Intelligent adverse drug reaction prediction method and device based on multi-source data fusion

Abstract

The embodiment of the invention discloses an intelligent prediction method and device for adverse drug reaction based on multi-source data fusion, relates to the technical field of drug-gene-adverse drug reaction neural networks, and aims to solve the problem of low ADR prediction performance due to lack of a more effective feature analysis method and design of a more efficient model in the prior art. The method comprises the following steps: acquiring multi-source data; performing standardization processing on the multi-source data, constructing drug descriptor features, and obtaining a drug similarity vector set; constructing adverse drug reaction descriptor features, and obtaining an adverse drug reaction similarity vector set; and training the target adverse drug reaction intelligent prediction model based on the drug similarity vector set and the adverse drug reaction similarity vector set, and after training is completed, predicting an adverse reaction generated when a target drug is applied to a target individual based on the target adverse drug reaction intelligent prediction model.

Description

Technical Field

[0001] The present application relates to the technical field of drug-gene-adverse drug reaction neural network, and in particular to a multi-source data fusion method and device for intelligent prediction of adverse drug reactions. Background Art

[0002] Adverse drug reactions (ADRs) are an important issue in clinical treatment, which may cause serious health risks or even lead to death. The current collection of pharmacogenomics data from clinical observations has increased significantly, making it possible to predict ADRs based on existing experience. Faced with a large amount of complex and diverse pharmacogenomics data, if the rules are efficiently discovered, the prediction of unknown ADRs will play a vital role in clinical pharmacovigilance. In recent years, computer technology with ultra-high computing power and artificial intelligence analysis capabilities and related algorithms have played an increasingly significant role in assisting clinical ADR prediction and analysis, but how to effectively process these massive data and mine the objective rules contained in them for accurate prediction remains a huge challenge. Existing studies have used source data such as the LINCS L1000 project, the STITCH database, and the Comparative Toxicogenomics Database (CTD) to build models and have achieved certain success. However, these methods still have some shortcomings in terms of effective feature extraction and optimal model structure design. In addition, although these public source databases provide a rich genomic information basis for ADR prediction, the huge amount of genetic data also poses considerable challenges to feature processing and model design. Therefore, it is necessary to develop more effective feature analysis methods and design more efficient models to further improve the performance of ADR prediction. Summary of the invention

[0003] The main purpose of this application is to provide a method and device for intelligent prediction of adverse drug reactions based on multi-source data fusion, aiming to solve the problem that the prior art lacks more effective feature analysis methods and designs more efficient models, resulting in low performance in predicting ADRs.

[0004] To achieve the above objectives, the technical solutions adopted in the embodiments of the present application are as follows:

[0005] In a first aspect, the present application provides a method for intelligently predicting adverse drug reactions by fusion of multi-source data, comprising the following steps:

[0006] Acquire multi-source data; wherein the multi-source data includes comparative toxicological genomics data, LINCS L1000 data, PubChem data, medical subject thesaurus data and SIDER data; the comparative toxicological genomics data is used to interpret drug-gene interactions, the LINCS L1000 data is used to interpret gene expression signals, the PubChem data is used to interpret drug SMILES, the medical subject thesaurus data is used to interpret gene-disease interactions, and the SIDER data is used to interpret drug-side effect pairs;

[0007] After standardizing the multi-source data, constructing drug descriptor features to obtain a drug similarity vector set; constructing adverse drug reaction descriptor features to obtain a drug adverse reaction similarity vector set;

[0008] The target drug adverse reaction intelligent prediction model is trained based on the drug similarity vector set and the drug adverse reaction similarity vector set. After the training is completed, the adverse reaction generated when the target drug is used for the target individual is predicted based on the target drug adverse reaction intelligent prediction model.

[0009] As some optional methods of the present application, the drug similarity vector set includes: chemical structure similarity of two drugs, drug similarity based on drug-gene interaction and drug similarity of gene expression difference after drug perturbation;

[0010] The adverse drug reaction similarity vector set includes: semantic similarities of different disease ontologies and a gene-disease association relationship set.

[0011] As some optional methods of the present application, the chemical structure similarity values ​​of the two drugs are obtained based on the following steps:

[0012] By using the drug Compound ID provided in SIDER, the SMILES string representation of each drug structure was downloaded from PubChem in batches, and then converted into a topological fingerprint using the Rdkit tool in Python; the topological fingerprint is generated based on the topological structure and rotation angle of the four-membered ring in the molecule to describe the stereochemistry and interaction of the molecule; and the Tanimoto coefficient is used as a metric to obtain the drug d i and drugs j Chemical structure similarity values ​​of two drugs ( Represents the similarity between drugs in the drug chemical structure space, where cs(i,j) represents the chemical space structure of two different drugs).

