A model for predicting the risk of medication during lactation and its use

Abstract

The application discloses a model for predicting lactation medication risk, wherein when the model is used to predict the lactation medication risk, human trials are not needed for the to-be-tested drug, the lactation risk can be predicted based on the molecular structure of the to-be-tested drug, and when the lactation medication risk is predicted, the ROC_AUC value of 10-fold cross-validation reaches 0.80, indicating that the model has excellent prediction performance and generalization ability. The model can provide a rapid lactation medication risk evaluation tool for doctors and pharmacists, assist in formulating a safe medication scheme, avoid unnecessary weaning, reduce the adverse reaction risk of breastfed infants, and improve the medication safety level of mothers and infants. Meanwhile, the model can be combined to accurately evaluate the early safety of a new drug, identify and avoid potential lactation medication risks, guide molecular optimization in the new drug research and development process, and reduce the research and development cost and time.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and toxicology, and in particular to a model for predicting the risks of medication use during lactation and its application. Background Technology

[0002] Breast milk, as the best source of nutrition for infants, has unparalleled advantages. It is not only rich in essential nutrients needed for infant growth but also plays a role in immune defense and promoting development. Multiple studies have shown that breastfeeding significantly reduces infant mortality and improves overall infant health by mitigating the risk of various diseases, including respiratory infections, asthma, type 1 and type 2 diabetes, and gastrointestinal disorders such as severe diarrhea, Crohn's disease, and ulcerative colitis. For breastfeeding mothers, breastfeeding aids postpartum recovery, reduces the risk of infection and bleeding, and significantly lowers the incidence of breast cancer, ovarian cancer, cardiovascular disease, and osteoporosis. Therefore, the World Health Organization (WHO) and several pediatric health organizations recommend exclusive breastfeeding for the first six months of life.

[0003] However, the safety of breastfeeding cannot be ignored. Existing research shows that more than half of breastfeeding mothers need to use medication during the breastfeeding period, which means that many infants may be exposed to various drugs through breast milk. Breast milk, as a potential carrier of exogenous substances such as drugs, may pass these substances to the infant, thereby affecting the infant's development, immune system, and endocrine system. Data from the U.S. Poison Control Center from 2001 to 2017 shows that 76,416 calls were received regarding breastfeeding medication, including 2,319 actual drug exposure reports. Of these drug exposure events, 51.4% (1,192 cases) of infants experienced adverse reactions, mainly manifested as drowsiness, irritability, rash, and gastrointestinal symptoms. 466 cases (20.1%) required medical treatment, including 38 cases (8.2%) requiring intensive care and 53 cases (11%) requiring hospitalization. Although most cases had a good prognosis, 222 cases still experienced adverse reactions of varying degrees, including 8 severe adverse reactions and 1 death. Therefore, accurately assessing the transport and exposure levels of drugs in breast milk is of great significance to ensuring the safety of breastfed infants.

[0004] Currently, the main methods for assessing the risk of drug exposure in infants are relative infant dose (RID) and milk / plasma ratio (M / P). RID is calculated by dividing the amount of drug the infant receives from breast milk (mg / kg / day) by the mother's dose (mg / kg / day). This standardized weight method helps determine how much of the mother's medication is ingested by the infant. Generally, a RID <10% is considered relatively safe for breastfed infants. The M / P ratio reflects the ratio of drug concentration in breast milk to drug concentration in maternal plasma. A higher M / P ratio (≥1) indicates a higher concentration of drug in breast milk, while a lower M / P ratio (<1) indicates that only a small amount of drug is transferred to breast milk.

[0005] However, due to ethical restrictions, it is difficult to directly determine the RID or M / P value of a drug through clinical trials. To address this issue, several predictive models based on quantitative structure-activity relationship (QSAR) have been developed to predict the M / P ratio of a drug based on its structure and physicochemical properties, thereby assessing the risk of drug exposure in breastfeeding infants. However, the M / P ratio itself has inherent defects in assessing the risk of drug use during lactation: (1) Uncontrollable factors such as laboratory conditions, lactation period, and measurement methods affect the calculation error of the M / P ratio, thus affecting the reliability of the model; (2) Most models only consider the physicochemical properties of drugs, while ignoring in vivo biotransformation, metabolic kinetics, and the toxicity of the drugs themselves. The passive diffusion mechanism alone cannot fully explain the drug transport process, so the prediction results may deviate from the actual situation of drugs in the body; (3) The M / P ratio itself cannot truly reflect the level of drug exposure in infants. For example, when the maternal blood drug concentration is high, even if the M / P ratio is <1, the blood drug concentration in breast milk may reach a significant level. For drugs with a large volume of distribution, even if the M / P ratio is as high as 8, the actual drug concentration in breast milk may still be low due to the extremely low maternal blood drug concentration. Therefore, relying solely on the prediction model of the M / P ratio to assess the risk of drug exposure in infants has obvious limitations.

