Rapid screening method for cytochrome p450 enzyme inhibitors based on qsar model

Abstract

The application is a QSAR model-based cytochrome P450 enzyme inhibitor rapid screening method, which can quickly predict whether a compound has an inhibitory effect on cytochrome P450 enzyme. Cytochrome P450 enzyme is a hemisulfate protein, and the application takes cytochrome P450 enzyme as the research object, and uses machine learning to predict the activity results of the inhibitory effect of five different subtypes of cytochrome P450 enzyme. Compared with the random prediction baseline, the classification performance of the QSAR model in each subtype is significantly improved, and the PR-AUC is increased by about 80%. The application takes support vector machine and random forest model as the basis, and provides a new QSAR model construction method for rapidly screening compounds that inhibit cytochrome P450 enzyme.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of environmental toxicology and machine learning, specifically involving the application of machine learning methods such as support vector machines and random forests to rapidly screen compounds that inhibit cytochrome P450 enzymes, with the inhibition of cytochrome P450 enzymes as the toxicity endpoint. Background Technology

[0002] Biotoxicity refers to the adverse biological reactions triggered by exposure to certain toxic substances (such as heavy metals, nanomaterials, pesticides, etc.). These toxic reactions can interfere with the normal physiological functions of organisms, and their toxic effects can manifest as molecular-level oxidative stress damage, metabolic pathway disorders, apoptosis or necrosis, abnormal immune system responses, etc., ultimately leading to organ dysfunction, chronic pathological changes, and even individual death. Currently, with the widespread application of novel compounds in various fields, research on biotoxicity has attracted high attention from multiple disciplines. In the field of materials science, researchers focus on the toxicity mechanisms induced by novel materials (especially nanomaterials), striving to provide a theoretical basis for setting biosafety thresholds and environmentally friendly design of materials by exploring the toxicity caused by exposure to novel materials. In the field of pharmacy, because drugs may have toxic side effects while exerting therapeutic effects, researchers focus on the mechanism research of the balance between drug efficacy and toxicity to achieve the clinical translation of efficacy-toxicity balance. In the environmental field, researchers use methods such as pollutant exposure experiments to assess the environmental and eco-health risk effects of pollutants, deeply analyze their ecotoxicological mechanisms, and provide theoretical support for pollution control and environmental regulation. Therefore, a systematic study of biotoxicity not only helps to reveal the mechanisms of action of toxic substances, but also provides a scientific basis for risk assessment and regulatory policy formulation in related fields. As new pollutants are constantly being discovered, their toxic mechanisms remain unexplored; therefore, in-depth research on biotoxicity has become a key link in promoting multidisciplinary integration and technological innovation.

[0003] Cytochrome P450 (CYP) enzymes are a class of heme-containing monooxygenases widely found in various organisms, including humans, animals, plants, and microorganisms. They are characterized by diverse functions and high evolutionary conservation. These enzymes catalyze various oxidation reactions during molecular oxygen reduction cleavage, a typical reaction being: R–H + O₂ + NAD(P) ⇌ H⁺ + H⁺. + →R–OH+NAD(P) ++H₂O can convert inactive C–H bonds into more polar hydroxyl groups. Besides hydroxylation, P450 enzymes can also catalyze various transformation reactions such as double bond epoxidation, N / S oxidation, dealkylation, and C–O / C–C bond formation. CYP enzymes are widely involved in the synthesis and metabolism of endogenous substances such as cholesterol and hormones, and play an important role in the transformation and detoxification of exogenous substances (such as drugs, environmental pollutants, and natural toxins). Therefore, P450 enzymes hold an important position in environmental toxicology and biotransformation research. This study selected data from the PubChem database exploring the inhibitory effects of organic compounds on human Cytochrome P450 enzymes (CYPs) as a dataset. This dataset tested five human CYP450 enzyme subtypes (CYP1A2, CYP2C9, CYP2C19, CYP2D6, and CYP3A4) to assess the potential interference of compounds with different structures on the metabolic system, thereby providing theoretical support and a data foundation for compound toxicity prediction and metabolic safety assessment.

