Method for predicting rat pharmacokinetic data based on mouse pharmacokinetic data and neural networks

Abstract

This invention discloses a method for predicting rat pharmacokinetic data based on mouse pharmacokinetic data and neural networks, belonging to the field of drug development and pharmacokinetic technology. This method collects time-dependent drug concentration data in plasma, liver, kidney, and spleen of mice after drug administration, and constructs an input feature matrix by combining this data with rat physiological parameters. After normalization and sample partitioning, an improved BP neural network is trained using the Levenberg-Marquardt algorithm, and model optimization is performed using MSE, R, and MAE. The optimal model is then input with mouse pharmacokinetic data and rat physiological parameters corresponding to unknown rat samples, outputting independent drug-time concentration data for rat plasma, liver, kidney, and spleen, and calculating PK parameters. This invention can predict drug-time concentration data in plasma and liver, kidney, and spleen tissues of animals not yet tested, reducing animal experiments, shortening the cycle, and lowering costs, providing a new tool for subsequent extrapolation PK and individualized analysis.

Description

Technical Field

[0001] This invention relates to the fields of drug development and pharmacokinetics, specifically to a method for predicting rat pharmacokinetics data based on mouse pharmacokinetics data and neural networks. Background Technology

[0002] Pharmacokinetics (PK) is a core component of preclinical drug research, used to quantitatively describe the dynamic processes of drug absorption, distribution, metabolism, and excretion (ADME) in vivo. It provides crucial evidence for candidate drug screening, dosage design, safety evaluation, and clinical translation. Rodents (mice and rats) are the most commonly used model animals for preclinical pharmacokinetic studies. Mice are low-cost, easy to raise, and require small sample sizes, making them suitable for high-throughput initial screening. Rats are closer to humans in physiological structure, metabolic enzyme systems, and drug distribution characteristics, making them a core validation model for drugs entering preclinical trials.

[0003] In actual R&D processes, pharmacokinetic experiments are typically conducted in mice to obtain preliminary data before proceeding to rat experiments. Traditional species extrapolation relies heavily on physiological pharmacokinetic (PBPK) models, compartmental models, or empirical formulas, which have significant limitations: Strong model assumptions and poor adaptability: PBPK requires a large amount of prior data on physiological parameters, tissue partition coefficients, and metabolic enzyme activities, resulting in high parameter acquisition costs and long cycles, and limited applicability to small molecule drugs and innovative structural compounds; Single prediction dimension: Most methods can only predict plasma drug-time curves, making it difficult to simultaneously obtain concentrations in target tissues / toxic target organs such as the liver, kidney, and spleen, thus failing to comprehensively assess tissue accumulation and safety risks; Low data utilization: High temporal resolution and multi-tissue concentration data generated in mouse experiments are not fully utilized, and directly conducting rat experiments results in a waste of animals, time, and funds; Insufficient prediction accuracy: Traditional linear or simple nonlinear fitting cannot characterize the complex ADME nonlinear characteristics of drugs in different species, leading to large extrapolation errors and reducing the reliability of R&D decisions.

[0004] Back Propagation (BP) neural networks, as the most widely used feedforward neural networks, employ the Levenberg-Marquardt (LM) optimization algorithm. They have fast convergence speed and high fitting accuracy, making them suitable for processing small sample, high-dimensional, and nonlinear pharmacokinetic data.

