Adverse drug reaction prediction method based on comment text information enhancement

Abstract

The invention relates to the technical field of bioinformatics, and discloses an adverse drug reaction prediction method based on comment text information enhancement, and the method comprises the following steps: constructing an adverse drug reaction dictionary; extracting adverse drug reaction information based on the comment text; integration of adverse drug reaction information; integrating the drug characteristic information of the public database; calculating the similarity; constructing a heterogeneous graph; building a graph convolutional network model; model training and evaluation; predicting adverse reaction cases of drugs; and analyzing the occurrence mechanism of adverse drug reaction. According to the adverse drug reaction prediction method based on comment text information enhancement, diversified data such as drug comments, clinical test data, actual medication experience and feedback are comprehensively utilized, patient features are added to enrich the relationship between drugs and adverse reactions, more comprehensive drug safety assessment is provided, and the prediction accuracy is improved. In addition, compared with other models, the GCN model has the advantages that the training and prediction are quicker, and the calculation difficulty can be reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of bioinformatics, and in particular to a method for predicting adverse drug reactions based on enhanced text information of patient comments. Background Art

[0002] At present, adverse drug reactions (ADRs) have become a hot topic of concern in the medical community and the public, and the issue of drug safety has received increasing attention from the whole society. Due to the limitations of the number of test populations and test cycles during drug development, new drugs with potential adverse drug reactions may enter the market. Therefore, how to judge and predict adverse drug reactions has great theoretical and practical value.

[0003] (1) Data sources for adverse drug reaction research

[0004] The current adverse drug reaction research based on data analysis mainly comes from three types of sources: drug knowledge base, social network comment text and monitoring database.

[0005] Research based on drug knowledge base mainly relies on knowledge related to adverse drug reactions, drug chemical structure information, drug-drug interaction information, drug target information, protein-protein interaction information, drug side effect information, drug-disease association information, gene expression spectrum, etc., covering drug-related knowledge such as chemistry, phenotype, genome and cell network. Nowadays, databases such as Pubchem, DrugBank, CTD, TTD, SIDER, etc. have integrated a large amount of drug-related structured data, providing a rich data foundation for discovering adverse drug reactions. For example, the drug side effect frequency data set in SIDER and ADReCS was used to train the drug-side effect collaborative propagation model, and good results were achieved.

[0006] Research based on social network comment text uses a large amount of unconstrained comments, blogs, microblogs and other social media data publicly available on the Internet to identify comment information related to drug use, such as reasons for medication, methods of medication, descriptions of adverse reactions, etc., and extracts, classifies and analyzes these comment information based on natural language processing technology and machine learning algorithms to conduct knowledge discovery of adverse drug reactions. For example, Zhao Mingzhen et al. used the principle of information entropy to identify disease and adverse drug reaction entities, and proposed a method for calculating the correlation of biological test questions based on the Skip-gram model to seek associated proteins of potential adverse drug reactions.

[0007] Among the studies based on monitoring databases, FAERS is often used as a data source for drug adverse reaction signal mining because of its rich data, openness and transparency, and unified terminology and format. For example, Yang Li, Xia Runan, and others obtained the adverse reaction signals of ixazomib and axitinib by mining the FAERS database.

[0008] (2) Methods for predicting adverse drug reactions

[0009] Using computational methods such as data mining and machine learning to analyze massive amounts of biomedical data and achieve accurate predictions of adverse drug reactions has become a popular research method in this field. However, data mining methods are often used to discover adverse reactions that have already occurred, and cannot predict potential adverse reactions. Machine learning methods predict adverse drug reactions and perform well on specific data sets. For example, Xue De Du Yao used the changes in gene expression before and after illness and before and after medication as explanatory variables, and used the random forest algorithm to predict drug side effects, which was significantly improved compared to the traditional K nearest neighbor algorithm. However, the performance of traditional machine learning models depends largely on the quality and quantity of training data, and their generalization ability may be insufficient.

[0010] To this end, some scholars have used the method of joint deep learning and drug protein link prediction to construct a multi-message transmission neural network to predict adverse drug reactions. Some scholars have also proposed a collaborative propagation model of a high-order similarity network for the frequency of adverse drug reactions. Wu Juhua et al. developed a knowledge graph embedded convolutional neural network ADR prediction model, which predicts ADR by integrating drug characteristic information such as drug targets, indications, enzymes and carrier proteins. This model predicts all types of ADRs in a unified manner, reducing the redundant workload of developing prediction models for each ADR separately in previous studies, and improving prediction efficiency and accuracy. The above methods all use data from public databases for prediction, but ignore the newly emerging adverse drug reactions in actual medication and the differences between patients in actual medication. These data can enrich the information during prediction, improve the accuracy of adverse drug reaction prediction, and promptly discover new risks, which can help provide patients with safer and more personalized treatment plans in clinical practice.

