An ai-based physiological pharmacokinetic prediction method

Abstract

The present application relates to the technical field of drug research and development, in particular to a physiological pharmacokinetics prediction method based on AI, comprising the following steps: performing data set construction and preprocessing, carrying out molecular feature extraction and fusion, implementing molecular feature screening and dimension reduction, executing machine learning model construction and optimization, interfacing with a classic PBPK model to form an AI-PBPK model, and carrying out PK parameter prediction performance evaluation. The physiological pharmacokinetics prediction method based on AI effectively solves the technical problems of low prediction accuracy of drug PK parameters, poor model adaptability, low parameter search efficiency, and high molecular feature dimension prone to overfitting in the prior art through multi-dimensional molecular feature fusion, feature screening and dimension reduction, and automatic model construction and optimization.

Description

Technical Field

[0001] This invention relates to the field of drug development technology, specifically to an AI-based method for predicting physiological pharmacokinetics. Background Technology

[0002] Drug development involves target discovery and validation, lead compound screening and optimization, preclinical trials of candidate compounds, and finally, clinical trials. In clinical trials, pharmacokinetic (PK) information is crucial for determining the efficacy and safety of compounds; many candidate compounds fail in clinical trials due to a lack of favorable PK properties. Therefore, identifying compounds with good human pharmacokinetic properties before entering clinical trials can significantly improve the success rate of clinical trials, reduce the risk of failure, and lower the time and financial costs of drug development.

[0003] Generally, key pharmacokinetic parameters include clearance (CL), steady-state apparent volume of distribution (Vss), apparent permeability (Papp), area under the plasma concentration-time curve (AUC), and peak concentration (Cmax). CL, Vss, and Papp are often derived through in vitro experiments combined with mechanistic derivation. For clearance, in vitro experiments combined with IVIVE (invitro-invivoextrapolation) is the most commonly used conversion pathway. IVIVE obtains the intrinsic clearance of liver microsomes or hepatocytes through in vitro metabolic experiments, introduces a scaling factor to obtain the in vivo liver clearance, and then combines this with a mechanical liver model to predict in vivo plasma clearance. This is a classic method for predicting drug clearance, but it has a narrow applicability and low prediction accuracy. Parameters such as area under the plasma concentration-time curve (AUC) and peak concentration (Cmax) are often obtained using physiological pharmacokinetic simulations (PBPK). PBPK modeling is based on human physiological structure (blood flow, volume, metabolic enzyme activity, etc. in various tissues and organs) and drug physicochemical properties (solubility, lipid solubility, metabolic rate, etc.). It simulates the entire ADME process of drugs in the human body through mathematical equations, outputting concentration-time curves of the drug in various tissues and organs. The core logic is "physiological parameters + drug parameters → mathematical simulation → PK prediction". This method is an important tool for drug PK prediction, but it requires a large number of drug-specific parameters, facing difficulties in parameter acquisition and long cycles in the early stages of drug development. With the development of artificial intelligence technology, PBPK models combined with artificial intelligence (AI+PBPK) have become a new direction of exploration. However, due to the influence of data quality and scale, and model architecture selection, prediction bias is large, and the accuracy is insufficient when the initial prediction is not corrected. The superposition of machine learning prediction bias and PBPK model bias leads to a significant decrease in the prediction accuracy of the final drug PK parameters (AUC, Cmax). At the same time, existing AI+PBPK methods suffer from either overly simplistic or overly complex characterization of compound molecules. Overly simplistic characterization prevents the model from effectively learning molecular information, resulting in insufficient prediction accuracy; overly complex characterization easily leads to model overfitting when data is insufficient, causing it to fail in predicting new molecules. Meanwhile, the prediction performance of existing AI+PBPK methods heavily relies on manual model selection and hyperparameter adjustment, which limits the efficiency and accuracy of drug PK parameter prediction and cannot meet the actual needs of rapid and accurate screening of candidate compounds in the early stages of drug development.

[0004] Traditional methods combine computational compound molecular descriptors with in vitro experimental data to improve the prediction accuracy of the area under the plasma concentration-time curve (AUC) after oral administration in rats. This approach compares the accuracy of a purely computational scheme based on compound molecular descriptors using a machine learning model with that of predictions using in vitro experimental data and the traditional IVIVE method. Finally, it is proposed that introducing in vitro data as an explanatory variable into the machine learning model can significantly improve the prediction accuracy.

[0005] This approach directly utilizes the chemical structure of small molecules to predict their complete pharmacokinetic (PK) plasma concentration-time curves in rats. A two-step machine learning framework is constructed: First, feature vectors are extracted from the compound using molecular descriptors, and key PK parameters are predicted using an XGBoost model. Second, the predicted CL, Vdss, and specific sampling time points from the first step are used as input features, and the XGBoost network directly predicts the plasma drug concentration at the corresponding time points. This approach eliminates dependence on in vitro data, performing high-throughput predictions solely based on molecular SMILES; however, the chemical structural similarity between the molecule and the training set data has a decisive impact on performance.

