A pharmacokinetics automatic modeling method and system based on a unified mechanism skeleton and machine learning assistance

Abstract

This invention relates to an automatic pharmacodynamic modeling method and system based on a unified mechanism framework and machine learning assistance, belonging to the field of pharmacodynamic analysis technology. It solves the problem in existing technologies of lacking a method for constructing pharmacodynamic models that both conforms to physiological mechanisms and is applicable to most pharmacodynamic models. The method includes: constructing a unified mechanism pharmacodynamic model framework; acquiring target pharmacodynamic study data; inputting the target pharmacodynamic study data into a trained structure recognition model to predict the driving type, mechanism of action, and initial modeling parameters of the target pharmacodynamic study; obtaining candidate structures for the target pharmacodynamic study based on the predicted driving type and mechanism of action; performing parameter fitting on each candidate structure under the unified mechanism pharmacodynamic model framework based on the target pharmacodynamic study data and initial modeling parameters; determining the optimal candidate structure based on the parameter fitting, thus obtaining the pharmacodynamic model of the target pharmacodynamic study. This significantly improves modeling efficiency, result interpretability, and the clarity of physiological mechanisms.

Description

Technical Field

[0001] This invention relates to the field of pharmacodynamic analysis technology, and in particular to an automatic pharmacodynamic modeling method and system based on a unified mechanism framework and machine learning assistance. Background Technology

[0002] Pharmacodynamic models are used to describe the quantitative relationship between drug exposure and biological effects, and can be further used for pharmacodynamic mechanism analysis, parameter estimation, dosage regimen optimization, exposure-effect analysis, population difference assessment, clinical translational research, and disease progression evaluation. These models are widely used in new drug development, preclinical pharmacodynamic evaluation, population pharmacology analysis, model-guided dosing, and mechanism explanation.

[0003] Currently, the existing technologies related to this invention can be mainly categorized into the following two types.

[0004] The first category is the current modeling approach, which relies primarily on manual modeling based on researchers' experience and following a quantitative pharmacology workflow. In this category of existing techniques, researchers typically develop models around several common pharmacodynamic model families, and then combine these with population data for parameter estimation, covariate analysis, model diagnosis, simulation, and dosage optimization.

[0005] Researchers need to select an appropriate structure from candidate model families such as direct-effect models, indirect-response models, effect-compartment / biological-phase models, and TMDD models, based on the specific drug, endpoint, and research question. They then perform combination, parameter estimation, covariate screening, goodness-of-fit evaluation, visualization diagnostics, and simulation analysis. For different projects, drugs, and endpoints, it is often necessary to rebuild the model structure, reset the parameter system, and reorganize the modeling process.

[0006] This type of method has the following problems: (1) There are many families of pharmacodynamic models. When faced with different pharmacodynamic phenomena and data characteristics, it is often necessary to iterate and optimize the model repeatedly. When changing the scenario, it is usually necessary to reorganize the model structure, parameter system and implementation process. The overall process is cumbersome and time-consuming.

[0007] (2) When multiple mechanisms are superimposed, the workload of model combination and iterative optimization increases significantly, the modeling complexity rises rapidly, and it may even lead to failure to complete the task on schedule.

[0008] (3) It has high professional requirements for modelers, relies heavily on rich PK / PD experience, mathematical modeling ability and software implementation ability, and has a high threshold for use.

[0009] (4) The results are highly dependent on the modeler’s personal experience and habits, and the standardization is insufficient, which is not conducive to the formation of an automated modeling system that can be used in batches.

[0010] The second category is artificial intelligence or machine learning-assisted solutions. In recent years, artificial intelligence and machine learning methods have begun to be used to assist in PK / PD modeling. The typical implementation of this type of existing technology is to directly learn response patterns from time series data using statistical learning models or neural network models, or to use machine learning methods to assist in covariate selection, parameter prediction, and model search, focusing on prediction accuracy, time series coding capabilities, irregular sampling adaptability, or partial automation capabilities.

[0011] This approach has the following problems: (1) The output is often a predicted value, classification label or implicit feature, lacking a clear mechanistic structure. It is difficult to directly give pharmacologically significant parameter results. For researchers and reviewers, the interpretability is insufficient.

[0012] (2) It is difficult to naturally connect with the structure selection and parameter fitting process of traditional mechanism models.

[0013] (3) Although using neural network models can improve the fitting effect, the final model has poor interpretability and is difficult to be accepted by drug regulatory authorities. Therefore, it is also difficult for companies to use it to support the development of innovative drugs.

[0014] In summary, there are currently two main approaches: artificial pharmacodynamic models for population pharmacokinetic / pharmacodynamic modeling and machine learning or deep learning methods. While each can address some issues, current pharmacodynamic modeling practices lack a highly efficient method for constructing pharmacodynamic models that both conforms to physiological mechanisms and is applicable to the vast majority of pharmacodynamic models. This results in high barriers to entry and high costs for pharmacodynamic models, hindering their application. Summary of the Invention

[0015] Based on the above analysis, the embodiments of the present invention aim to provide an automatic pharmacodynamic modeling method and system based on a unified mechanistic framework and machine learning assistance, in order to solve the problem of the lack of existing pharmacodynamic model construction methods that both conform to physiological mechanisms and are applicable to the vast majority of pharmacodynamic models.

[0016] On one hand, embodiments of the present invention provide an automatic pharmacodynamic modeling method based on a unified mechanistic framework and machine learning assistance, comprising the following steps: Construct a unified framework for mechanism-based pharmacodynamics; Input the target efficacy study data into the trained structure recognition model to predict the driving type, mechanism of action, and initial modeling parameters of the target efficacy study; Candidate structures for target efficacy studies are obtained based on the predicted driving type and mechanism of action; Based on the target efficacy study data and the initial modeling parameters, parameter fitting is performed on each candidate structure under the framework of the unified mechanism pharmacodynamic model; the optimal candidate structure is determined according to the parameter fitting, and the pharmacodynamic model of the target efficacy study is obtained.

[0017] Based on the above method, a further improvement is made to obtain the trained structure recognition model in the following way: A training sample set was constructed by collecting multiple historical pharmacodynamic models and corresponding research data. A multi-task neural network model is constructed; the multi-task neural network model is used to predict the driving type, mechanism of action, and modeling parameters of a drug efficacy model based on research data; The neural network model is trained on the training sample set based on the multi-task training loss to obtain a trained structure recognition model.

[0018] Based on further improvements to the above methods, candidate structures for target efficacy studies are obtained based on predicted driving types and mechanisms of action, including: Choose the one with the higher prediction probability Each driver type is a candidate driver type; For each element in the predicted mechanism result, if the value of the element is greater than the upper limit of the uncertainty interval, the value of the element is set to 1; if the value of the element is less than the lower limit of the uncertainty interval, the value of the element is set to 0. For each element in the uncertainty interval of the mechanism of action result, the candidate mechanism of action result is obtained by transforming the element to 0 or 1; By combining candidate action mechanisms and candidate driver types, candidate structures are obtained.