[0011] (3) Analysis of adverse drug reaction prediction results

[0012] Previous drug adverse reaction prediction and verification were limited to using the drug adverse reaction associations in the drug adverse reaction database to verify the rationality of the prediction results, or conducting case analysis on the constructed model based on literature reports, indicating that the model had good prediction performance, but no follow-up research was conducted on the predicted adverse drug reaction mechanism. This study constructed a drug target-adverse reaction target PPI network and enrichment analysis methods to explain and verify the prediction results at the biological level, improve the entire prediction and verification process of existing research, and make the prediction results more scientific and reliable.

[0013] In summary, there are studies on adverse drug reactions based on various relevant databases, but the prediction of potential adverse drug reactions usually uses drug compound information, ignoring individual differences in the medication process. Moreover, most existing studies identify adverse drug reactions that have been manifested and reported, and do not add the identified information to the prediction work. They only achieve the discovery of known adverse drug reactions, which may ignore actual medication experience and rare adverse reaction information. Most technologies only predict the adverse drug reactions that may be caused by drugs, and only by understanding the causes of adverse drug reactions can we avoid the occurrence of adverse drug reactions. This paper integrates multiple data sources to add more information on the association between drugs and adverse reactions, achieve more accurate prediction effects, predict the possibility of adverse drug reactions that have not appeared, and provide a reference for the clinical application of drugs. In terms of the verification of adverse drug reaction prediction results, this paper focuses on the interaction between drug targets and the adverse reaction mechanisms caused by these interactions, that is, how drugs affect the expression of drug target genes, thereby affecting their interacting proteins, leading to changes in related pathways, and ultimately causing adverse drug reactions. Summary of the invention

[0014] The purpose of the present invention is to provide a method for predicting adverse drug reactions based on enhanced comment text information, which comprehensively utilizes drug review websites, clinical trial data, and drug data sources, and comprehensively utilizes diverse data such as user actual experience and feedback, adds patient characteristics, and enriches the relationship between drugs and adverse reactions, providing a more comprehensive drug safety assessment. In addition, compared with other models, the GCN model is faster in training and prediction, which can reduce the difficulty of calculation.

[0015] To achieve the above object, the present invention provides a method for predicting adverse drug reactions based on review text information enhancement, comprising the following steps:

[0016] Step 1, construction of adverse drug reaction dictionary:

[0017] Galeano's dataset was used as the benchmark dataset, the preferred terms of adverse drug reactions and their synonyms in MedDRA ontology were used as the vocabulary source, and the adverse reaction terms were encoded by direct coding to construct the adverse drug reaction dictionary.

[0018] Step 2: Extraction of adverse drug reaction information based on review text:

[0019] Collect adverse drug reaction review text data from drug review websites using a python crawler;

[0020] Use Java program and the constructed adverse drug reaction dictionary to implement word segmentation, stop word removal and entity extraction for the adverse drug reaction information data based on the review text, and obtain the adverse drug reaction association based on the review text;

[0021] Step 3, Integration of adverse drug reaction information:

[0022] Screen the drugs in the benchmark data set in FAERS, remove invalid or incomplete records, and obtain the adverse reaction data of these drugs in the report form; integrate different adverse reactions of the same drug to obtain drug-adverse reaction association information based on the FAERS database;

[0023] The extracted adverse drug reaction associations based on review text, adverse drug reaction associations based on the FAERS database, and related medication information are merged with the data of the benchmark dataset to obtain a drug-adverse reaction association matrix that integrates the review and FAERS database information;

[0024] Step 4: Integration of drug feature information from public databases:

[0025] The SMILES chemical formula, target and other information of the drugs in the dataset were obtained from DrugBank, the indication information of the drugs was obtained from the TTD database, and the hierarchical structure information of the adverse reactions in the dataset (including the upper and lower hierarchical relationships and structure numbers of adverse reactions) was obtained from the ADReCS database, which served as the basis for the subsequent calculation of the similarity of drugs and adverse reactions. Finally, five associations of drug-molecular structure, drug-target, drug-indication, adverse reaction hierarchical relationship, drug-adverse reaction and patient characteristic information were obtained.

[0026] Step 5, calculation of similarity:

[0027] ① Comprehensive drug similarity:

[0028] First, the SMILES string is converted into a molecular fingerprint. For the fingerprints of two drug molecules, their intersection and union are calculated. The similarity of the drug-molecule structure is calculated using the Tanimoto coefficient to obtain the drug molecule structure similarity matrix.