[0006] However, the above solution has the following drawbacks: Low prediction accuracy: The molecular feature representation is too simple and cannot fully extract molecular information; there is no professional feature screening and dimensionality reduction step, and high-dimensional features are prone to model overfitting; the machine learning prediction bias and the PBPK model bias are superimposed, resulting in insufficient prediction accuracy of the final PK parameters, and the prediction effect is extremely poor for non-classical chemical compounds.

[0007] Low modeling efficiency: Model selection and hyperparameter adjustment rely on manual trial and error, which is cumbersome; page parameter entry, data preprocessing, parameter substitution and other steps all require manual intervention, which is time-consuming and labor-intensive, and cannot meet the needs of rapid screening in drug development.

[0008] Narrow applicability: It performs poorly in evaluating compounds that are not metabolized by CYP450, transporter-mediated, or complex non-classical chemical compounds, and its model adaptability and generalization ability are insufficient.

[0009] It requires a high level of professional experience: the selection of model structure, adjustment of algorithm hyperparameters, error correction and other aspects all require professionals with knowledge in multiple fields such as quantitative pharmacology and machine learning. The high technical application threshold is not conducive to large-scale promotion.

[0010] In summary, an AI-based method for predicting physiological pharmacokinetics is needed to address the aforementioned issues. Summary of the Invention

[0011] The purpose of this invention is to provide an AI-based method for predicting physiological pharmacokinetics to solve the problems mentioned in the background art.

[0012] To achieve the above objectives, the present invention provides the following technical solution: This invention proposes an AI-based method for predicting physiological pharmacokinetics, comprising the following steps: S1. Construct and preprocess the dataset, collect relevant data on the physicochemical, ADME, and PK properties of compounds to build a database, divide the dedicated training set and test set, collect external independent test sets related to the PK properties of compounds, and complete the standardization and preprocessing of all data. S2. Conduct molecular feature fusion extraction, extract multiple types of structure-based molecular fingerprints and calculate molecular physicochemical descriptors, combine graph neural networks to obtain various molecular property parameters, and integrate them to form a multi-dimensional fused molecular feature set; S3. Implement feature selection and dimensionality reduction. Use a specified feature selection algorithm to perform targeted dimensionality reduction on the fused molecular feature set and select a fully relevant subset of features that match the molecular characteristic prediction task. S4. Perform machine learning model construction and optimization, build a basic model set based on an automated machine learning framework, build models using a multi-layer base model stacking ensemble strategy, and complete the automated selection, combination and hyperparameter optimization of models through k-fold cross-validation to determine and retain the optimal machine learning model; S5. Complete the construction of the AI-PBPK model fusion, substitute the key parameters of physicochemical, ADME, and PK predicted by the optimal machine learning model into the classic PBPK model, standardize the configuration of the physiological parameter module of the classic PBPK model, realize the algorithm fusion and automatic parameter docking of the machine learning model and the classic PBPK model, and form an integrated AI-PBPK model. S6. Conduct performance evaluation of PK core parameter prediction. Input standardized parameters into the AI-PBPK model to obtain the predicted values ​​of PK core parameters. Use specified evaluation indicators to verify the model performance. Compare and verify the performance of clearance parameter prediction with the traditional IVIVE method. On an external independent test set, compare and verify the performance of blood drug concentration-time curve area and peak drug concentration parameter prediction with the traditional PBPK method.

[0013] Preferably, the implementation process of step S1 is as follows: Key physicochemical, ADME, and PK parameter data of compounds were collected from open-source databases and public literature to construct a dataset of compound-related parameters. A dataset containing 1352 molecules was selected for the human body clearance rate prediction task, and it was divided into a training set containing 1217 molecules and a test set containing 135 molecules in a 9:1 ratio. A dataset containing 1272 molecules was selected for the human apparent permeability coefficient prediction task, and it was divided into a training set containing 1145 molecules and a test set containing 127 molecules in a 9:1 ratio. A dataset containing 1308 molecules was selected for the human homeostatic apparent distribution volume prediction task. It was divided into a training set containing 1178 molecules and a test set containing 130 molecules in a 9:1 ratio. We obtained an in vitro intrinsic clearance dataset containing the intrinsic clearance of human liver microsomes, intrinsic clearance of human hepatocytes, free drug fraction in plasma, and whole blood plasma partition ratio of 135 molecules in the test set, which was used for comparison and validation with the traditional IVIVE method. A dataset of 38 molecules from 2018 to the present was collected from publicly available literature as a new external test set. All datasets were subjected to standardized preprocessing operations, including missing value completion and data normalization.

[0014] Preferably, the implementation process of step S2 is as follows: extract molecular access system fingerprints, atom pair fingerprints, and topological torsion fingerprints; use Open Babel to directly calculate various descriptors of the molecule, including physicochemical properties, charge properties, topological properties, and three-dimensional conformational properties; and fuse the molecular access system fingerprints, atom pair fingerprints, topological torsion fingerprints, and molecular descriptors to form a multi-dimensional fused molecular feature set.

