Methods, systems, equipment, media, and products for constructing and predicting oral drug BDDCS classification models.

Abstract

This invention discloses a method, system, device, medium, and product for constructing and predicting a BDDCS classification model for oral drugs. The method includes: acquiring a BDDCS dataset and dividing it into a training set and an external test set using a random stratified sampling method; calculating the physicochemical descriptors and molecular fingerprints of the drugs to obtain a feature set; filtering the feature data within the training set and selecting core features; constructing a conditional stacking model that includes two decision processes: solubility classification and metabolic degree classification. The conditional stacking model can be either a solubility-first stacking model or a metabolic degree-first stacking model; training the conditional stacking model using the feature-selected training set and using the macro-average F1 score as the optimization metric; and evaluating the classification performance of the trained conditional stacking model through 10-fold cross-validation and an external test set. The conditional stacking framework proposed in this invention can improve classification accuracy.

Description

Technical Field

[0001] This invention relates to the field of biomedical big data processing technology, specifically to a method, system, device, medium, and product for constructing and predicting a BDDCS classification model for oral drugs. Background Technology

[0002] The Biopharmaceutics Drug Disposition Classification System (BDDCS), proposed by Benet et al. in 2005, can be used to predict the primary elimination pathway of drugs, assess the impact of efflux and uptake transporters on the absorption of orally administered drugs, and provide practical guidance for understanding transporter-enzyme interactions, drug-drug interactions, and the food effect after oral administration. The BDDCS uses solubility and extent of metabolism (EoM) as its core classification criteria: solubility is based on FDA standards, specifically whether the approved maximum oral dose (MSD) of a drug can be completely dissolved in 250 mL of aqueous medium with a pH range of 1.0-7.5 at 37°C; EoM is inferred from the fraction of unchanged drug excreted in urine (fe), with fe ≤ 30% indicating extensive metabolism and fe ≥ 70% indicating poor metabolism. Based on these two dimensions, drugs are classified into four categories: Category 1 (high solubility, extensive metabolism), Category 2 (low solubility, extensive metabolism), Category 3 (high solubility, poor metabolism), and Category 4 (low solubility, poor metabolism).

[0003] In recent years, machine learning has become an important method for establishing quantitative structure-property relationships (QSPR), enabling rapid computer simulation prediction of candidate drug properties before compound synthesis. This provides a new technical path for developing models that predict BDDCS categories based on molecular structure. Existing studies have utilized algorithms such as random forest (RF), support vector machine (SVM), and neural networks, combined with molecular descriptors and physicochemical properties, to construct multiple BDDCS prediction models.

[0004] However, existing BDDCS prediction studies have limitations: First, the overall accuracy reported in the literature is generally low (0.55-0.63); second, the datasets have significant class imbalance problems, especially poor prediction performance for Class 4 compounds; and third, most models lack mechanistic interpretability. Summary of the Invention

[0005] The purpose of this invention is to propose a method, system, device, medium and product for constructing and predicting oral drug BDDCS classification models, in order to solve the technical problems of lack of systematic feature optimization, poor model adaptability and insufficient exploitation of the value of BDDCS datasets in existing methods, and to improve the prediction accuracy and model generalization ability of drug treatment classification.

[0006] To achieve the above objectives, the technical method adopted by the present invention is as follows: Firstly, a method for constructing a BDDCS classification model for oral drugs is provided, including the following steps: S1. Stratified sampling: Obtain the original BDDCS dataset, exclude non-oral administration compounds and substances not applicable to molecular descriptor calculations, and divide them into training set and external test set according to their category proportion using a random stratified sampling method. S2. Molecular descriptor calculation: Based on the SMILES representation of compounds, calculate the physicochemical descriptors and molecular fingerprints of orally administered drugs to obtain a feature set; S3. Feature selection: Missing value imputation, standardization, low variance and high correlation filtering are performed in the training set. The Boruta algorithm is used to select core features under 10-fold cross-validation. S4. Model Construction: Construct a conditional stacking model that includes two decision processes: solubility classification and metabolic degree classification. The conditional stacking model can be either a solubility-first stacking model or a metabolic degree-first stacking model. The conditional stacking model decomposes the prediction classification task into a sequential decision process with conditional dependencies. S5. Model Training and Evaluation: The constructed conditional stacked model is trained using the training set after feature selection, and the macro-average F1 score is used as the optimization metric. The classification performance of the trained conditional stacked model is evaluated through 10-fold cross-validation and an external test set.

