A toxicity prediction model for predicting agranulocytosis induced by non-chemotherapeutic drugs and its application

Abstract

The present invention discloses a toxicity prediction model for predicting agranulocytosis induced by non-chemotherapeutic drugs. The toxicity prediction model uses the best molecular descriptor combination of the drug to be tested as input information, and inputs it into the integrated model of the trained BalancedBaggingClassifier combined with the gradient boosting decision tree of the base classifier, thereby obtaining the toxicity prediction result of the drug to be tested. The toxicity prediction model can accurately predict the toxicity of agranulocytosis induced by non-chemotherapeutic drugs, reduce the risk of adverse reactions for agranulocytosis in new drug research and development, reduce research and development costs, and accelerate the research and development process; and the toxicity prediction model can evaluate the risk of agranulocytosis caused by existing non-chemotherapeutic drugs, newly listed unknown adverse reaction drugs or preclinical research / clinical research compounds, provide medication warnings for clinicians, help improve patient medication safety, and improve medical quality and patient quality of life.

Description

Technical Field

[0001] The present invention relates to the technical field of computational chemistry and toxicology, and in particular to a toxicity prediction model for predicting agranulocytosis induced by non-chemotherapeutic drugs and its application. Background Art

[0002] Agranulocytosis (granulocytosis) usually refers to severe neutropenia, that is, the absolute neutrophil count in the blood is less than 0.5×10 9 It is reported that drugs are the main cause of agranulocytosis, accounting for 70-97% of all cases. Drug-induced agranulocytosis can be divided into two categories: one is related to cytotoxic chemotherapy drugs, called chemotherapy-induced agranulocytosis; the other is related to non-chemotherapeutic drugs, called idiopathic drug-induced agranulocytosis, also known as non-chemotherapeutic drug-induced agranulocytosis.

[0003] Chemotherapy-induced granulocytosis occurs frequently in cancer patients, so the patient's blood cell count can be closely monitored for timely intervention. For patients using cytotoxic drugs with a high risk of granulocytosis, they are usually treated with granulocyte colony-stimulating factor (G-CSF) or granulocyte-macrophage colony-stimulating factor (GM-CSF) after medication to prevent the occurrence of neutropenia, reduce the degree of neutropenia, shorten the duration of granulocytosis, and accelerate the recovery of granulocytes, thereby reducing the risk of concurrent infection and fever; however, in contrast, non-chemotherapeutic drug-induced granulocytosis usually lacks early intervention and prevention measures. And non-chemotherapeutic drug-induced granulocytosis is a serious and life-threatening adverse drug reaction. Such patients may not have obvious symptoms at first, but then they will rapidly develop severe infections, such as severe pneumonia, sepsis, or septic shock; because this type of adverse drug reaction is relatively rare, the medical field has limited understanding of its mechanism and incidence, making it difficult for medical staff to predict in advance which patients may experience such adverse reactions, and there is no clear method to prevent them.

[0004] Currently, a variety of non-chemotherapeutic drugs have been reported to cause agranulocytosis, some of which are high-risk drugs, including methimazole, ticlopidine, clozapine, sulfasalazine, trimethoprim-sulfamethoxazole, and metamizole. However, the exact mechanism of non-chemotherapeutic drug-induced agranulocytosis has not yet been fully elucidated. Based on clinical observations and laboratory research results, two possible mechanisms have been proposed: (1) drug-related antibody-mediated destruction of circulating neutrophils, such as drugs such as β-lactam antibiotics and propylthiouracil; (2) direct toxic effects on bone marrow granulocyte precursors, such as drugs such as chlorpromazine and clozapine. Although the mechanism of agranulocytosis induced by some non-chemotherapeutic drugs has been understood, the mechanism of agranulocytosis for most non-chemotherapeutic drugs is still difficult to clearly elucidate. In clinical practice, the metabolites of some drugs react rapidly after formation, resulting in low antibody titers of the metabolites, which are therefore difficult to capture in tests. In addition, the influence of genetic polymorphisms may also result in the formation of drug metabolites only under disease-related conditions in specific individuals, resulting in a low and difficult to predict incidence of non-chemotherapeutic drug-induced agranulocytosis.

[0005] To date, no systematic in vivo and in vitro evaluation models have been established to study non-chemotherapeutic drug-induced agranulocytosis. Due to the above limitations and the strict population inclusion criteria of drug clinical trials (from phase I to phase III), non-chemotherapeutic drug-induced agranulocytosis has in most cases only been noticed during adverse reaction monitoring after drug marketing.

[0006] Therefore, there is an urgent need to develop new models to evaluate the toxicity of non-chemotherapeutic drugs-induced agranulocytosis, which can be used to guide the drug molecule design stage and evaluate the risk of causing agranulocytosis in early clinical trials of drugs. Summary of the invention

[0007] The purpose of the present invention is to overcome the above-mentioned deficiencies of the prior art and to provide a toxicity prediction model for predicting non-chemotherapeutic drug-induced agranulocytosis and its application.