[0019] Further improvements to the above method, after combining candidate action mechanisms and candidate driving types to obtain candidate structures, include the following steps before parameter fitting: Extract the scores with higher overall scores The candidate structures were selected as the final candidate structures; The comprehensive score of a candidate structure is the sum of the predicted probability of the candidate driving type in the candidate structure and the original predicted probability of the element with a value of 1 in the candidate action mechanism.

[0020] Based on further improvements to the above method, the objective function for parameter fitting is: ; Where K represents the number of efficacy endpoints in the target efficacy study. This represents the normalized root mean square error at the k-th efficacy endpoint. Indicates a sparse penalty term. This indicates a stability penalty term. , and This represents the weighting coefficient.

[0021] Based on the further improvement of the above method, the normalized root mean square error is calculated using the following formula: ; in, This represents the number of observation points for the j-th subject at the k-th efficacy endpoint, and M represents the number of subjects. This represents the observation value of the j-th subject at the k-th efficacy endpoint at the i-th time point. This represents the parameter vector to be fitted under the current candidate structure. This represents the predicted value of the j-th subject at the k-th efficacy endpoint at the i-th time point, obtained from the current fitting parameters. It represents the standard deviation of the observed values ​​at the k-th efficacy endpoint.

[0022] Based on the further improvement of the above method, the stability penalty term is calculated using the following formula: ; in, Indicates the number of parameters to be fitted. Let represent the j-th parameter to be fitted; K represents the number of efficacy endpoints in the target efficacy study. M represents the number of observation points for the j-th subject at the k-th efficacy endpoint, and M represents the number of subjects. This represents the predicted value of the j-th subject at the i-th time point of the k-th efficacy endpoint, given the current candidate structure and current fitting parameters. This represents the predicted value of the j-th subject at the (i-1)-th time point of the k-th efficacy endpoint, given the current candidate structure and current fitting parameters. This represents the predicted value of the j-th subject at the (i+1)-th time point of the k-th efficacy endpoint, given the current candidate structure and current fitting parameters.

[0023] Based on the further improvement of the above method, the sparsity penalty term is calculated using the following formula: ; Where Q represents the number of parameters. This represents the q-th active parameter in the current candidate structure. This represents the normalized weight of the q-th active parameter.

[0024] Based on further improvements to the above method, the optimal candidate structure is determined according to parameter fitting, resulting in a pharmacokinetic model for the target efficacy study, including: For each candidate structure, find the set of fitting parameters that minimizes the objective function as the optimal parameters for that candidate structure; Calculate the normalized root mean square error of each candidate structure under its optimal parameters, and select the candidate structure with the smallest normalized root mean square error as the optimal candidate structure. The optimal candidate structure and its corresponding optimal parameters constitute the pharmacokinetic model for the study of the target drug efficacy.

[0025] On the other hand, embodiments of the present invention provide an automated pharmacodynamic modeling system based on a unified mechanistic framework and machine learning assistance, comprising: The skeleton construction module is used to build a unified mechanism pharmacodynamic model skeleton; The model structure prediction module is used to input the target efficacy study data into the trained structure recognition model to predict the driving type, mechanism of action, and initial modeling parameters of the target efficacy study; and to obtain candidate structures for the target efficacy study based on the predicted driving type and mechanism of action. The parameter fitting module, based on the target efficacy study data and the initial modeling parameters, performs parameter fitting on each candidate structure under the framework of the unified mechanism pharmacodynamic model; the optimal candidate structure is determined according to the parameter fitting, and the pharmacodynamic model of the target efficacy study is obtained.

[0026] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: 1. By constructing a unified mechanistic framework and combining intelligent structure recognition and parameter fitting, the traditional process of building pharmacodynamic models that relies on human experience and repeated trial and error is transformed into an automated, standardized, and repeatable intelligent recommendation process, which significantly improves modeling efficiency, interpretability of results, and clarity of physiological mechanisms.

[0027] 2. The traditional paradigm of pharmacodynamic and biological modeling—selecting model structure, estimating parameters, and iteratively optimizing—is transformed into a new, efficient, and high-quality paradigm that establishes a unified model framework and intelligently recommends model structure and estimated parameters.

[0028] 3. Organize the various PD mechanisms that were originally scattered across multiple model families into a unified mechanism framework.

[0029] 4. Under a unified framework, the entire process of pharmacodynamic model structure identification, parameter fitting, result screening and result output is completed automatically, reducing the reliance of traditional pharmacodynamic modeling on human experience and repeated trial and error.

[0030] 5. By introducing sparse constraints and stability constraints through a unified objective function, unnecessary erroneous activation of mechanisms and abnormal parameter combinations are reduced.

[0031] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0032] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 This is a flowchart of an embodiment of the pharmacodynamic automatic modeling method based on a unified mechanism framework and machine learning assistance according to an embodiment of the present invention; Figure 2 This is a pharmacodynamic model structure diagram of the unified mechanism framework in the embodiments of the present invention; Figure 3 This is a schematic diagram illustrating the effect of a model in an embodiment of the present invention; Figure 4 This is a block diagram of the pharmacodynamic automatic modeling system based on a unified mechanism framework and machine learning assistance, according to an embodiment of the present invention. Detailed Implementation

[0033] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0034] A specific embodiment of the present invention discloses an automatic pharmacodynamic modeling method based on a unified mechanistic framework and machine learning assistance, such as... Figure 1 As shown, it includes the following steps: S1. Construct a unified mechanism-based pharmacodynamic model framework; S2. Input the target efficacy study data into the trained structure recognition model to predict the driving type, mechanism of action, and initial modeling parameters of the target efficacy study; S3. Based on the predicted driving type and mechanism of action, candidate structures for target efficacy studies are obtained; S4. Based on the target efficacy study data and the initial modeling parameters, perform parameter fitting on each candidate structure under the unified mechanism pharmacodynamic model framework; determine the optimal candidate structure based on the parameter fitting to obtain the pharmacodynamic model for the target efficacy study.

[0035] Compared with existing technologies, the pharmacodynamic automatic modeling method based on a unified mechanism framework and machine learning assistance provided in this embodiment incorporates pharmacodynamic models into a unified expression framework by constructing a unified mechanism pharmacodynamic model framework. This solves the problems of existing technologies where there are many families of pharmacodynamic models, and models often need to be repeatedly selected, reconstructed, and compared for different pharmacodynamic phenomena. Changing the scenario usually requires reorganizing the model structure, parameter system, and implementation process, resulting in a cumbersome and time-consuming overall process. At the same time, it provides a foundation for the final model structure and parameter results to have a clear physiological mechanism. Based on the trained structure recognition model, it predicts the driving type and mechanism of action of the target pharmacodynamic research, thereby transforming the traditional experience-based structure selection process into a computable and repeatable one. This automated process eliminates reliance on human experience in structure selection. Candidate structures for target pharmacodynamic studies are obtained based on predicted driving types and mechanisms of action. Using data from these studies, parameters are fitted to each candidate structure within a unified mechanistic framework. The candidate structure with the optimal parameter fit is then selected as the pharmacodynamic model for the target pharmacodynamic study. This organizes various pharmacodynamic (PD) mechanisms, previously scattered across multiple model families, into a unified framework. The approach to pharmacodynamic and biological modeling shifts from the traditional paradigm of selecting model structures, estimating parameters, and iteratively optimizing to a new, efficient, and high-quality paradigm of establishing a unified model framework and intelligently recommending model structures and estimating parameters. This reduces the dependence of traditional pharmacodynamic modeling on human experience and repeated trial and error. By combining machine learning results with mechanistic modeling, the model achieves a clear mechanistic structure, pharmacologically meaningful parameter outputs, and high interpretability, facilitating understanding and validation by researchers and reviewers.