[0004] Quantitative structure-activity relationship (QSAR) models are a class of predictive models that establish a quantitative relationship between the structural features of a compound and its biological activity using mathematical and statistical methods. In cheminformatics, the chemical structural information of a compound can first be converted into textual encoding (such as the string "SMILES"). Then, the descriptor of the corresponding compound is calculated using the "SMILES" string, thereby transforming the compound's structural information into numerical features that can be input into the model.

[0005] Previous studies have attempted to use machine learning methods to predict the inhibitory activity of compounds on cytochrome P450 enzymes. However, existing models are relatively limited in type, with a trade-off between stability and generalization ability. Furthermore, the feature dimensions used are too high, with some models introducing more than 1900 molecular descriptors, which may cause information redundancy, increase computational complexity, and adversely affect model performance. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the technical problem this invention aims to solve is to provide a rapid screening method for cytochrome P450 enzyme inhibitors based on the QSAR model. This method can efficiently and stably predict whether a compound has an inhibitory effect on cytochrome P450 enzymes with a small number of descriptors, thereby quickly determining whether the compound has biotoxicity.

[0007] The technical solution adopted by the present invention to solve the aforementioned technical problem is as follows:

[0008] This invention provides a rapid screening method for cytochrome P450 enzyme inhibitors based on the QSAR model, comprising the following steps:

[0009] Step 1: Data Collection and Preprocessing

[0010] The data with Toxity ID 1851 in the PubChem database was used as the compound dataset. This compound dataset contains compound data and biotoxicity panel IDs. The biotoxicity panel IDs are numbered 1-5, corresponding to five human CYP450 enzyme subtypes. Data with the value "Inconclusion" for "PUBCHEM_ACTIVITY_OUTCOME" were removed, as were data of metal compounds and compounds containing free unpaired electrons.

[0011] Step 2: Descriptor Calculation and Processing

[0012] Using Python version 3.12, the padelpy package was used to calculate KlekotaRoth fingerprints, MACCS fingerprints, and PubChem fingerprints, and preliminary descriptor screening was performed. The preliminary screening steps and criteria were as follows:

[0013] 1) Delete descriptors that have more than 10% unassigned values ​​across all samples;

[0014] 2) If a descriptor has the same value in more than 95% of the samples, it is considered to lack discriminative power and is removed.

[0015] 3) Calculate the variance of each descriptor across all samples. If the variance of a descriptor is less than 0.1, it is discarded.

[0016] 4) Based on the Pearson correlation coefficient, evaluate the linear correlation between any two descriptors. If the correlation between two descriptors is greater than 0.9, retain one representative descriptor and delete redundant terms.

[0017] The third step is to build a screening model.

[0018] For each subtype, feature importance is calculated for the initially screened descriptors based on the RF model to obtain the importance ranking of the descriptors. Then, descriptors are progressively selected for input into the model according to feature importance from high to low, with the number of input descriptors gradually increasing by one each time. The different numbers of input descriptors constitute multiple sub-models. The overall architecture of the sub-models adopts the SVM model and / or the RF model. The performance of each sub-model and the number of input features are comprehensively compared. Finally, the sub-model with fewer input descriptors while ensuring prediction performance is selected as the optimal model and incorporated into the final modeling system. The optimal model selected is different for different subtypes. Thus, the optimal model for each of the five subtypes and the important descriptors selected by the corresponding optimal model are obtained.

[0019] The important descriptors were combined with the compound dataset according to the SMILES identifier string to obtain five sub-datasets for the five human CYP450 enzyme subtypes.

[0020] For each subset of data, the optimal models for the five subtypes are trained using grid search and random undersampling, resulting in the five optimal models after training.