[0015] Preferably, the implementation process of step S3 is as follows: using a feature selection algorithm, configuring the regularization operation parameters of the algorithm for each molecular property prediction task, inputting the multi-dimensional fused molecular feature set corresponding to the training set into the configured algorithm, compressing the feature coefficients through L1 regularization constraints, and selecting the non-zero coefficient feature subset that has a strong predictive effect on each target variable, eliminating redundant features in the multi-dimensional fused molecular feature set, and using the fully correlated feature subset as the input feature set for subsequent machine learning model training, completing the feature dimensionality reduction process and recording the feature selection results.

[0016] Preferably, the implementation process of step S4 is as follows: a machine learning model is built based on an automated machine learning framework, a generalized linear model, a tree model, a classic machine learning model, and an artificial neural network model are set as the basic model set, and the parameter range of each basic model is configured. A multi-layer model ensemble strategy is adopted to build a machine learning model. The output results of each basic model are fused and the fused results are input into the stacked ensemble model. The model training and ensemble operation are completed layer by layer. The training set is subjected to k-fold cross-validation. During the validation process, the model combination and hyperparameter configuration are iteratively optimized. The model with the best performance during the validation process is selected as the final optimal machine learning model. The test set is input into the optimal machine learning model to complete the preliminary performance evaluation.

[0017] Preferably, the implementation process of step S5 is as follows: The clearance rate, steady-state apparent volume of distribution, and apparent permeability coefficient predicted by the optimal machine learning model are automatically substituted into the classic PBPK model through the algorithm interface, replacing the drug-specific parameters obtained from animal experiments in the classic PBPK model. Based on preset parameters such as population physiological parameters, tissue and organ blood flow, tissue and organ volume, clinical trial drug dosage, and drug administration duration, the physiological parameter module of the classic PBPK model is configured in a standardized manner. The algorithm integration interface enables the fusion of machine learning models and classic PBPK models, achieving precise docking between input feature parameters and model input nodes, and forming an integrated AI-PBPK model that can directly predict pharmacokinetic parameters.

[0018] Preferably, in step S6, the coefficient of determination, average conversion error, and absolute average conversion error are used as evaluation indicators to verify the model performance. The implementation process is as follows: set up For observed attribute values, The predicted value is n, where n is the sample size, calculated with base 10. and Calculate the coefficient of determination: ; In the formula, The average value of the observed attribute; Calculate the average conversion error: ; Calculate the absolute average conversion error: ; Substitute the observed and predicted values ​​of the clearance rate, steady-state apparent distribution volume, and apparent permeability coefficient into the above formulas to calculate the determination coefficient, average reduction error, and absolute average reduction error for each parameter, thus completing the performance verification of the machine learning model.

[0019] Preferably, the implementation process for completing the comparative verification with the traditional IVIVE method and the external independent verification in step S6 is as follows: The AI-PBPK model was used to predict the human plasma clearance rate of 135 molecules. The in vitro inherent clearance rate data of 135 molecules were used to calculate the human plasma clearance rate using the traditional IVIVE dilution scaling method and the unbound scaling model. Automatically calculate the coefficient of determination, average conversion error, and absolute average conversion error for the three methods, and complete the comparative verification with the same dataset. The 38 compounds in the external validation set were input into the AI-PBPK model to automatically predict the area under the blood drug concentration-time curve and the core parameters of peak drug concentration PK. The error range between the predicted and measured values ​​of each parameter was calculated to complete the external independent validation of the model's generalization ability. Integrate the predicted values ​​of PK core parameters, model fitting performance indicators, comparison results with the same dataset, and external independent validation error range to form a complete report and output it.

[0020] Preferably, the specific implementation of the multi-layer base model stacking integration strategy in step S4 is as follows: The input feature set is set to be connected to the first basic model layer. The basic model layer includes classic machine learning models, tree models, generalized linear models, and artificial neural networks. Each basic model operates independently and outputs prediction results. The feature splicing layer is set up to splice all the prediction results of the base model layer to form a spliced ​​feature set and pass it to the stacked model layer. The stacked model layer performs secondary training on the concatenated feature set and outputs the results. A weight allocation layer is set to weight the output results of the stacked model layer, and finally outputs the model prediction value.

[0021] Preferably, it also includes the step of building a standardized front-end operation page, which automates the standardized parameter input and submission process through the front-end page, and synchronizes the parameters entered on the front-end page to the back-end database in real time, so that the back-end system automatically triggers the full process calculation from S1 to S6.