[0007] Furthermore, it also includes interpreting feature contributions by combining SHAP analysis.

[0008] Further, in step S2, the physicochemical descriptor includes two-dimensional and three-dimensional descriptors calculated based on MOE software and Mordred tools, Abraham solvation parameters obtained from the database, and MSD extracted from the BDDCS dataset; the molecular fingerprint is a fingerprint feature generated based on PaDEL-Descriptor.

[0009] Further, in step S4, the conditional stacking model includes a two-layer ensemble model. The first-layer ensemble model consists of multiple base learners and one meta-learner. The second-layer ensemble model includes two sub-ensemble models, each consisting of multiple base learners and one meta-learner. If the conditional stacking model is a solubility-first stacking model, the first-layer ensemble model is used to classify solubility, and the two sub-ensemble models of the second-layer ensemble model classify the degree of metabolism for drugs classified into high solubility and low solubility categories, respectively. If the conditional stacking model is a metabolism-first stacking model, the first-layer ensemble model is used to classify the degree of metabolism, and the two sub-ensemble models of the second-layer ensemble model classify the solubility for drugs classified into high metabolism and low metabolism categories, respectively. When training the conditional stacking model, the training of the lower-layer ensemble model depends on the prediction results of the upper-layer ensemble model.

[0010] Furthermore, in step S5, the classification performance evaluation metrics of the model include accuracy, macro average F1 score, macro average AUC score, and recall rate for each class.

[0011] Secondly, a BDDCS classification and prediction method for oral drugs is provided, including: The conditional stacking model, constructed using the previously described method for constructing an oral drug BDDCS classification model, is used to perform BDDCS classification on the oral drugs to be classified.

[0012] Thirdly, a BDDCS classification model construction system for oral drugs is provided, including: Stratified sampling module: used to acquire the original BDDCS dataset, exclude non-oral administration compounds and substances not applicable to molecular descriptor calculations, and divide them into training set and external test set according to their category proportion using a random stratified sampling method; Molecular descriptor calculation module: used for compound-based SMILES representation, calculating the physicochemical descriptors and molecular fingerprints of orally administered drugs to obtain feature sets; Feature selection module: used to perform missing value imputation, standardization, low variance and high correlation filtering in the training set, and uses the Boruta algorithm to select core features under 10-fold cross-validation; Model building module: Used to build a conditional stacking model that includes two decision processes: solubility classification and metabolic degree classification. The conditional stacking model is either a solubility-first stacking model or a metabolic degree-first stacking model. The conditional stacking model decomposes the prediction classification task into a sequential decision process with conditional dependencies. Model training and evaluation module: Used to train the constructed conditional stacked model with the training set after feature selection, and use the macro average F1 score as the optimization index; evaluate the classification performance of the trained conditional stacked model through 10-fold cross-validation and external test set.

[0013] Fourthly, an electronic device is provided, comprising: a processor, and a memory coupled to the processor; the memory storing computer programs or instructions; The processor is used to execute the computer program or instructions so that the electronic device performs the oral drug BDDCS classification model construction method or oral drug BDDCS classification prediction method as described above.

[0014] Fifthly, a computer-readable storage medium is provided, on which a computer program is stored, which, when the computer program or instructions are run on a computer, causes the computer to execute the oral drug BDDCS classification model construction method or the oral drug BDDCS classification prediction method as described above.

[0015] In a sixth aspect, a computer program product is provided, including a computer program or instructions that, when run on a computer, cause the computer to execute the oral drug BDDCS classification model construction method or the oral drug BDDCS classification prediction method as described above.

[0016] The present invention has the following beneficial effects: This invention solves the model bias problem caused by sample imbalance by accurately screening the BDDCS dataset and implementing random stratified sampling, thereby improving data quality and representativeness. At the same time, the system integrates physicochemical descriptors and molecular fingerprints, and through multi-stage feature screening, it reduces dimensionality and collinearity while retaining key information, thereby improving the efficiency of characterizing the complex properties of drug molecules.