[0005] Currently, there is no method based on backpropagation (BP) neural networks to directly predict in vivo pharmacokinetic data from mouse pharmacokinetic data to multi-tissue pharmacokinetic data in rats. Existing technologies cannot stably output high-precision pharmacokinetic data and PK parameters from plasma and multiple tissues while reducing the use of rats, shortening the experimental cycle, and lowering R&D costs. Therefore, developing an efficient, accurate, and universal method for extrapolating pharmacokinetic data from rodent species is of great value for improving the efficiency of preclinical drug development and promoting the welfare of laboratory animals. Summary of the Invention

[0006] The technical problem to be solved by this invention is to address the deficiencies of the prior art by providing a method for predicting rat pharmacokinetic data based on mouse pharmacokinetic data and neural networks. An effective prediction model is established, which is based on in vivo mouse pharmacokinetic data and combined with rat physiological parameters. The model uses a BP neural network to complete sample screening, effective data amplification and nonlinear mapping learning, and quickly predicts in vivo blood drug concentration-time data and PK parameters in multiple tissues such as rat plasma, liver, kidney and spleen, effectively solving the current technical problems.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] A method for predicting rat pharmacokinetic data based on mouse pharmacokinetic data and neural networks includes the following steps:

[0009] Step S1: Constructing a mouse drug-time curve dataset: First, identify the drug, monitor and collect drug-time data of mouse plasma, liver, kidney and spleen tissues after drug injection in vivo, associate and store the collected time data and the corresponding tissue drug concentration data, and form a mouse drug-time curve dataset in chronological order; simultaneously collect the corresponding rat physiological parameters and construct the input feature matrix.

[0010] Step S2: Construct and initialize an improved BP neural network adapted for multi-tissue species extrapolation of pharmacokinetics: Build a three-layer hierarchical constraint BP neural network, which includes a pharmacokinetics feature input layer, a multi-tissue weight hierarchical hidden layer, and a multi-tissue concentration-splitting output layer.

[0011] Step S3, Dataset Preprocessing and Sample Splitting: The input and output data are normalized to the [0,1] interval using min-max normalization to eliminate dimensional differences; the normalized data is then divided into training, validation, and test sets.

[0012] Step S4: Improve BP neural network training and hierarchical weight optimization: The fusion features of drug-time data of mouse plasma, liver, kidney and spleen tissues and rat physiological parameters are used as input, and the measured drug-time data of rat plasma, liver, kidney and spleen tissues are used as output labels; during training, the weights are updated by hierarchical backpropagation, and loss weighting coefficients are added to the liver and kidney toxicity target organ branches to prioritize the constraint of target organ concentration prediction bias.

[0013] Step S5: Multi-dimensional pharmacokinetic-guided model evaluation and optimal model selection: Simultaneously calculate the global mean square error (MSE), tissue-specific correlation coefficient (R), and tissue-specific mean absolute error (MAE); set a weighted comprehensive fitting score for hepatotoxic and nephrotoxic tissues, and prioritize the network with higher goodness of fit to toxic organs as the optimal improved BP neural network; among them, the mean square error (MSE) is used to measure the overall deviation between the predicted and measured values; the correlation coefficient (R) is used to reflect the degree of fit between the predicted and measured curves of drug-time in each tissue; and the mean absolute error (MAE) is used to reflect the average deviation level of the pharmacokinetic concentration prediction results of a single tissue.

[0014] Step S6: In vivo pharmacokinetic data prediction in rats: Input the drug-time data of mice corresponding to unknown rat samples and the physiological parameters of rats into the optimal improved BP neural network. The branch outputs independent drug-time concentration data of rat plasma, liver, kidney, and spleen. Based on the concentration curves of each tissue, the elimination half-life is automatically calculated in batches. The core PK parameters are: clearance rate (CL), area under the curve (AUC), steady-state apparent volume of distribution (Vss), and maximum concentration (Cmax).

[0015] Furthermore, in step S1, the mouse drug-time curve dataset includes drug concentration data of plasma, liver, kidney, and spleen tissues at multiple time points within 0-8 hours after drug administration. Specifically, data is collected every 0.5 hours, with 13 drug concentration data points collected each time. The tissue drug concentration data is then filtered to obtain 106 drug concentration-time data points. The same drug is administered 20 times to obtain 20 sets of samples.

[0016] Furthermore, the rat physiological parameters include body weight, blood flow, distribution coefficient, total surface area of ​​tissue cells, interstitial volume, and tissue volume parameters.