[0022] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention's AI-based physiological pharmacokinetic prediction method effectively solves the technical problems of low accuracy in drug PK parameter prediction, poor model adaptability, low parameter search efficiency, and easy overfitting due to high molecular feature dimensionality in existing technologies through multi-dimensional molecular feature fusion, feature screening and dimensionality reduction, and automated model construction and optimization. It achieves accurate prediction of key ADME parameters such as CL, Vss, and Papp, as well as PK parameters such as AUC and Cmax. In terms of clearance rate prediction, AI-PBPK significantly outperforms the traditional IVIVE method, especially in the context of 135 compounds. On the external validation set, the prediction accuracy R² reached 0.374. Regarding overall pharmacokinetic characteristics prediction, on an external validation set consisting of 38 compounds, for pharmacokinetic parameters such as AUC and Cmax, the error range between the predicted and observed values ​​was 0.25 to 4 times for 55.3% of the compounds, and 0.25 to 4 times for 71.1% of the compounds. This demonstrates that AI-PBPK can provide new insights into the pharmacokinetic prediction of compounds in the drug discovery stage, and is helpful for the screening of early candidate compounds, facilitating its widespread application. Attached Figure Description

[0023] Figure 1 A flowchart illustrating the construction process of the AI-PBPK method of the present invention is shown; Figure 2 A schematic diagram of the base model stacking integration strategy of the AI-PBPK of the present invention is shown. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Example 1, please refer to Figures 1 to 2 This invention proposes an AI-based method for predicting physiological pharmacokinetics, comprising the following steps: S1. Construct and preprocess the dataset, collect relevant data on the physicochemical, ADME, and PK properties of compounds to build a database, divide the dedicated training set and test set, collect external independent test sets related to the PK properties of compounds, and complete the standardization and preprocessing of all data. It should also be noted that the implementation process of step S1 is as follows: Key physicochemical, ADME, and PK parameter data of compounds were collected from open-source databases and public literature to construct a dataset of compound-related parameters. A dataset containing 1352 molecules was selected for the human body clearance rate prediction task, and it was divided into a training set containing 1217 molecules and a test set containing 135 molecules in a 9:1 ratio. A dataset containing 1272 molecules was selected for the human apparent permeability coefficient prediction task, and it was divided into a training set containing 1145 molecules and a test set containing 127 molecules in a 9:1 ratio. A dataset containing 1308 molecules was selected for the human homeostatic apparent distribution volume prediction task. It was divided into a training set containing 1178 molecules and a test set containing 130 molecules in a 9:1 ratio. We obtained an in vitro intrinsic clearance dataset containing the intrinsic clearance of human liver microsomes, intrinsic clearance of human hepatocytes, free drug fraction in plasma, and whole blood plasma partition ratio of 135 molecules in the test set, which was used for comparison and validation with the traditional IVIVE method. A dataset of 38 molecules from 2018 to the present was collected from publicly available literature as a new external test set. All datasets were subjected to standardized preprocessing operations, including missing value completion and data normalization.

[0026] S2. Conduct molecular feature fusion extraction, extract multiple types of structure-based molecular fingerprints and calculate molecular physicochemical descriptors, combine graph neural networks to obtain various molecular property parameters, and integrate them to form a multi-dimensional fused molecular feature set; It should also be noted that the implementation process of step S2 is as follows: extract molecular access system fingerprints, atom pair fingerprints, and topological torsion fingerprints; use Open Babel to directly calculate various descriptors of the molecule, including physicochemical properties, charge properties, topological properties, and three-dimensional conformational properties; and fuse the molecular access system fingerprints, atom pair fingerprints, topological torsion fingerprints, and molecular descriptors to form a multi-dimensional fused molecular feature set.

[0027] S3. Implement feature selection and dimensionality reduction. Use a specified feature selection algorithm to perform targeted dimensionality reduction on the fused molecular feature set and select a fully relevant subset of features that match the molecular characteristic prediction task. It should also be noted that the implementation process of step S3 is as follows: Using a feature selection algorithm, the regularization operation parameters of the algorithm are configured for each molecular property prediction task. The multi-dimensional fused molecular feature set corresponding to the training set is input into the configured algorithm. The feature coefficients are compressed by L1 regularization constraint, and a subset of non-zero coefficient features with strong predictive power for each target variable is selected. Redundant features in the multi-dimensional fused molecular feature set are removed. The fully correlated feature subset is used as the input feature set for subsequent machine learning model training. Feature dimensionality reduction processing is completed and the feature selection results are recorded.

[0028] S4. Perform machine learning model construction and optimization, build a basic model set based on an automated machine learning framework, build models using a multi-layer base model stacking ensemble strategy, and complete the automated selection, combination and hyperparameter optimization of models through k-fold cross-validation to determine and retain the optimal machine learning model; It should also be noted that the implementation process of step S4 is as follows: A machine learning model is built based on an automated machine learning framework. A generalized linear model, a tree model, a classic machine learning model, and an artificial neural network model are set as the basic model set. The parameter ranges of each basic model are configured. A multi-layer model ensemble strategy is used to construct the machine learning model. The output results of each basic model are fused, and the fused results are input into the stacked ensemble model. Model training and ensemble operations are completed layer by layer. K-fold cross-validation is performed on the training set. During the validation process, the model combination and hyperparameter configuration are iteratively optimized. The model with the best performance during the validation process is selected as the final optimal machine learning model. The test set is input into the optimal machine learning model to complete the preliminary performance evaluation.