[0029] Then, the intersection and union of the targets or treatment indications of the drugs are calculated respectively, and the drug-target similarity and drug-indication similarity are calculated respectively according to the Jaccard coefficient to obtain the drug-target and indication similarity matrices;

[0030] The characteristics of the patients taking the medicine are also used as one of the bases for calculating the similarity of the medicine. The missing values ​​and outliers of data such as gender, age, disease condition, and duration of medication are processed. Gender is used as a numerical variable, and the duration of medication and age are converted into categorical variables by partitioning. The disease condition (disease type) is classified according to the disease type. The above feature types are encoded using one-hot encoding, and Word2Vec is used to construct the features of the disease condition. All the above features (including one-hot encoded features and features generated by Word2Vec) are spliced ​​together to form a feature vector of a patient. The features of multiple patients form a feature matrix. The feature matrix is ​​converted into a one-dimensional vector through dimensionality reduction technology, representing the patient characteristics of a drug. The similarity of drugs based on patient characteristics is calculated based on cosine similarity to obtain a similarity matrix of drugs based on patient characteristics.

[0031] Finally, a comprehensive drug similarity matrix is ​​obtained through the weighted similarity fusion algorithm;

[0032] ②Similarity of adverse reactions:

[0033] The similarity between adverse reactions is calculated based on the directed acyclic graph DAG. For each adverse reaction, a directed acyclic graph DAG based on the adverse reaction hierarchy is constructed, which contains semantic descriptors related to the adverse reaction, A. Its DAG can be expressed as DAG(A) = {TA, EA}, where TA is the set of ancestor nodes of A including itself, and EA is the set of all edges in the DAG;

[0034] The contribution of descriptor t to “A” in DAG(A) is calculated by the formula. Then, the contributions of all descriptors in DAG(A) are summed to obtain the semantic value of “A”. The adverse reaction s is calculated based on the semantic value. i and j The semantic similarity between them is as follows to obtain the adverse reaction similarity matrix;

[0035] Step 6, construction of heterogeneous graph:

[0036] Based on the comprehensive drug similarity matrix, adverse reaction similarity matrix and drug-adverse reaction association matrix, a heterogeneous graph G is constructed;

[0037] Step 7: Building the graph convolutional network model:

[0038] The constructed heterogeneous graph is introduced into the GCN encoder to extract drug and disease embeddings respectively, and the association in the test dataset is set to 0. The encoder can be expressed as:

[0039]

[0040] Among them, L en is the total number of coding layers, is the degree matrix, is the self-loop adjacency matrix of matrix A after adding self-connection, H l is the node feature matrix of the lth layer, H (l+1) is the node feature matrix of the l+1th layer, (W en ) l is the learning weight matrix of layer l, using ReLU as the activation function;

[0041] The obtained node representation is decoded using the inner product decoder to obtain the final embedding of drugs and adverse reactions:

[0042]

[0043] Where E d ∈R m×d For the final embedding of the drug, E a ∈R n×d For the final embedding of adverse reactions, a l is a weight initialized to 1 / (l+1);

[0044] Introduce sigmoid as the activation function into the GCN decoder, and use the inner product decoder to get the predicted correlation matrix:

[0045] A′=sigmoid(E d ·W.E. a )

[0046] Where A′ ij It is a drug i and adverse reactionsa j The associated prediction score, W∈R d×d is a trainable parameter matrix;

[0047] The model optimization part uses weighted cross entropy loss to measure the loss between the original matrix and the reconstructed association matrix, and uses the Adam algorithm to optimize the loss;

[0048] Step 8, model training and evaluation:

[0049] Ten-fold cross validation was used to train and evaluate model performance. The performance of the graph convolutional network (GCN) model was evaluated using indicators such as accuracy, F1 score, precision, recall, ROC curve, and AUC value.

[0050] Ten-fold cross validation randomly divides the entire data set into ten subsets, uses one of the datasets as the test set, and combines the remaining nine datasets as the training set. The training set is used to train the model, and the test set is used to evaluate the performance of the model. The above steps are repeated until each subset is used as a test set once. In all ten folds, the average of the model performance indicators is calculated and used as the final result.

[0051] Step 9, prediction of adverse drug reactions:

[0052] Taking drugs for treating epilepsy as an example, the adverse reactions that may be caused by the drugs are predicted in the trained model, and the drug-adverse reaction association pairs with a prediction value greater than 0.7 are selected as the final prediction results, indicating the risk of adverse reactions that may be caused by the medication;

[0053] Step 10, analysis of the mechanism of adverse drug reactions:

[0054] For the top ten drug-adverse reaction associations in the case analysis, drug target information was obtained from DrugBank, drug target-related genes were obtained from STRING, and adverse reaction-related genes were obtained from GeneCards. The intersection of drug target-target-related genes-adverse reaction-related genes was processed to obtain intersection genes. Cytoscape was used to construct a PPI interaction network diagram, GO and KEGG analysis were performed on the intersection genes, and the biological processes and potential influencing pathways in the target genes were further analyzed; the mechanism of drug-induced adverse reactions was revealed, and the rationality of the prediction results was verified.