[0036] By constructing a unified mechanistic framework and combining intelligent structure recognition and parameter fitting, the traditional process of building pharmacodynamic models that relies on human experience and repeated trial and error is transformed into an automated, standardized, and repeatable intelligent recommendation process, which significantly improves modeling efficiency, interpretability of results, and clarity of physiological mechanisms.

[0037] During implementation, a unified mechanism pharmacodynamic model framework (PD Union Model) is constructed. Based on the unified mechanism framework, a continuous kinetic equation system that can cover multiple pharmacodynamic mechanisms is established as a common basis for subsequent structure identification and parameter fitting.

[0038] It should be noted that the pharmacodynamics of this embodiment is population-based.

[0039] Specifically, the unified mechanism pharmacodynamic model framework provided in this embodiment is not a simple splicing of several independent models, but rather a unified relationship between drug exposure, receptor action, signal transduction and pharmacodynamic output pre-established within the same kinetic matrix, so that different pharmacodynamic scenarios can be expressed in the same structure.

[0040] Specifically, the unified mechanism pharmacodynamic model framework includes an absorption and system exposure module, a receptor module, a drug input module, a first delay module, a master pharmacodynamic function module, a master pharmacodynamic state module, a downstream pharmacodynamic state module, a second delay module, a feedback module, a rhythm modulation module, a biological phase module, a direct effect module, a disease state module, and a second PD axis module.

[0041] The absorption and systemic exposure module characterizes the process by which a drug enters the body and forms systemic exposure. The following equations represent absorption and systemic exposure: ; ; in, Indicates the amount of drug in the absorption chamber. denoted by , where t represents the first-order absorption rate constant, and t represents the time variable. Indicates the drug concentration in the central compartment. Indicates the volume distribution parameters of the central chamber. This represents the first-order elimination rate constant of the drug in the central compartment. This represents the rate constant of drug-receptor binding. Indicates the concentration of free receptors or free targets. This represents the dissociation rate constant of the drug-receptor complex.

[0042] This set of equations describes the continuous kinetic process of a drug entering the body and forming systemic exposure after administration. The absorption chamber equation (Equation (1)) represents the drug's initial entry into the absorption phase and its subsequent transfer from the absorption chamber at a first-order rate. The central chamber equation (Equation (2)) represents the drug's distribution and clearance within the body after entering the central chamber, and its further participation in receptor binding. This equation set is constructed because the process from drug administration to systemic exposure can typically be characterized by a continuous input, continuous elimination, and continuous binding kinetic relationship. By unifying the upstream processes of absorption, systemic exposure, and receptor binding into the same set of equations, this approach aligns with physiological realities and existing pharmacokinetic representation methods, providing a consistent exposure basis for subsequent drug input, delayed drug delivery, and drug effect formation.

[0043] Equation (1) reflects the absorption, input, and transport processes after drug administration. Equation (2) reflects drug exposure under the influence of input, clearance, and binding.

[0044] The receptor module is used to characterize changes in the drug-target binding and dissociation complex, as well as the receptor's own turnover process. The following set of equations is used for characterization: ; ; This set of equations describes the formation and elimination of receptors, as well as the processes of drug binding, dissociation, and complex changes with receptors. This mathematical construction of the equations expresses the physiological interaction between drugs and receptors, and the changes in the receptors themselves, in accordance with existing pharmacokinetic principles. By explicitly incorporating the free receptor and complex states into a unified mechanistic framework, a driving foundation is laid for subsequent drug action. Equation (3) reflects receptor formation, disappearance, and occupancy. Equation (4) represents the dynamic changes of the bound target.

[0045] in, This represents the rhythm modulation function at the receptor generation end, used to calculate the generation value of free receptor or free target concentration at time t. This represents the receptor disappearance rate constant. Indicates drug-receptor complex concentration. This represents the rate constant of disappearance of the drug-receptor complex.

[0046] The drug input module is used to characterize the upstream factors that drive downstream drug efficacy. In implementation, it employs... The drug-driven type at time t includes the following five types: 、 、 、 and .

[0047] in, This represents the level of free receptors or free targets at time t. This represents the level of the drug-receptor complex at time t. This represents the receptor occupancy rate at time t. This represents the unoccupancy rate of the receptor at time t. This represents the central compartment drug concentration at time t. These are the five most common drug-driven types, covering the vast majority of drug-driven scenarios.

[0048] The first delay module describes the hysteresis and smoothing process as the drug action signal is transmitted to the pharmacodynamic layer. Specifically, it is characterized by the following equation: ; ; ; This set of equations describes the hysteresis process as the drug action signal gradually propagates from the upstream driver to the pharmacological layer. The reason for using a chain-like delay equation is that there is a continuous transmission and gradual establishment time process between drug action and pharmacological response. By setting multiple delay states, this invention can more naturally characterize the non-instantaneous changes before the formation of pharmacological effect.

[0049] in, This indicates the drug concentration in the first transfer chamber of the first delay module, which, together with the latter two transfer chambers, describes the hysteresis before the drug reaches the pharmacodynamic layer; in the biological phase model, As the concentration in the effect room. This represents the transmission rate constant of the first delay module. This represents the biological phase equilibrium rate constant.

[0050] Indicates the switch for the biological phase module. This indicates that the biophase module is enabled. This indicates that the biophase module is closed.

[0051] This indicates the drug concentration in the second transfer chamber of the first delay module. This indicates the drug concentration in the third transfer chamber of the first delay module, which is the effect chamber concentration under normal conditions, and the drug concentration that achieves its effect after hysteresis.

[0052] The principal pharmacokinetics function module is used to map the effect-site drug concentration to the effective action intensity of the principal pharmacokinetics state. The specific function is represented as follows: ; ; in, This represents the actual driving force of the principal drug efficacy state function, taken in normal mode. , taken in biophase mode , This represents the maximum effect parameter of the main pharmacodynamic state function or direct effect function. This represents the half-maximal effect parameter of the main pharmacodynamic state function or direct effect function. The Hill coefficient represents the state function of the main drug efficacy. This represents the drug concentration in the first transfer chamber of the first delay module at time t. This represents the drug concentration in the third transfer chamber of the first delay module at time t.

[0053] This represents the principal pharmacodynamic state action function, used to calculate the intensity of the principal pharmacodynamic state action at time t. In classical pharmacodynamics, the Emax-Hill relationship is commonly used to describe the nonlinear relationship between drug action intensity and driving force. By adjusting... EC50 and The formula employs three types of parameters to characterize situations with weak or strong effects, as well as situations with early or late onset of action and flat or steep curves, thus adapting to differences in the amplitude and sensitivity of effects under different pharmacodynamic states. Therefore, this invention uses this formula to transform a uniform driving force into an adjustable and interpretable main pharmacodynamic state effect intensity, enabling different pharmacodynamic manifestations to be uniformly expressed through parameter variations within the same mechanistic framework.