[0013] As some optional methods of the present application, the drug similarity based on drug-gene interaction refers to:

[0014] Drug-gene interactions are represented by One-Hot encoding, which converts each unique gene interaction into a binary vector; drug d i and drugs j Based on similarity of drug-gene interactions The Jaccard index is calculated using the formula:

[0015]

[0016] Among them, GT i and GT j Respectively represent the drug d i and d j A collection of genes that interact with each other.

[0017] As some optional methods of the present application, the drug similarity of the gene expression difference after drug perturbation refers to:

[0018] Drugs i and d j Similarity based on differences in gene expression after drug perturbation Calculated by the cosine similarity formula, the formula is:

[0019]

[0020] in, and represents the gene expression feature vector after the perturbation of drug i or j, and Representation vector and The norm of .

[0021] As some optional methods of the present application, the semantic similarity of the different disease ontologies is obtained by the following steps:

[0022] For each adverse drug reaction, a directed acyclic graph was constructed based on its hierarchical descriptor in MeSH, where the node represents the node of the adverse reaction in MeSH and the edge represents the relationship between the current node and its ancestor; the adverse reaction s is represented as a graph DAG s =(s,N s ,E s ), where N s represents the set of all ancestor nodes including the adverse reaction s node itself, E sIt represents the set of all edges from parent nodes to child nodes. In order to encode the semantics of disease ontology terms in a measurable format, the study defines the semantic value SV(A) of disease A as DAG A The total contribution of all diseases in to the semantics of disease A, where disease terms close to disease A contribute more to its semantics; for two different disease ontologies A and B, the semantic similarity S(A,B) between A and B is defined as:

[0023]

[0024] As some optional methods of the present application, the gene-disease association set is obtained by the following steps:

[0025] The compound-gene relationship table CTD_genes_diseases.csv.gz was downloaded from CTD as a collection of gene-disease associations in the knowledge graph, containing 107,911,805 gene-disease association records; the gene-disease associations in CTD include expert-extracted and inferred associations; the extracted associations are curated and extracted by CTD curators from published literature, or exported from the OMIM database using the mim2gene file in the NCBI gene database; there are three types of evidence for direct gene-disease associations: M markers, mechanism, and therapeutic evidence; M markers refer to specific variations in gene sequences; mechanism evidence includes how gene variations affect the structure and function of proteins, and how these changes interfere with normal cells or physiological processes; therapeutic evidence refers to the improvement of disease symptoms or prevention of disease occurrence through treatment targeting specific genes or their products.

[0026] As some optional methods of the present application, the target drug adverse reaction intelligent prediction model is based on the original drug feature combination Drug, feature cross V FCs , Side combination of original features of adverse drug reactions is trained through three branches respectively;

[0027] Drug and Side use two linear sub-networks with the same structure for learning, while V FCs A convolutional neural network sub-network is used for learning; the architecture of the linear sub-network consists of two fully connected layers, a batch normalization layer, and an activation function layer; that is:

[0028] O m,k =Dropout (p) (Relu(FC n1 (CAT(x m ,c k )))),

[0029] x′ m,k =Linear(Om,k ),

[0030] Among them, x′ m,k In LSN d and LSN s where represent the potential representations of m drugs and m adverse drug reactions with k different similarity features, i.e., the output of LSN. d and LSN s Represents two linear subnetworks with the same structure, FC (*) represents a fully connected layer, where * represents the number of neurons, and Dropout represents a Dropout layer with probability p. Linear and Relu represent linear function and rectified linear unit activation function respectively, and CAT connects the given feature vector. Through this step, the linear embedding vector emb of drugs and adverse reactions can be obtained. drugs and emb adrs .

[0031] As some optional methods of the present application, the target drug adverse reaction intelligent prediction model is to drugs 、emb cross and emb ADRs After concatenation, the vectors are used as input vectors to be input into a multi-label classifier for classification; the classifier consists of two fully connected layers, two activation function layers, and a BN layer; and finally a vector is output, in which a number greater than 0 indicates the association between the drug and the ADR, and a larger number indicates a greater possibility of association; the output vector satisfies the following relationship:

[0032]

[0033] Among them, FC (*) represents a fully connected layer, where * represents the number of neurons, and Dropout represents a Dropout layer with probability p. Linear and Relu represent linear function and rectified linear unit activation function respectively, and CAT connects the given feature vector.