[0055] Furthermore, in step 2, all drug review texts on the drug review website are collected, and the collected information includes the review text and the reviewer's gender, age, medical condition, duration of medication, etc.

[0056] Furthermore, in step 5, the calculation formula is as follows:

[0057] Tanimoto coefficient:

[0058]

[0059] Where A and B are the molecular fingerprints of drugs a and b respectively;

[0060] Jaccard coefficient:

[0061]

[0062] A and B are the target (indication) sets of drugs a and b, respectively;

[0063] Cosine Similarity:

[0064]

[0065] Where A and B are the patient feature vectors of drugs a and b respectively.

[0066] Furthermore, in step 5, the calculation formula is as follows:

[0067] Directed Acyclic Graph DAG:

[0068]

[0069] where θ is the semantic contribution factor of the edge connecting node t with its child node t.

[0070] Furthermore, the heterogeneous graph G in step 6 is expressed as:

[0071]

[0072] Where S d , S a are the drug comprehensive similarity matrix and adverse reaction similarity matrix, respectively. d ,~S a are the matrices of drug and adverse reaction similarity matrices after Laplace transformation, and A is the association matrix between drugs and adverse reactions.

[0073] Furthermore, the number of encoder layers in step 7 is 2. In order to make full use of the embedding of drug and adverse reaction nodes, the importance weights of different GCN layers are adaptively and dynamically adjusted, and the attention mechanism is added in the process of node feature representation.

[0074] Furthermore, the calculation formula of the evaluation index in step 8 is as follows:

[0075] Use accuracy:

[0076]

[0077] Among them, TP is the number of samples correctly predicted as positive by the model; TN is the number of samples correctly predicted as negative by the model; FP is the number of samples incorrectly predicted as positive by the model; FN is the number of samples incorrectly predicted as negative by the model.

[0078] Accuracy:

[0079]

[0080] Among them, TP is the number of samples correctly predicted as positive by the model; FP is the number of samples incorrectly predicted as positive by the model.

[0081] Recall:

[0082]

[0083] Among them, TP is the number of samples correctly predicted by the model as positive; FN is the number of samples incorrectly predicted by the model as negative.

[0084] F1 score:

[0085]

[0086] Among them, Precision is the precision rate and Recall is the recall rate.

[0087] ROC curve:

[0088] For different thresholds, the TPR and FPR of the model are calculated, and points are plotted in a Cartesian coordinate system with FPR as the horizontal axis and TPR as the vertical axis. All calculated points are connected to form a ROC curve;

[0089]

[0090] Among them, TP is the number of samples correctly predicted as positive by the model; TN is the number of samples correctly predicted as negative by the model; FP is the number of samples incorrectly predicted as positive by the model; FN is the number of samples incorrectly predicted as negative by the model.

[0091] The AUC value is the area under the ROC curve.

[0092] The advantages and positive effects of the adverse drug reaction prediction method based on review text information enhancement described in the present invention are:

[0093] 1. Comprehensively utilize drug review websites, clinical trial data, and drug data sources, and utilize diverse data such as user actual experience and feedback, add patient characteristics, and enrich the relationship between drugs and adverse reactions, providing a more comprehensive drug safety assessment. In addition, compared with other models, the GCN model is faster in training and prediction, which can reduce the difficulty of calculation.

[0094] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] Figure 1 It is a schematic diagram of an embodiment of a method for predicting adverse drug reactions based on review text information enhancement of the present invention. DETAILED DESCRIPTION

[0096] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0097] Unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0098] Example

[0099] The process of drug adverse reaction prediction method integrating drug reviews and FAERS database is as follows Figure 1 shown.

[0101] 1. Extraction and integration of adverse drug reaction information based on review text and FAERS

[0102] 1.1 Data Preparation

[0103] Constructing the adverse drug reaction dictionary: Galeano's dataset is used as the benchmark dataset, which includes 760 drugs and 995 adverse drug reactions. The dictionary is constructed based on the preferred terms and synonyms of adverse drug reactions in the MedDRA ontology, and the adverse reaction terms are encoded by direct coding. A unique identifier is assigned to each term and synonym, and stored as a text file to construct the adverse drug reaction dictionary;

[0104] Collection of adverse drug reaction information based on review text: We crawled the drug review website WebMD, which has a large amount of Internet user data, through a self-written python crawler program. The drug review texts in all data sets on the entire website were collected, and the collected information included the review texts and the gender, age, illness, and duration of medication of the reviewers. We used a self-written java program and a constructed adverse drug reaction dictionary to implement word segmentation, stop word removal, and entity extraction on the WebMD review texts, and obtained drug-adverse reaction association information based on the review texts.

[0105] Drugs in the benchmark data set were screened in FAERS, and after invalid or incomplete records were removed, the adverse reaction data of these drugs in the report forms were obtained, and different adverse reactions of the same drug were integrated to obtain drug-adverse reaction association information based on the FAERS database.