[0017] Furthermore, the pharmacokinetic feature input layer integrates mouse time-series concentration features and rat physiological features for joint encoding. The input features include administration duration, drug concentrations in mouse plasma, liver, kidney, and spleen, and rat body weight, blood flow, partition coefficient, total surface area of ​​tissue cells, interstitial volume, and tissue volume parameters.

[0018] Furthermore, in step S2, a constraint function is added to the pharmacokinetic core evaluation index of the multi-tissue weighted hierarchical hidden layer. The multi-tissue weighted hierarchical hidden layer adopts the S-shaped logarithmic activation function, and the number of nodes is determined by a nonlinear empirical formula that combines first-order elimination kinetics with temporal fluctuation correction.

[0019] Furthermore, the constraint function constrains the drug concentration time-series fitting error, pharmacokinetic parameter prediction error, and multi-tissue weighted fluctuation error, replacing the single fitting loss. The specific function structure is as follows:

[0020]

[0021] in, This represents the total loss value of the model. To account for time-series concentration loss, mean squared error (MSE) was used to calculate the deviation between the predicted and actual drug concentrations at corresponding time points in mice and rats. To account for pharmacokinetic parameter losses, prediction biases are calculated for core pharmacokinetic parameters such as clearance rate, apparent volume of distribution, and half-life. The tissue-specific weight decay constraint loss is defined; α, β, and γ are preset weight coefficients, which correspond to the loss contribution ratios of concentration fitting, parameter prediction, and weight regularization, respectively, and the sum of the three is 1.

[0022] The nonlinear empirical formula is as follows:

[0023]

[0024] Where N is the initial number of candidate nodes in the multi-organization weighted hierarchical hidden layer; int is the floor function that rounds the calculation result down to the integer; T is the total number of time points for pharmacokinetic time series sampling; k el is the drug elimination rate constant; CL is the clearance rate.

[0025] Furthermore, the multi-tissue concentration-based output layer is configured with independent output neuron branches for four tissues: plasma, liver, kidney, and spleen, with each branch configured with a dedicated output bias term.

[0026] Furthermore, the improved BP neural network is trained using the Levenberg-Marquardt algorithm; the training objective of the Levenberg-Marquardt algorithm is... The number of verification checks is set to 200.

[0027] Furthermore, the dataset preprocessing process in step S3 includes sample screening, pharmacokinetic time series data amplification, and removal of abnormal concentration points.

[0028] Furthermore, in step S3, the dataset samples are randomly divided into 70% training set, 15% validation set, and 15% test set.

[0029] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0030] (1) The rat pharmacokinetic data prediction method based on mouse pharmacokinetic data and neural network provided by the present invention collects drug-time data of mouse plasma, liver, kidney and spleen tissues, and integrates the drug-time data with rat body weight, blood flow, distribution coefficient, total surface area of ​​tissue cells, interstitial volume and tissue volume parameters to construct an input feature matrix. By improving the BP neural network, the species extrapolation mapping relationship between mouse pharmacokinetic data and rat pharmacokinetic data is established. Therefore, a small number of mouse experiments can replace large-scale rat in vivo experiments, reduce the number of animals used, shorten the experimental cycle and reduce manpower and financial investment.

[0031] (2) The rat pharmacokinetic data prediction method based on mouse pharmacokinetic data and neural network provided by the present invention sets up an improved BP neural network including a pharmacokinetic feature input layer, a multi-tissue weighted hierarchical hidden layer and a multi-tissue concentration branch output layer. The multi-tissue concentration branch output layer is set up with independent output neuron branches corresponding to rat plasma, liver, kidney and spleen respectively. During the training process, loss weighting coefficients are added to the liver and kidney toxicity target organ branches. It can simultaneously output rat plasma, liver, kidney and spleen multi-tissue concentration data, covering pharmacodynamic target organs and toxicity target organs, and supporting a more comprehensive safety evaluation.