[0054] The principal pharmacological state module describes the generation and disappearance of the principal pharmacological state on the first PD axis. It is characterized as follows: ; This equation is used to describe the first PD axis. The dynamic changes. In normal spindle mode, The "generating end-disappearing end" structure is used for characterization, that is, using This represents the rhythm modulation function at the generation end of the main drug effect state, with This represents its vanishing end and allows for the drug action function. and feedback function Co-modulation of the generation process. The reason for adopting this form is that many pharmacodynamic endpoints can be physiologically understood as a dynamic equilibrium formed by the combined effects of continuous generation and continuous disappearance of a certain endogenous biological factor, physiological function, or response state. Classical PD models often use turnover models or indirect response models to characterize such processes. Drugs, feedback, and rhythms usually do not generate new states out of thin air, but rather change the level of pharmacodynamics by altering the original generation or disappearance ends.

[0055] When the disease state module is enabled, the equation switches to a "natural progression-drug killing" structure, i.e., using... The natural increase in disease state quantity or tumor burden is represented by... This indicates the effective inhibition or killing effect of the drug. This notation is used because disease progression and tumor growth endpoints are better represented by a balance between growth and inhibition. This is achieved by introducing a gating parameter into the same equation. and This invention enables the main drug efficacy state to represent both the classic cyclical drug efficacy endpoint and disease progression or tumor burden endpoints, thereby accommodating different physiological mechanisms under a unified mechanistic framework.

[0056] in, This represents the principal pharmacodynamic state, the core state of the first drug action axis. It describes the formation process of the principal pharmacodynamic state endpoint and allows drug action, feedback regulation, and rhythm to jointly influence its generation. For example, when studying the hypoglycemic effect of a drug, the pharmacodynamic state can be blood glucose concentration; when studying the antiplatelet effect of a drug, the pharmacodynamic state can be platelet concentration.

[0057] Indicates the direct effect module switch. This indicates that the direct effects module is enabled. This indicates that the direct effects module is off. This indicates the switch for the disease status module; 1 means the module is on, and 0 means the module is off. This represents the rhythm modulation function at the generation end of the main pharmacological state, used to calculate the generated value of the main pharmacological state at time t. This represents the rate constant of the disappearance of the main drug effect state. The natural progression rate constant representing the disease state quantity or tumor burden. Let t represent the effective kill rate at time t.

[0058] The downstream pharmacodynamic state module describes the transmission of pharmacodynamic state from the main pharmacodynamic state to the downstream pharmacodynamic state, and is characterized as follows: ; in, Indicates by The downstream pharmacodynamic state driven by this process Indicates the switch for disease status modules. Indicates the direct effect module switch. This represents the rate constant of the disappearance of the main drug's active effect. This represents the rate constant of the disappearance of the downstream drug effect state. Indicates the direct effect mode used for driving The main drug efficacy status.

[0059] ; in, This represents the average generation rate at the main drug effect state generation end, when the direct effects module is enabled. Driven directly in the above form .

[0060] The second delay module describes the delayed propagation process of downstream pharmacodynamic changes before feedback occurs. Specifically, it is characterized by the following equation: ; ; ; in, This represents the drug concentration in the first transfer chamber of the second delay module, and the drug concentration in the latter two transfer chambers. , Together, they describe the lag before feedback is generated. This indicates the drug concentration in the second transfer chamber of the second delay module. This represents the drug concentration in the third transfer chamber of the second delay module and is the input to the first PD axis feedback function.

[0061] This represents the transmission rate constant of the second delay module.

[0062] The feedback module is used to describe the feedback regulation process formed by downstream pharmacodynamic changes and returned to the main pharmacodynamic state generation end.

[0063] The feedback module uses the first PD axis feedback function. The function calculates the feedback quantity (feedback intensity).

[0064] .

[0065] This represents the maximum effect parameter of the first PD axis feedback function. This represents the half-maximal effect parameter of the first PD axis feedback function. This indicates the drug concentration in the three transfer chambers of the first PD axis and the second delay module at time t. That is, the drug concentration involved in the feedback mechanism. This represents the Hill coefficient of the first PD axis feedback function.

[0066] The rhythm modulation module describes the periodic fluctuations at the receptor generation end and the main pharmacological state generation end. Specifically, it is represented by the following equation: ; ; in, This represents the rhythm modulation function at the receptor generation end. This represents the rhythm modulation function at the generation end of the main pharmacokinetic state. Both achieve rhythmic changes in the receptor layer or the main pharmacokinetic state layer by superimposing a periodic function on the average generation rate.

[0067] This represents the average generation rate at the receptor-generating end. Indicates the amplitude of the receptor-generating rhythm. Indicates the receptor-generating end rhythm phase, Indicates the rhythm period parameter, This represents the average generation rate at the origin of the main drug effect state. This indicates the amplitude of the rhythm at the generation end of the main drug effect state. This indicates the rhythm phase at the generation end of the main drug effect state.

[0068] The biophase module is used to describe biophase drug action site distribution lag or effect chamber driven processes.

[0069] The opening or closing of the biological phase module is controlled by gating parameters. Characterization, if the biophase module is activated, then the original... The first delay chain input of the control is switched to be controlled by control.

[0070] The direct effects module describes scenarios where upstream driving factors directly determine drug efficacy. Specifically, the direct effects module employs... Function description driven The main drug efficacy status.

[0071] The disease state module describes the natural progression of disease states and the regulatory process of drugs on disease states, or the natural increase of tumor burden and the inhibition or killing process of tumor burden by drugs. The following auxiliary quantities are used for characterization: ; ; in, This function represents the inhibitory or killing effect of a drug on disease state quantity or tumor burden, and is used to calculate the magnitude of the inhibitory or killing effect of a drug on disease state quantity or tumor burden. This indicates the effective kill rate after considering drug resistance or effect decay. This represents the maximum effect parameter of the inhibition or killing function. The half-maximum effect parameter represents the function of inhibition or killing. Hill coefficient represents the function of inhibition or killing. This indicates a parameter representing drug resistance or effect attenuation.

[0072] The second PD axis module is used to describe a second independent pharmacodynamic branch parallel to the main axis and its feedback chain. Specifically, it is described using the following equations: ; ； ; ; ; in, This indicates the main pharmacodynamic status of the second PD axis. This indicates the downstream pharmacodynamic status of the second PD axis. Indicates the second PD axis switch. This represents the average generation rate of the principal states of the second PD axis. This represents the drug action function on the second PD axis, used to calculate the intensity of the drug's effect on the principal pharmacodynamic state on the second PD axis at time t. This represents the second PD-axis feedback function, used to calculate the intensity of the second PD-axis feedback at time t. This represents the rate constant of the disappearance of the principal states of the second PD axis. This represents the rate constant of the disappearance of the downstream state of the second PD axis. This represents the transmission rate constant of the second PD axis feedback chain. This indicates the drug concentration in the first transfer chamber of the second PD axis feedback module. This indicates the drug concentration in the second transfer chamber of the second PD axis module. This indicates the drug concentration in the third transfer chamber of the second PD axis module, which is the drug concentration that ultimately provides feedback.