[0017] This invention constructs a conditional stacking model comprising two decision-making processes: solubility classification and metabolic degree classification. The conditional stacking model can be either a solubility-first stacking model or a metabolic degree-first stacking model. This model decomposes the prediction and classification task into a sequential decision-making process with conditional dependencies, incorporating BDDCS decision logic for collaborative training. This addresses the weak generalization ability of single models and significantly improves classification accuracy. Furthermore, the system introduces SHAP analysis to visualize feature contributions, resolving the low reliability of black-box models and improving the clinical interpretability and practicality of the results. Therefore, it provides efficient and reliable technical support for early drug development and screening, dosage form design optimization, and rational drug use in clinical practice. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a diagram illustrating the architecture of various oral drug BDDCS classification models provided in this invention, including: a four-class baseline model; an ensemble model, including a binary random forest model and a general stacked model; and a conditional stacked model, comprising two architectures: solubility-first and metabolism-first. In the diagram, classification 1, classification 2, classification 3, and classification 4 correspond to class 1 (high solubility, extensive metabolism), class 2 (low solubility, extensive metabolism), class 3 (high solubility, poor metabolism), and class 4 (low solubility, poor metabolism). Figure 2 This is a flowchart of the method for constructing an oral drug BDDCS classification model provided in this embodiment of the invention; Figure 3 These are ROC curves of the conditional stacking models provided in this embodiment of the invention, where (a) and (b) are ROC curves of the solubility-first stacking model and the metabolism-first stacking model, respectively; in the figure, category 1, category 2, category 3 and category 4 correspond to category 1 (high solubility, extensive metabolism), category 2 (low solubility, extensive metabolism), category 3 (high solubility, poor metabolism) and category 4 (low solubility, poor metabolism). Figure 4 This is a graph showing the relative contributions of base learners in the solubility-first stacking model provided in this embodiment of the invention. Figure 5 This is a SHAP bee colony graph showing the feature importance and effect of the solubility node (SOL) random forest model in the solubility priority stacking model provided in this embodiment of the invention; Figure 6 This is a SHAP bee colony graph showing the feature importance and effect of the high-solubility metabolic node (SOL-HIGH) random forest model in the solubility-priority stacking model provided in this embodiment of the invention. Figure 7 This is a SHAP beehive graph showing the feature importance and effect of the low-solubility metabolic node (SOL-LOW) random forest model in the solubility-priority stacking model provided in this embodiment of the invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0021] Existing methods suffer from technical problems such as a lack of systematic feature optimization, poor model adaptability, insufficient interpretability, and failure to fully exploit the value of the BDDCS dataset. To address these issues, this invention proposes a conditional stacking framework, reconstructing the BDDCS prediction task into a process of sequential classification along two dimensions: solubility and metabolic degree. Through multi-level feature selection and conditional ensemble modeling, it improves classification accuracy, interpretability, and preclinical applicability. Performance comparisons were conducted with a four-class classification model, an independent binary classification model based on solubility / metabolic degree, and a general stacking method. The technical solution of this invention is further described below with specific embodiments.

[0022] This invention provides a method for constructing a BDDCS classification model for oral drugs, such as... Figure 2 As shown, it includes the following steps: S1. Stratified Sampling: Obtain the original BDDCS dataset, exclude non-oral administration compounds and substances not applicable to molecular descriptor calculations, and divide them into training set and external test set according to their category proportions using a random stratified sampling method.

[0023] Specifically, based on the updated BDDCS dataset released by Benet et al., data filtering was performed: First, drugs administered via non-oral routes and compounds unsuitable for molecular descriptor calculations needed to be excluded, including simple inorganic salts (such as potassium chloride, ferrous sulfate, and lithium carbonate) and metal-containing compounds (such as vitamin B1). 12 Platinum-based drugs and peptides, as well as compounds classified as BDDCS 0.

[0024] In this embodiment, the final dataset is obtained by random stratified sampling at a ratio of 4:1, which randomly divides the drugs into a training set (858 types) and an external test set (215 types) to ensure that the proportion distribution of BDDCS categories in the training set and the external test set is consistent with that of the original dataset.

[0025] In this embodiment, the training set is used for feature selection, model development, and hyperparameter optimization through cross-validation, while the external test set is used specifically for independent model performance evaluation.