[0032] (3) The rat pharmacokinetic data prediction method based on mouse pharmacokinetic data and neural network provided by the present invention normalizes mouse drug-time data and rat physiological parameters and divides samples. It uses the Levenberg-Marquardt algorithm to train an improved BP neural network and evaluates and screens the trained network through mean squared error (MSE), correlation coefficient (R), and mean absolute error (MAE). After obtaining the optimal improved BP neural network, it directly outputs rat drug-time concentration data and further calculates... The system can quickly output rat prediction results by inputting mouse data, and it is easy to operate and highly versatile, without the need for complex PB / PK modeling and parameter fitting. Attached Figure Description

[0033] Figure 1 This is a schematic diagram illustrating the design concept of the rat pharmacokinetic data prediction method based on mouse pharmacokinetic data and neural networks of the present invention.

[0034] Figure 2 For the present invention Figure 1 The drug-time curves in mouse blood tissues involved;

[0035] Figure 3 For the present invention Figure 1 Drug-time curves of mouse liver tissue involved;

[0036] Figure 4 For the present invention Figure 1Drug-time curves of mouse kidney tissue involved;

[0037] Figure 5 For the present invention Figure 1 Drug-time curves of mouse spleen tissue involved;

[0038] Figure 6 For the present invention Figure 1 The relevant rat physiological parameters are shown in the diagram.

[0039] Figure 7 This is a flowchart of the dataset preprocessing process of the present invention;

[0040] Figure 8 This is a schematic diagram of the BP neural network structure of the present invention;

[0041] Figure 9 This is a curve showing the comparison between measured and predicted values ​​of rat blood tissue in this invention;

[0042] Figure 10 This is a curve showing the comparison between measured and predicted values ​​of rat liver tissue in this invention;

[0043] Figure 11 This is a curve showing the comparison between measured and predicted values ​​of rat kidney tissue in this invention;

[0044] Figure 12 This is a curve showing the comparison between the measured and predicted values ​​of rat spleen tissue in this invention. Detailed Implementation

[0045] 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.

[0046] Methods for predicting rat pharmacokinetic data based on mouse pharmacokinetic data and neural networks, such as Figure 1 As shown, it includes the following steps:

[0047] Step S1: Construct a mouse drug-time curve dataset. The mouse drug-time curve is in the form of... Figure 2-5 As shown: First, the drug is identified, and drug-time data of mouse plasma, liver, kidney, and spleen tissues are collected in vivo after drug injection. The collected time data and corresponding tissue drug concentration data are associated and stored, and a mouse drug-time curve dataset is formed in chronological order. Simultaneously, the physiological parameters of rats corresponding to the drug are collected to construct the input feature matrix.

[0048] Step S2: Construct and initialize an improved BP neural network adapted for multi-tissue species extrapolation of pharmacokinetics: Build a three-layer hierarchical constraint BP neural network, which includes a pharmacokinetics feature input layer, a multi-tissue weight hierarchical hidden layer, and a multi-tissue concentration-splitting output layer.

[0049] Step S3, Dataset Preprocessing and Sample Splitting: The input and output data are normalized to the [0,1] interval using min-max normalization to eliminate dimensional differences; the normalized data is then divided into training, validation, and test sets.

[0050] Step S4: Improve BP neural network training and hierarchical weight optimization: The fusion features of drug-time data of mouse plasma, liver, kidney and spleen tissues and rat physiological parameters are used as input, and the measured drug-time data of rat plasma, liver, kidney and spleen tissues are used as output labels; during training, the weights are updated by hierarchical backpropagation, and loss weighting coefficients are added to the liver and kidney toxicity target organ branches to prioritize the constraint of target organ concentration prediction bias.