[0021] The screening model includes a discrimination module and five optimal models after training. The discrimination module identifies the toxicity endpoint marker of the input inhibitor and enters the corresponding optimal model according to the toxicity endpoint marker to obtain the screening result of whether the compound has biotoxicity.

[0022] Furthermore, during training, the contribution of each important descriptor to the prediction results of the optimal model is evaluated by combining the permutation importance assessment. The top-k structural features with the highest importance are then selected and ranked. Subsequently, based on the substructure correspondence file provided by the official fingerprint type, the specific structural fragments represented by each high-importance descriptor are analyzed. Compounds with bridging bicyclic structures, trisubstituted central atom structures, five-membered ring structures, aromatic structures, and macromolecular structures containing more than 16 carbons are more likely to inhibit cytochrome P450 enzymes.

[0023] Furthermore, four metrics—AUC-ROC, PR-AUC, Log Loss, and Brier Score—were used as standards for evaluating model performance. AUC-ROC measures the model's ability to distinguish between positive and negative classes, PR-AUC measures the model's performance in positive class recognition, Log Loss measures the log loss of the predicted probability from the true label, and Brier Score measures the mean squared error between the predicted probability and the actual label.

[0024] Furthermore, the performance of the model is judged by the comprehensive performance of the four indicators.

[0025] Furthermore, the optimal models for the five subtypes are as follows: CYP2C19 selects an SVM model with 60 descriptors; CYP2D6 selects an RF model with 60 descriptors; CYP3A4 selects an SVM model with 60 descriptors; CYP1A2 selects an RF model with 50 descriptors; and CYP2C9 selects an SVM model with 70 descriptors.

[0026] The present invention also protects a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, can implement the steps of the method.

[0027] Compared with the prior art, the beneficial effects of the present invention are:

[0028] (1) The method of the present invention is for rapid screening of the inhibitory effects of compounds on five subtypes of cytochrome P450. It can achieve rapid screening of the activity prediction of the inhibitory effects of compounds on cytochrome P450 enzymes. In the examples, the model prediction accuracy is good and the average AUC-ROC value of the test set is 0.85.

[0029] (2) This invention creatively uses interpretable fingerprint descriptors to predict the biotoxicity of compounds and explores the influence of compound structure on biotoxicity. The fingerprint descriptors used have clear structural correspondences, with each descriptor representing a specific molecular substructure, thus enabling the model to have structural interpretability. Through systematic analysis of the relationship between key substructural features and cytochrome P450 enzyme inhibitory activity, the identification and summarization of inhibitory-related structural fragments can be achieved. Based on this, in the absence of biological experimental data, the molecular structural information of compounds can be used to quickly and preliminarily determine whether they have the potential to inhibit CYP450 enzymes, thereby significantly improving compound screening efficiency and reducing experimental costs.

[0030] (5) This invention employs two representative machine learning algorithms, Support Vector Machine and Random Forest, and selects molecular fingerprint descriptors with clear structural information to construct screening models for the major CYP450 isoenzymes (1A2, 2C9, 2C19, 2D6, and 3A4). The aim is to improve the interpretability of the model while ensuring good predictive performance, thereby achieving efficient screening of CYP450 inhibitors and uncovering key structural features to provide a theoretical basis for subsequent drug design and toxicity risk assessment.

[0031] The average number of descriptors in the input model of this invention is 60, while models constructed in other literature typically use hundreds or even thousands of descriptors. Using fewer descriptors significantly reduces model complexity. Furthermore, the descriptors used in this model have strong interpretability. Unlike other models that only have predictive capabilities, the model of this invention systematically analyzes the compound structure to quickly and preliminarily determine whether a compound has the potential to inhibit CYP450 enzymes based solely on its molecular structure information. This avoids the drawbacks of existing technologies that use a large number of purely mathematical descriptors (such as combinatorial topological indices), lacking clear physicochemical meaning, limiting model interpretability, and hindering the revelation of the potential mechanisms between compound structure and enzyme inhibitory activity. Attached Figure Description

[0032] Figure 1 The figure shows the evaluation results of the optimal model performance indicators for cytochrome CYP2C19.