[0073] ; ; in, This represents the maximum effect parameter of the drug action function on the second PD axis. This represents the half-maximal effect parameter of the drug action function along the second PD axis. This represents the maximum effect parameter of the second PD-axis feedback function. This represents the half-maximal effect parameter of the second PD-axis feedback function. This indicates the drug concentration in the third transfer chamber of the second PD axis feedback module at time t. Hill coefficient represents the drug action function on the second PD axis. This represents the Hill coefficient of the second PD axis feedback function.

[0074] It should be noted that the biological phase module, direct effect module, disease state module, feedback module, rhythm modulation module, and second PD axis module can be turned on or off to form different mechanisms of action. These six modules are variable-state mechanism modules. The remaining modules cannot be turned off, but their effects can be adjusted by adjusting parameters.

[0075] When the biological phase module is turned off, it means that the driving force of the main drug effect is no longer taken from the effect-site concentration. Instead, it reverts to the state at the end of the first delay chain in normal mode. supply.

[0076] When the direct effects module is turned off, it indicates the downstream pharmacodynamic status. No longer formed by the direct effect pathway Driven by the main drug effect state It is passed down downstream.

[0077] When the disease status module is closed, it indicates the status of the main drug efficacy. Instead of using the combined effects of natural disease progression or tumor growth and drug killing, it reverts to the ordinary active drug efficacy turnover equation.

[0078] When the feedback module is turned off, the feedback action function is zero, that is... Thus, changes in downstream drug efficacy no longer modulate the generation end of the main drug efficacy state.

[0079] When the rhythm modulation module is turned off, the periodic amplitude terms in both the receptor generation and main drug effect generation ends are set to zero, i.e., ... and .

[0080] When the second PD axis module is closed, mathematically This indicates that the parallel second PD axis does not participate in the current modeling process, meaning that the master state of the second PD axis will no longer be introduced. Downstream status and its feedback chain .

[0081] Based on the above equations (1) to (27), the unified mechanism framework has systematically organized the core mechanisms such as absorption and exposure, target binding and target turnover, drug action input, signal transmission hysteresis, main drug effect formation, downstream drug effect, feedback, rhythm modulation, and the second PD axis within the same continuous kinetic matrix. Different special cases are manifested in the unified mechanism framework through different driving types, different action terms, different parameter combinations, and different organizational methods. This preserves the physiological interpretability of the mechanism model and provides a common basis for subsequent automatic structure identification and unified parameter fitting. The structure identification results and parameter fitting results also need to be interpreted within this unified mechanism framework; therefore, the final model structure and parameter results have a clear physiological mechanism basis. It should be noted that the gating parameters satisfy... .

[0082] During implementation, the structure of the unified mechanism pharmacodynamic model framework can be as follows: Figure 2 As shown.

[0083] After drug administration, it first enters the absorption chamber. The administered dose is Dose, the bioavailability is F, and the drug amount at time 0 is... The amount of medicine in the room is Absorbing indoor drugs at a rate The volume of the entry is The central chamber contains a drug dosage of [missing information]. Central compartment drug concentration This corresponds to the absorption and system exposure module. In the central compartment, the drug further binds to the receptor, dissociates, and forms a complex. Meanwhile, receptor Continuous generation and disappearance. Drug action input. This can be determined by the free receptor, the complex, receptor occupancy, receptor non-occupancy, or the central compartment drug concentration. The upstream driving signal passes through the first delay chain. Gradually delivered to the main pharmacological layer; when the biophase module is activated, it switches to... This represents the effect-site concentration. Principal pharmacodynamic function. Unified driving force Mapping to the intensity of principal drug efficacy. Principal drug efficacy status on the first PD axis. Used to describe the formation of the primary pharmacodynamic endpoint, and is influenced by drug action, feedback regulation, and rhythm modulation. Downstream pharmacodynamic state. Depend on It is passed downstream to form and further enters the second delay chain. A feedback signal is generated. Feedback function. Depend on Drive formation and return modulation The generation end thus forms the first PD axis feedback closed loop.

[0084] When the direct effects module is enabled, drug action drives downstream states through the direct effects pathway; when the disease state module is enabled... Switch to disease progression or tumor growth dynamics.

[0085] The second PD axis is composed of , and feedback chain The composition is used to describe another set of pharmacodynamic endpoints and their feedback processes that run parallel to the main axis.

[0086] In conclusion, Figure 2 The unified mechanistic framework shown organizes mechanisms such as absorption and exposure, receptor action, drug input, delayed delivery, main pharmacokinetics, downstream pharmacokinetics, feedback regulation, rhythm, biological phase, direct effects, disease state modules, and the second PD axis within the same continuous kinetic system.

[0087] The trained structure recognition model is trained on the framework of the unified mechanism pharmacodynamic model and is used to predict the corresponding driving type, mechanism of action and initial values ​​of the parameters to be fitted based on pharmacodynamic research data.

[0088] The efficacy study data includes PK time series, PD time series, and dose data of the target population.

[0089] It should be noted that the drive type refers to the drug drive type obtained by the drug input module. .

[0090] The mechanism of action refers to the activation or deactivation of six variable mechanism of action modules: the biological phase module, the direct effect module, the disease state module, the feedback module, the rhythm modulation module, and the second PD axis module.

[0091] The structure recognition model outputs the probability of belonging to each driving type, the probability of each action mechanism module being activated, and the initial values ​​of the parameters to be fitted.

[0092] Specifically, the trained structure recognition model is obtained using the following method: A training sample set was constructed by collecting multiple historical pharmacodynamic models and corresponding research data. A multi-task neural network model is constructed; the multi-task neural network model is used to predict the driving type, mechanism of action, and modeling parameters of a drug efficacy model based on research data; The neural network model is trained on the training sample set based on the multi-task training loss to obtain a trained structure recognition model.

[0093] In practice, to obtain a well-trained structure recognition model, multiple historical pharmacodynamic models and corresponding research data are collected to construct a training sample set containing data (dose, pharmacokinetic concentration at each time point, pharmacodynamic status, and pharmacodynamic status) and ground truth labels (driving type, mechanism of action, and modeling parameter values). If the sample size of some historical models is small, Monte Carlo simulations can be used to generate corresponding data for the models, constructing simulated samples to expand the training sample set.

[0094] During implementation, one-hot encoding is used to encode driver types and construct driver type labels.

[0095] Regarding the mechanism of action, the on / off status of the six modules—biological phase module, direct effect module, disease state module, feedback module, rhythm modulation module, and second PD axis module—can be encoded into a vector containing six elements. An element value of 1 indicates that the corresponding module is on, and an element value of 0 indicates that the corresponding module is off, thus constructing corresponding labels.

[0096] The modeling parameters are listed below under "Parameters to be Fitted". In practice, all modeling parameters can be encoded into a vector, with each element corresponding to a modeling parameter, thus obtaining the labels for the modeling parameters.