[0026] In the final dataset, there are 423 species (39.4%) in BDDCS class 1, 411 species (38.3%) in BDDCS class 2, 167 species (15.6%) in BDDCS class 3, and 72 species (6.7%) in BDDCS class 4, showing a significant class imbalance.

[0027] S2. Molecular descriptor calculation: Based on the SMILES representation of the compound, calculate the physicochemical descriptor and molecular fingerprint of the oral drug to obtain the feature set.

[0028] In this embodiment, all molecular descriptors were calculated based on the SMILES representation of the compounds. Based on the SMILES representation, two-dimensional and three-dimensional descriptors were calculated using MOE software and Mordred tools, resulting in 2087 computable descriptors. Furthermore, Abraham solvation parameters (A, B, S, E, V) were obtained from the UFZ-LSER database, and MSD data were referenced from the original BDDCS dataset by Benet et al. The predicted pKa values ​​for water solubility (logS), lipophilicity (logP), and acidity / basicity were obtained from the DrugBank database. Based on the predicted pKa values, the distribution coefficient logD under physiological pH conditions was calculated. 7.4 The compounds were then categorized into acidic, basic, neutral, or amphoteric substances based on their ionization behavior, and converted into binary variables using one-heat encoding. Finally, combining the aforementioned physicochemical descriptors, the initial feature pool yielded a total of 2108 variables.

[0029] In this embodiment, the molecular fingerprint was generated using the PaDEL-Descriptor tool, which generated various molecular fingerprints, including basic fingerprints, extended fingerprints, E-state fingerprints, topological graph fingerprints, and MACCS (Molecular Access System) fingerprints.

[0030] S3. Feature selection: Missing value imputation, standardization, low variance and high correlation filtering are performed in the training set. The Boruta algorithm is used to select core features under 10-fold cross-validation.

[0031] This embodiment first constructs multiple different molecular feature combinations, each of which is formed by pairing a preprocessed molecular descriptor with a single type of molecular fingerprint. Then, the optimal combination is selected from these combinations and determined as the core feature set for subsequent model development. Subsequently, the selected core feature set is integrated to form the final candidate feature pool, which serves as the basis for the subsequent feature selection process.

[0032] It should be noted that, to avoid information leakage, all feature selections in this embodiment were performed only within the training set. First, molecular descriptors with missing values ​​(such as maximum dosage and Abraham solvation parameters) were imputed using the median, based solely on the training set data, and all continuous descriptors were Z-score standardized. Subsequently, invalid variables were removed through low-variance filtering (variance < 0.01), reducing the number of descriptors from 2108 to 1884. Finally, to reduce multicollinearity, pairwise correlations between descriptors were calculated, and one feature was removed from each highly correlated feature pair (correlation |r| > 0.95), further optimizing the number of descriptors to 1025. Based on the preliminary model performance, this embodiment selected a 1024-bit extended fingerprint as the molecular fingerprint for subsequent model development, thus forming a combined feature set containing 2049 features.

[0033] In the feature selection process, this embodiment uses the Boruta algorithm, executed synchronously with model training within a 10-fold cross-validation framework, with a maximum of 50 iterations. It should be noted that all feature selections are performed independently from the aforementioned combined feature set containing 2049 features, and the number of features selected at each node is not cumulative. The number of features selected by different model strategies is as follows: the four-class baseline model and the general stacked model each retain 97 features; in the binary random forest model, 162 and 92 features were selected for solubility prediction and metabolic level prediction, respectively; in the solubility-first stacked model, 162 features were selected at the solubility node, and 54 and 63 features were selected for metabolic level prediction in the high-solubility subgroup and low-solubility subgroup, respectively; in the metabolic level-first stacked model, 92 features were selected at the metabolic level node, and subsequently 125 and 28 features were selected for downstream solubility prediction in the high-metabolic-level subgroup and low-metabolic-level subgroup, respectively.

[0034] S4. Model Construction: Construct a conditional stacking model that includes two decision processes: solubility classification and metabolic degree classification. The conditional stacking model can be either a solubility-first stacking model or a metabolic degree-first stacking model. The conditional stacking model decomposes the prediction classification task into a sequential decision process with conditional dependencies.