[0033] Figure 2The figure shows the evaluation results of the optimal model performance index for cytochrome CYP2D6.

[0034] Figure 3 The figure shows the evaluation results of the optimal model performance index for cytochrome CYP3A4.

[0035] Figure 4 The figure shows the evaluation results of the optimal model performance index for cytochrome CYP1A2.

[0036] Figure 5 The figure shows the evaluation results of the optimal model performance indicators for cytochrome CYP2C9. Detailed Implementation

[0037] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but should not be considered as limiting the invention. It should be noted that those skilled in the art can make various adjustments and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0038] This invention provides a rapid screening method for cytochrome P450 enzyme inhibitors based on the QSAR model, comprising the following steps:

[0039] (1) Collect the data with Toxity ID 1851 from the PubChem database as the dataset;

[0040] (2) The KlekotaRoth fingerprint, MACCS fingerprint and PubChem fingerprint of the compound were calculated using the padelpy package;

[0041] (3) The inhibitory effects of five different subtypes of cytochrome P450 enzymes were predicted using support vector machine and random forest models. For each subtype, the feature importance of the descriptors after initial screening was calculated based on the RF model to obtain the importance ranking of the descriptors. Then, the descriptors were selected step by step according to the feature importance from high to low, and the number of descriptors input into the model was gradually increased by one each time. The number of different input descriptors constituted multiple sub-models. The overall architecture of the sub-model adopted the SVM model and / or the RF model. The performance of each sub-model and the number of input features were compared. Finally, the sub-model with fewer input descriptors while ensuring prediction performance was selected as the optimal model and included in the final modeling system. The optimal model selected for different subtypes was different. Thus, the optimal model for each of the five subtypes and the important descriptors selected by the corresponding optimal model were obtained.

[0042] (4) The important descriptors were combined with the compound dataset according to the SMILES identifier string to obtain five sub-datasets for the five human CYP450 enzyme subtypes.

[0043] For each subset of data, the optimal models for the five subtypes are trained using grid search and random undersampling, resulting in the five optimal models after training.

[0044] The screening model includes a discrimination module and five trained optimal models. The discrimination module identifies the toxicity endpoint markers of the input inhibitors and selects the corresponding optimal model based on these markers to obtain the screening results regarding whether the compound exhibits biotoxicity. The optimal model is denoted as the QSAR model.

[0045] The performance of the constructed QSAR models was evaluated based on four metrics, confirming that the models have good activity discrimination capabilities.

[0046] Example 1

[0047] 1. Data collection and preprocessing

[0048] This invention uses data with a Toxity ID of 1851 from the PubChem (Organic Small Molecule Bioactivity Database) as a compound dataset (with toxicity endpoint identifiers) to construct the model. Considering the determinism of the data, data with a "PUBCHEM_ACTIVITY_OUTCOME" value of "Inconclusion" were deleted from the original data, as well as data of metal compounds and compounds containing free unpaired electrons.

[0049] The database contains a wide range of organic compounds, including but not limited to drug molecules. While the initial purpose of investigating inhibition of cytochrome P450 enzymes was for drug development, the data can also be used to predict environmental exposure toxicity.

[0050] 2. Descriptor Calculation and Processing

[0051] Using Python version 3.12, the padelpy package was used to calculate KlekotaRoth fingerprints, MACCS fingerprints, and PubChem fingerprints, and preliminary descriptor screening was performed. The preliminary screening steps and criteria were as follows:

[0052] (1) Delete descriptors that have more than 10% unassigned values ​​across all samples;

[0053] (2) If a descriptor has the same value in more than 95% of the samples, it is considered to lack discriminative power and is removed.