[0097] Build a neural network model, such as a convolutional neural network. Train the network using multi-task learning. These multi-task tasks include driver type prediction, mechanism prediction, and parameter prediction.

[0098] Constructing the loss function for multi-task training: ; in, The driving type loss can be calculated using the cross-entropy loss function to determine the loss between the prediction result and the driving type label. This represents the loss due to the mechanism of action. The BCEWithLogits loss function can be used to calculate the loss between the predicted mechanism of action and the mechanism of action label. The parameter prediction loss can be calculated using MaskedMSE (mean squared error loss) during implementation. , and This represents the weighting coefficients, which sum to 1.

[0099] The gradient is calculated based on the total loss, and the parameters of the neural network model are updated by backpropagation. Training stops when the preset number of training iterations is reached or the change in total loss is less than the preset accuracy, and a trained structure recognition model is obtained.

[0100] Acquire target efficacy study data and input them into a trained structure recognition model to predict the driving type, mechanism of action, and modeling parameter values ​​of the target efficacy study.

[0101] Furthermore, based on the predicted driving type and mechanism of action, candidate structures for target efficacy studies are obtained, specifically including: Candidate structures for target efficacy studies were obtained based on the predicted driving type and mechanism of action, including: Choose the one with the higher prediction probability Each driver type is a candidate driver type; For each element in the predicted mechanism result, if the value of the element is greater than the upper limit of the uncertainty interval, the value of the element is set to 1; if the value of the element is less than the lower limit of the uncertainty interval, the value of the element is set to 0. For each element in the uncertainty interval of the mechanism of action result, the candidate mechanism of action result is obtained by transforming the element to 0 or 1; Candidate action mechanisms and candidate driver types are combined to obtain candidate structures. During implementation, these structures are sorted by driver type probability, and the top-ranked structures are selected. Each driver type is a candidate driver type.

[0102] For each element in the predicted mechanism result, if the value of the element is greater than the upper limit of the uncertainty interval, the value of the element is set to 1; if the value of the element is less than the lower limit of the uncertainty interval, the value of the element is set to 0. For each element in the uncertainty interval of the mechanism of action result, the candidate mechanism of action result is obtained by transforming the element to 0 or 1.

[0103] During implementation, the uncertainty interval can be set to [0.3, 0.7]. In the prediction results of the action mechanism, the module corresponding to the element with a probability greater than 0.7 is definitely in the open state, and the element is set to 1. The module corresponding to the element with a probability less than 0.3 is definitely in the closed state, and the element is set to 0.

[0104] Elements within the uncertain range cannot be definitively identified in a specific state. Therefore, different states are represented by setting the element to 1 or 0.

[0105] For example, the mechanism of prediction is as follows: {0.2,0.5,0.1,0.6,0.96,0.21}, set the values ​​of elements greater than the upper limit of the uncertain interval to 1, and the values ​​of elements less than the lower limit of the uncertain interval to 0, to get {0,0.5,0,0.6,1,0}.

[0106] By transforming elements 0.5 and 0.6 to 0 or 1 respectively, and combining different transformations, a total of 4 candidate mechanisms of action were obtained.

[0107] By combining candidate action mechanisms and candidate driver types, candidate structures are obtained.

[0108] In practice, to improve computational efficiency, after combining candidate action mechanisms and candidate driving types to obtain candidate structures, the following steps are included before parameter fitting: Extract the scores with higher overall scores The candidate structures were selected as the final candidate structures.

[0109] In practice, the comprehensive score of a candidate structure is the sum of the predicted probability of the candidate driving type in the candidate structure and the original predicted probability of the element with a value of 1 in the candidate action mechanism.

[0110] For example, the candidate driver type of a certain candidate structure is The corresponding predicted probability is 0.87, the candidate action mechanism is {0,0,0,1,1,0}, and the corresponding original action mechanism is {0.2,0.5,0.1,0.6,0.96,0.21}. Therefore, the comprehensive score of this candidate structure is: 0.87+0.6+0.96.

[0111] Choose the one with the higher overall score Each candidate structure is then used for subsequent parameter fitting.

[0112] and The number can be set according to the computational efficiency and accuracy requirements. A higher number can be set if efficiency requirements are low and accuracy requirements are high. and Otherwise, reduce and The value of .

[0113] Based on the target efficacy research data and the predicted initial modeling parameters, parameter fitting is performed on each candidate structure under the framework of the unified mechanism pharmacodynamic model. That is, under a given candidate structure, a set of optimal parameters is found so that the model prediction curve is as close as possible to the observed data.

[0114] By using the predicted modeling parameter values ​​as the starting point for the parameters to be fitted, the optimal parameters can be found more quickly.

[0115] Under each candidate structure, a population joint fitting task is constructed: all subjects' observational data at each time point (including the observed pharmacodynamic concentration (PK) and pharmacodynamic endpoint (PD) at each time point, and their corresponding dosing information) are simultaneously incorporated into the same optimization problem, and the errors at each observation point are summarized according to the objective function definition to form the population layer fitting objective corresponding to that candidate structure. This step clarifies the optimization object targeted by subsequent parameter search and refinement.

[0116] The parameters to be fitted include: , , , , , , , (In normal mode) (Biophase model) (When the feedback module is enabled) , (When the rhythm modulation module is enabled) (When the rhythm modulation module is enabled) (When the rhythm modulation module is enabled) , (When the rhythm modulation module is enabled) (When the rhythm modulation module is enabled) , , (When the disease status module is enabled) (When the disease status module is enabled) (When the second PD axis module is turned on) (When the second PD axis module is turned on) (When the second PD axis module is turned on) (When the second PD axis module is turned on) (When the direct effects module is enabled) (When the direct effects module is enabled) (When the direct effects module is enabled) (When the feedback module is enabled) (When the feedback module is enabled) (When the feedback module is enabled) (When the disease status module is enabled) (When the disease status module is enabled) (When the disease status module is enabled) (When the second PD axis module is turned on) (When the second PD axis module is turned on) (When the second PD axis module is turned on) (Second PD axis module activated) (When the second PD axis module is turned on) (When the second PD axis module is turned on).

[0117] In practice, for each candidate structure, the modeling parameter values ​​are used as the initial values ​​of the parameters to be fitted to the candidate structure. The PD value is calculated by simultaneously solving the aforementioned equations (1) to (27) (Formula (10)). If a second independent pharmacodynamic branch exists, calculate it simultaneously. (Value). Error is calculated based on the calculated PD and observed PD, and the fitting parameters are optimized.

[0118] At time t=0, the initial values ​​of the absorption chamber and the central chamber when the drug is administered orally or via absorption: ; ; Initial values ​​of the absorption chamber and central chamber during intravenous bolus administration: ; ;in, Indicates the dosage.

[0119] Initial value of receptor layer ; .

[0120] The initial value of the first delay module is aligned with the input: .

[0121] The initial value of the feedback delay chain of the second delay module and Alignment: The initial value of the feedback delay chain of the second PD axis module and... Alignment: .

[0122] Initial values ​​for PD and its downstream components are: , ; in, The initial baseline value representing the efficacy status of the main drug. The initial baseline value represents the downstream efficacy status.