[0029] It should also be noted that the specific implementation of the multi-layer base model stacking integration strategy in step S4 is as follows: The input feature set is set to be connected to the first basic model layer. The basic model layer includes classic machine learning models, tree models, generalized linear models, and artificial neural networks. Each basic model operates independently and outputs prediction results. The feature splicing layer is set up to splice all the prediction results of the base model layer to form a spliced ​​feature set and pass it to the stacked model layer. The stacked model layer performs secondary training on the concatenated feature set and outputs the results. A weight allocation layer is set to weight the output results of the stacked model layer, and finally outputs the model prediction value.

[0030] S5. Complete the construction of the AI-PBPK model fusion, substitute the key parameters of physicochemical, ADME, and PK predicted by the optimal machine learning model into the classic PBPK model, standardize the configuration of the physiological parameter module of the classic PBPK model, realize the algorithm fusion and automatic parameter docking of the machine learning model and the classic PBPK model, and form an integrated AI-PBPK model. It should also be noted that the implementation process of step S5 is as follows: The clearance rate, steady-state apparent volume of distribution, and apparent permeability coefficient predicted by the optimal machine learning model are automatically substituted into the classic PBPK model through the algorithm interface, replacing the drug-specific parameters obtained from animal experiments in the classic PBPK model. Based on preset parameters such as population physiological parameters, tissue and organ blood flow, tissue and organ volume, clinical trial drug dosage, and drug administration duration, the physiological parameter module of the classic PBPK model is configured in a standardized manner. The algorithm integration interface enables the fusion of machine learning models and classic PBPK models, achieving precise docking between input feature parameters and model input nodes, and forming an integrated AI-PBPK model that can directly predict pharmacokinetic parameters.

[0031] S6. Conduct performance evaluation of PK core parameter prediction. Input standardized parameters into the AI-PBPK model to obtain the predicted values ​​of PK core parameters. Use specified evaluation indicators to verify the model performance. Compare and verify the performance of clearance parameter prediction with the traditional IVIVE method. On an external independent test set, compare and verify the performance of blood drug concentration-time curve area and peak drug concentration parameter prediction with the traditional PBPK method.

[0032] It should also be noted that step S6 uses the coefficient of determination, average conversion error, and absolute average conversion error as evaluation indicators to verify model performance. The implementation process is as follows: set up For observed attribute values, The predicted value is n, where n is the sample size, calculated with base 10. and Calculate the coefficient of determination: ; In the formula, The average value of the observed attribute; Calculate the average conversion error: ; Calculate the absolute average conversion error: ; Substitute the observed and predicted values ​​of the clearance rate, steady-state apparent distribution volume, and apparent permeability coefficient into the above formulas to calculate the determination coefficient, average reduction error, and absolute average reduction error for each parameter, thus completing the performance verification of the machine learning model.

[0033] It should also be noted that the implementation process for completing the comparison verification with the traditional IVIVE method and the external independent verification in step S6 is as follows: The AI-PBPK model was used to predict the human plasma clearance rate of 135 molecules. The in vitro inherent clearance rate data of 135 molecules were used to calculate the human plasma clearance rate using the traditional IVIVE dilution scaling method and the unbound scaling model. Automatically calculate the coefficient of determination, average conversion error, and absolute average conversion error for the three methods, and complete the comparative verification with the same dataset. The 38 compounds in the external validation set were input into the AI-PBPK model to automatically predict the area under the blood drug concentration-time curve and the core parameters of peak drug concentration PK. The error range between the predicted and measured values ​​of each parameter was calculated to complete the external independent validation of the model's generalization ability. Integrate the predicted values ​​of PK core parameters, model fitting performance indicators, comparison results with the same dataset, and external independent validation error range to form a complete report and output it.

[0034] It also includes the steps of building a standardized front-end operation page, which automates the standardized parameter input and submission process through the front-end page, and synchronizes the parameters entered on the front-end page to the back-end database in real time. The back-end system automatically triggers the full process calculation from S1 to S6.