[0051] Step S5: Multi-dimensional pharmacokinetic-guided model evaluation and optimal model selection: Simultaneously calculate the global mean square error (MSE), tissue-specific correlation coefficient (R), and tissue-specific mean absolute error (MAE); set a weighted comprehensive fitting score for hepatotoxic and nephrotoxic tissues, and prioritize the network with higher goodness of fit to toxic organs as the optimal improved BP neural network; among them, the mean square error (MSE) is used to measure the overall deviation between the predicted and measured values; the correlation coefficient (R) is used to reflect the degree of fit between the predicted and measured curves of drug-time in each tissue; and the mean absolute error (MAE) is used to reflect the average deviation level of the pharmacokinetic concentration prediction results of a single tissue.

[0052] Step S6: In vivo pharmacokinetic data prediction in rats: Input the drug-time data of mice corresponding to unknown rat samples and the physiological parameters of rats into the optimal improved BP neural network. The branch outputs independent drug-time concentration data of rat plasma, liver, kidney, and spleen. Based on the concentration curves of each tissue, the elimination half-life is automatically calculated in batches. The core PK parameters are: clearance rate (CL), area under the curve (AUC), steady-state apparent volume of distribution (Vss), and maximum concentration (Cmax).

[0053] Furthermore, in step S1, the mouse drug-time curve dataset includes drug concentration data of plasma, liver, kidney, and spleen tissues at multiple time points within 0-8 hours after drug administration. Specifically, data is collected every 0.5 hours, with 13 drug concentration data points collected each time. The tissue drug concentration data is then filtered to obtain 106 drug concentration-time data points. The same drug is administered 20 times to obtain 20 sets of samples.

[0054] Furthermore, the rat physiological parameters include body weight, blood flow, distribution coefficient, total surface area of ​​tissue cells, interstitial volume, and tissue volume parameters, such as... Figure 6 As shown.

[0055] Furthermore, the pharmacokinetic feature input layer integrates mouse time-series concentration features and rat physiological features for joint encoding. The input features include administration duration, drug concentrations in mouse plasma, liver, kidney, and spleen, and rat body weight, blood flow, partition coefficient, total surface area of ​​tissue cells, interstitial volume, and tissue volume parameters.

[0056] Furthermore, in step S2, a constraint function is added to the pharmacokinetic core evaluation index of the multi-tissue weighted hierarchical hidden layer. The multi-tissue weighted hierarchical hidden layer adopts the S-shaped logarithmic activation function, and the number of nodes is determined by a nonlinear empirical formula that combines first-order elimination kinetics with temporal fluctuation correction.

[0057] Furthermore, the constraint function constrains the drug concentration time-series fitting error, pharmacokinetic parameter prediction error, and multi-tissue weighted fluctuation error, replacing the single fitting loss. The specific function structure is as follows:

[0058]

[0059] in, This represents the total loss value of the model. To account for time-series concentration loss, mean squared error (MSE) was used to calculate the deviation between the predicted and actual drug concentrations at corresponding time points in mice and rats. To account for pharmacokinetic parameter losses, prediction biases are calculated for core pharmacokinetic parameters such as clearance rate, apparent volume of distribution, and half-life. The tissue-specific weight decay constraint loss is defined; α, β, and γ are preset weight coefficients, which correspond to the loss contribution ratios of concentration fitting, parameter prediction, and weight regularization, respectively, and the sum of the three is 1.

[0060] The nonlinear empirical formula is as follows:

[0061]

[0062] Where N is the initial number of candidate nodes in the multi-organization weighted hierarchical hidden layer; int is the floor function that rounds the calculation result down to the integer; T is the total number of time points for pharmacokinetic time series sampling; k el is the drug elimination rate constant; CL is the clearance rate.

[0063] Furthermore, the multi-tissue concentration-based output layer is configured with independent output neuron branches for four tissues: plasma, liver, kidney, and spleen, with each branch configured with a dedicated output bias term.