[0008] The first object of the present invention is to provide a toxicity prediction model for predicting agranulocytosis induced by non-chemotherapeutic drugs.

[0009] The second object of the present invention is to provide application of the above toxicity prediction model in predicting the toxicity of non-chemotherapeutic drugs induced agranulocytosis.

[0010] A third object of the present invention is to provide a method for predicting the toxicity of non-chemotherapeutic drugs inducing agranulocytosis.

[0011] The fourth object of the present invention is to provide a toxicity prediction system for predicting agranulocytosis induced by non-chemotherapeutic drugs.

[0012] In order to achieve the above object, the present invention is implemented by the following scheme:

[0013] A toxicity prediction model for predicting non-chemotherapeutic drug-induced agranulocytosis, wherein the toxicity prediction model uses the best molecular descriptor combination of the drug to be tested as input information, which is input into an integrated model of a trained BalancedBaggingClassifier combined with a base classifier gradient boosting decision tree, and the integrated model of the trained BalancedBaggingClassifier combined with a base classifier gradient boosting decision tree outputs a toxicity label of the drug to be tested;

[0014] The value of the toxicity label is 0 or 1; when the toxicity label of the drug to be tested is 0, the drug to be tested does not have the toxicity of inducing agranulocytosis; when the toxicity label is 1, the drug to be tested has the toxicity of inducing agranulocytosis; the drug to be tested is a non-chemotherapeutic drug;

[0015] The method for obtaining the optimal molecular descriptor combination of the drug to be tested is as follows: obtaining the chemical structural formula of the drug to be tested and performing preprocessing to obtain the chemical structural formula of the drug to be tested after the preprocessing, calculating the molecular descriptor of the drug to be tested, deleting the null value in the molecular descriptor of the drug to be tested, and then performing standardization to obtain the standardized molecular descriptor of the drug to be tested; deleting the molecular descriptors with a variance of 0 in the molecular descriptors of the drug to be tested after the standardization, calculating the correlation between every two molecular descriptors in the molecular descriptors of the drug to be tested after the standardization, deleting one of the two molecular descriptors with a correlation greater than 0.9, and obtaining the optimal molecular descriptor combination of the drug to be tested;

[0016] The pretreatment is strong base protonation, strong acid deprotonation, removal of inorganic counter ions and addition of explicit hydrogen ions;

[0017] The trained integrated model of BalancedBaggingClassifier combined with base classifier gradient boosting decision tree is an integrated model of BalancedBaggingClassifier combined with base classifier gradient boosting decision tree trained using a training set containing positive data and negative data;

[0018] The positive data are non-chemotherapeutic drugs that can cause agranulocytosis, and the negative data are non-chemotherapeutic drugs after agranulocytosis-related drugs are removed from the SIDER database.

[0019] Preferably, the drug to be tested is a non-chemotherapeutic drug with an average standardized distance in [0,1]; the average standardized distance is calculated by Euclidean Applicability Domain 1.0 software in combination with a training set containing positive data and negative data.

[0020] Preferably, the preprocessing is performed using MOE software.

[0021] Preferably, the molecular descriptor of the drug to be tested is a Discovery Studio molecular descriptor and / or an RDKit molecular descriptor of the drug to be tested.

[0022] More preferably, the molecular descriptor of the drug to be tested is the Discovery Studio molecular descriptor and the RDKit molecular descriptor of the drug to be tested.

[0023] Further preferably, the Discovery Studio molecular descriptor of the drug to be tested is calculated by DiscoveryStudio software.

[0024] Further preferably, the RDKit molecular descriptor of the drug to be tested is calculated by the rdkit package of Python software.

[0025] Preferably, the normalization process is a Z-score normalization process.

[0026] Preferably, the correlation is a Pearson correlation.

[0027] Preferably, the optimal molecular descriptor combination of the drug to be tested includes 237 molecular descriptors, and the 237 molecular descriptors are as follows:

[0028]

[0029]

[0030] Preferably, the training and verification using a training set and a test set containing positive data and negative data are specifically as follows: the toxicity label of the positive data in the training set and the test set is assigned a value of 1, and the toxicity label of the negative data in the training set and the test set is assigned a value of 0, and the training set is input into an integrated model of BalancedBaggingClassifier combined with a base classifier gradient boosting decision tree for ten-fold cross validation to obtain a trained integrated model of BalancedBaggingClassifier combined with a base classifier gradient boosting decision tree.

[0031] Preferably, the negative data are non-chemotherapeutic drugs with ATC code of L01 in the SIDER database excluding agranulocytosis-related drugs.

[0032] More preferably, the removing of agranulocytosis-related drugs from the SIDER database specifically comprises: deleting drugs whose side effect descriptions contain agranulocytosis, neutropenia, Febrile neutropenia, leukopenia, White blood cell count decreased, Neutrophil count increased, White blood cell count low, Neutrophil count decreased, Absolute neutrophil count decreased, White blood cell count increased, Blood count abnormal, White blood cell count abnormal, Band neutrophil count increased or Blood neutrophil count decreased in the SIDER database.