[0035] Example 2, please refer to Figures 1 to 2 In practical applications, this invention provides an AI-based method for predicting physiological pharmacokinetics, specifically comprising the following steps: The flowchart for method construction is as follows Figure 1 As shown, it includes seven core steps: dataset construction and preprocessing, molecular feature fusion and extraction, feature screening and dimensionality reduction, machine learning model construction and optimization, AI-PBPK model fusion construction, and PK parameter prediction performance evaluation. AI-PBPK covers a wide range of compound-related parameters. For parameters CL, Vss, and Papp, the sampling machine learning framework predicts them. Key PK parameters such as AUC and Cmax are calculated using the predictions of CL, Vss, and Papp as inputs and the classic PBPK model. Step 1: Data Collection The learning data for the physicochemical properties of compounds and key ADME parameters in the AI-PBPK machine learning model are all from open-source databases and public literature. For the task of predicting clearance rate in the human body, a dataset containing 1352 molecules collected from publicly available literature was used for the task. It was divided into a training set containing 1217 molecules and a test set containing 135 molecules at a ratio of 9:1. A dataset containing 1272 molecules was selected for the human apparent permeability coefficient prediction task, and it was divided into a training set containing 1145 molecules and a test set containing 127 molecules in a 9:1 ratio. A dataset containing 1308 molecules was selected for the human homeostatic apparent distribution volume prediction task. It was divided into a training set containing 1178 molecules and a test set containing 130 molecules in a 9:1 ratio. We obtained an in vitro intrinsic clearance dataset containing the intrinsic clearance of human liver microsomes, intrinsic clearance of human hepatocytes, free drug fraction in plasma, and whole blood plasma partition ratio of 135 molecules in the test set, which was used for comparison and validation with the traditional IVIVE method. A dataset of 38 molecules from 2018 to the present was collected from publicly available literature as a new external test set. All datasets were subjected to standardized preprocessing operations, including missing value completion and data normalization. Step 2: Feature Extraction Molecular fingerprints are a representation that enables computational programs to capture key molecular information, playing a crucial role in computer-aided drug development. AI-PBPK employs three structure-based molecular fingerprinting methods: extracting molecular access system fingerprints, atom-pair fingerprints, and topological torsion fingerprints. It uses Open Babel to directly calculate various descriptors of molecules, including physicochemical properties, charge properties, topological properties, and three-dimensional conformational properties. It then fuses molecular access system fingerprints, atom-pair fingerprints, topological torsion fingerprints, and molecular descriptors to form a multi-dimensional fused molecular feature set. In addition to molecular structure-based representations, RDKit molecular descriptors are used to directly calculate the chemical properties of the entire molecule and are also added as features to the model. Specifically, the steps are as follows: Extract the molecular access system fingerprint. Each bit of the fingerprint is 1 if the compound molecule has a specified substructure, and 0 otherwise. Atom pair fingerprints are extracted, which use any two atoms in the molecule and their atom types, local connectivity environment and topological distance as coding units to characterize the overall topological structure features of the molecule; The topological torsion fingerprint is extracted. This fingerprint uses the atomic sequence consisting of four consecutive connected atoms in the molecule as the basic coding segment to encode the local skeleton connection pattern and structural change characteristics of the molecule. Open Babel is used to directly calculate various descriptors of molecules, including physicochemical properties, charge properties, topological properties, and three-dimensional conformational properties. By fusing molecular access system fingerprints, atom pair fingerprints, topological twist fingerprints, and molecular descriptors, a multi-dimensional fused molecular feature set is formed. Step 3: Feature Selection Regularization parameters are configured for the algorithms of each molecular property prediction task, and the multi-dimensional fused molecular feature set corresponding to the training set is input into the configured algorithm; By compressing feature coefficients using L1 regularization constraints, a subset of non-zero coefficient features with strong predictive power for each target variable is selected, and redundant features in the multi-dimensional fusion molecular feature set are eliminated; The fully relevant feature subset is used as the input feature set for subsequent machine learning model training to complete feature dimensionality reduction and record the feature selection results.

[0036] Step 4: Machine Learning Model Building and Optimization A machine learning model is built based on an automated machine learning framework. A generalized linear model, a tree model, a classical machine learning model, and an artificial neural network model are set as the basic model set. The parameter range of each basic model is configured. A multi-layer model ensemble strategy is used to build the machine learning model. The output results of each basic model are fused and the fused result is input into the stacked ensemble model. The model training and ensemble operation are completed layer by layer. k-fold cross-validation is performed on the training set. During the validation process, the model combination and hyperparameter configuration are iteratively optimized. The model with the best performance during the validation process is selected as the final optimal machine learning model. The test set is input into the optimal machine learning model to complete the preliminary performance evaluation. In one specific embodiment, due to the limitation of data volume, complex deep neural networks are likely to overfit. In contrast, machine learning models with fewer parameters, such as random forests, decision trees, and artificial neural networks (ANNs), tend to perform better on ADME and PK property prediction tasks. However, the performance of these methods is closely related to the model type and the choice of hyperparameters, often requiring manual fine-tuning. Meanwhile, considering that building a combined model from multiple base models can often improve the final performance, the range of parameter search has grown exponentially. To achieve the best model combination and optimal parameter settings, AI-PBPK adopts an automated machine learning method based on H2O AutoML. This method uses a multi-layer base model stacking ensemble strategy, which passes the prediction results output by the base model as input to a higher-level stacked model. like Figure 2 As shown, the first layer has multiple basic models, whose outputs are connected and then passed to the next layer, which itself consists of multiple stacked models. These stacked models then act as the base models for another layer; The generalized linear model, tree model, classical machine learning model and artificial neural network model are set as the base model set, and the parameter range of each base model is set. Then, the model stacking and parameter search are performed automatically based on H2O AutoML. In addition, k-fold cross-validation is performed on the training set to continuously optimize and find the best model combination and parameters; The model that performs best during the validation process will be retained and further evaluated on the test set; Step 5: AI-PBPK Model Fusion Construction The AI-PBPK model of this invention is obtained by deep fusion of a machine learning model and a classic PBPK model. The fusion logic is as follows: 1. The key ADME parameters such as drug CL, Vss, and Papp predicted by the machine learning model are automatically substituted into the classic PBPK model, replacing the drug-specific parameters obtained through animal experiments in the traditional PBPK model, thus solving the problem of PBPK model parameters depending on animal experiments. 2. Based on preset parameters such as population physiological parameters, tissue and organ blood flow / volume, clinical trial drug dosage, and drug administration duration, the physiological parameter module of the classic PBPK model is configured in a standardized manner; 3. The system automatically completes the algorithm fusion and parameter docking of the machine learning model and the PBPK model to form an integrated AI-PBPK model. This model retains the simulation advantages of the PBPK model based on physiological mechanisms and integrates the accurate prediction capabilities of machine learning, achieving complementary advantages. Step 6: Model Evaluation 1. Machine learning model performance evaluation: For regression tasks such as CL, Vass, and Papp prediction, the coefficient of determination (COP) is used to evaluate the performance of the model. Average Fold Error (AFE) and Absolute Average Fold Error (AAFE) are used as metrics to evaluate model fit performance. set up This represents the observed attribute value with a base of 10. This represents the predicted value with a base of 10. The calculation method for the above indicators is shown in Formula 1-3: 2. Comparison with traditional IVIVE methods on the same dataset: 135 molecular datasets were input into the model of this invention. At the same time, the traditional IVIVE dilution scaling method and the method without scaling were used to predict CLp on the same dataset. The system automatically compared the R², AFE, and AAFE indices of the three methods to verify the performance advantages of the model of this invention. 3. External independent validation: Input the external PK dataset of 38 compounds into the AI-PBPK model. The system automatically predicts core PK parameters such as AUC and Cmax, calculates the error range between the predicted and measured values, and validates the generalization ability of the model.