[0006] Therefore, there is an urgent need to develop a new method that can overcome the limitations of traditional pharmacokinetic parameters such as the M / P ratio, directly predict the clinical drug use risks during lactation, provide medical staff with an evidence-based risk assessment tool that can directly guide clinical decision-making, and ensure the safety of medication for mothers and infants. Summary of the Invention

[0007] The purpose of this invention is to overcome the above-mentioned shortcomings of the prior art and provide a model for predicting the risks of medication use during lactation and its application.

[0008] The primary objective of this invention is to provide a model for predicting the risks of medication use during lactation.

[0009] A second objective of this invention is to provide the application of the above-mentioned model in predicting the risks of medication use during lactation.

[0010] A third objective of this invention is to provide a method for predicting the risks of medication use during lactation.

[0011] A fourth objective of this invention is to provide a system for predicting the risks of medication use during lactation.

[0012] To achieve the above objectives, the present invention is implemented through the following solution: This invention claims protection for a model for predicting the risk of medication use during lactation, wherein the model takes a combination of molecular descriptors of the drug to be tested as input features and inputs them into a trained machine learning model, and the trained machine learning model outputs a lactation risk label for the drug to be tested. The risk label for medication use during lactation is 0 or 1; when the risk label for medication use during lactation is 0, the drug to be tested is predicted to be a low-risk drug during lactation; when the risk label for medication use during lactation is 1, the drug to be tested is predicted to be a high-risk drug during lactation. The combination of molecular descriptors for the drug under test is the Mordred molecular descriptor, RDKit molecular descriptor, and / or MACCS molecular fingerprint of the drug under test. The trained machine learning model is a machine learning model trained using a training set containing positive and negative data; the machine learning model is a random forest model, gradient boosting decision tree model, adaptive boosting model, lightweight gradient boosting machine model, and / or extreme gradient boosting model; the positive data is high-risk drugs during lactation, and the negative data is low-risk drugs during lactation.

[0013] Preferably, the Mordred molecular descriptor of the drug to be tested is calculated using the Mordred toolkit in Python software; The RDKit molecular descriptor of the drug to be tested was calculated using the rdkit toolkit in Python software; The MACCS molecular fingerprint of the drug to be tested was calculated using the rdkit toolkit in Python software.

[0014] More preferably, the version number of the mordred toolkit is 1.2.0; and the version number of the rdkit toolkit is 2022.09.5.

[0015] Preferably, the molecular descriptor combination of the drug to be tested is standardized before being input into the trained machine learning model.

[0016] More preferably, the standardization process is z-score standardization.

[0017] The z-score standardization process is performed based on the training set data of the trained machine learning model during the training process; the training set data during the training process is stored in the trained machine learning model.

[0018] Preferably, the molecular descriptor combination of the drug to be tested includes 35 molecular descriptors, as shown below: .

[0019] Preferably, the trained machine learning model is a machine learning model trained and evaluated using a training set containing positive and negative data through 10-fold cross-validation.

[0020] Preferably, the machine learning model is a random forest model, a gradient boosting decision tree model, an adaptive boosting model, a lightweight gradient boosting machine model, and / or a limit gradient boosting model.

[0021] More preferably, the machine learning model is a gradient boosting decision tree model.

[0022] More preferably, the hyperparameters of the gradient boosting decision tree model before training are configured as follows: n_estimators=196, learning_rate=0.35000000000000003, criterion="friedman_mse", max_depth=2, max_features=8, subsample=0.59.

[0023] Preferably, the positive data refers to L4 and / or L5 drugs as defined in the "Hale's Medications & Mothers' Milk 2025-2026" guidelines; The negative data refers to L1 and / or L2 level drugs in the "Hale's Medications & Mothers' Milk 2025-2026" guidelines.

[0024] More preferably, the "Hale's Medications & Mothers' Milk 2025-2026" guideline is the 21st edition of the "Hale's Medications & Mothers' Milk 2025-2026" guideline.

[0025] The model shown in this invention can predict the risk of medication use during lactation without conducting human trials on the drug. It can predict the risk of lactation based on the molecular structure of the drug. Furthermore, when predicting the risk of medication use during lactation, the ROC_AUC value of 10-fold cross-validation reached 0.80, indicating that the model has excellent predictive performance and generalization ability.