[0033] Preferably, the positive data include the chemical structural formula and / or PubChem CID code of non-chemotherapeutic drugs that can cause agranulocytosis; the negative data include the chemical structural formula and / or PubChem CID code of non-chemotherapeutic drugs after excluding agranulocytosis-related drugs in the SIDER database.

[0034] Preferably, the BalancedBaggingClassifier combined with the integrated model of the base classifier gradient boosting decision tree also outputs a toxicity probability value of the drug to be tested; the toxicity probability value is 0-1; the toxicity probability value is output by the predict_proba function embedded in the BalancedBaggingClassifier.

[0035] The present invention also claims the use of any of the above-mentioned toxicity prediction models in predicting the toxicity of non-chemotherapeutic drugs induced agranulocytosis.

[0036] The present invention also claims a method for predicting the toxicity of non-chemotherapeutic drugs induced agranulocytosis, using any of the above-mentioned toxicity prediction models for prediction.

[0037] The present invention also claims a toxicity prediction system for predicting non-chemotherapeutic drug-induced agranulocytosis, comprising a data acquisition module, a toxicity prediction module and a result output module;

[0038] The data acquisition module is used to obtain the best molecular descriptor combination of the drug to be tested;

[0039] The toxicity prediction module is any of the above-mentioned toxicity prediction models, which uses the optimal molecular descriptor combination of the drug to be tested obtained by the data acquisition module as an input variable to obtain the toxicity prediction result of the drug to be tested;

[0040] The toxicity prediction result of the drug to be tested is a toxicity label of the drug to be tested, and the value of the toxicity label is 0 or 1;

[0041] The result output module is used to output the toxicity prediction result of the drug to be tested obtained by the toxicity prediction module;

[0042] The drug to be tested is a non-chemotherapeutic drug; the method for obtaining the optimal molecular descriptor combination of the drug to be tested is as shown above.

[0043] Preferably, the toxicity prediction result also includes a toxicity probability value; the toxicity probability value is 0-1.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] The present invention provides a toxicity prediction model for predicting agranulocytosis induced by non-chemotherapeutic drugs. The optimal molecular descriptor combination of a drug to be tested is used as input information and input into an integrated model of a trained BalancedBaggingClassifier combined with a base classifier gradient boosting decision tree. The integrated model of the trained BalancedBaggingClassifier combined with the base classifier gradient boosting decision tree outputs a toxicity label of the drug to be tested. If the toxicity label of the drug to be tested is 0, the drug to be tested does not have the toxicity of inducing agranulocytosis; if the toxicity label of the drug to be tested is 1, the drug to be tested has the toxicity of inducing agranulocytosis. The toxicity prediction model can accurately predict the agranulocytosis-induced toxicity of non-chemotherapeutic drugs, can perform effective toxicity screening in the early drug design stage, and help reduce the risk of adverse reactions to agranulocytosis in new drug development, reduce research and development costs, and accelerate the drug development process; and the toxicity prediction model can be used to evaluate the risk of agranulocytosis in existing non-chemotherapeutic drugs, newly marketed drugs with unknown adverse reactions, or preclinical research / clinical research compounds, provide important medication warnings for clinicians, help improve patient medication safety, reduce the incidence of adverse reactions, and improve medical quality and patient quality of life. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a flow chart for constructing a toxicity prediction model for predicting non-chemotherapeutic drug-induced agranulocytosis in Example 1. DETAILED DESCRIPTION

[0047] The present invention is further described in detail below in conjunction with the accompanying drawings and specific examples of the specification. The examples are only used to explain the present invention and are not used to limit the scope of the present invention. The test methods used in the following examples are conventional methods unless otherwise specified; the materials and reagents used are reagents and materials that can be obtained from commercial channels unless otherwise specified.

[0048] Example 1 A toxicity prediction model for predicting non-chemotherapeutic drug-induced agranulocytosis

[0049] 1. Model Construction Method

[0050] A flowchart of the construction of a toxicity prediction model for predicting non-chemotherapeutic drug-induced agranulocytosis is shown in Figure 1 As shown, the specific steps include:

[0051] 1. Collection and preprocessing of drug datasets

[0052] Positive data set: Through existing literature review, we collected non-chemotherapeutic drugs (toxic drugs) known to induce agranulocytosis, and obtained the PubChem CID codes and chemical structures of the toxic drugs;

[0053] Existing literature reviews include: [1] Andersohn F, Konzen C, Garbe A E. Systematic review: agranulocytosis induced by nonchemotherapy drugs. [J]. Annals of Internal Medicine, 2007, 146(9): 657-65. DOI: doi: 10.7326 / 0003-4819-146-9-200705010-00009.