[0034] In a second aspect, the present application provides an intelligent prediction device for adverse drug reactions based on multi-source data fusion, comprising:

[0035] A data acquisition module, used to acquire multi-source data; wherein the multi-source data includes comparative toxicological genomics data, LINCS L1000 data, PubChem data, medical subject word list data and SIDER data;

[0036] Constructing a feature module, which is used to construct a drug descriptor feature after standardizing the multi-source data to obtain a drug similarity vector set; constructing a drug adverse reaction descriptor feature to obtain a drug adverse reaction similarity vector set;

[0037] The prediction module is used to train the target drug adverse reaction intelligent prediction model based on the drug similarity vector set and the drug adverse reaction similarity vector set. After the training is completed, the adverse reaction generated when the target drug is used for the target individual is predicted based on the target drug adverse reaction intelligent prediction model.

[0038] Compared with the prior art, the beneficial effects of this application are:

[0039] This application provides a new deep learning architecture, the Drug-Gene-ADR neutral network (DGANet), which can effectively integrate pharmacogenomics source data from multiple public databases, mainly including chemical-gene interactions (CGIs), gene-disease associations (GDAs), and changes in gene expression caused by drug disturbances. Based on convolutional neural networks (CNNs), the information associated with potential drug-gene-drug adverse reactions is expanded to form a new deep learning architecture with learning cross-features, DGANet, which aims to improve the accuracy and efficiency of adverse drug reactions (ADRs) prediction. In addition, based on existing genetic data, new genomic features are proposed to further enhance the model's ability to predict unknown ADRs, and verify the effectiveness of DGANet for new ADR discovery. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A summary diagram of data sources involved in the embodiments of the present application;

[0041] Figure 2 A data processing flow chart of an embodiment of the present application;

[0042] Figure 3 The overall structure diagram of DAGNet involved in the embodiment of the present application;

[0043] Figure 4 This is a structural diagram of the convolutional neural sub-network involved in an embodiment of the present application. DETAILED DESCRIPTION

[0044] It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0045] This application proposes a prediction model based on deep learning technology, which uses a combination of a convolutional neural network VGGNet (Visual Geometry Group Network) framework and a multilayer perceptron (Multilayer Perceptron, MLP) to extract features from multi-source pharmacogenomics data and predict adverse reactions that may be caused by drugs. This application names the entire model as a drug-gene-adverse drug reaction neural network (Drug-Gene-ADR neutralnetwork, DGANet). DGANet mainly integrates three representation features of drugs (drug similarity based on drug chemical structure, drug similarity based on drug-gene interaction, and drug similarity based on gene expression differences after drug perturbation) and representation features of adverse drug reactions. It learns the potential relationship between drugs, genes, and adverse drug reactions through a multi-layer convolutional neural network, predicts possible adverse reactions of drugs, and gives the probability of occurrence of various adverse reactions that may be caused by current drugs.

[0046] This application selects multiple public databases as data sources, the data sources are summarized as follows Figure 1 It includes the Comparative Toxicogenomics Database (CTD), LINCSL1000, PubChem, the US National Library of Medicine's Medical Subject Headings (MeSH), and SIDER (Side Effect Resource) to ensure the breadth and reliability of research source data.

[0047] Among the public data sources selected in this application, CTD is a selected database that focuses on describing cross-species chemical gene / protein interactions and the association between chemical and genetic diseases, aiming to reveal the impact of susceptibility and environmental factors on diseases. The database is manually organized and annotated by professional bioinformaticians to ensure the high quality and accuracy of the data. Its content covers more than 52,638,316 toxicogenomics relationships, involving more than 17,100 chemicals, 54,300 genes, 6,100 phenotypes, 7,270 diseases and 202,000 exposure events. CTD uses MeSH and OMIM (Online Mendelian Inheritance in Man) terms to construct a comprehensive disease vocabulary to better understand the potential mechanisms of the effects of specific drugs on the human body.