[0035] This embodiment constructs three types of models (such as...). Figure 1A system comparison was conducted (as shown). First, a four-class baseline model was used, employing five commonly used algorithms: Random Forest (RF), Extreme Gradient Boosting (XGBoost), Category Boosting (CatBoost), K-Nearest Neighbors (KNN), and Support Vector Machine (SVM) for direct four-class prediction. All models underwent feature standardization before training, and class weighting was used to address data imbalance. Next, an ensemble model was constructed to evaluate whether integrating the baseline classifiers could improve prediction performance. This included a general stacked model and a binary random forest model. The general stacked model used RF, XGBoost, CatBoost, SVM, and KNN as base learners, and logistic regression as the meta-learner. The out-of-bag prediction results generated during cross-validation were used as input features for this meta-learner, ultimately generating four-class BDCS prediction results. The binary random forest model predicted solubility and metabolic degree in parallel, then combined the categories according to the BDCS definition, using RF as the base algorithm for both. To integrate the BDCS definition structure into the modeling process, two conditional stacked models—one prioritizing solubility and the other prioritizing metabolic degree—were further developed. The prediction task was decomposed into a multi-stage decision-making process with conditional dependencies, and each stage was trained separately.

[0036] The conditional stacked model comprises two ensemble models. The first ensemble model consists of multiple base learners and one meta-learner. The second ensemble model comprises two sub-ensemble models, each consisting of multiple base learners and one meta-learner. If the conditional stacked model is a solubility-first stacked model, the first ensemble model is used to classify solubility, and the two sub-ensemble models of the second ensemble model classify the metabolic extent of drugs classified into high and low solubility categories, respectively. If the conditional stacked model is a metabolism-first stacked model, the first ensemble model is used to classify metabolic extent, and the two sub-ensemble models of the second ensemble model classify the solubility of drugs classified into high and low metabolism categories, respectively. In this embodiment, the conditional stacked model is illustrated by using RF, XGBoost, and CatBoost as base learners and logistic regression as the meta-learner for all three sub-ensemble models. When training the conditional stacked model, the training of the lower ensemble model depends on the subset of data corresponding to the prediction results of the upper ensemble model.

[0037] S5. Model Training and Evaluation: The constructed conditional stacked model is trained using the training set after feature selection, and the macro-average F1 score is used as the optimization metric. The classification performance of the trained conditional stacked model is evaluated through 10-fold cross-validation and an external test set.

[0038] It should be noted that all models in this embodiment are implemented in the Python 3.9 environment, hyperparameters are tuned through Bayesian optimization, and the macro-average F1 score is used as the main evaluation metric. Training is completed under the 10-fold cross-validation framework to enhance the robustness of model selection and reduce variance.

[0039] After training is completed, this embodiment evaluates all models through cross-validation and independent external test sets. The main performance indicators include accuracy, macro average F1 score, macro average AUC and recall for each class, with special attention paid to a few BDDCS classes.

[0040] Evaluation metrics include accuracy: ; Area under the ROC curve (macroeconomic average AUC): ; Macro-F1 score: ; Recall rates for each category: ; Where: i is the category index; TP is the true positive instance, the number of samples that are actually positive and correctly predicted as positive; FP is the false positive instance, the number of samples that are actually negative but incorrectly predicted as positive; FN is the false negative instance, the number of samples that are actually positive but incorrectly predicted as negative; TN is the true negative instance, the number of samples that are actually negative and correctly predicted as negative.

[0041] Table 1 summarizes the predictive performance of all models evaluated through 10-fold cross-validation and independent external test set validation.

[0042] Table 1. Performance comparison of different models on independent test sets

[0043] The results show that in the four-class baseline model, the test set accuracy ranges from 0.60 (SVM) to 0.68 (RF / XGBoost), the macro-average F1 score ranges from 0.52 (KNN) to 0.62 (RF), and the macro-average AUC ranges from 0.81 (KNN) to 0.85 (CatBoost). The general stacked model shows a moderate improvement over the baseline model, with a test set accuracy of 0.68, a macro-average F1 score of 0.6, and a macro-average AUC of 0.86. The binary random forest model, which predicts solubility and metabolic degree in parallel, achieves a test set accuracy of 0.65, a macro-average F1 score of 0.57, and a macro-average AUC of 0.84.