[0037] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based method for predicting physiological pharmacokinetics, characterized in that, Includes the following steps: S1. Construct and preprocess the dataset, collect relevant data on the physicochemical, ADME, and PK properties of compounds to build a database, divide the dedicated training set and test set, collect external independent test sets related to the PK properties of compounds, and complete the standardization and preprocessing of all data. S2. Conduct molecular feature fusion extraction, extract multiple types of structure-based molecular fingerprints and calculate molecular physicochemical descriptors, combine graph neural networks to obtain various molecular property parameters, and integrate them to form a multi-dimensional fused molecular feature set; S3. Implement feature selection and dimensionality reduction. Use a specified feature selection algorithm to perform targeted dimensionality reduction on the fused molecular feature set and select a fully relevant subset of features that match the molecular characteristic prediction task. S4. Perform machine learning model construction and optimization, build a basic model set based on an automated machine learning framework, build models using a multi-layer base model stacking ensemble strategy, and complete the automated selection, combination and hyperparameter optimization of models through k-fold cross-validation to determine and retain the optimal machine learning model; S5. Complete the construction of the AI-PBPK model fusion, substitute the key parameters of physicochemical, ADME, and PK predicted by the optimal machine learning model into the classic PBPK model, standardize the configuration of the physiological parameter module of the classic PBPK model, realize the algorithm fusion and automatic parameter docking of the machine learning model and the classic PBPK model, and form an integrated AI-PBPK model. S6. Conduct performance evaluation of PK core parameter prediction. Input standardized parameters into the AI-PBPK model to obtain the predicted values ​​of PK core parameters. Use specified evaluation indicators to verify the model performance. Compare and verify the performance of clearance parameter prediction with the traditional IVIVE method. On an external independent test set, compare and verify the performance of blood drug concentration-time curve area and peak drug concentration parameter prediction with the traditional PBPK method.

2. The AI-based method for predicting physiological pharmacokinetics according to claim 1, characterized in that, The implementation process of step S1 is as follows: Key physicochemical, ADME, and PK parameter data of compounds were collected from open-source databases and public literature to construct a dataset of compound-related parameters. A dataset containing 1352 molecules was selected for the human body clearance rate prediction task, and it was divided into a training set containing 1217 molecules and a test set containing 135 molecules in a 9:1 ratio. A dataset containing 1272 molecules was selected for the human apparent permeability coefficient prediction task, and it was divided into a training set containing 1145 molecules and a test set containing 127 molecules in a 9:1 ratio. A dataset containing 1308 molecules was selected for the human homeostatic apparent distribution volume prediction task. It was divided into a training set containing 1178 molecules and a test set containing 130 molecules in a 9:1 ratio. We obtained an in vitro intrinsic clearance dataset containing the intrinsic clearance of human liver microsomes, intrinsic clearance of human hepatocytes, free drug fraction in plasma, and whole blood plasma partition ratio of 135 molecules in the test set, which was used for comparison and validation with the traditional IVIVE method. A dataset of 38 molecules from 2018 to the present was collected from publicly available literature as a new external test set. All datasets were subjected to standardized preprocessing operations, including missing value completion and data normalization.

3. The AI-based method for predicting physiological pharmacokinetics according to claim 1, characterized in that, The implementation process of step S2 is as follows: extract molecular access system fingerprints, atom pair fingerprints, and topological torsion fingerprints; use OpenBabel to directly calculate various descriptors of the molecule, including physicochemical properties, charge properties, topological properties, and three-dimensional conformational properties; and fuse the molecular access system fingerprints, atom pair fingerprints, topological torsion fingerprints, and molecular descriptors to form a multi-dimensional fused molecular feature set.