[0048] LINCS L1000 is one of the Library of Integrated Network-Based Cellular Signatures (LINCS) projects supported by the National Institutes of Health (NIH). It aims to collect and analyze the response data of human cells to various biological and chemical perturbations through high-throughput technology to promote the understanding of biological functions. The database contains approximately 1,000,000 expression profiles, mainly by treating 99 cell lines with 32,855 small molecules and measuring the expression profiles of 978 landmark genes on 384-well plates, while using deep learning methods to predict the expression of the remaining 12,328 genes. The data preprocessing of LINCS L1000 is divided into five levels, the first level (Luminex Bead File, LXB): raw, unprocessed flow cytometry data from the Luminex scanner. An LXB file is generated for each well of the 384-well plate, and each file contains the fluorescence intensity value of each observed analyte in the well. Level 2 (Gene Expression, GEX): Gene expression values ​​for every 1000 genes after deconvolution from Luminex beads. Level 3 (Quantile-2NormalizedData, Q2NORM): Gene expression profiles of directly measured landmark transcripts and inferred genes. Normalization is performed using invariant set scaling and quantile normalization. Level 4 (Z-Score Calculations, Z-SCORES): Features of differentially expressed genes are calculated by robust Z-scores of each profile relative to the control (PC (plate control) as the blank control population; VC (vehicle control) as the vehicle control). Level 5 (Signature, SIG) consists of the results of repeated experimental groups, usually treated 3 times per group, ultimately forming a single differential expression vector derived from the weighted average of a single repeated experiment, providing a valuable resource for studying cellular responses.

[0049] PubChem is an open chemical database maintained by the National Center for Biotechnology Information (NCBI) of the NIH. It contains information on millions of chemical substances from a wide range of sources, not only small molecule drugs, but also larger molecules such as nucleotides and carbohydrates, providing information on chemical structure, biological activity, patents, health, safety and toxicity data, etc. PubChem is easy to associate with multiple public databases such as Drugbank, STITCH, SIDER, etc., which enhances the availability and practicality of data.

[0050] MeSH is an authoritative and standardized medical terminology vocabulary compiled by the National Library of Medicine (NLM) of the United States for indexing and retrieving biomedical literature. MeSH organizes terms in a tree structure, including terms and qualifiers, organized hierarchically, covering 16 categories of biomedical terms, especially disease terms, organized hierarchically for easy understanding and accurate retrieval, and is a key tool for medical information retrieval.

[0051] SIDER is a database of marketed drugs and their adverse reactions, also maintained by NCBI of NIH, to support the study of drug side effects and the evaluation of drug safety. SIDER integrates data from multiple sources, including U.S. Food and Drug Administration (FDA) label information, drug instructions, clinical trial results, and scientific literature, and provides information such as drug name, active ingredient, dosage form, route of administration, and description, frequency, and severity of side effects, which is of great value for drug development, clinical practice, and drug regulation. The SIDER 4.1 version used in this application contains 1,430 drugs, 5,880 side effects, and 139,756 drug-side effect pairs, providing valuable resources for the identification, evaluation, and management of drug side effects. By integrating information from these databases, this application study was able to comprehensively explore the drug mechanism of action, side effects, and their complex relationship with disease.

[0052] like Figure 2 The data preprocessing process shown in the figure, in the data preprocessing link, this application constructs drug descriptors and adverse drug reaction descriptors, the specific details are as follows:

[0053] Constructing drug descriptors: By using the drug Compound ID provided in SIDER, the SMILES (Simplified Molecular Input Line Entry System) string representation of each drug structure can be downloaded from PubChem in batches, and then converted into a topological fingerprint using the Rdkit tool in Python. The topological fingerprint is generated based on the topological structure and rotation angle of the four-membered ring in the molecule, which can effectively describe the stereochemistry and interaction of the molecule. In order to quantify the chemical structure similarity between two drugs, this application uses the Tanimoto coefficient as a metric, and defines the chemical structure similarity of two drugs as: It represents the similarity between drugs in the drug chemical structure space, where cs(i,j) represents the chemical space structure of two different drugs.

[0054] Secondly, in order to obtain drug similarity based on drug-gene interactions, this application downloaded the compound-gene relationship table CTD_chem_gene_ixns.csv.gz (version: updated on February 28, 2024) from CTD. The data set contains 2,676,084 compound-gene interaction records. CTD extracts specific compound-gene or protein interactions from published literature. Most of these interactions are binary, but also include more complex nested events. Each chemical-gene interaction has a certain degree of limitation, such as increase, decrease, influence or no influence. Drug-gene interactions are represented by One-Hot encoding, that is, each unique gene interaction is converted into a binary vector. Drug d i and d j Based on similarity of drug-gene interactions The Jaccard index is calculated using the formula:

[0055]

[0056] Among them, GT i and GT j Respectively represent the drug d i and d j A collection of genes that interact with each other.