[0026] This invention also claims protection for the application of any of the above-described models in predicting the risks of medication use during lactation.

[0027] The present invention also claims protection for a method for predicting the risks of medication use during lactation, the method being to make predictions using any of the models described above.

[0028] This invention also claims protection for a prediction system for predicting the risks of medication use during lactation, including a data acquisition module, a prediction module, and a result output module; The data acquisition module is used to acquire the combination of molecular descriptors of the drug to be tested; The prediction module is any of the models described above, and uses the combination of molecular descriptors of the drug to be tested obtained by the data acquisition module as input variables to obtain the risk label of the drug to be tested during lactation. The combination of molecular descriptors for the drug under test is the Mordred molecular descriptor, RDKit molecular descriptor, and / or MACCS molecular fingerprint of the drug under test. The risk label for medication use during lactation is 0 or 1; when the risk label for medication use during lactation is 0, the drug to be tested is predicted to be a low-risk drug during lactation; when the risk label for medication use during lactation is 1, the drug to be tested is predicted to be a high-risk drug during lactation. The result output module is used to output the breastfeeding medication risk label of the drug to be tested obtained by the prediction module.

[0029] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a model for predicting the risk of medication use during lactation. The model uses a combination of molecular descriptors of the drug under test as input features, which are then fed into a trained machine learning model to obtain a lactation risk label for the drug. The lactation risk label can be 0 or 1. When the lactation risk label is 0, the drug is predicted as a low-risk drug for lactation; when the lactation risk label is 1, the drug is predicted as a high-risk drug for lactation. When using this model to predict lactation risk, no human trials are required for the drug; the prediction can be achieved based on the molecular structure of the drug. Furthermore, the ROC_AUC value of 10-fold cross-validation for predicting lactation risk reached 0.80, indicating that the model has excellent predictive performance and generalization ability. This model provides doctors and pharmacists with a rapid tool for assessing the risks of medication use during breastfeeding, assisting in the development of safe medication plans, avoiding unnecessary weaning, reducing the risk of adverse reactions in breastfed infants, and improving the safety of medication use for both mother and child. Furthermore, the model can accurately evaluate the early safety of new drugs, identify and mitigate potential risks of medication use during breastfeeding, guide molecular optimization in the drug development process, and reduce development costs and time. Detailed Implementation

[0030] The present invention will be further described in detail below with reference to specific embodiments. These embodiments are only used to explain the present invention and are not intended to limit the scope of the present invention. Unless otherwise specified, the experimental methods used in the following embodiments are conventional methods; the materials and reagents used are commercially available unless otherwise specified.

[0031] Example 1: Screening of Molecular Descriptors and Molecular Fingerprints I. Experimental Methods 1. Drug Dataset Acquisition The drug dataset was constructed based on the "Hale's Medications & Mothers' Milk 2025-2026" (21st edition) guide, as follows: In the Hale's Medications & Mothers' Milk 2025-2026 (21st edition) guidelines, L1 drugs (the safest, referring to drugs that have been used safely by a large number of breastfeeding mothers and have not been shown to have adverse effects on infants) and L2 drugs (relatively safe, referring to drugs that have been used in studies by some breastfeeding mothers and have not been found to have significant adverse reactions or have minor risks but whose benefits outweigh the risks) are classified as low-risk drugs. L4 drugs (potentially harmful, referring to drugs that have evidence that may have adverse effects on infants or have significant risks, and it is recommended to suspend breastfeeding and closely monitor the infant's condition when using them) and L5 drugs (contraindicated, referring to drugs that have evidence that have significant harm to infants and whose risks significantly outweigh the potential benefits, and which are prohibited for use by breastfeeding women) are classified as high-risk drugs.

[0032] Next, from the low-risk and high-risk drug groups, biological drugs (protein or RNA drugs), mixtures, inorganic compounds, and organometallic compounds were removed. The molecular structures of the remaining drugs were standardized using the "wash" function of the MOE software, while drugs with molecular weights <30 or >1000 were also removed. Duplicate drugs were then removed using InChIKey to obtain a high-quality drug dataset, including high-quality low-risk and high-risk drugs. The high-quality low-risk drug group includes 212 drugs (32 L1 drugs and 180 L2 drugs), and the high-quality high-risk drug group includes 179 drugs (126 L4 drugs and 53 L5 drugs).

[0033] Next, the high-quality drug dataset was randomly divided into a training set (313 drugs) and a test set (78 drugs) in an 8:2 ratio. The number of drugs in the training set and the test set is shown in Table 1, the specific drug information in the training set is shown in Table 2, and the specific drug information in the test set is shown in Table 3.