[0054] [2]Garbe,E.,Non-chemotherapy drug-induced agranulocytosis.Expertopinion on drug safety 2007,6,(3),323-335.doi.org / 10.1517 / 14740338.6.3.323;

[0055] [3]Emmanuel Andrès,Frédéric Maloisel.Idiosyncratic drug-inducedagranulocytosis or acute neutropenia[J].Current Opinion in Hematology,2008,15(1):15-21.DOI:10.1097 / MOH.0b013e3282f15fb9.;

[0056] [4]Andrès,Emmanuel,Mourot-Cottet R.Non-chemotherapy drug-inducedneutropenia-an update[J].Expert Opinion on Drug Safety,2017:14740338.2017.1376645.DOI:10.1080 / 14740338.2017.1376645.;

[0057] [5]Andrès,Villalba,Zulfiqar,et al.State of Art of Idiosyncratic Drug-Induced Neutropenia or Agranulocytosis,with a Focus on Biotherapies[J].Journal of Clinical Medicine,2019,8(9):1351.DOI:10.3390 / jcm8091351.。

[0058] Negative data set: In the SIDER database, delete drugs with side effect descriptions containing agranulocytosis, neutropenia, febrile neutropenia, leukopenia, white blood cell count decreased, neutrophil count increased, white blood cell count low, neutrophil count decreased, absolute neutrophil count decreased, white blood cell count increased, blood count abnormal, white blood cell count abnormal, band neutrophil count increased, or blood neutrophil count decreased. Then delete drugs with ATC code of L01 to obtain non-chemotherapeutic drugs (non-toxic drugs) not related to agranulocytosis, and obtain the PubChem CID code and chemical structure formula of non-toxic drugs.

[0059] Quality control was performed on the positive data set and the negative data set respectively. The quality control included: (1) eliminating biopharmaceuticals, mixtures, inorganic compounds and organometallic compounds in the positive data set and the negative data set; (2) using the "wash" function in the MOE software (version number 2015.10) to perform strong base protonation, strong acid deprotonation, removal of inorganic counterions and addition of explicit hydrogen ions on the positive data set and the negative data set respectively; (3) eliminating chemical structural formulas with relative molecular mass <30 or relative molecular mass >1000 in the positive data set and the negative data set; (4) using the InChiKey of the compound to deduplicate the positive data set and the negative data set respectively.

[0060] A quality control positive data set containing the chemical structural formulas of 219 drugs and a quality control negative data set containing the chemical structural formulas of 540 drugs were obtained, with a total of 759 drug chemical structural formulas, of which the toxicity label of the drug in the positive data set was 1, and the toxicity label of the drug in the negative data set was 0; the chemical structural formulas of the 759 drugs were randomly divided into a training set (607) and a test set (152) according to a quantity ratio of 4:1; the data set information of the training set and the test set is shown in Table 1.

[0061] Table 1 Dataset information

[0062]

[0063]

[0064] 2. Calculation of molecular descriptors and molecular fingerprints and screening of optimal feature subsets

[0065] (1) Calculation of molecular descriptors and molecular fingerprints

[0066] The MOE molecular descriptors (209 bits) and Discovery Studio molecular descriptors (DS molecular descriptors, 440 bits) of the training set and test set obtained in step 1 were calculated using MOE software (version number 2022.02) and Discovery Studio software (version number 2019), respectively, and the RDKit molecular descriptors (196 bits) of the training set and test set obtained in step 1 were calculated using the rdkit package (version number 2022.9.5) of Python software; and the PaDEL-Descriptor software (version number 2.21) was used to calculate the 12 types of molecular fingerprints of the training set and test set obtained in step 1; the molecular descriptors of the training set and test set and the types and bits of molecular fingerprints are shown in Table 2.

[0067] Table 2 Types and digits of molecular descriptors and molecular fingerprints

[0068]

[0069] Preprocessing of molecular descriptors and molecular fingerprints: Delete missing values ​​(null values) in the molecular descriptors and molecular fingerprints of the test set and training set, and then use the StandardScaler function of the sklearn.preprocessing module in the Python software to standardize the molecular descriptors after deleting the missing values ​​according to the Z-score standardization method; calculate the variance of each molecular descriptor and molecular fingerprint, and delete the molecular descriptors and molecular fingerprints with a variance of 0; and calculate the Pearson correlation coefficient between every two molecular descriptors and molecular fingerprints. For two molecular descriptors or molecular fingerprints with a Pearson correlation coefficient (correlation) greater than 0.9, delete the molecular descriptor or molecular fingerprint with a lower variance to obtain the molecular descriptors and molecular fingerprints of the preprocessed training set, as well as the molecular descriptors and molecular fingerprints of the preprocessed test set;

[0070] The processing method of preprocessed molecular descriptors and molecular fingerprints is recorded as Feature Preprocessing.

[0071] (2) Screening of the best feature subset

[0072] The Balanced Random Forest (BRF) algorithm is used as the optimal feature subset screening model to perform feature selection on the molecular descriptors of the preprocessed training set and test set and the molecular fingerprints of the preprocessed training set and test set obtained in step (1) to obtain the optimal feature subset (optimal molecular descriptors and molecular fingerprints).