[0057] Finally, drug similarity based on differences in gene expression after drug perturbation was obtained from the LINCS L1000 project. This project analyzed the changes in gene expression after cells were treated with each small molecule compound at different doses and time points. In the GSE92742 dataset, the "distil_ss" variable quantified the magnitude of differential expression of landmark genes, representing the intensity of gene expression for each experiment. This application study ignored the differences in cell type, dose, or time point, and selected the "distil_ss" value with the largest gene expression difference as the characteristic representation of each drug. The characteristics of a drug are defined as a vector of continuous values, each value representing the direction and magnitude of differential expression between the control sample and the compound-treated sample. The differential characteristics of gene expression after drug perturbation were calculated using the Characteristic Direction (CD) method. Drug d i and d j Similarity based on differences in gene expression after drug perturbation Calculated by the cosine similarity formula, the formula is:

[0058]

[0059] By integrating data from three aspects, namely, drug chemical structure, drug-gene interaction, and gene expression differences after drug perturbation, a multi-dimensional drug descriptor feature was constructed, providing a comprehensive and in-depth perspective for drug similarity analysis. This not only helps to understand the mechanism of action of drugs, but also provides new ideas and methods for drug repositioning and new drug development.

[0060] Constructing new adverse drug reaction descriptors: In order to construct more comprehensive and efficient adverse drug reaction descriptors, this application study adopted two methods: semantic similarity of adverse drug reactions based on MeSH hierarchical descriptors and gene-disease association. First, for each adverse drug reaction, the study constructed a directed acyclic graph (DAG) based on its hierarchical descriptor in MeSH, where the node represents the node of the adverse reaction in MeSH (that is, its medical descriptor or medical term), and the edge represents the relationship between the current node and its ancestor. Specifically, the adverse reaction s can be represented as a graph DAG s =(s,N s ,E s ), where N s represents the set of all ancestor nodes including the adverse reaction s node itself, E s Then it represents the set of all edges from parent nodes to child nodes. In order to encode the semantics of disease ontology terms in a measurable format, the study defines the semantic value SV(A) of disease A as DAG AThe total contribution of all diseases in to the semantics of disease A, where disease terms close to disease A contribute more to its semantics. For two different disease ontologies A and B, the semantic similarity S(A,B) between A and B is defined as:

[0061]

[0062] This formula not only considers the position of the disease ontology in the DAG, but also reflects its semantic relationship with its ancestor words, making the calculation of semantic similarity closer to clinical practical significance.

[0063] In addition, the compound-gene relationship table CTD_genes_diseases.csv.gz (version: updated on February 28, 2024) was downloaded from CTD as a collection of gene-disease associations in the knowledge graph, containing 107,911,805 gene-disease association records. The gene-disease associations in CTD include expert-extracted and inferred associations. The extracted associations are curated and extracted by CTD curators from published literature, or exported from the OMIM database using the mim2gene file in the NCBI gene database. There are three types of evidence for direct gene-disease associations: M markers (genetic markers directly associated with the disease), mechanisms (involving the understanding of how gene variants lead to the biological basis of the disease), and treatments (the connection between gene variants and diseases can be verified through therapeutic interventions). M markers usually refer to specific variations in gene sequences, such as single nucleotide polymorphisms (SNPs), insertions / deletions (indels), or copy number variations (CNVs), which are significantly associated with disease risk or phenotype. Mechanistic evidence includes how gene variations affect the structure and function of proteins and how these changes interfere with normal cells or physiological processes, leading to disease. Therapeutic evidence refers to the fact that treatments targeting specific genes or their products can improve disease symptoms or prevent the occurrence of diseases, providing strong evidence for the direct link between genes and diseases.

[0064] In addition to direct associations, CTD also includes partially inferred associations that are established through chemical-gene interactions curated by CTD. Inference is performed based on an inference score of the network topology consisting of genes, diseases, and one or more chemicals used for inference. The inference score reflects the degree of similarity between the CTD chemical-gene-disease network and a similar scale-free random network and is calculated as the log-transformed product of two commonly used neighbor statistics for assessing functional relationships between proteins in protein-protein interaction networks. The inference score takes into account the connectivity of genes and diseases and the number of chemicals used for inference, as well as the connectivity of each chemical. The higher the inference score, the more likely the network is considered to have atypical connectivity. Many biological networks, such as disease and metabolic networks, have been shown to be scale-free random networks.