4. The AI-based method for predicting physiological pharmacokinetics according to claim 1, characterized in that, The implementation process of step S3 is as follows: Using a feature selection algorithm, the regularization operation parameters of the algorithm are configured for each molecular property prediction task. The multi-dimensional fused molecular feature set corresponding to the training set is input into the configured algorithm. The feature coefficients are compressed by L1 regularization constraint, and a subset of non-zero coefficient features with strong predictive power for each target variable is selected. Redundant features in the multi-dimensional fused molecular feature set are removed. The fully correlated feature subset is used as the input feature set for subsequent machine learning model training. Feature dimensionality reduction processing is completed and the feature selection results are recorded.

5. The AI-based method for predicting physiological pharmacokinetics according to claim 1, characterized in that, The implementation process of step S4 is as follows: A machine learning model is built based on an automated machine learning framework. A generalized linear model, a tree model, a classic machine learning model, and an artificial neural network model are set as the basic model set. The parameter ranges of each basic model are configured. A multi-layer model ensemble strategy is used to construct the machine learning model. The output results of each basic model are fused, and the fused results are input into the stacked ensemble model. Model training and ensemble operations are completed layer by layer. K-fold cross-validation is performed on the training set. During the validation process, the model combination and hyperparameter configuration are iteratively optimized. The model with the best performance during the validation process is selected as the final optimal machine learning model. The test set is input into the optimal machine learning model to complete the preliminary performance evaluation.

6. The AI-based method for predicting physiological pharmacokinetics according to claim 1, characterized in that, The implementation process of step S5 is as follows: The clearance rate, steady-state apparent volume of distribution, and apparent permeability coefficient predicted by the optimal machine learning model are automatically substituted into the classic PBPK model through the algorithm interface, replacing the drug-specific parameters obtained from animal experiments in the classic PBPK model. Based on preset parameters such as population physiological parameters, tissue and organ blood flow, tissue and organ volume, clinical trial drug dosage, and drug administration duration, the physiological parameter module of the classic PBPK model is configured in a standardized manner. The algorithm integration interface enables the fusion of machine learning models and classic PBPK models, achieving precise docking between input feature parameters and model input nodes, and forming an integrated AI-PBPK model that can directly predict pharmacokinetic parameters.

7. The AI-based method for predicting physiological pharmacokinetics according to claim 1, characterized in that, In step S6, the coefficient of determination, average conversion error, and absolute average conversion error are used as evaluation indicators to verify the model performance. The implementation process is as follows: set up For observed attribute values, The predicted value is n, where n is the sample size, calculated with base 10. and Calculate the coefficient of determination: ； In the formula, The average value of the observed attribute; Calculate the average conversion error: ； Calculate the absolute average conversion error: ； Substitute the observed and predicted values ​​of the clearance rate, steady-state apparent distribution volume, and apparent permeability coefficient into the above formulas to calculate the determination coefficient, average reduction error, and absolute average reduction error for each parameter, thus completing the performance verification of the machine learning model.

8. The AI-based method for predicting physiological pharmacokinetics according to claim 1, characterized in that, The implementation process for completing the comparison verification with the traditional IVIVE method and the external independent verification in step S6 is as follows: The AI-PBPK model was used to predict the human plasma clearance rate of 135 molecules. The in vitro inherent clearance rate data of 135 molecules were used to calculate the human plasma clearance rate using the traditional IVIVE dilution scaling method and the unbound scaling model. Automatically calculate the coefficient of determination, average conversion error, and absolute average conversion error for the three methods, and complete the comparative verification with the same dataset. The 38 compounds in the external validation set were input into the AI-PBPK model to automatically predict the area under the blood drug concentration-time curve and the core parameters of peak drug concentration PK. The error range between the predicted and measured values ​​of each parameter was calculated to complete the external independent validation of the model's generalization ability. Integrate the predicted values ​​of PK core parameters, model fitting performance indicators, comparison results with the same dataset, and external independent validation error range to form a complete report and output it.

9. The AI-based method for predicting physiological pharmacokinetics according to claim 5, characterized in that, The specific implementation of the multi-layer base model stacking integration strategy in step S4 is as follows: The input feature set is set to be connected to the first basic model layer. The basic model layer includes classic machine learning models, tree models, generalized linear models, and artificial neural networks. Each basic model operates independently and outputs prediction results. The feature splicing layer is set up to splice all the prediction results of the base model layer to form a spliced ​​feature set and pass it to the stacked model layer. The stacked model layer performs secondary training on the concatenated feature set and outputs the results. A weight allocation layer is set to weight the output results of the stacked model layer, and finally outputs the model prediction value.

10. The AI-based method for predicting physiological pharmacokinetics according to claim 1, characterized in that, It also includes the steps of building a standardized front-end operation page, which automates the standardized parameter input and submission process through the front-end page, and synchronizes the parameters entered on the front-end page to the back-end database in real time. The back-end system automatically triggers the full process calculation from S1 to S6.

Images