[0064] Furthermore, the improved BP neural network is trained using the Levenberg-Marquardt algorithm; the training objective of the Levenberg-Marquardt algorithm is... The number of verification checks is set to 200.

[0065] Furthermore, the dataset preprocessing process in step S3 includes sample screening, pharmacokinetic time series data amplification, and removal of abnormal concentration points.

[0066] Furthermore, in step S3, the dataset samples are randomly divided into 70% training set, 15% validation set, and 15% test set.

[0067] To better illustrate the specific implementation methods, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0068] The data preprocessing flowchart involved in the embodiments of the present invention is as follows: Figure 7 As shown: First, the drug to be studied should be identified. Taking doxorubicin as an example, the experimental subjects should be determined as ICR mice. Then, in vivo experiments should be conducted by injecting 10 mg / kg of doxorubicin via the tail vein. Blood drug concentration data in the jugular vein, liver, kidney, and spleen of the mice should be continuously monitored for 8 hours. The data should be stored in an Excel spreadsheet, with the time column as the first column. The blood drug concentration data should then be arranged sequentially according to blood, liver, kidney, and spleen, resulting in 106 drug-time points. The experiment should be repeated at least 20 times to obtain 20 sets of samples. If a larger sample size is required, the data can be expanded to 200 or 1000 sets.

[0069] The schematic diagram of the BP neural network structure involved in the embodiments of the present invention is shown below. Figure 8 As shown: the data input consists of five quantities: time, blood, liver, kidney, and spleen. Therefore, the network input and output are set to 5 neurons. During training, a complete training sample set must be maintained, and optimization is achieved by adjusting parameters such as the learning rate. In the algorithm of this invention, the network training accuracy is set to 0.0001, the maximum number of iterations is 1000, the learning rate is 0.1, and the algorithm training time range is 18s < t < 25s.

[0070] like Figure 9-12 The image shows the CT curves of rats in various tissues predicted by this invention: A new set of mouse data is input into the program as a test set, and CT curves of the rat's blood, liver, kidney, and spleen are obtained. The predicted values ​​(red curve) and actual values ​​(blue curve) show a significant consistency in their trends. The final training results are shown below. Figure 9 , Figure 10 , Figure 11 , Figure 12 As shown.

[0071] Table 1

[0072]

[0073] Table 1 shows the fitness index of the BP neural network model involved in the embodiments of the present invention in four tissues: the goodness of fit of all tissues is above 0.9, among which the blood data has the highest goodness of fit, R 2 =0.99, and the numerical concentration difference between each group was between 0.0142 and 1.904, indicating that the trained model has high accuracy in predicting data from mice to rats.

[0074] Table 2

[0075]

[0076] Table 2 shows a comparison between the actual monitored PK parameters and the predicted PK parameters from the neural network model in rats involved in the embodiments of the present invention: In four tissues of rats—blood, liver, kidney, and spleen—the actual detected PK parameters and the PK parameters predicted by the neural network model, covering the half-life (t... 1 / 2 The core pharmacokinetic parameters included clearance rate (CL), area under the curve (AUC), apparent volume of distribution (Vss), and peak concentration (Cmax). Overall, the model-predicted PK parameters for each tissue closely matched the actual monitored values ​​in rats. There were no significant numerical deviations in key pharmacokinetic parameters for different organs such as blood, liver, kidney, and spleen. Whether it was the half-life reflecting drug elimination rate, clearance rate reflecting drug clearance capacity, or the AUC, Vss, and peak concentration Cmax representing drug exposure, the predicted and measured values ​​showed completely consistent trends.