[0065] The overall network structure of the model is as follows Figure 3 As shown, the input of the model includes a drug similarity vector set and a drug adverse reaction similarity vector set. The drug similarity set Drug is represented as in Indicates the similarity between drugs in the drug chemical structure space, represents the similarity of drug-drug gene variation intensity space after drug perturbation, Represents the similarity of drugs in the drug-gene interaction space. The drug adverse reaction similarity set Side is expressed as in Indicates the similarity of adverse reactions-adverse reactions in the disease tree structure space of MESH, Represents the similarity of adverse reaction-adverse reaction in the gene-adverse reaction association space.

[0066] In order to capture the nonlinear relationship between different features of drugs and adverse reactions, the combined features of drugs and adverse drug reactions, namely feature crossover, are constructed. The combined features of drugs and adverse drug reactions are defined as V FCs , which can be expressed as in Represents a tensor product.

[0067] Drug original feature combination Drug, feature cross V FCs , Side, and the drug adverse reaction original feature combination are trained through three branches. Specifically, Drug and Side use two linear subnetworks (LinearSubnetworks, LSN) with the same structure. d and LSN s ) to learn, and V FCsA Convolutional Neural Networks Subnetwork (CSN) is used for learning. The architecture of the linear subnetwork consists of two fully connected layers, a batch normalization (BN) layer, and an activation function layer. The specific formula is as follows:

[0068] O m,k =Dropout (p) (Relu(FC n1 (CAT(x m ,c k )))),

[0069] x′ m,k =Linear(O m,k ),

[0070] Among them, x′ m,k In LSN d and LSN s where FC represents the potential representation of m drugs and m adverse drug reactions with k different similarity features, i.e., the output of LSN. (*) represents a fully connected layer, where * represents the number of neurons, and Dropout represents a Dropout layer with probability p. Linear and Relu represent linear function and rectified linear unit activation function respectively, and CAT connects the given feature vector. Through this step, the linear embedding vector emb of drugs and adverse reactions can be obtained. drugs and emb adrs .

[0071] The convolutional neural network subnetwork CSN refers to the VGGNet approach. Each convolution block consists of Conv2d (convolution) + BatchNorm2d (batch normalization) + ReLU (activation layer). Considering the experimental resources and time issues, the number of layers and parameters were simplified, and 6 convolution layers were constructed. The layers were connected by a normalization layer, and the hidden layers were connected by the ReLU activation function. A 2×2 convolution kernel was used, the step size was set to 2, and the number of channels was set to 32. After each sampling, the height and width of the feature matrix were reduced to half of the original, and the final output was named emb. cross , the CSN structure is as follows Figure 4 shown.

[0072] The emb obtained from the above steps drugs 、emb cross and emb ADRsAfter concatenation, the vectors are used as input vectors to be input into the multi-label classifier for classification. The classifier consists of two fully connected layers, two activation function layers, and a BN layer. Finally, a vector is output, in which numbers greater than 0 represent the association between the drug and the ADR, and the larger the number, the greater the possibility of the association. The output is expressed as the formula:

[0073]

[0074] Finally, the loss function determines the ability of the model to handle tasks in deep learning. Multi-label classification tasks often face problems such as class imbalance, label sparsity, and implicit class correlation. In order to solve these problems, this application adopts ZLPR (zero-bounded log-sum-exp & pairwise rank-based) proposed by Su et al. as the loss function of DGANet to measure the error between the predicted value and the true value. ZLPR generalizes the "softmax + cross entropy" scheme to multi-label classification scenarios without the need to specifically adjust class weights and thresholds. Compared with the loss based on binary relevance (BR), ZLPR can better capture the label correlation and the ranking relationship between positive and negative categories. Compared with the loss based on label ranking (LR), ZLPR can adaptively determine the number of target categories and enhance the label ranking ability of the model. ZLPR combines the advantages of LR and BR, achieves efficient spatiotemporal performance, retains the category correlation information in the original data, and is suitable for situations where the number of categories is uncertain. The formula of the ZLPR loss function is as follows. In the experiment, the learning rate is set to 0.005:

[0075]

[0076] Among them, s i and j Indicates that these two variables represent the output score of the model for the i or j category, Ω pos and Ω pos Represents the positive label set and the negative label set.

[0077] By adopting the ZLPR loss function, multi-label classification tasks can be handled more effectively and the performance and robustness of the model can be improved.

[0078] Through the above steps, a multimodal feature fusion model was constructed, which can effectively capture the complex relationship between drugs and adverse reactions, and provide a new method for the prediction and management of adverse drug reactions.