[0073] The feature selection methods include (1) filtering method: using the F-test to select molecular descriptors with significant correlation with toxicity labels (P < 0.05), and using the mutual information method (MI) to calculate the mutual information value between the molecular descriptors and the toxicity labels, and selecting molecular descriptors with mutual information values ​​> 0; (2) embedding method: retaining molecular descriptors with importance (weight coefficient) ≥ 0.003; (3) recursive feature elimination combined with cross-validation (RFECV) method.

[0074] Using the molecular descriptors and molecular fingerprints of the preprocessed training set obtained in step (1), a ten-fold cross validation was performed based on the BRF algorithm according to different combinations (as shown in Table 3) to obtain the toxicity labels of the training set data of different combinations output by the BRF algorithm.

[0075] The true positive (True Positive, TP), true negative (True Negative, TN), false positive (False Positive, FP) and false negative (False Negative, FN) of different combinations when verified based on the BRF algorithm are recorded, and the BRF performance index corresponding to different combinations is calculated according to Formula I to Formula IV.

[0076] BRF performance indexes include: area under the ROC curve (ROC_AUC), accuracy (ACC), sensitivity (SEN), specificity (SPE) and Matthews correlation coefficient (MCC); true positive refers to the positive data correctly identified in the training set, true negative refers to the negative data correctly identified in the training set, false positive refers to the positive data incorrectly identified in the training set, and false negative refers to the negative data incorrectly identified in the training set.

[0077] Different molecular descriptors, molecular descriptor combinations and molecular fingerprints and the corresponding BRF performance indices are shown in Table 3.

[0078] Table 3 Different molecular descriptors, molecular descriptor combinations and molecular fingerprints and their corresponding BRF performance indices

[0079]

[0080]

[0081] Formula 1: ACC = (TP + TN) / (TP + TN + FP + FN);

[0082] Formula II: SEN = TP / (TP+FN);

[0083] Formula III: SPE = TN / (TN + FP);

[0084] Formula IV: (TP×TN-FP×FN) / {(FP+TN)(FP+TP)(FN+TN)(FN+TP)} 1 / 2 .

[0085] Based on the performance index shown in Table 3, it can be determined that the molecular descriptor set of DS+RDKit after the preprocessing shown in step (1) can be used to construct a machine learning model to obtain the model with the best performance, and the Matthews correlation coefficient (MCC) of the model is 0.5263.

[0086] Therefore, the best feature subset (best molecular descriptor) is the molecular descriptor combination after the DS+RDKit molecular descriptor is preprocessed, which contains 237 DS+RDKit molecular descriptor information; the 237 DS+RDKit molecular descriptor information is shown in Table 4.

[0087] Table 4 237 DS+RDKit molecular descriptor information

[0088]

[0089]

[0090] The best feature subset as shown in Table 4 (including 237 DS+RDKit molecular descriptors of each non-chemotherapy drug in the training set) in the preprocessing training set obtained in step (1) is obtained to obtain the training set model optimization data; the best feature subset as shown in Table 4 (including 237 DS+RDKit molecular descriptors of each non-chemotherapy drug in the test set) in the preprocessing test set obtained in step (1) is obtained to obtain the test set model optimization data.

[0091] 3. Establishment of integrated machine learning model

[0092] Nine different ensemble learning models are set, including Balanced Random Forest Classifier (BRF, the basic classifier is Random Forest RF) and Easy Ensemble Classifier (EEC, the basic classifier is AdaBoost) and 7 Balanced Bagging Classifier (BBC) combined with base classifiers: Extreme Gradient Boosting Tree (XGBoost), Gradient Boosting Decision Tree (GBDT), Lightweight Gradient Boosting Machine (LightGBM), Support Vector Machine (SVM), Multilayer Perceptron (MLP), k-Nearest Neighbor (KNN) and Logistic Regression (LR).

[0093] Among them, BalancedRandomForestClassifier, EasyEnsembleClassifier and BalancedBaggingClassifier are implemented by the BalancedRandomForestClassifier function, EasyEnsembleClassifier function and BalancedBaggingClassifier function of the ensemble module of the imbalance-learn package (version number 0.11.0); GBDT, SVM, MLP, KNN and LR are implemented by "GradientBoostingClassifier", "SVC", "MLPClassifier", "KNeighborsClassifier" and "LogisticRegression" in the Scikit-learn (version number 1.1.3) open source framework respectively; LightGBM is implemented by "LGBMClassifier" in the lightgbm package (version number 3.3.5); XGBoost is implemented by "XGBClassifier" in the xgboost package (version number 1.6.1).

[0094] The training set model optimization data obtained in step 2 was combined with each ensemble learning model for ten-fold cross validation, and the performance indices (ROC_AUC, ACC, SEN, SPE and MCC) of different ensemble learning models for ten-fold cross validation were calculated; the test set model optimization data obtained in step 2 was combined with each ensemble learning model for external validation, and the performance indices of different ensemble learning models for external validation were calculated; when each ensemble learning model was performing ten-fold cross validation, the Hyperopt package (version number 0.2.7) in Python was used to automatically optimize the algorithm hyperparameters of the ensemble learning model.

