the cross-relationship matrix consists of {0, 1}.

5. The prediction method according to claim 1, characterized in that, The updated node features are obtained through the following feature fusion formula: ， in, Let h represent the set of neighboring nodes of the target node i. i This indicates the node i that needs to be updated. This indicates that the node characteristics are being updated.

6. The prediction method according to claim 1, characterized in that, In step 3), the relationship between the drug and adverse reactions is determined based on the value of the feature matrix re_A_Dr[x][y] of the drug and adverse reactions after training.

7. An adverse reaction prediction system based on drug symmetric meta-pathways, characterized in that, The system is constructed based on the method according to any one of claims 1-6, characterized in that the system comprises: The model training module is used to collect data on the interrelationships between drug Dr, protein P, disease Di, and adverse reaction A, and to train a basic model for predicting adverse drug reactions based on symmetric metapaths; and The prediction module is used to predict the relationship between drugs and adverse reactions using a trained basic model for drug adverse reaction prediction based on symmetric metapaths.

Images