[0034] Table 1. Detailed information about the training and testing sets.

[0035] Table 2 Specific drug information in the training set

[0036] Table 3 Drug Information for the Test Set

[0037] 2. Calculation of molecular descriptors and molecular fingerprints, and selection of feature subsets. The Mordred package (version 1.2.0) in Python was used to calculate the Mordred descriptors (including 1613-bit two-dimensional features) for each drug in the training and test sets. The rdkit package (version 2022.09.5) in Python was used to calculate the RDKit descriptors (208-bit) for each drug in the training and test sets. The MACCS molecular fingerprints (including 166-bit features) for each drug in the training and test sets were also calculated using the rdkit package (version 2022.09.5). The molecular descriptors and molecular fingerprints for each drug in the training and test sets, as well as their types and bit lengths, are shown in Table 4.

[0038] Table 4. Molecular descriptors and molecular fingerprint types and bit lengths for each drug in the training and testing sets.

[0039] The molecular descriptors and molecular fingerprints (hereinafter collectively referred to as features) of each drug in the training set are divided into different groups based on their single feature (Mordred 2D, RDKit, or MACCS), two feature combinations (Mordred 2D+RDKit, Mordred 2D+MACCS, or RDKit+MACCS), and three feature combinations (Mordred 2D+RDKit+MACCS). The following processing is performed on each group in sequence: (1) Preprocessing of molecular descriptors and molecular fingerprints: Molecular descriptors and molecular fingerprints with missing values ​​in each drug group in the training set were deleted. Then, the StandardScaler function of the sklearn.preprocessing module in Python software was used to standardize the molecular descriptors after deleting missing values ​​according to the z-score standardization method. The variance of each molecular descriptor and molecular fingerprint was calculated, and molecular descriptors and molecular fingerprints with a variance of 0 were deleted. The Pearson correlation coefficient between each pair of molecular descriptors and molecular fingerprints was calculated. For two molecular descriptors or molecular fingerprints with a Pearson correlation coefficient > 0.9, one of the molecular descriptors or molecular fingerprints was deleted to obtain the preprocessed molecular descriptors and molecular fingerprints of each drug group in the training set.

[0040] (2) Feature filtering: Four feature selection strategies are set, including: 1) Mutual Information Method (MI): Calculate the mutual information score between each feature and the risk of medication use during lactation, and retain features with a score > 0; 2) Embedded tree-based feature selection (ETB): The random forest model is trained with features as input and the risk of medication use during lactation as the prediction target, and features with importance scores ≥0.0048 are selected during the training process. 3) Recursive feature elimination with cross-validation (RFECV): Combines a random forest classifier to iteratively remove the least important features and introduces a cross-validation mechanism to prevent overfitting. The features that are retained are the RFECV features. 4) Genetic algorithm (GA): Set the population size to 50, the number of iterations to 40, the crossover rate to 0.5, and the mutation rate to 0.2. Combine the genetic algorithm to perform feature selection and obtain GA features.

[0041] Based on the molecular descriptors and molecular fingerprints of each drug group in the preprocessed training set obtained in step (1), the four feature selection strategies shown above are used sequentially for feature screening of each group to obtain the features of each drug group in the training set after screening according to each feature selection strategy.

[0042] 3. Combine the model to select the best feature subset Using the features selected by each drug group in the training set obtained in step 2 according to each feature selection strategy as input, and based on the random forest model (RF), with the risk of lactation use of each drug in the training set as the prediction target, 10-fold cross-validation is performed to train the random forest model, and the performance of the random forest model is recorded, including the area under the ROC curve (ROC_AUC), accuracy (ACC), sensitivity (SEN), specificity (SPE), and Matthews correlation coefficient (MCC); where ACC, SEN, SPE, and MCC are calculated according to formulas I to IV respectively; Formula I: ACC = (TP + TN) / (TP + TN + FP + FN); Formula II: SEN = TP / (TP + FN); Formula III: SPE = TN / (TN + FP); Formula IV: MCC = (TP × TN - FP × FN) / {(FP + TN)(FP + TP)(FN + TN)(FN + TP)} 1 / 2 ; True positive (TP) represents positive data correctly identified in the training set, true negative (TN) represents negative data correctly identified in the training set, false positive (FP) represents drugs that are actually low-risk but are incorrectly predicted as high-risk by the model, and false negative (FN) represents drugs that are actually high-risk but are incorrectly predicted as low-risk by the model.

[0043] II. Experimental Results The performance results of the random forest model trained using features selected from each drug group in the training set according to each feature selection strategy are shown in Table 5.