[0095] The performance indexes of each ensemble learning model are shown in Table 5.

[0096] Table 5 Performance index of ensemble learning model

[0097]

[0098] Based on the performance index of each ensemble learning model shown in Table 5, it can be determined that when the ensemble learning model of BalancedBaggingClassifier (BCC) combined with the base classifier gradient boosted decision tree (GBDT) is combined with the best feature subset obtained in step 2 (237 DS+RDKit molecular descriptors shown in Table 4) for the prediction of agranulocytosis-induced toxicity, both the ten-fold cross validation of the training set and the external validation of the test set have the best model performance (MCC optimal).

[0099] Therefore, the best feature subset obtained in step 2 (237 DS+RDKit molecular descriptors shown in Table 4) was combined with the integrated learning model of BalancedBaggingClassifier (BCC) combined with the base classifier gradient boosting decision tree (GBDT) to establish a toxicity prediction model for predicting non-chemotherapeutic drug-induced agranulocytosis; the algorithm hyperparameters of the best basic classifier model were: n_estimators = 198, learning_rate = 0.5, criterion = "squared_error", max_depth = 15, max_features = 30, subsample = 0.7000000000000001.

[0100] 4. Definition of the application domain of the toxicity prediction model for predicting non-chemotherapeutic drug-induced agranulocytosis

[0101] The Euclidean Applicability Domain 1.0 software was used to calculate the average standardized distance of each non-chemotherapy drug molecular descriptor in the training set model optimization data obtained in step 2 (including 237 DS+RDKit molecular descriptors of each non-chemotherapy drug in the training set); the boundary region was constructed using the average standardized distance of each non-chemotherapy drug molecular descriptor in the training set model optimization data to obtain the application domain [0, 1] of the toxicity prediction model.

[0102] Then, based on the average standardized distance of the training set model optimization data, the Euclidean ApplicabilityDomain1.0 software was used to calculate step 2 to obtain the average standardized distance of the molecular descriptors of each non-chemotherapy drug in the test set model optimization data (including 237 DS+RDKit molecular descriptors of each non-chemotherapy drug in the test set); the calculated average standardized distances of each non-chemotherapy drug in the test set model optimization data were all between 0 and 1, and were all within the application domain of the toxicity prediction model, indicating that the results of external validation using the test set model optimization data are reliable.

[0103] 2. Toxicity prediction model

[0104] 1. Model composition

[0105] A toxicity prediction model for predicting non-chemotherapeutic drug-induced agranulocytosis takes the best feature subset of the drug to be tested (including 237 DS+RDKit molecular descriptor information shown in step 1) as input variables, and inputs the input variables into an integrated learning model of a trained BalancedBaggingClassifier (BCC) combined with a base classifier gradient boosted decision tree (GBDT). The integrated learning model of the trained BBC combined with the base classifier GBDT outputs a toxicity prediction result of the drug to be tested, and the toxicity prediction result includes a toxicity probability value and a toxicity label; the toxicity probability value is output by a predict_proba function embedded in the BalancedBaggingClassifier, and the toxicity label is output by a predict function embedded in the BalancedBaggingClassifier.

[0106] The drug to be tested is a non-chemotherapeutic drug; the toxicity label of the drug to be tested is 0 or 1, when the toxicity label is 0, the drug to be tested does not have the toxicity of inducing agranulocytosis; when the toxicity label is 1, the drug to be tested has the toxicity of inducing agranulocytosis; the toxicity probability value of the drug to be tested is 0-1, the closer the toxicity probability value is to 1, the higher the credibility that the drug to be tested has the toxicity of inducing agranulocytosis; the closer the toxicity probability value is to 0, the higher the credibility that the drug to be tested does not have the toxicity of inducing agranulocytosis.

[0107] The integrated learning model of the trained BBC combined with the base classifier GBDT is trained using the training set model optimization data shown in step 1, as follows:

[0108] The toxicity label of the positive data in the training set model optimization data is assigned to 1, and the toxicity label of the negative data in the training set model optimization data is assigned to 0. The training set model optimization data is input into the BBC combined with the base classifier GBDT integrated learning model for ten-fold cross validation to obtain the trained BBC combined with the base classifier GBDT integrated learning model.

[0109] 2. Usage

[0110] Obtain the best feature subset of the drug to be tested (including the 237 DS+RDKit molecular descriptor information shown in Table 4 in step 1), and use Euclidean Applicability Domain 1.0 software combined with the best feature subset of the drug to be tested to calculate the average standardized distance of the molecular descriptor of the drug to be tested (based on the training set model optimization data);

[0111] (1) The average standardized distance of the molecular descriptors of the drug to be tested is between [0, 1], and the best feature subset of the drug to be tested is input as an input variable into the toxicity prediction model shown in step 1, and the toxicity prediction model outputs the toxicity probability value and toxicity label of the drug to be tested;

[0112] The toxicity label of the drug to be tested is 0, which means that the drug to be tested does not have the toxicity of inducing agranulocytosis; the toxicity label of the drug to be tested is 1, which means that the drug to be tested has the toxicity of inducing agranulocytosis.