[0106] Integration of adverse drug reaction entities: The extracted drug-adverse reaction associations based on review texts, drug-adverse reaction associations based on the FAERS database, and related medication information are merged with the data of the benchmark dataset to obtain a drug-adverse reaction association matrix that integrates reviews and FAERS database information.

[0107] 2. Integration of drug characteristic information from public databases

[0108] The SMILES chemical formula, target and other information of the drugs in the data set were obtained from DrugBank, the indication information of the drugs was obtained from the TTD database, and the hierarchical structure information of the adverse reactions in the data set was obtained from the ADReCS database, including the upper and lower hierarchical relationships and structure numbers of the adverse reactions, which served as the basis for the subsequent calculation of the similarity of drugs and adverse reactions. Finally, five associations were obtained, including drug-molecular structure, drug-target, drug-indication, adverse reaction hierarchical relationship, and drug-adverse reaction, as well as patient characteristic information.

[0109] 3. Calculation of Similarity

[0110] The similarity calculation is divided into two parts: comprehensive drug similarity calculation and adverse reaction similarity calculation.

[0111] First, the SMILES string is converted into a molecular fingerprint. For the fingerprints of two drug molecules, their intersection and union are calculated. The similarity of the drug-molecule structure is calculated using the Tanimoto coefficient to obtain the drug molecule structure similarity matrix.

[0112] Then, the intersection and union of the targets of the drugs or the indications for treatment are calculated respectively, and the drug-target similarity and drug-indication similarity are calculated according to the Jaccard coefficient to obtain the drug-target and indication similarity matrices.

[0113] Secondly, the characteristics of the patients taking the medicine are used as one of the bases for calculating the similarity of the medicines. The similarity of the medicines based on the patient characteristics is calculated according to the cosine similarity, and the similarity matrix of the medicines based on the patient characteristics is obtained. Among them, the calculation of the patient feature vector: processing the missing values ​​and abnormal values ​​of data such as gender, age, disease condition, and duration of medication, using gender as a numerical variable, converting the duration of medication and age into categorical variables by partitioning, and classifying the disease condition (disease type) according to the disease type, using unique hot encoding for the above feature types, and using Word2Vec for feature construction for the disease condition, splicing all the above features (including unique hot encoded features and features generated by Word2Vec) together to form a patient's feature vector, and the features of multiple patients form a feature matrix. The feature matrix is ​​converted into a one-dimensional vector through dimensionality reduction technology, representing the patient characteristics of a drug.

[0114] Through the weighted similarity fusion algorithm, a comprehensive drug similarity matrix is ​​obtained.

[0115] The similarity between adverse reactions is calculated based on the directed acyclic graph (DAG). For each adverse reaction, a directed acyclic graph (DAG) based on the adverse reaction hierarchy is constructed, which contains semantic descriptors related to the adverse reaction, A. Its DAG can be expressed as DAG (A) = {TA, EA}, where TA is the set of ancestor nodes of A including itself, and EA is the set of all edges in the DAG.

[0116] The contribution of descriptor t to "A" in DAG(A) is calculated by the formula. Then, the contributions of all descriptors in DAG(A) are summed to obtain the semantic value of "A". Calculation of adverse reactions s based on semantic values i and j The semantic similarity between them is as follows to obtain the adverse reaction similarity matrix S semantic The calculation formula is as follows:

[0117] Tanimoto coefficient:

[0118]

[0119] Where A and B are the molecular fingerprints of drugs a and b respectively;

[0120] Jaccard coefficient:

[0121]

[0122] A and B are the target (indication) sets of drugs a and b, respectively;

[0123] Cosine Similarity:

[0124]

[0125] Where A and B are the patient feature vectors of drugs a and b respectively;

[0126] Directed Acyclic Graph (DAG):

[0127]

[0128] where θ is the semantic contribution factor of the edge connecting node t with its child node t.

[0129] 4. Construction of Heterogeneous Graph

[0130] Based on the comprehensive drug similarity matrix, adverse reaction similarity matrix and drug-adverse reaction association matrix, a heterogeneous graph G is constructed, which is expressed as:

[0131]

[0132] Where S d , S a are the drug comprehensive similarity matrix and adverse reaction similarity matrix, respectively. d ,~S a are the matrices of drug and adverse reaction similarity matrices after Laplace transformation, and A is the association matrix between drugs and adverse reactions.

[0133] 5. Construction of graph convolutional network model

[0134] The constructed heterogeneous graph is introduced into the GCN encoder to extract drug and disease embeddings respectively, and the association in the test dataset is set to 0. The encoder can be expressed as:

[0135]

[0136] Among them, L en is the total number of coding layers, is the degree matrix, is the self-loop adjacency matrix of matrix A after adding self-connection, H l is the node feature matrix of the lth layer, H (l+1) is the node feature matrix of the l+1th layer, (W en ) l is the learning weight matrix of the lth layer, using ReLU as the activation function. The number of encoder layers in this model is 2. In order to make full use of the embedding of drug and adverse reaction nodes, the importance weights of different GCN layers are adaptively and dynamically adjusted, and the attention mechanism is added in the process of node feature representation.