[0044] Table 5 Performance results of the random forest model

[0045] The results showed that when using the MCC index to evaluate the features selected according to the feature selection strategies for each drug and group in the training set to train the random forest model, the 43 feature subsets obtained after screening using the ETB strategy with two feature combinations (Mordred+RDKit) achieved an MCC of 0.46 and an ROC_AUC of 0.77 when combined with the random forest model to predict the risk of medication use during lactation. The 35 feature subsets obtained after screening using the ETB strategy with three feature combinations (Mordred 2D+RDKit+MACCS) also achieved an MCC of 0.46 and an ROC_AUC of 0.77 when combined with the random forest model to predict the risk of medication use during lactation.

[0046] Both combinations achieved a 98% reduction in dimensionality compared to the original number of molecular descriptors and molecular fingerprint types (1987) in the training set, and both demonstrated superior performance in predicting the risk of medication use during lactation.

[0047] The 43 feature subsets obtained after filtering two feature combinations (Mordred+RDKit) according to the ETB strategy are shown in Table 6; the 35 feature subsets obtained after filtering three feature combinations (Mordred 2D+RDKit+MACCS) according to the ETB strategy are shown in Table 7.

[0048] Table 6. Feature subset containing 43 features obtained after filtering by Mordred+RDKit according to the ETB strategy.

[0049] Table 7. Feature subsets containing 35 features obtained after filtering Mordred 2D+RDKit+MACCS using the ETB strategy.

[0050] Example 2: Establishment of a model for predicting the risk of medication use during lactation I. Establishment of a model for predicting the risk of medication use during lactation 1. Experimental Methods Five different ensemble learning models were set up, including: Random Forest (RF), Gradient Boosting Decision Trees (GBDT), Adaptive Boosting (AdaBoost), Light Gradient Boosting Machine (LightGBM), and eXtreme Gradient Boosting (XGBoost).

[0051] Each ensemble learning model is implemented based on an open-source Python framework: RF, AdaBoost, and GBDT are implemented using RandomForestClassifier, AdaBoostClassifier, and GradientBoostingClassifier respectively through scikit-learn (version 1.5.2); LightGBM (version 3.3.5) and XGBoost (version 1.6.1) are implemented using dedicated packages LGBMClassifier and XGBClassifier respectively.

[0052] During the training process, the hyperparameters of each ensemble learning model were tuned using a Bayesian optimization algorithm based on a tree-structured Parzen Estimator (TPE). The tuning process was implemented using the Hyperopt package in Python (version 0.2.7), with the goal of maximizing the performance of the Matthews correlation coefficient (MCC) of 10-fold cross-validation. For hyperparameter optimization, specifically including: for RF, optimizing n_estimators (number of trees), max_depth (maximum depth of trees), max_features (maximum number of features to consider during splitting), and criterion (split quality evaluation criterion, such as Gini or Entropy); For GBDT, optimize n_estimators (number of trees), learning_rate (learning rate, which controls the contribution weight of each tree), criterion (splitting quality evaluation criterion, such as friedman_mse or squared_error), max_depth (maximum depth of the tree), max_features (maximum number of features to consider when splitting), and subsample (subsample ratio). For AdaBoost, optimize n_estimators (number of base learners), max_depth (maximum depth of the tree), learning_rate (learning rate), and algorithm (learning algorithm, such as SAMME or SAMME.R). For LightGBM, optimize boosting_type, n_estimators, learning_rate, max_depth, num_leaves, colsample_bytree, subsample, reg_alpha, reg_lambda, and min_child_samples. For XGBoost, optimize n_estimators (number of trees), learning_rate (learning rate), max_depth (maximum depth of the tree), subsample (subsample ratio), colsample_bytree (column sampling ratio), booster (booster type), colsample_bynode (node ​​column sampling ratio), gamma (minimum loss reduction), reg_lambda (L2 regularization coefficient), and min_child_weight (minimum sample weight of leaf nodes).

[0053] Next, 43 features (Mordred+RDKit_ETB_43) of each drug in the training and test sets of Example 1, as shown in Table 6, were obtained. These 43 features were used as input features for each ensemble learning model (the 43 features were z-score standardized before input). The lactation risk of each drug in the training set was used as the prediction target for each ensemble learning model. Ten-fold cross-validation was performed on the training set, combining the actual lactation risk of each drug in the training set, to obtain the trained ensemble learning model. The model was then tested using the test set, and the ROC_AUC, ACC, SEN, SPE, and MCC values ​​of each ensemble learning model during the ten-fold cross-validation and testing process were recorded. During the ten-fold cross-validation, the Hyperopt package (version 0.2.7) in Python was used to automatically optimize the hyperparameters of each ensemble learning model. The output of each ensemble learning model is a risk probability value, which is [0, 1]. If the risk probability value is ≥0.5, the risk label for medication use during lactation is 1, indicating a positive result (high-risk medication during lactation); if the risk probability value is <0.5, the risk label for medication use during lactation is 0, indicating a negative result (low-risk medication during lactation).