[0044] The solubility-first stacking model first classifies solubility, then uses local stacking within each subset to predict metabolic extent, achieving the best overall performance. Its test set accuracy is 0.72, the macro-average F1 score is 0.64, and the macro-average AUC is 0.85, outperforming all baseline and unconditional ensemble models. Particularly for minority classes (class 4), the conditional stacking strategy significantly improves prediction stability (ROC-AUC > 0.70). The metabolic extent-first model performs slightly worse than the solubility-first stacking model, but is still superior to the four-class baseline model, general stacking, and binary random forest model. Therefore, in practical implementation, the solubility-first stacking model is the preferred choice.

[0045] Figure 3 The figures are ROC curves for the conditional stacking models, where (a) and (b) are ROC curves for the solubility-first stacking model and the metabolism-first stacking model, respectively.

[0046] S6. Combine SHAP analysis to interpret feature contributions.

[0047] To explain model behavior and identify key drivers of classification decisions, the SHAP additive interpretation method is used to perform interpretability analysis on the best-performing model.

[0048] In this embodiment, for tree-based models, the TreeSHAP algorithm is chosen to calculate accurate feature attribution values. The coefficients of the logistic regression meta-learner are analyzed for the solubility-first stacking model (see...). Figure 4 Random Forest was assigned the highest positive weights at all three levels of nodes, with values ​​of 2.47 (solubility node SOL), 2.60 (high-solubility metabolism node SOL-HIGH), and 1.54 (low-solubility metabolism node SOL-LOW), respectively. XGBoost and CatBoost showed positive weights in all prediction tasks, and their combined weights at the low-solubility metabolism node (XGBoost 1.39, CatBoost 1.41) were close to the weight of Random Forest at the same node (1.54).

[0049] Node-specific SHAP analysis was then performed. Figure 5 , Figure 6 and Figure 7This study aims to elucidate the distinct feature contributions of each base learner within the solubility-first model. For the solubility node, the Random Forest model is primarily driven by h_logD (logD value at pH 7.0 calculated by MOE), with lower h_logD values ​​corresponding to positive SHAP, indicating an increased likelihood of compounds being classified into low-solubility groups (BDDCS classes 2 and 4). Conversely, XGBoost and CatBoost are dominated by MSD, with higher MSD values ​​corresponding to positive SHAP values. The directional trends of MSD and h_logD in the Random Forest are generally opposite to those observed in the Gradient Boosting model.

[0050] In the high-solubility subgroup, metabolic predictions by the random forest were primarily driven by h_logD and logP_db (logP predicted by DrugBank), followed by GCUT_SLOGP_0 (MOE logP GCUT descriptor), logD7.4, and SlogP (MOE logP descriptor). Higher values ​​of these lipophilic descriptors consistently increased the probability of compounds being classified into the low-metabolism category (Category 3). Conversely, in the low-solubility subgroup, metabolic predictions by the random forest were dominated by SlogP and SdssC (a Mordred E state descriptor), and higher values ​​of these features favored classification into the low-metabolism category (Category 4, the positive category). Lower values ​​of the Abraham A descriptor contributed to the positive SHAP value. The boosting models exhibit different feature priorities: XGBoost is primarily dominated by PEOE_VSA_FPNEG (polar-negative van der Waals surface integral), SdssC, and h_logD, while CatBoost heavily relies on h_logD, SdssC, and SlogP. Consistent with the solubility node, the SHAP values ​​produced by these models often show an opposite trend to those of Random Forest across all nodes.

[0051] This invention also provides a BDDCS classification prediction method for oral drugs, comprising: For oral drugs to be classified, extract the core features screened in step S3. Based on the core characteristics of the oral drugs to be classified, a conditional stacking model is constructed using the oral drug BDDCS classification model construction method described above to perform BDDCS classification of the oral drugs.

[0052] This invention also provides a system for constructing an oral drug BDDCS classification model, comprising: Stratified sampling module: used to acquire the original BDDCS dataset, exclude non-oral administration compounds and substances not applicable to molecular descriptor calculations, and divide them into training set and external test set according to their category proportion using a random stratified sampling method; Molecular descriptor calculation module: used for compound-based SMILES representation, calculating the physicochemical descriptors and molecular fingerprints of orally administered drugs to obtain feature sets; Feature selection module: used to perform missing value imputation, standardization, low variance and high correlation filtering in the training set, and uses the Boruta algorithm to select core features under 10-fold cross-validation; Model training module: used to build a conditional stacking model that includes two decision processes: solubility classification and metabolic degree classification. The conditional stacking model is either a solubility-first stacking model or a metabolic degree-first stacking model. The conditional stacking model decomposes the prediction classification task into a sequential decision process with conditional dependencies. Model evaluation and optimization module: The classification performance of the conditional stacked model is evaluated through 10-fold cross-validation and external test set. The macro-average F1 score is used as the main optimization index, and the final conditional stacked model is saved.