[0113] (2) When the average normalized distance of the molecular descriptors of the drug to be tested is greater than 1, the drug to be tested cannot be predicted using the toxicity prediction model shown in step 1.

[0114] Example 2 Prediction of toxicity of non-chemotherapeutic drugs induced agranulocytosis

[0115] 1. Toxicity prediction of the hypoglycemic drug tolbutamide

[0116] 1. Experimental methods

[0117] The PubChem number of tolbutamide is determined to be 5505, and the Smiles code of tolbutamide is obtained through the PubChem website; the Smiles code of tolbutamide is input into the MOE software, and the chemical structure of tolbutamide is subjected to strong base protonation, strong acid deprotonation, inorganic counterions removal, and explicit hydrogen ions addition using the "wash" function of the MOE software; then the DS molecular descriptor and RDKit molecular descriptor of tolbutamide are calculated respectively using the rdkit package of the Discovery Studio software and the Python software, and the DS molecular descriptor and RDKit molecular descriptor of tolbutamide are subjected to the Z-score standardization method using the StandardScaler function of the sklearn.preprocessing module in the Python software, to obtain the standardized tolbutamide molecular descriptor. The best feature subset of tolbutamide (including 237 DS+RDKit molecular descriptor information shown in Table 4 in Example 1) is obtained from the standardized tolbutamide molecular descriptor.

[0118] According to the method of using the toxicity prediction model shown in step 2 of Example 1, the toxicity of tolbutamide-induced agranulocytosis is predicted in combination with the best feature subset of tolbutamide, and the toxicity prediction result of tolbutamide by the toxicity prediction model shown in step 2 of Example 1 is obtained.

[0119] 2. Experimental results

[0120] According to the method of using the toxicity prediction model shown in step 2 of Example 1, when the toxicity of tolbutamide-induced agranulocytosis was predicted in combination with the best feature subset of tolbutamide, the average standardized distance of the tolbutamide molecular descriptor calculated by the Euclidean ApplicabilityDomain 1.0 software was 0.097, which was between [0, 1].

[0121] The toxicity prediction model shown in step 2 of Example 1 has a toxicity probability value of 0.9966 in the toxicity prediction result of tolbutamide, and the toxicity label is 1. The toxicity prediction result shows that tolbutamide has the toxicity of inducing agranulocytosis.

[0122] The drug instructions for the hypoglycemic drug tolbutamide also point out that the drug has the adverse reaction of granulocytopenia, indicating that the predicted results are consistent with the actual results.

[0123] 2. Toxicity prediction for a newly launched drug (atypical antipsychotic Caplyta)

[0124] 1. Experimental methods

[0125] The active ingredient of the atypical antipsychotic drug Caplyta is lumateperone. Tolbutamide in step 1 is replaced with lumateperone, and the other processes remain unchanged to obtain the toxicity prediction result of lumateperone by the toxicity prediction model shown in step 2 of Example 1.

[0126] 2. Experimental results

[0127] When the toxicity prediction model shown in step 2 of Example 1 was used to predict the toxicity of the atypical antipsychotic Caplyta (active ingredient: Lumateperone), the average standardized distance of the Lumateperone molecular descriptor calculated by the Euclidean Applicability Domain 1.0 software was 0.127, which was between [0, 1].

[0128] The toxicity prediction model shown in step 2 of Example 1 has a toxicity probability value of 0.5969 for the toxicity prediction result of the atypical antipsychotic Caplyta (active ingredient: Lumateperone), and the toxicity label is 1. The toxicity prediction result shows that the atypical antipsychotic Caplyta (active ingredient: Lumateperone) has the toxicity of inducing agranulocytosis.

[0129] The product instructions for the atypical antipsychotic Caplyta (active ingredient: Lumateperone) indicate that agranulocytosis is one of the adverse reactions of the drug, and the U.S. Food and Drug Administration (FDA) Adverse Event Reporting System (FAERS) database also records cases of agranulocytosis caused by Lumateperone, indicating that the predicted results are consistent with the actual results.

[0130] 3. Toxicity prediction for a newly launched drug with unknown adverse reactions (Resentiro)

[0131] 1. Experimental methods

[0132] Tolbutamide in step 1 was replaced by Resmetirom, and the other treatments remained unchanged to obtain the toxicity prediction result of Resmetirom using the toxicity prediction model shown in step 2 of Example 1.

[0133] 2. Experimental results

[0134] When the toxicity prediction model shown in step 2 of Example 1 was used to predict the toxicity of resteril, the average standardized distance of the resteril molecular descriptor calculated by the Euclidean Applicability Domain 1.0 software was 0.276, which was between [0, 1].

[0135] The toxicity prediction model shown in step 2 of Example 1 has a toxicity probability value of 0.4594 in the toxicity prediction result of resteril, and the toxicity label is 0. The toxicity prediction result shows that resteril does not have the toxicity of inducing agranulocytosis.