[0054] Next, 35 features (Mordred+RDKit+MACCS_ETB_35) of each drug in the training and test sets of Example 1, as shown in Table 7, were obtained. These 35 features were used as input features for each ensemble learning model (the 35 features were z-score standardized before input). The lactation risk of each drug in the training set was used as the prediction target for each ensemble learning model. Combined with the actual lactation risk of each drug in the training set, 10-fold cross-validation was performed on the training set to obtain the trained ensemble learning model. The model was then tested using the test set, and the ROC_AUC, ACC, SEN, SPE, and MCC values ​​of each ensemble learning model during 10-fold cross-validation and testing were recorded. During 10-fold cross-validation, the Hyperopt package (version 0.2.7) in Python was used to automatically optimize the hyperparameters of each ensemble learning model. The output of each ensemble learning model is a risk probability value, which is [0, 1]. If the risk probability value is ≥0.5, the risk label for medication use during lactation is 1, indicating a positive result (high-risk medication during lactation); if the risk probability value is <0.5, the risk label for medication use during lactation is 0, indicating a negative result (low-risk medication during lactation).

[0055] 2. Experimental Results The results of five ensemble learning models trained and tested using drugs from the training and testing sets are shown in Table 8.

[0056] Table 8 Results of five ensemble learning models trained and tested using drugs from training and testing sets.

[0057] The results show that when using Gradient Boosting Decision Tree (GBDT) in combination with the 35 features shown in Table 7 of Example 1 to predict the risk of medication use during lactation, it exhibits excellent results in both 10-fold cross-validation on the training set and independent testing on the test set. In the 10-fold cross-validation on the training set, the ROC_AUC reached 0.80 and the MCC reached 0.52. In the test set, the ROC_AUC reached 0.79 and the accuracy reached 0.73, the sensitivity reached 0.76, the specificity reached 0.71, and the MCC reached 0.46.

[0058] Therefore, using the training set shown in Example 1, a gradient boosting decision tree (GBDT) is established by combining the 35 features of the drugs in the training set as shown in Table 7 of Example 1 with the 35 features of the drugs in the training set as input to the GBDT (the 35 features are input after z-score standardization). The risk of medication use during lactation of the drugs in the training set is used as the prediction target for training and testing with the test set. The GBDT model for predicting the risk of medication use during lactation stores the training set data during training (the levels of the 35 features of each drug in Table 7 of the training set are used as a benchmark, and the 35 features of the drug to be tested are z-score standardized). The hyperparameters of the GBDT model for predicting the risk of medication use during lactation are configured as follows: n_estimators=196, learning_rate=0.35000000000000003, criterion="friedman_mse", max_depth=2, max_features=8, subsample=0.59.

[0059] II. Definition of the application domain of the GBDT model for predicting medication risks during lactation 1. Experimental Methods Based on the 35 features of each drug in the training set as shown in Table 7 of Example 1, the standardized Euclidean distance of each drug in the training set was calculated using EuclideanApplicability Domain 1.0 software, and the boundary region was constructed using the standardized Euclidean distance of each drug in the training set to obtain the application domain of the lactation medication risk prediction model (the standardized Euclidean distance threshold is 1.0).

[0060] Next, based on the 35 features of each drug in the test set as shown in Table 7 of Example 1, the standardized Euclidean distance of each drug in the test set was calculated using Euclidean Applicability Domain 1.0 software. When the standardized Euclidean distance of the drugs in the test set is ≤1.0, it indicates that the drug is within the application domain of the lactation medication risk prediction model, and the model prediction results are reliable. When the standardized Euclidean distance of the drugs in the test set is >1.0, it indicates that the compound is outside the application domain of the lactation medication risk prediction model, and the prediction results need to be interpreted with caution.

[0061] 2. Experimental Results The application domain results evaluation showed that among the 78 drugs in the test set, 77 drugs (98.7%) were within the defined application domain, and only one drug had a standardized distance score of more than 1.0, which was outside the application domain.

[0062] This indicates that the structural features of the vast majority of compounds in the test set fall within the chemical space defined by the model, and the GBDT model used to predict the risk of medication use during lactation has a high predictive reliability for 98.7% of the drugs in the test set.