[0137] The obtained node representation is decoded using the inner product decoder to obtain the final embedding of drugs and adverse reactions:

[0138]

[0139] Where E d ∈R m×d For the final embedding of the drug, E a ∈R n×d The final embedding for adverse reactions. l is a weight initialized to 1 / (l+1).

[0140] Introduce sigmoid as the activation function into the GCN decoder, and use the inner product decoder to get the predicted correlation matrix:

[0141] A′=sigmoid(E d ·W.E. a ) (9)

[0142] Where A′ ij It is a drug i and adverse reactionsa j The associated prediction score, W∈R d×d is a trainable parameter matrix.

[0143] The model optimization part uses weighted cross entropy loss to measure the loss between the original matrix and the reconstructed association matrix, and uses the Adam algorithm to optimize the loss.

[0144] 6. Model training and evaluation

[0145] Choose 10-fold cross validation to train and evaluate model performance, and use indicators such as accuracy, F1 score, precision, recall, ROC curve, and AUC value to evaluate the performance of the graph convolutional network (GCN) model. The 10-fold cross validation randomly divides the entire dataset into ten subsets (folds), uses one dataset as the test set, and combines the remaining nine datasets as the training set. The training set is used to train the model, and the test set is used to evaluate the performance of the model. Repeat the above steps until each subset is used as a test set. In all ten folds, the average of the model performance indicators is calculated and used as the final result.

[0146] The evaluation index calculation formula is as follows:

[0147] Accuracy:

[0148]

[0149] Among them, TP is the number of samples correctly predicted as positive by the model; TN is the number of samples correctly predicted as negative by the model; FP is the number of samples incorrectly predicted as positive by the model; FN is the number of samples incorrectly predicted as negative by the model.

[0150] Precision:

[0151]

[0152] Among them, TP is the number of samples correctly predicted as positive by the model; FP is the number of samples incorrectly predicted as positive by the model.

[0153] Recall:

[0154]

[0155] Among them, TP is the number of samples correctly predicted by the model as positive; FN is the number of samples incorrectly predicted by the model as negative.

[0156] F1 Score:

[0157]

[0158] Among them, Precision is the precision rate and Recall is the recall rate.

[0159] ROC curve:

[0160] For different thresholds, calculate the TPR and FPR of the model. Plot the points in a Cartesian coordinate system with FPR as the horizontal axis and TPR as the vertical axis. Connect all the calculated points to form a ROC curve.

[0161]

[0162] Among them, TP is the number of samples correctly predicted as positive by the model; TN is the number of samples correctly predicted as negative by the model; FP is the number of samples incorrectly predicted as positive by the model; FN is the number of samples incorrectly predicted as negative by the model.

[0163] The AUC value is the area under the ROC curve.

[0164] 7. Prediction of adverse drug reactions

[0165] Taking drugs for treating epilepsy as an example, the possible adverse reactions of the drugs are predicted in the trained model, and the drug-adverse reaction association pairs with prediction values ​​greater than 0.7 are selected as the final prediction results, indicating the risk of adverse reactions that may be caused by medication.

[0166] 8. Analysis of the mechanism of adverse drug reactions

[0167] For the top ten drug-adverse reaction associations in the case analysis, drug target information was obtained from DrugBank, drug target-related genes were obtained from STRING, adverse reaction-related genes were obtained from GeneCards, and drug target-target-related genes-adverse reaction-related genes were intersected to obtain intersection genes. Cytoscape was used to construct a PPI interaction network diagram, and GO and KEGG analysis was performed on the intersection genes. The biological processes and potential impact pathways in the target genes were further analyzed. The mechanism of drug-induced adverse reactions was revealed and the rationality of the prediction results was verified. By analyzing the drug's target, the related targets of adverse reactions, biological pathways, etc., we can deeply understand the mechanism of adverse reactions, explain the causes of adverse reactions at the molecular level, reveal the drug's mode of action in the human body and its impact on biological systems, and help evaluate the safety of drugs. This will promote the continuous improvement of drugs and further evaluate the risks and benefits of drugs.

[0168] Therefore, the present invention adopts the above-mentioned adverse drug reaction prediction method based on review text information enhancement, comprehensively utilizes drug review websites, clinical trial data and drug data sources, comprehensively utilizes diverse data such as user actual experience and feedback, adds patient characteristics, enriches the relationship between drugs and adverse reactions, and provides a more comprehensive drug safety assessment. In addition, compared with other models, the GCN model is faster in training and prediction, which can reduce the difficulty of calculation.