[0123] If there is a second PD axis: , ; This represents the initial baseline value of the second PD axis master state. The initial baseline value represents the downstream state of the second PD axis.

[0124] It should be noted that the structure identification stage outputs several candidate structures to narrow down the model's structure search space. To reduce the risk of overfitting and improve the stability of parameter estimation, given that the candidate structures are determined, the effective magnitude of the mechanisms already included is reduced during the parameter fitting stage. L1 regularization constraints are applied to the gating parameters in the candidate structures, causing the strength of mechanisms lacking data support to shrink to near zero during the fitting process.

[0125] In implementation, existing fitting algorithms can be used for parameter fitting, and the objective function for fitting is: ; Where K represents the number of efficacy endpoints in the target efficacy study. This represents the normalized root mean square error at the k-th efficacy endpoint. Indicates a sparse penalty term. This indicates a stability penalty term. , and This represents the weighting coefficient.

[0126] During implementation, the normalized root mean square error is calculated using the following formula: ; in, Let M represent the number of observation points for the j-th subject at the k-th efficacy endpoint, and M represent the number of subjects.

[0127] This represents the observation value of the j-th subject at the k-th efficacy endpoint at the i-th time point. This represents the parameter vector to be fitted under the current candidate structure. This represents the predicted value of the j-th subject at the k-th efficacy endpoint at the i-th time point, obtained from the current fitting parameters. It represents the standard deviation of the observed values ​​at the k-th efficacy endpoint.

[0128] Specifically, the sparsity penalty term is calculated using the following formula: ; Where Q represents the number of parameters. This represents the q-th active parameter in the current candidate structure. This represents the normalized weight of the q-th active parameter.

[0129] Action parameters refer to the parameters in the action mechanism module used to characterize the actual strength of the module's action. By uniformly applying L1 regularization constraints to the action parameters, the action strength of the action mechanism module, which lacks data support, shrinks to near zero during the parameter fitting process.

[0130] Specifically, the operating parameters include: (Disease Status Module) (Feedback Module) (Rhythm Modulation Module) (Corresponding to the rhythm modulation module) (Second PD axis module) and (Second PD axis module).

[0131] Specifically, the stability penalty term is calculated using the following formula: ; in, Indicates the number of parameters to be fitted. Let represent the j-th parameter to be fitted. K represents the number of efficacy endpoints in the target efficacy study. M represents the number of observation points for the j-th subject at the k-th efficacy endpoint, and M represents the number of subjects. This represents the predicted value of the j-th subject at the i-th time point of the k-th efficacy endpoint, given the current candidate structure and current fitting parameters. This represents the predicted value of the j-th subject at the (i-1)-th time point of the k-th efficacy endpoint, given the current candidate structure and current fitting parameters. This represents the predicted value of the j-th subject at the (i+1)-th time point of the k-th efficacy endpoint, given the current candidate structure and current fitting parameters.

[0132] The first term penalizes non-physiological negative parameter values; the second term penalizes anomalous reversals and excessive oscillations in the prediction curve at adjacent time points. This term constrains anomalous parameter combinations and unstable dynamic behaviors by imposing constraints on non-physiological parameters and the local second-order difference of the curve.

[0133] Specifically, the optimal candidate structure is determined based on parameter fitting, resulting in a pharmacokinetic model for the target efficacy study, including: For each candidate structure, find the set of fitting parameters that minimizes the objective function as the optimal parameters for that candidate structure; Calculate the normalized root mean square error of each candidate structure under its optimal parameters, and select the candidate structure with the smallest normalized root mean square error as the optimal candidate structure. The optimal candidate structure and its corresponding optimal parameters constitute the pharmacokinetic model for the study of the target drug efficacy.

[0134] For each candidate structure, find the one that makes the objective function... The smallest set of parameters This set of parameters This is considered to be the optimal parameter for the candidate structure, balancing fitting accuracy, structural simplicity, and dynamic stability.

[0135] Calculate the normalized root mean square error (NRMSE) of each candidate structure under its optimal parameters. Use the minimum NRMSE as the criterion for optimal structure. Select the candidate structure with the minimum NRMSE as the optimal structure. The optimal structure and its corresponding optimal parameters constitute a pharmacokinetic model, thereby obtaining the pharmacokinetic model for the target pharmacodynamic study.

[0136] Simulation results from 80 scenarios, including single-pathogenic and multi-pathogenic (PD) scenarios, demonstrate that the pharmacodynamic automatic modeling method based on a unified mechanistic framework and machine learning assistance in this embodiment correctly predicts the structure in over 90% of the scenarios, indicating accurate structure identification. The NRMSE is less than 0.2, indicating accurate fitting results. PD Union can accurately simulate the vast majority of clinical PD scenarios and fulfills the requirement for automated pharmacodynamic model building. The pharmacodynamic process of a common oral medication (such as an antihypertensive drug) is used as an example to illustrate the effect. In this scenario, drug exposure increases after administration, and pharmacodynamic indicators such as blood pressure initially decrease after a short period, then gradually recover to near baseline as drug exposure decreases. The corresponding mechanism modules mainly include: absorption and systemic exposure module, receptor module, drug input module, first delay module, main pharmacodynamic function module and main pharmacodynamic state module, downstream pharmacodynamic state module, and the second delay module (which is active). The drug effect is negative and does not involve the feedback module. The rhythm modulation module, biological phase module, direct effect module, disease state module, and second PD axis module are all active. The model's performance in this scenario is as follows: Figure 3 As shown. Figure 3 (a) in the figure is a time series fitting plot, with time on the horizontal axis and response value on the vertical axis; the blue line represents the model prediction value and the hollow dots represent the observed value; this plot is used to judge the model's ability to capture the overall dynamic process (such as the downward-upward trend, peak and valley positions and amplitude changes); in the example, the predicted curve and the observed trend are basically consistent, indicating that the dynamic structure is well matched.

[0137] Figure 3In the example, (b) is an observation-prediction scatter plot: the horizontal axis represents the predicted value, and the vertical axis represents the observed value; the diagonal line represents the ideal state of "observation = prediction"; the closer the scatter points are to the diagonal line, the smaller the point-to-point error; in the example, the scatter points are roughly distributed along the diagonal line, reflecting that the model has good numerical consistency across all observation points.

[0138] (c) Residuals-Time Plot: The horizontal axis represents time, and the vertical axis represents the residuals (observations minus predictions); the horizontal zero line represents no bias. This plot is designed to examine systematic biases (such as persistent positive or negative bias) and time-related errors. In the example, the residuals fluctuate randomly around the zero line, with no obvious unidirectional drift, indicating that the model has no significant systematic mismatch.

[0139] PDUnion demonstrates good predictive ability for drug efficacy scenarios. The time series fitting curve is consistent with the overall observation points, and most of the observation-prediction scatter points are distributed near the consistency diagonal. The residuals are randomly distributed around 0 without any obvious systematic shift. Combined with the low NRMSE=0.117, this indicates that the model can accurately simulate the dynamic changes in drug efficacy under typical oral antihypertensive drug scenarios.