[0063] Example 3: A prediction system for predicting medication risks during lactation. A prediction system for predicting the risks of medication use during lactation includes a data acquisition module, a prediction module, and a result output module; The data acquisition module is used to acquire the molecular descriptor of the drug to be tested, and to perform z-score normalization based on the molecular descriptor of the training set described in Example 2 (the GBDT model used to predict the risk of medication use during lactation in Example 2 stores the training set data during the training process), and then obtain the normalized molecular descriptor combination of the drug to be tested; wherein the normalized molecular descriptor combination contains the 35 features shown in Table 7 of Example 2. The prediction module is the GBDT model shown in Example 2 for predicting the risk of medication use during lactation. It uses the standardized molecular descriptor combination of the drug to be tested obtained by the data acquisition module as the input variable to obtain the risk label of the drug to be tested during lactation. The GBDT model used to predict the risk of medication use during lactation outputs a risk probability value, which is [0, 1]. If the risk probability value is ≥0.5, the risk label of the drug to be tested during lactation is 1, and it is judged as positive (high-risk drug during lactation); if the risk probability value is <0.5, the risk label of the drug to be tested during lactation is 0, and it is judged as negative (low-risk drug during lactation).

[0064] The result output module is used to output the breastfeeding medication risk label of the drug to be tested obtained by the prediction module.

[0065] Example 4: Application of a predictive system for predicting medication risks during lactation I. Risk Prediction for Amoxicillin Use During Lactation 1. Experimental Methods The PubChem number for amoxicillin was determined to be 33613. The SMILES code for amoxicillin was obtained from the PubChem website. Input the SMILES code of amoxicillin into the MOE software and use the "wash" function of the MOE software to preprocess the chemical structure of amoxicillin. Next, the RDKit molecular descriptor, MACCS molecular fingerprint, and Mordred molecular descriptor of amoxicillin were calculated using the rdkit package and Mordred in Python software. Using the StandardScaler function of the sklearn.preprocessing module in Python software, the molecular descriptor of amoxicillin was z-score standardized based on the molecular descriptor of the training set described in Example 2 (the GBDT model used to predict the risk of medication use during lactation in Example 2 contains training set data from the training process). Next, the standardized molecular descriptor combination of amoxicillin (35 features shown in Table 7 of Example 1) was obtained and input into the prediction system shown in Example 3. The risk label and risk probability value of lactation medication output by the prediction system shown in Example 3 were recorded. Simultaneously, the standardized Euclidean distance of the amoxicillin drug molecule descriptor was calculated using Euclidean Applicability Domain 1.0 software.

[0066] 2. Experimental Results When the prediction system shown in Example 3 was used to predict the risk of using the antibacterial drug amoxicillin, the output risk label for use during lactation was 0, and the corresponding risk probability value was 0.07; indicating that the risk of using amoxicillin during lactation is very low.

[0067] The LRC of amoxicillin in the "Hale's Medications & Mothers' Milk 2025-2026" (21st edition) is L1 (the safest), which is consistent with the prediction results of the prediction system shown in Example 3; at the same time, the normalized Euclidean distance of the amoxicillin molecular descriptor is 0.09 < 1.0, indicating that its prediction results are reliable.

[0068] II. Risk Prediction for the Use of Amiodarone, an Antiarrhythmic Drug, During Lactation 1. Experimental Methods Following the method described in "I. Prediction of the Risk of Amoxicillin Use During Lactation", amoxicillin was replaced with amiodarone (PubChem No. 2157). The remaining steps were exactly the same. The prediction was performed using the prediction system shown in Example 3, and the output of the lactation risk label, risk probability value and standardized Euclidean distance were recorded.

[0069] 2. Experimental Results When using the prediction system shown in Example 3 to predict the risk of amiodarone, an antiarrhythmic drug, the output risk label for medication use during lactation was 1, with a corresponding risk probability value of 0.96; indicating that the risk of using amiodarone during lactation is very high.

[0070] Amiodarone has an LRC of L5 (disabled) in the Hale's Medications & Mothers' Milk 2025-2026 (21st edition) guide, which is consistent with the prediction results of the prediction system shown in Example 3; at the same time, the normalized Euclidean distance of the amiodarone molecule descriptor is 0.13 < 1.0, indicating that its prediction results are reliable.