[0169] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for predicting adverse drug reactions based on enhanced review text information. It is characterized in that The following steps are involved: Step 1, construction of adverse drug reaction dictionary: Galeano's dataset was used as the benchmark dataset, the preferred terms of adverse drug reactions and their synonyms in MedDRA ontology were used as the vocabulary source, and the adverse reaction terms were encoded by direct coding to construct the adverse drug reaction dictionary. Step 2: Extraction of adverse drug reaction information based on review text: The review data of the drugs in the dataset are collected from the drug review website through a python crawler program; Use Java program and the constructed adverse drug reaction dictionary to implement word segmentation, stop word removal and entity extraction of adverse drug reaction information data based on review text, and obtain drug-adverse reaction association information based on review text; Step 3, Integration of adverse drug reaction information: Screen the drugs in the benchmark data set in FAERS, remove invalid or incomplete records, obtain the adverse reaction data of these drugs in the report form, integrate different adverse reactions of the same drug, and obtain drug-adverse reaction association information based on the FAERS database; The extracted drug-adverse reaction associations based on review texts, drug-adverse reaction associations based on the FAERS database, and related medication information are merged with the data of the benchmark dataset to obtain a drug-adverse reaction association matrix that integrates the review and FAERS database information. Step 4: Integration of drug feature information from public databases: The SMILES chemical formula and target information of the drugs in the dataset were obtained from DrugBank, the indication information of the drugs was obtained from the TTD database, and the hierarchical structure information of the adverse reactions in the dataset was obtained from the ADReCS database, including the hierarchical relationship and structure number of the adverse reactions, which served as the basis for the subsequent calculation of the similarity of drugs and adverse reactions. Finally, five associations were obtained, including drug-molecular structure, drug-target, drug-indication, adverse reaction hierarchical relationship, and drug-adverse reaction, as well as patient characteristic information. Step 5, calculation of similarity: ① Comprehensive drug similarity: First, the SMILES string is converted into a molecular fingerprint. For the fingerprints of two drug molecules, their intersection and union are calculated. The similarity of the drug-molecule structure is calculated using the Tanimoto coefficient to obtain the drug molecule structure similarity matrix. Then, the intersection and union of the targets or treatment indications of the drugs are calculated respectively, and the drug-target similarity and drug-indication similarity are calculated respectively according to the Jaccard coefficient to obtain the drug-target and indication similarity matrices; Calculate the similarity of drugs based on patient characteristics according to cosine similarity, and obtain a similarity matrix of drugs based on patient characteristics; Finally, a comprehensive drug similarity matrix is ​​obtained through the weighted similarity fusion algorithm; ②Similarity of adverse reactions: The similarity between adverse reactions is calculated based on the directed acyclic graph (DAG). For each adverse reaction, a directed acyclic graph (DAG) based on the adverse reaction hierarchical structure is constructed to calculate the adverse reaction similarity matrix. Step 6, construction of heterogeneous graph: Based on the comprehensive drug similarity matrix, adverse reaction similarity matrix and drug-adverse reaction association matrix, a heterogeneous graph G is constructed; Step 7: Building the graph convolutional network model: The constructed heterogeneous graph is introduced into the GCN encoder to extract drug and disease embeddings respectively, and the association in the test dataset is set to 0. The encoder is expressed as: Among them, L en is the total number of coding layers, is the degree matrix, is the self-loop adjacency matrix of matrix A after adding self-connection, H l is the node feature matrix of the lth layer, H (l+1) is the node feature matrix of the l+1th layer, (W en ) l is the learning weight matrix of layer l, using ReLU as the activation function; The obtained node representation is decoded using the inner product decoder to obtain the final embedding of drugs and adverse reactions: Where E d ∈R m×d For the final embedding of the drug, E a ∈R n×d For the final embedding of adverse reactions, a l is a weight initialized to 1 / (l+1); Introduce sigmoid as the activation function into the GCN decoder, and use the inner product decoder to get the predicted correlation matrix: A′=sigmoid(E d ·W·E a ) Where A′ ij It is a drug i and adverse reactionsa j The associated prediction score, W∈R d×d is a trainable parameter matrix; The model optimization part uses weighted cross entropy loss to measure the loss between the original matrix and the reconstructed association matrix, and uses the Adam algorithm to optimize the loss; Step 8, model training and evaluation: Choose 10-fold cross validation to train and evaluate model performance, and use accuracy, F1 score, precision, recall, ROC curve and AUC value indicators to evaluate the performance of the graph convolutional network GCN model; Ten-fold cross validation randomly divides the entire data set into ten subsets, uses one of the datasets as the test set, and combines the remaining nine datasets as the training set. The training set is used to train the model, and the test set is used to evaluate the performance of the model. The above steps are repeated until each subset is used as a test set once. In all ten folds, the average of the model performance indicators is calculated and used as the final result. Step 9, prediction of adverse drug reactions: Taking drugs for treating epilepsy as an example, the adverse reactions that may be caused by the drugs are predicted in the trained model, and the drug-adverse reaction association pairs with a prediction value greater than 0.7 are selected as the final prediction results, indicating the risk of adverse reactions that may be caused by the medication; Step 10, analysis of the mechanism of adverse drug reactions: For the top ten drug-adverse reaction associations in terms of accuracy in the case analysis, drug target information was obtained from DrugBank, drug target-related genes were obtained from STRING, and adverse reaction-related genes were obtained from GeneCards. The intersection of drug target-target-related genes-adverse reaction-related genes was processed to obtain intersection genes. Cytoscape was used to construct a PPI interaction network diagram, GO and KEGG analysis was performed on the intersection genes, and the biological processes in the target genes and potential influencing pathways were further analyzed.