[0054] (3) Calculate the variance of each descriptor across all samples (i.e., the degree of numerical variation of different compounds on the descriptor). If the variance of a descriptor is less than 0.1, it indicates that it varies little among samples and has limited information contribution, so it should be removed.

[0055] (4) Based on the Pearson correlation coefficient, evaluate the linear correlation between any two descriptors. If the correlation between two descriptors is greater than 0.9, retain one representative descriptor and delete redundant terms to reduce the impact of feature collinearity on model training.

[0056] This invention relates to five cytochrome P450 enzymes, containing a total of 63,619 data points; the descriptor of the model is considered, so that the model not only has predictive ability, but also has good interpretability.

[0057] 3. Model Building

[0058] (1) Dataset partitioning:

[0059] The dataset was divided into training set: validation set: test set in a ratio of 3:1:1 using stratified sampling.

[0060] (2) Further filter the descriptors based on the predicted value y:

[0061] All training data is put into five-fold cross-validation. The training set is divided into five folds. Each time, one fold is selected as the test set and the rest are used as the training set. Undersampling is performed on the training subset of the current fold, and hyperparameter search for five-fold cross-validation is performed. The best parameters found are used to retrain the random forest model to calculate the feature importance. 60 inputs are fed into the model, and then the hyperparameter model is trained. Finally, the feature importance of the five folds is averaged and sorted.

[0062] (3) Model building:

[0063] For each subtype, feature importance is calculated for the initially screened descriptors based on the RF model to obtain the importance ranking of the descriptors. Then, descriptors are progressively selected for input into the model according to feature importance from high to low, with the number of input descriptors gradually increasing by one each time. The different numbers of input descriptors constitute multiple sub-models. The overall architecture of the sub-models adopts the SVM model and / or the RF model. The performance of each sub-model and the number of input features are comprehensively compared. Finally, the sub-model with fewer input descriptors while ensuring prediction performance is selected as the optimal model and incorporated into the final modeling system. The optimal model selected is different for different subtypes. Thus, the optimal model for each of the five subtypes and the important descriptors selected by the corresponding optimal model are obtained.

[0064] The important descriptors were combined with the compound dataset according to the SMILES identifier string to obtain five sub-datasets for the five human CYP450 enzyme subtypes.

[0065] For each subset of data, the optimal models for the five subtypes are trained using grid search and random undersampling, resulting in the five optimal models after training.

[0066] The screening model includes a discrimination module and five optimal models after training. The discrimination module identifies the toxicity endpoint marker of the input inhibitor and enters the corresponding optimal model according to the toxicity endpoint marker to obtain the screening result of whether the compound has biotoxicity.

[0067] During training, the contribution of each important descriptor to the prediction result of the optimal model is evaluated by combining the permutation importance assessment, and the top-k structural features with the highest importance are selected and ranked.

[0068] 4. Model Performance Evaluation

[0069] Four metrics—AUC-ROC, PR-AUC, Log Loss, and Brier Score—are used to evaluate the model's performance. AUC-ROC measures the model's ability to distinguish between positive and negative classes; PR-AUC measures the model's performance in positive class identification; Log Loss measures the logarithmic loss between the predicted probability and the true label; and Brier Score measures the mean squared error between the predicted probability and the actual label.

[0070] The four classification performance metrics of the five optimal models were visualized. The AUC-ROC metric is plotted as the ROC curve, with a curve closer to the top left corner being better. The PR-AUC metric corresponds to the PR curve, with a curve closer to the top right corner being better. The Log Loss metric is represented by a scatter plot of prediction confidence versus Log Loss. For positive samples (red in the figure), as the x-axis value increases, the Log Loss approaches 0, indicating that the model has high confidence and can correctly predict positive samples with low loss. For negative samples (blue in the figure), as the x-axis value decreases, the Log Loss approaches 0, similarly indicating that the model has high confidence and can correctly predict negative samples with low loss. The Brier Score metric corresponds to the calibration curve; the closer the curve is to the diagonal, the more reliable the model's predicted probability. The classification performance metrics of the five optimal models are as follows: Figures 1 to 5 As shown.