[0071] III. Risk Prediction for Breastfeeding Use of Remibrutinib, a Newly Marketed Drug with Unknown Risks During Lactation 1. Experimental Methods Following the method described in "I. Prediction of the Risk of Amoxicillin Use During Lactation", amoxicillin was replaced with remibutinib (PubChem No. 118107483). The remaining steps were exactly the same. The prediction was performed using the prediction system shown in Example 3, and the output of the lactation risk label, risk probability value and standardized Euclidean distance were recorded.

[0072] 2. Experimental Results When using the prediction system shown in Example 3 to predict the risk of remibrutinib use during lactation, the output risk label for lactation use was 1, with a corresponding risk probability value of 1.00; this indicates that the risk of remibrutinib use during lactation is very high; at the same time, the standardized Euclidean distance of the remibrutinib molecular descriptor is 0.18 < 1.0, indicating that its prediction results are reliable.

[0073] Remibtinib is a drug used to treat chronic spontaneous urticaria (SSO). It is used to treat adult patients with SSO who still have symptoms after using H1 antihistamines. Some of these patients are breastfeeding women. Therefore, for breastfeeding patients with severe SSO after childbirth who need to use remibtinib, it is recommended to suspend breastfeeding or choose a safer alternative treatment.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, other variations or modifications can be made based on the above description and ideas, and it is neither necessary nor possible to exhaustively describe all implementation methods here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A model for predicting the risk of medication during lactation, characterized in that, The model uses the combination of molecular descriptors of the drug under test as input features, which are then input into a trained machine learning model. The trained machine learning model outputs the risk label of the drug under test for use during lactation. The risk label for medication use during lactation is 0 or 1; when the risk label for medication use during lactation is 0, the drug to be tested is predicted to be a low-risk drug during lactation. When the risk label for medication use during lactation is 1, the drug to be tested is predicted to be a high-risk drug during lactation. The combination of molecular descriptors for the drug under test is the Mordred molecular descriptor, RDKit molecular descriptor, and / or MACCS molecular fingerprint of the drug under test. The trained machine learning model is a machine learning model trained using a training set containing positive and negative data; the machine learning model is a random forest model, gradient boosting decision tree model, adaptive boosting model, lightweight gradient boosting machine model, and / or extreme gradient boosting model; the positive data is high-risk drugs during lactation, and the negative data is low-risk drugs during lactation.

2. The model according to claim 1, characterized in that, The Mordred molecular descriptor of the drug to be tested was calculated using the Mordred toolkit in Python software; The RDKit molecular descriptor of the drug to be tested was calculated using the rdkit toolkit in Python software; The MACCS molecular fingerprint of the drug to be tested was calculated using the rdkit toolkit in Python software.

3. The model according to claim 1, characterized in that, The molecular descriptor combination of the drug under test is standardized before being input into the trained machine learning model.

4. The model according to claim 1, characterized in that, The molecular descriptor combination of the drug to be tested includes 35 molecular descriptors, which are shown below: 。 5. The model according to claim 1, characterized in that, The machine learning model is a gradient boosting decision tree model.

6. The model according to claim 5, characterized in that, The hyperparameters of the gradient boosting decision tree model before training are configured as follows: n_estimators=196, learning_rate=0.35000000000000003, criterion="friedman_mse", max_depth=2, max_features=8, subsample=0.

59.

7. The model according to claim 1, characterized in that, The positive data refers to L4 and / or L5 drugs as defined in the "Hale's Medications & Mothers' Milk 2025-2026" guidelines. The negative data refers to L1 and / or L2 level drugs in the "Hale's Medications & Mothers' Milk 2025-2026" guidelines.

8. The application of the model described in any one of claims 1 to 7 in predicting the risks of medication use during lactation.

9. A method for predicting the risks of medication use during lactation, characterized in that, The method involves making predictions using the model described in any one of claims 1 to 7.

10. A prediction system for predicting the risks of medication use during lactation, characterized in that, It includes a data acquisition module, a prediction module, and a result output module; The data acquisition module is used to acquire the combination of molecular descriptors of the drug to be tested; The prediction module is the model described in any one of claims 1 to 7, which uses the combination of molecular descriptors of the drug to be tested obtained by the data acquisition module as input variables to obtain the risk label of the drug to be tested during lactation. The combination of molecular descriptors for the drug under test is the Mordred molecular descriptor, RDKit molecular descriptor, and / or MACCS molecular fingerprint of the drug under test. The risk label for medication use during lactation is 0 or 1; when the risk label for medication use during lactation is 0, the drug to be tested is predicted to be a low-risk drug during lactation. When the risk label for medication use during lactation is 1, the drug to be tested is predicted to be a high-risk drug during lactation. The result output module is used to output the breastfeeding medication risk label of the drug to be tested obtained by the prediction module.

Images