[0079] This application study systematically compared the effects of different pharmacogenomic cross-feature combinations in assisting the prediction of adverse reactions, and comprehensively evaluated the performance of the proposed new model. Experimental results show that the model proposed in this application study has the best predictive ability under all test conditions when using drug-gene interaction (CGI) as a feature. It is particularly noteworthy that when combined with the use of drug chemical structure (CS) information, the model achieved the best performance of AUROC (Area Under the Receiver Operating Characteristic Curve)) of 92.30% and AUPRC (Area Under the Precision-Recall Curve) of 91.80%, proving the optimality of this innovative feature combination.

[0080] In addition, this application innovatively uses the MESH disease concept to measure the semantic similarity between adverse reactions and combines it with the gene-disease association (GDA). This method is better than relying solely on GDA. This improvement not only improves the model's ability to understand complex biomedical data, but also provides a more accurate basis for subsequent analysis. More importantly, the neighborhood similarity embedded in known drug-adverse reaction associations further improves the overall performance of the model. Specifically, this strategy increased the AUROC by 0.38%-0.77% and the AUPRC by 0.62%-0.85%. These improvements demonstrate the importance of considering existing knowledge for improving prediction accuracy.

[0081] Finally, a class-balanced sampling technique was implemented within the DGANet framework to construct a balanced dataset for training and validation. This method effectively alleviates the negative impact of sample inequality and ensures consistency and reliability between evaluation indicators such as AUROC and AUPRC. Even when working on a relatively small dataset, and regardless of whether additional neighborhood similarity information based on Drug Side-effect Association (DSA) is introduced, DGANet can still achieve an increase of 4.6% and 5.06% in the AUROC indicator, respectively, showing strong generalization ability and potential application value.

[0082] Therefore, through carefully designed feature engineering and algorithm optimization, this application has successfully developed a new model that can efficiently and accurately predict adverse drug reactions, laying a solid foundation for promoting the development of personalized medicine.

[0083] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A multi-source data fusion method for intelligent prediction of adverse drug reactions, characterized in that: The following steps are involved: Acquire multi-source data; wherein the multi-source data includes comparative toxicological genomics data, LINCS L1000 data, PubChem data, medical subject thesaurus data and SIDER data; the comparative toxicological genomics data is used to interpret drug-gene interactions, the LINCS L1000 data is used to interpret gene expression signals, the PubChem data is used to interpret drug SMILES, the medical subject thesaurus data is used to interpret gene-disease interactions, and the SIDER data is used to interpret drug-side effect pairs; After standardizing the multi-source data, constructing drug descriptor features to obtain a drug similarity vector set; constructing adverse drug reaction descriptor features to obtain a drug adverse reaction similarity vector set; The target drug adverse reaction intelligent prediction model is trained based on the drug similarity vector set and the drug adverse reaction similarity vector set. After the training is completed, the adverse reaction generated when the target drug is used for the target individual is predicted based on the target drug adverse reaction intelligent prediction model.

2. The intelligent prediction method for adverse drug reactions based on multi-source data fusion according to claim 1, characterized in that: The drug similarity vector set includes: chemical structure similarity of two drugs, drug similarity based on drug-gene interaction and drug similarity based on gene expression difference after drug perturbation; The adverse drug reaction similarity vector set includes: semantic similarities of different disease ontologies and a gene-disease association relationship set.

3. The intelligent prediction method for adverse drug reactions based on multi-source data fusion according to claim 2 is characterized in that: The chemical structure similarity values ​​of the two drugs are obtained based on the following steps: By using the drug Compound ID provided in SIDER, the SMILES string representation of each drug structure was downloaded from PubChem in batches, and then converted into a topological fingerprint using the Rdkit tool in Python; the topological fingerprint is generated based on the topological structure and rotation angle of the four-membered ring in the molecule to describe the stereochemistry and interaction of the molecule; and the Tanimoto coefficient is used as a metric to obtain the drug d i and drugs j Chemical structure similarity values ​​of two drugs 4. The intelligent prediction method for adverse drug reactions based on multi-source data fusion according to claim 2 is characterized in that: The drug similarity based on drug-gene interaction refers to: Drug-gene interactions are represented by One-Hot encoding, which converts each unique gene interaction into a binary vector; drug d i and drugs j Based on similarity of drug-gene interactions The Jaccard index is calculated using the formula: Among them, GT i and GT j Respectively represent the drug d i and d j A collection of genes that interact with each other.

5. The intelligent prediction method for adverse drug reactions based on multi-source data fusion according to claim 2 is characterized in that: The drug similarity of the gene expression difference after drug perturbation refers to: Drugs i and d j Similarity based on differences in gene expression after drug perturbation Calculated by the cosine similarity formula, the formula is: in, and represents the gene expression feature vector after drug i or j is disturbed, and Representation vector and The norm of .