[0071] To more comprehensively evaluate the model performance, this study introduces a non-learning baseline model (Random Guessing Model) as a performance lower bound reference. This model does not rely on any input features and only makes random predictions based on the prior distribution of positive classes in the training set, which is used to measure the model performance without any learning ability. The random model has theoretical expected values ​​under various indicators, as follows: (1) AUC-ROC random model: the theoretical value is 0.5, which means that positive and negative samples are completely indistinguishable and are only equivalent to the random sorting of a coin toss; (2) PR-AUC random model: equal to the proportion ρ of positive classes in the training set. real , representing the expected accuracy of being identified as positive under random selection; (3) Log Loss stochastic model: equal to the information entropy of the label distribution, i.e. -ρ real log(ρ real )-(1-ρ real log(1-ρ) real ), representing the average loss of a model with no discriminative power; (4) BrierScore stochastic model: equal to ρ real log(1-ρ real ), representing the minimum mean squared error when the model consistently outputs a fixed probability (i.e., a positive class proportion). In this invention, all model evaluation metrics report comparison results with this random benchmark to verify whether the actual model possesses discriminative ability beyond random guessing in real data. Based on the evaluation results of the four categories of metrics, the optimal models are selected: CYP2C19 selects the SVM model with 60 descriptors; CYP2D6 selects the RF model with 60 descriptors; CYP3A4 selects the SVM model with 60 descriptors; CYP1A2 selects the RF model with 50 descriptors; and CYP2C9 selects the SVM model with 70 descriptors.

[0072] The five figures above illustrate the performance evaluation results of the models constructed for five different cytochrome P450 isoenzymes (CYP2C19, CYP2D6, CYP3A4, CYP1A2, and CYP2C9). As shown in the figures, the mean AUC-ROC of each model on the test set reaches 0.85, indicating that the model possesses good overall classification ability. Meanwhile, the PR-AUC value is significantly higher than the baseline proportion of positive samples in the dataset, further illustrating that the model has strong discrimination and recall capabilities when handling imbalanced data, especially in identifying active compounds. Furthermore, the Log Loss and Brier Score indicators are both at low levels, indicating that the model has high confidence and calibration performance in output probability prediction. In summary, all models demonstrate good performance in terms of accuracy, robustness, and reliability, and can effectively support the task of predicting CYP450 enzyme inhibitory activity.

[0073] 5. Discussion of Compound Structure

[0074] To explore the structure corresponding to the fingerprint descriptor

[0075] This invention utilizes fingerprint descriptors used in model training, combining a permutation importance method to evaluate the contribution of each descriptor to the model's prediction results. From this, the top 60 most important structural features are selected based on a random forest ranking. Subsequently, according to the substructure correspondence file provided by the official fingerprint type, the specific structural fragments represented by each high-importance descriptor are analyzed. Analysis of these key structural fragments reveals that organic compounds containing bridging bicyclic structures, trisubstituted central atom structures, five-membered rings, aromatic rings, and polymeric skeletons containing more than 16 carbon atoms are significantly more likely to inhibit cytochrome P450 enzymes. This result indicates that the extracted structural features not only possess statistical significance but also clear chemical interpretability, providing theoretical support for subsequent structure optimization and toxicity screening.

[0076] In the second step of descriptor calculation and processing, descriptors were screened using various methods such as matching value, variance, and correlation to ensure that the descriptors input into the model in the subsequent model construction were most representative. Based on the above results, in the third step of model construction, the descriptors with the highest importance were selected for model construction to achieve the best classification effect.