2. A method for predicting adverse drug reactions based on enhanced comment text information according to claim 1, Features: In step 2, the drug reviews in the data set on the drug review website are collected, and the collected information includes the review text and the reviewer's gender, age, medical condition, and duration of medication.

3. A method for predicting adverse drug reactions based on enhanced comment text information according to claim 1, It is characterized in that In step 5, the calculation formula is as follows: Tanimoto coefficient: Where A and B are the molecular fingerprints of drugs a and b respectively; Jaccard coefficient: Among them, A and B are the targets of drugs a and b respectively, that is, the set of indications.

4. A method for predicting adverse drug reactions based on enhanced comment text information according to claim 1, It is characterized in that In step 5, ①, patient characteristic information is also used as one of the bases for calculating drug similarity. Missing values ​​and outliers in gender, age, disease condition, and medication duration data are processed. Gender is used as a numerical variable. Medication duration and age are converted into categorical variables by partitioning. Disease conditions or types of diseases are classified according to disease types. One-hot encoding is used for the above feature types, and Word2Vec is used for feature construction of disease conditions. All features, including one-hot encoded features and features generated by Word2Vec, are concatenated to form a feature vector of a patient. The features of multiple patients form a feature matrix. The feature matrix is ​​converted into a one-dimensional vector through dimensionality reduction technology, representing the patient characteristics of a drug. The similarity of drugs based on patient characteristics is calculated based on cosine similarity to obtain a similarity matrix of drugs based on patient characteristics.

5. A method for predicting adverse drug reactions based on enhanced comment text information according to claim 4, It is characterized in that The cosine similarity is calculated as follows: Where A and B are the patient feature vectors of drugs a and b respectively.

6. A method for predicting adverse drug reactions based on enhanced comment text information according to claim 1, It is characterized in that In step 5, the calculation formula is as follows: Directed Acyclic Graph DAG: where θ is the semantic contribution factor of the edge connecting node t with its child node t.

7. A method for predicting adverse drug reactions based on enhanced comment text information according to claim 1, It is characterized in that The heterogeneous graph G in step 6 is represented as: Where S d , S a are the drug comprehensive similarity matrix and adverse reaction similarity matrix, respectively. d ,~S a are the matrices of drug and adverse reaction similarity matrices after Laplace transformation, and A is the association matrix between drugs and adverse reactions.

8. A method for predicting adverse drug reactions based on enhanced comment text information according to claim 1, It is characterized in that The number of encoder layers in step 7 is 2. In order to make full use of the embedding of drug and adverse reaction nodes, the importance weights of different GCN layers are adaptively and dynamically adjusted, and the attention mechanism is added in the process of node feature representation.

9. A method for predicting adverse drug reactions based on enhanced comment text information according to claim 1, It is characterized in that The calculation formula for the evaluation index in step 8 is as follows: Use accuracy: Among them, TP is the number of samples correctly predicted as positive by the model; TN is the number of samples correctly predicted as negative by the model; FP is the number of samples incorrectly predicted as positive by the model; FN is the number of samples incorrectly predicted as negative by the model; Accuracy: Among them, TP is the number of samples correctly predicted as positive by the model; FP is the number of samples incorrectly predicted as positive by the model; Recall: Among them, TP is the number of samples correctly predicted by the model as positive; FN is the number of samples incorrectly predicted by the model as negative; F1 score: Among them, Precision is the precision rate, Recall is the recall rate; ROC curve: For different thresholds, the TPR and FPR of the model are calculated, and points are plotted in a Cartesian coordinate system with FPR as the horizontal axis and TPR as the vertical axis. All calculated points are connected to form a ROC curve; Among them, TP is the number of samples correctly predicted as positive by the model; TN is the number of samples correctly predicted as negative by the model; FP is the number of samples incorrectly predicted as positive by the model; FN is the number of samples incorrectly predicted as negative by the model; The AUC value is the area under the ROC curve.

Images