[0053] This invention also provides an oral drug BDDCS classification prediction system, which, based on the oral drug BDDCS classification model construction system provided in the foregoing embodiments, further includes: The classification module is used to perform BDDCS classification of oral drugs to be classified using the constructed conditional stacking model.

[0054] It should be understood that the functional unit modules in the various embodiments of the present invention can be concentrated in one processing unit, or each unit module can exist physically separately, or two or more unit modules can be integrated into one unit module, and can be implemented in hardware or software.

[0055] This invention also provides an electronic device, including: a processor, and a memory coupled to the processor; the memory stores computer programs or instructions. The processor is used to execute the computer program or instructions so that the electronic device performs the oral drug BDDCS classification model construction method or oral drug BDDCS classification prediction method as described in the foregoing embodiments.

[0056] The aforementioned electronic devices can be various computing devices such as desktop computers, laptops, handheld computers, and cloud servers. These electronic devices typically include (but are not limited to) a processor and memory.

[0057] The processor can be a central processing unit (CPU), or other types of general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. This processor serves as the control core of the electronic device, connecting various components of the device through various interfaces and lines.

[0058] The memory is used to store computer programs. The processor implements various functions of the electronic device by running or executing the computer programs stored in the memory and by calling the data stored in the memory.

[0059] Memory typically includes a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function, etc.; the data storage area may store data created based on the use of the electronic device, etc. In addition, memory may include high-speed random access memory, or non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart memory cards (SMCs), secure digital cards (SD cards), flash memory cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0060] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when the computer program or instructions are run on a computer, causes the computer to execute the oral drug BDDCS classification model construction method or the oral drug BDDCS classification prediction method as described in the foregoing embodiments.

[0061] This invention also provides a computer program product, including a computer program or instructions, which, when run on a computer, cause the computer to execute the oral drug BDDCS classification model construction method or the oral drug BDDCS classification prediction method as described in the foregoing embodiments.

[0062] The computer program may exist in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include any entity or device capable of carrying the computer program code, such as recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media.

[0063] In summary, this study establishes a conditional stacking framework for predicting BDDCS classification. The proposed conditional stacking model represents a practical and interpretable method for improving BDDCS prediction, thereby enhancing the applicability of computer simulation screening in early drug development.

[0064] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A method for constructing a BDDCS classification model for oral drugs, characterized in that, Includes the following steps: S1. Stratified sampling: Obtain the original BDDCS dataset, exclude non-oral administration compounds and substances not applicable to molecular descriptor calculations, and divide them into training set and external test set according to their category proportion using a random stratified sampling method. S2. Molecular descriptor calculation: Based on the SMILES representation of compounds, calculate the physicochemical descriptors and molecular fingerprints of orally administered drugs to obtain a feature set; S3. Feature selection: Missing value imputation, standardization, low variance and high correlation filtering are performed in the training set. The Boruta algorithm is used to select core features under 10-fold cross-validation. S4. Model Construction: Construct a conditional stacking model that includes two decision processes: solubility classification and metabolic degree classification. The conditional stacking model can be either a solubility-first stacking model or a metabolic degree-first stacking model. The conditional stacking model decomposes the prediction classification task into a sequential decision process with conditional dependencies. The conditional stacking model comprises a two-layer ensemble model. The first-layer ensemble model consists of multiple base learners and one meta-learner. The second-layer ensemble model comprises two sub-ensemble models, each consisting of multiple base learners and one meta-learner. If the conditional stacking model is a solubility-first stacking model, the first-layer ensemble model is used to classify solubility, and the two sub-ensemble models of the second-layer ensemble model classify the degree of metabolism for drugs classified into high solubility and low solubility categories, respectively. If the conditional stacking model is a metabolism-first stacking model, the first-layer ensemble model is used to classify the degree of metabolism, and the two sub-ensemble models of the second-layer ensemble model classify the solubility for drugs classified into high metabolism and low metabolism categories, respectively. When training a conditional stacked model, the training of the lower-level ensemble model depends on the prediction results of the upper-level ensemble model; S5. Model Training and Evaluation: The constructed conditional stacking model is trained using the training set after feature selection, and the macro-average F1 score is used as the optimization index. The classification performance of the trained conditional stacked model was evaluated using 10-fold cross-validation and an external test set.