[0077] Table 3

[0078]

[0079] Table 3 shows the absolute error statistics of the PK parameters involved in the embodiments of the present invention: it quantitatively presents the absolute errors of the prediction results of the five PK parameters in each tissue, clearly reflecting the deviation level and accuracy differences of the model prediction data. The data shows that the overall absolute error range of all parameters is 0.05~1.783, indicating a relatively low overall error level; among them, the half-life index has an error range of only 0.011~0.703, making it the index with the smallest error and best stability among all pharmacokinetic parameters. Looking at individual tissues, the absolute errors of the PK parameters in the spleen are generally low, while the errors in the kidney, liver, and blood are slightly higher, but none show significant deviations. This further demonstrates that the model has controllable errors and good stability in multi-tissue drug metabolism prediction, and the model prediction results have good accuracy.

[0080] Those skilled in the art should understand that, unless otherwise specified, the meanings of the technical and scientific terms used herein are consistent with the general understanding of the relevant technical field. Furthermore, terms defined in general dictionaries should be understood in the context of the technical background in this field and should not be interpreted in an overly idealized or formalistic manner divorced from practical application scenarios.

[0081] The above embodiments have described in detail the main concept, technical solution, and technical effects of the present invention. It should be noted that the above content is merely illustrative and not intended to limit the scope of protection of the present invention. Any equivalent modifications, substitutions, or optimizations based on the present invention without departing from its core principles are within the scope of the present invention.

Claims

1. A method for predicting rat pharmacokinetic data based on mouse pharmacokinetic data and neural networks, characterized in that, Includes the following steps: Step S1: Constructing a mouse drug-time curve dataset: First, identify the drug, monitor and collect drug-time data of mouse plasma, liver, kidney and spleen tissues after drug injection in vivo, associate and store the collected time data and the corresponding tissue drug concentration data, and form a mouse drug-time curve dataset in chronological order; simultaneously collect the corresponding rat physiological parameters and construct the input feature matrix. Step S2: Construct and initialize an improved BP neural network adapted for multi-tissue species extrapolation of pharmacokinetics: Build a three-layer hierarchical constraint BP neural network, which includes a pharmacokinetics feature input layer, a multi-tissue weight hierarchical hidden layer, and a multi-tissue concentration-splitting output layer. Step S3, Dataset Preprocessing and Sample Splitting: The input and output data are normalized to the [0,1] interval using min-max normalization to eliminate dimensional differences; the normalized data is then divided into training, validation, and test sets. Step S4: Improve BP neural network training and hierarchical weight optimization: Use the fusion features of drug-time data of mouse plasma, liver, kidney and spleen tissues and rat physiological parameters as input, and use the measured drug-time data of rat plasma, liver, kidney and spleen tissues as output labels; During training, weights are updated via hierarchical backpropagation, and loss weighting coefficients are added to the target organ branches for hepatotoxicity and nephrotoxicity to prioritize constrain the prediction bias of target organ concentration. Step S5: Multi-dimensional pharmacokinetic-guided model evaluation and optimal model selection: Simultaneously calculate the global mean square error (MSE), tissue-specific correlation coefficient (R), and tissue-specific mean absolute error (MAE); set a weighted comprehensive fitting score for hepatotoxic and nephrotoxic tissues, and prioritize the network with higher goodness of fit to toxic organs as the optimal improved BP neural network; among them, the mean square error (MSE) is used to measure the overall deviation between the predicted and measured values; the correlation coefficient (R) is used to reflect the degree of fit between the predicted and measured curves of drug-time in each tissue; and the mean absolute error (MAE) is used to reflect the average deviation level of the pharmacokinetic concentration prediction results of a single tissue. Step S6: In vivo pharmacokinetic data prediction in rats: Input the drug-time data of mice corresponding to unknown rat samples and the physiological parameters of rats into the optimal improved BP neural network. The branch outputs independent drug-time concentration data of rat plasma, liver, kidney, and spleen. Based on the concentration curves of each tissue, the elimination half-life is automatically calculated in batches. The core PK parameters are: clearance rate (CL), area under the curve (AUC), steady-state apparent volume of distribution (Vss), and maximum concentration (Cmax).