[0136] Example 3 A toxicity prediction system for predicting agranulocytosis induced by non-chemotherapeutic drugs

[0137] A toxicity prediction system for predicting non-chemotherapeutic drug-induced agranulocytosis, comprising a data acquisition module, a toxicity prediction module and a result output module;

[0138] The data acquisition module is used to obtain the best feature subset of the drug to be tested (including 237 DS+RDKit molecular descriptor information shown in Table 4 of Example 1);

[0139] The toxicity prediction module is the toxicity prediction model shown in step 2 of Example 1, which uses the optimal feature subset of the drug to be tested obtained by the data acquisition module as an input variable to obtain the toxicity prediction result of the drug to be tested;

[0140] The toxicity prediction result of the drug to be tested is the toxicity probability value and toxicity label of the drug to be tested;

[0141] The result output module is used to output the toxicity prediction results of the drug to be tested obtained by the toxicity prediction module.

[0142] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the protection scope of the present invention. For ordinary technicians in this field, other different forms of changes or modifications can be made based on the above descriptions and ideas. It is not necessary and impossible to list all the implementation methods here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A toxicity prediction model for predicting non-chemotherapeutic drug-induced agranulocytosis, characterized in that: The toxicity prediction model uses the best molecular descriptor combination of the drug to be tested as input information, which is input into the integrated model of the trained BalancedBaggingClassifier combined with the base classifier gradient boosting decision tree. The integrated model of the trained BalancedBaggingClassifier combined with the base classifier gradient boosting decision tree outputs the toxicity label of the drug to be tested; The value of the toxicity label is 0 or 1; when the toxicity label of the drug to be tested is 0, the drug to be tested does not have the toxicity of inducing agranulocytosis; when the toxicity label is 1, the drug to be tested has the toxicity of inducing agranulocytosis; the drug to be tested is a non-chemotherapeutic drug; The method for obtaining the optimal molecular descriptor combination of the drug to be tested is as follows: obtaining the chemical structural formula of the drug to be tested and performing preprocessing to obtain the chemical structural formula of the drug to be tested after the preprocessing, calculating the molecular descriptor of the drug to be tested, deleting the null value in the molecular descriptor of the drug to be tested, and then performing standardization to obtain the standardized molecular descriptor of the drug to be tested; deleting the molecular descriptors with a variance of 0 in the molecular descriptors of the drug to be tested after the standardization, calculating the correlation between every two molecular descriptors in the molecular descriptors of the drug to be tested after the standardization, deleting one of the two molecular descriptors with a correlation greater than 0.9, and obtaining the optimal molecular descriptor combination of the drug to be tested; The pretreatment is strong base protonation, strong acid deprotonation, removal of inorganic counter ions and addition of explicit hydrogen ions; The trained integrated model of BalancedBaggingClassifier combined with base classifier gradient boosting decision tree is an integrated model of BalancedBaggingClassifier combined with base classifier gradient boosting decision tree trained using a training set containing positive data and negative data; The positive data are non-chemotherapeutic drugs that can cause agranulocytosis, and the negative data are non-chemotherapeutic drugs after agranulocytosis-related drugs are removed from the SIDER database.

2. The toxicity prediction model according to claim 1, characterized in that: The molecular descriptor of the drug to be tested is the Discovery Studio molecular descriptor and / or the RDKit molecular descriptor of the drug to be tested.

3. The toxicity prediction model according to claim 2, characterized in that: The DiscoveryStudio molecular descriptor of the drug to be tested is calculated by Discovery Studio software.

4. The toxicity prediction model according to claim 2, characterized in that: The RDKit molecular descriptor of the drug to be tested is calculated using the rdkit package of Python software.

5. The toxicity prediction model according to claim 1, characterized in that: The standardization process is a Z-score standardization process.

6. The toxicity prediction model according to claim 1, characterized in that: The optimal molecular descriptor combination of the drug to be tested consists of the following 237 molecular descriptors, which are as follows:

7. A method for predicting the toxicity of non-chemotherapeutic drugs induced agranulocytosis, characterized in that: The prediction is performed using the toxicity prediction model described in any one of claims 1 to 6.

8. A toxicity prediction system for predicting agranulocytosis induced by non-chemotherapeutic drugs, characterized in that: It includes a data acquisition module, a toxicity prediction module and a result output module; The data acquisition module is used to obtain the best molecular descriptor combination of the drug to be tested; The toxicity prediction module is a toxicity prediction model according to any one of claims 1 to 6, which uses the optimal molecular descriptor combination of the drug to be tested obtained by the data acquisition module as an input variable to obtain a toxicity prediction result of the drug to be tested; The toxicity prediction result of the drug to be tested is a toxicity label of the drug to be tested, and the value of the toxicity label is 0 or 1; The result output module is used to output the toxicity prediction result of the drug to be tested obtained by the toxicity prediction module; The drug to be tested is a non-chemotherapeutic drug; the method for obtaining the optimal molecular descriptor combination of the drug to be tested is as shown in claim 1.

Images