[0140] A specific embodiment of the present invention discloses an automatic pharmacodynamic modeling system based on a unified mechanistic framework and machine learning assistance, such as... Figure 4 As shown, it includes: The skeleton construction module is used to build a unified mechanism pharmacodynamic model skeleton; The model structure prediction module is used to input the target efficacy study data into the trained structure recognition model to predict the driving type, mechanism of action, and initial modeling parameters of the target efficacy study; and to obtain candidate structures for the target efficacy study based on the predicted driving type and mechanism of action. The parameter fitting module, based on the target efficacy study data and the initial modeling parameters, performs parameter fitting on each candidate structure under the framework of the unified mechanism pharmacodynamic model; the optimal candidate structure is determined according to the parameter fitting, and the pharmacodynamic model of the target efficacy study is obtained.

[0141] The above-described method and system embodiments are based on the same principles, and their related aspects can be referenced from each other to achieve the same technical effects. For specific implementation processes, please refer to the foregoing embodiments, which will not be repeated here.

[0142] Electronic device example: One specific implementation of this application discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the pharmacodynamic automatic modeling method based on a unified mechanism framework and machine learning-assisted method in the method embodiment.

[0143] Examples of readable storage media: One specific implementation of this application discloses a non-transitory computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the pharmacodynamic automatic modeling method based on a unified mechanism framework and machine learning-assisted modeling method in the method embodiment.

[0144] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0145] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for automatic pharmacokinetic modeling based on a unified mechanism skeleton and machine learning assistance, characterized in that, Includes the following steps: Construct a unified framework for mechanism-based pharmacodynamics; Input the target efficacy study data into the trained structure recognition model to predict the driving type, mechanism of action, and initial modeling parameters of the target efficacy study; Candidate structures for target efficacy studies are obtained based on the predicted driving type and mechanism of action; Based on the target efficacy study data and the initial modeling parameters, parameter fitting is performed on each candidate structure under the framework of the unified mechanism pharmacodynamic model; the optimal candidate structure is determined according to the parameter fitting, and the pharmacodynamic model of the target efficacy study is obtained.

2. The unified mechanism framework and machine learning assisted pharmacokinetics automatic modeling method according to claim 1, characterized in that, The trained structure recognition model is obtained using the following method: A training sample set was constructed by collecting multiple historical pharmacodynamic models and corresponding research data. A multi-task neural network model is constructed; the multi-task neural network model is used to predict the driving type, mechanism of action, and modeling parameters of a drug efficacy model based on research data; The neural network model is trained on the training sample set based on the multi-task training loss to obtain a trained structure recognition model.

3. The unified mechanism framework and machine learning assisted pharmacokinetics automatic modeling method according to claim 1, characterized in that, Candidate structures for target efficacy studies were obtained based on the predicted driving type and mechanism of action, including: Taking a predicted probability greater than a threshold as a criterion a driving type as a candidate driving type; For each element in the predicted mechanism result, if the value of the element is greater than the upper limit of the uncertainty interval, the value of the element is set to 1; if the value of the element is less than the lower limit of the uncertainty interval, the value of the element is set to 0. For each element in the uncertainty interval of the mechanism of action result, the candidate mechanism of action result is obtained by transforming the element to 0 or 1; By combining candidate action mechanisms and candidate driver types, candidate structures are obtained.

4. The unified mechanism framework and machine learning assisted pharmacokinetics automatic modeling method according to claim 3, characterized in that, After combining candidate action mechanisms and candidate driving types to obtain candidate structures, before performing parameter fitting, the following steps are also included: Extract the scores with higher overall scores The candidate structures were selected as the final candidate structures; The comprehensive score of a candidate structure is the sum of the predicted probability of the candidate driving type in the candidate structure and the original predicted probability of the element with a value of 1 in the candidate action mechanism.

5. The unified mechanism framework and machine learning assisted-based pharmacodynamics automatic modeling method according to claim 1, characterized in that, The objective function for parameter fitting is: ； wherein K represents the number of efficacy endpoints of the target efficacy study, denotes the normalized root mean square error of the kth efficacy endpoint, denotes the sparsity penalty term, denotes the stability penalty term, , and denotes the weight coefficient.

6. The unified mechanism framework and machine learning assisted pharmacokinetics automatic modeling method according to claim 5, characterized in that, The normalized root mean square error is calculated using the following formula: ； in, This represents the number of observation points for the j-th subject at the k-th efficacy endpoint, and M represents the number of subjects. This represents the observation value of the j-th subject at the k-th efficacy endpoint at the i-th time point. This represents the parameter vector to be fitted under the current candidate structure. This represents the predicted value of the j-th subject at the k-th efficacy endpoint at the i-th time point, obtained from the current fitting parameters. It represents the standard deviation of the observed values ​​at the k-th efficacy endpoint.

7. The automatic pharmacodynamic modeling method based on a unified mechanistic framework and machine learning assistance as described in claim 5, characterized in that, The stability penalty term is calculated using the following formula: ； in, Indicates the number of parameters to be fitted. Let represent the j-th parameter to be fitted; K represents the number of efficacy endpoints in the target efficacy study. M represents the number of observation points for the j-th subject at the k-th efficacy endpoint, and M represents the number of subjects. This represents the predicted value of the j-th subject at the i-th time point of the k-th efficacy endpoint, given the current candidate structure and current fitting parameters. This represents the predicted value of the j-th subject at the (i-1)-th time point of the k-th efficacy endpoint, given the current candidate structure and current fitting parameters. This represents the predicted value of the j-th subject at the (i+1)-th time point of the k-th efficacy endpoint, given the current candidate structure and current fitting parameters.

8. The automatic pharmacodynamic modeling method based on a unified mechanistic framework and machine learning assistance as described in claim 5, characterized in that, The sparsity penalty term is calculated using the following formula: ； Where Q represents the number of parameters. This represents the q-th active parameter in the current candidate structure. This represents the normalized weight of the q-th active parameter.

9. The automatic pharmacodynamic modeling method based on a unified mechanistic framework and machine learning assistance as described in claim 1, characterized in that, The optimal candidate structure is determined based on parameter fitting, resulting in a pharmacokinetic model for the target efficacy study, including: For each candidate structure, find the set of fitting parameters that minimizes the objective function as the optimal parameters for that candidate structure; Calculate the normalized root mean square error of each candidate structure under its optimal parameters, and select the candidate structure with the smallest normalized root mean square error as the optimal candidate structure. The optimal candidate structure and its corresponding optimal parameters constitute the pharmacokinetic model for the study of the target drug efficacy.

10. An automated pharmacodynamic modeling system based on a unified mechanistic framework and machine learning assistance, characterized in that, include: The skeleton construction module is used to build a unified mechanism pharmacodynamic model skeleton; The model structure prediction module is used to input the target efficacy study data into the trained structure recognition model to predict the driving type, mechanism of action, and initial modeling parameters of the target efficacy study; and to obtain candidate structures for the target efficacy study based on the predicted driving type and mechanism of action. The parameter fitting module, based on the target efficacy study data and the initial modeling parameters, performs parameter fitting on each candidate structure under the framework of the unified mechanism pharmacodynamic model; the optimal candidate structure is determined according to the parameter fitting, and the pharmacodynamic model of the target efficacy study is obtained.

Images