2. The method for predicting rat pharmacokinetic data based on mouse pharmacokinetic data and neural networks according to claim 1, characterized in that, In step S1, the mouse drug-time curve dataset includes drug concentration data of plasma, liver, kidney, and spleen tissues at multiple time points within 0-8 hours after drug administration. Specifically, data is collected every 0.5 hours, with 13 drug concentration data points collected each time. The tissue drug concentration data is then filtered to obtain 106 drug concentration-time data points. The same drug is administered 20 times to obtain 20 sets of samples.

3. The method for predicting rat pharmacokinetic data based on mouse pharmacokinetic data and neural networks according to claim 1, characterized in that, The rat physiological parameters include body weight, blood flow, distribution coefficient, total surface area of ​​tissue cells, interstitial volume, and tissue volume parameters.

4. The method for predicting rat pharmacokinetic data based on mouse pharmacokinetic data and neural networks according to claim 1, characterized in that, The pharmacokinetic feature input layer integrates mouse time-series concentration features and rat physiological features for joint encoding. The input features include administration duration, drug concentrations in mouse plasma, liver, kidney, and spleen, and rat body weight, blood flow, partition coefficient, total surface area of ​​tissue cells, interstitial volume, and tissue volume parameters.

5. The method for predicting rat pharmacokinetic data based on mouse pharmacokinetic data and neural networks according to claim 1, characterized in that, In step S2, a constraint function is added to the multi-tissue weighted hierarchical hidden layer to integrate the pharmacokinetic core evaluation index. The multi-tissue weighted hierarchical hidden layer adopts the S-type logarithmic activation function, and the number of nodes is determined by a nonlinear empirical formula that combines first-order elimination kinetics with time-series fluctuation correction.

6. The method for predicting rat pharmacokinetic data based on mouse pharmacokinetic data and neural networks according to claim 5, characterized in that, The constraint function constrains the drug concentration time-series fitting error, pharmacokinetic parameter prediction error, and multi-tissue weighted fluctuation error, replacing the single fitting loss. The specific function structure is as follows: in, This represents the total loss value of the model. To account for time-series concentration loss, mean squared error (MSE) was used to calculate the deviation between the predicted and actual drug concentrations at corresponding time points in mice and rats. To account for pharmacokinetic parameter losses, prediction biases are calculated for core pharmacokinetic parameters such as clearance rate, apparent volume of distribution, and half-life. The tissue-specific weight decay constraint loss is defined; α, β, and γ are preset weight coefficients, which correspond to the loss contribution ratios of concentration fitting, parameter prediction, and weight regularization, respectively, and the sum of the three is 1. The nonlinear empirical formula is as follows: Where N is the initial number of candidate nodes in the multi-organization weighted hierarchical hidden layer; int is the floor function that rounds the calculation result down to the integer; T is the total number of time points for pharmacokinetic time series sampling; k el is the drug elimination rate constant; CL is the clearance rate.

7. The method for predicting rat pharmacokinetic data based on mouse pharmacokinetic data and neural networks according to claim 1, characterized in that, The multi-tissue concentration-based output layer is configured with independent output neuron branches for four tissues: plasma, liver, kidney, and spleen, with each branch configured with a dedicated output bias term.

8. The method for predicting rat pharmacokinetic data based on mouse pharmacokinetic data and neural networks according to claim 1, characterized in that, The improved BP neural network is trained using the Levenberg-Marquardt algorithm; the training objective of the Levenberg-Marquardt algorithm is... The number of verification checks is set to 200.

9. The method for predicting rat pharmacokinetic data based on mouse pharmacokinetic data and neural networks according to claim 1, characterized in that, The dataset preprocessing process in step S3 includes sample screening, pharmacokinetic time series data amplification, and removal of abnormal concentration points.

10. The method for predicting rat pharmacokinetic data based on mouse pharmacokinetic data and neural networks according to claim 1, characterized in that, In step S3, the dataset samples are randomly divided into 70% training set, 15% validation set, and 15% test set.

Images