[0077] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A rapid screening method for cytochrome P450 enzyme inhibitors based on the QSAR model, characterized in that, Includes the following steps: Step 1: Data Collection and Preprocessing The data with Toxity ID 1851 in the PubChem database was used as the compound dataset. This compound dataset contains compound data and biotoxicity panel IDs. The biotoxicity panel IDs are numbered 1-5, corresponding to five human CYP450 enzyme subtypes. Data with "PUBCHEM_ACTIVITY_OUTCOME" value "Inconclusion" were removed, as were data of metal compounds and compounds containing free unpaired electrons. Step 2: Descriptor Calculation and Processing Using Python version 3.12, the padelpy package was used to calculate KlekotaRoth fingerprints, MACCS fingerprints, and PubChem fingerprints, and preliminary descriptor screening was performed. The preliminary screening steps and criteria were as follows: 1) Delete descriptors that have more than 10% unassigned values ​​across all samples; 2) If a descriptor has the same value in more than 95% of the samples, it is considered to lack discriminative power and is removed. 3) Calculate the variance of each descriptor across all samples. If the variance of a descriptor is less than 0.1, it is discarded. 4) Based on the Pearson correlation coefficient, evaluate the linear correlation between any two descriptors. If the correlation between two descriptors is greater than 0.9, retain one representative descriptor and delete redundant terms. The third step is to build a screening model. For each subtype, the feature importance of the initially screened descriptors is calculated based on the RF model to obtain the importance ranking of the descriptors. Then, the descriptors are selected step by step according to the feature importance from high to low, and the number of descriptors input into the model is gradually increased by one each time. The number of different input descriptors constitutes multiple sub-models. The overall architecture of the sub-model adopts the SVM model and / or the RF model. The performance of each sub-model and the number of input features are compared. Finally, the sub-model with fewer input descriptors while ensuring prediction performance is selected as the optimal model and incorporated into the final modeling system. Different subtypes select different optimal models, thus obtaining the optimal models for each of the five subtypes and the important descriptors selected for the corresponding optimal models; The important descriptors were combined with the compound dataset according to the SMILES identifier string to obtain five sub-datasets for the five human CYP450 enzyme subtypes. For each subset of data, the optimal models for the five subtypes are trained using grid search and random undersampling, resulting in the five optimal models after training. The screening model includes a discrimination module and five optimal models after training. The discrimination module identifies the toxicity endpoint marker of the input inhibitor and enters the corresponding optimal model according to the toxicity endpoint marker to obtain the screening result of whether the compound has biotoxicity. During training, the contribution of each important descriptor to the prediction results of the optimal model is evaluated by combining the permutation importance assessment. The top-k structural features with the highest importance are then selected and ranked. Subsequently, based on the substructure correspondence file provided by the official fingerprint type, the specific structural fragments represented by each high-importance descriptor are analyzed. Compounds with bridging bicyclic structures, trisubstituted central atom structures, five-membered ring structures, aromatic structures, and macromolecular structures containing more than 16 carbons are more likely to inhibit cytochrome P450 enzymes.

2. The method according to claim 1, characterized in that, Four metrics—AUC-ROC, PR-AUC, Log Loss, and BrierScore—are used as standards for evaluating model performance. AUC-ROC measures the model's ability to distinguish between positive and negative classes, PR-AUC measures the model's performance in positive class identification, Log Loss measures the log loss between the predicted probability and the true label, and BrierScore measures the mean squared error between the predicted probability and the actual label.

3. The method according to claim 1, characterized in that, The performance of the model is judged by the comprehensive performance of the four indicators.

4. The method according to claim 1, characterized in that, The optimal models for the five subtypes are as follows: CYP2C19 selects an SVM model with 60 descriptors; CYP2D6 selects an RF model with 60 descriptors; CYP3A4 selects an SVM model with 60 descriptors; CYP1A2 selects an RF model with 50 descriptors; and CYP2C9 selects an SVM model with 70 descriptors.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program can perform the steps of the method described in any one of claims 1-4.

Images