6. The intelligent prediction method for adverse drug reactions based on multi-source data fusion according to claim 2, characterized in that: The semantic similarity of the different disease ontologies is obtained by the following steps: For each adverse drug reaction, a directed acyclic graph was constructed based on its hierarchical descriptor in MeSH, where the node represents the node of the adverse reaction in MeSH and the edge represents the relationship between the current node and its ancestor; the adverse reaction s is represented as a graph DAG s =(s,N s , E s ), where N s represents the set of all ancestor nodes including the adverse reaction s node itself, E s It represents the set of all edges from parent nodes to child nodes. In order to encode the semantics of disease ontology terms in a measurable format, the study defines the semantic value SV(A) of disease A as DAG A The total contribution of all diseases in to the semantics of disease A, where the disease terms close to disease A contribute more to its semantics; for two different disease ontologies A and B, the semantic similarity S(A,B) between A and B is defined as:

7. The intelligent prediction method for adverse drug reactions based on multi-source data fusion according to claim 2, characterized in that: The gene-disease association set is obtained by the following steps: The compound-gene relationship table CTD_genes_diseases.csv.gz was downloaded from CTD as a collection of gene-disease associations in the knowledge graph, containing 107,911,805 gene-disease association records; gene-disease associations in CTD include expert-extracted and inferred associations; extracted associations are curated and extracted by CTD curators from published literature, or exported from the OMIM database using the mim2gene file in the NCBI gene database; there are three types of evidence for direct gene-disease associations: M markers, mechanism, and therapeutic evidence; M markers refer to specific variations in gene sequences; mechanism evidence includes how gene variations affect the structure and function of proteins, and how these changes interfere with normal cells or physiological processes; therapeutic evidence refers to the improvement of disease symptoms or prevention of disease occurrence through treatment targeting specific genes or their products.

8. The intelligent prediction method for adverse drug reactions based on multi-source data fusion according to claim 2, characterized in that: The target drug adverse reaction intelligent prediction model is based on the original drug feature combination Drug, feature cross V FCs , Side combination of original features of adverse drug reactions is trained through three branches respectively: Drug and Side use two linear sub-networks with the same structure for learning, while V FCs A convolutional neural network sub-network is used for learning; the architecture of the linear sub-network consists of two fully connected layers, a batch normalization layer, and an activation function layer; that is: About m,k =Dropout (p) (Relu(FC n1 (CAT(x m ,c k )))), x′ m，k =Linear(O m,k ), Among them, x′ m，k In LSN d and LSN s where represent the potential representations of m drugs and m adverse drug reactions with k different similarity features, i.e., the output of LSN; where LSN d and LSN s Represents two linear sub-networks with the same structure; FC(*) represents a fully connected layer, where * represents the number of neurons; Dropout represents a Dropout layer with probability p; Linear and Relu represent linear function and rectified linear unit activation function respectively, and CAT connects the given feature vector.

9. The intelligent prediction method for adverse drug reactions based on multi-source data fusion according to claim 2, characterized in that: The target drug adverse reaction intelligent prediction model is based on emb drugs 、emb cross and emb ADRs After concatenation, the vectors are used as input vectors to be input into a multi-label classifier for classification; the classifier consists of two fully connected layers, two activation function layers, and a BN layer; and finally a vector is output, in which a number greater than 0 indicates the association between the drug and the ADR, and a larger number indicates a greater possibility of association; the output vector satisfies the following relationship: Among them, FC(*) represents a fully connected layer, where * represents the number of neurons, Dropout represents a Dropout layer with probability p; Linear and Relu represent linear function and rectified linear unit activation function respectively, and CAT connects the given feature vector.

10. A multi-source data fusion intelligent prediction device for adverse drug reactions, characterized in that: include: A data acquisition module is used to acquire multi-source data; wherein the multi-source data includes comparative toxicological genomics data, LINCS L1000 data, PubChem data, medical subject word list data and SIDER data: Constructing a feature module, which is used to construct a drug descriptor feature after standardizing the multi-source data to obtain a drug similarity vector set; constructing a drug adverse reaction descriptor feature to obtain a drug adverse reaction similarity vector set; The prediction module is used to train the target drug adverse reaction intelligent prediction model based on the drug similarity vector set and the drug adverse reaction similarity vector set. After the training is completed, the adverse reaction generated when the target drug is used for the target individual is predicted based on the target drug adverse reaction intelligent prediction model.

Images