2. The method for constructing an oral drug BDDCS classification model according to claim 1, characterized in that, Also includes: S6. Combine SHAP analysis to interpret feature contributions.

3. The method for constructing an oral drug BDDCS classification model according to claim 1, characterized in that, In step S2, the physicochemical descriptor includes two-dimensional and three-dimensional descriptors calculated based on MOE software and Mordred tools, Abraham solvation parameters obtained from the database, and MSD extracted from the BDDCS dataset; the molecular fingerprint is a fingerprint feature generated based on PaDEL-Descriptor.

4. The method for constructing an oral drug BDDCS classification model according to claim 1, characterized in that, In step S5, the classification performance evaluation metrics of the model include accuracy, macro average F1 score, macro average AUC score, and recall for each class.

5. A BDDCS classification and prediction method for oral drugs, characterized in that, include: The oral drug BDDCS classification is performed using a conditional stacking model constructed by the method for constructing an oral drug BDDCS classification model as described in any one of claims 1-4.

6. A system for constructing a BDDCS classification model for oral drugs, characterized in that, include: Stratified sampling module: used to acquire the original BDDCS dataset, exclude non-oral administration compounds and substances not applicable to molecular descriptor calculations, and divide them into training set and external test set according to their category proportion using a random stratified sampling method; Molecular descriptor calculation module: used for compound-based SMILES representation, calculating the physicochemical descriptors and molecular fingerprints of orally administered drugs to obtain feature sets; Feature selection module: used to perform missing value imputation, standardization, low variance and high correlation filtering in the training set, and uses the Boruta algorithm to select core features under 10-fold cross-validation; Model building module: Used to build a conditional stacking model that includes two decision processes: solubility classification and metabolic degree classification. The conditional stacking model is either a solubility-first stacking model or a metabolic degree-first stacking model. The conditional stacking model decomposes the prediction classification task into a sequential decision process with conditional dependencies. The conditional stacking model comprises a two-layer ensemble model. The first-layer ensemble model consists of multiple base learners and one meta-learner. The second-layer ensemble model comprises two sub-ensemble models, each consisting of multiple base learners and one meta-learner. If the conditional stacking model is a solubility-first stacking model, the first-layer ensemble model is used to classify solubility, and the two sub-ensemble models of the second-layer ensemble model classify the degree of metabolism for drugs classified into high solubility and low solubility categories, respectively. If the conditional stacking model is a metabolism-first stacking model, the first-layer ensemble model is used to classify the degree of metabolism, and the two sub-ensemble models of the second-layer ensemble model classify the solubility for drugs classified into high metabolism and low metabolism categories, respectively. When training a conditional stacked model, the training of the lower-level ensemble model depends on the prediction results of the upper-level ensemble model; Model training and evaluation module: Used to train the constructed conditional stacked model with the training set after feature selection, and use the macro average F1 score as the optimization index; evaluate the classification performance of the trained conditional stacked model through 10-fold cross-validation and external test set.

7. An electronic device, characterized in that, include: A processor, and a memory coupled to the processor; the memory stores computer programs or instructions. The processor is used to execute the computer program or instructions to cause the electronic device to perform the oral drug BDDCS classification model construction method as described in any one of claims 1-4, or to perform the oral drug BDDCS classification prediction method as described in claim 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program or instructions are run on a computer, the computer performs the oral drug BDDCS classification model construction method as described in any one of claims 1-4, or the oral drug BDDCS classification prediction method as described in claim 5.

9. A computer program product, characterized in that, Includes a computer program or instructions that, when executed on a computer, cause the computer to perform the oral drug BDDCS classification model construction method as described in any one of claims 1-4, or to perform the oral drug BDDCS classification prediction method as described in claim 5.

Images