A small molecule drug dynamic analysis system of a spiking neural network

Abstract

This invention discloses a small molecule drug dynamic analysis system using a spiking neural network, relating to the field of drug analysis. The system includes: a management module for controlling the editing and sending of control commands for various functional modules and units; a data collection module for collecting known drug compound structural data and corresponding in vivo performance data; and a model building module for constructing a spiking graph neural network model architecture, using the drug compound structural data and corresponding in vivo performance data collected by the data collection module as training samples. The spiking computation of the biological neural network can better simulate the characteristics and dynamic processes of biological systems, thereby improving the prediction accuracy of complex drug molecule behavior. By dynamically adjusting the input data distribution during training, the spiking graph neural network possesses stronger generalization ability, captures error sources, and continuously adjusts the model structure and parameters from the model identification output, achieving dynamic optimization.

Description

Technical Field

[0001] This invention relates to the field of drug analysis technology, specifically to a dynamic analysis system for small molecule drugs using a spiking neural network. Background Technology

[0002] Small molecule drugs play an important role in the treatment of various diseases, especially cancer, infections and chronic diseases. The ability to quickly and accurately analyze the structure of small molecule drugs is of great significance for the development of new drugs. Spiking neural networks are a cutting-edge artificial intelligence technology that mimics the behavior of biological nervous systems. Compared with traditional artificial neural networks, they can better process temporal information and perform calculations with lower energy consumption, which makes them show potential application prospects in fields such as dynamic analysis of small molecule drug structures.

[0003] Modern drug discovery is a complex and time-consuming process involving a large amount of data analysis and structural prediction. Traditional drug design methods often rely on physical models and experience, making it difficult to efficiently select and optimize drugs.

[0004] Existing technologies typically rely on traditional machine learning models, such as decision trees or regression models, which are often insufficient for modeling complex structures. These methods struggle to effectively capture the nonlinear relationship between small molecule structures and their drug activity, fail to fully utilize graph structure data to process molecular features, and often simplify compounds into fixed feature vectors, making it difficult to capture the topological information of molecules and their dynamic changes.

[0005] Models often perform poorly when faced with new molecules, especially unseen compounds. Overfitting is a common problem during model training. After identifying errors, they often lack effective feedback and iterative update mechanisms, making it difficult to form a closed-loop system improvement solution. Summary of the Invention

[0006] In view of the above-mentioned shortcomings of the existing technology, the present invention provides a small molecule drug dynamic analysis system with spiking neural network, which can effectively solve the problems of the existing technology.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0008] This invention discloses a small molecule drug dynamic analysis system using a spiking neural network, comprising:

[0009] The management module is used to control the editing and sending of operation commands for each functional module and unit;

[0010] The data collection module is used to collect known drug compound structural data and corresponding in vivo biological performance data.

[0011] The model building module is used to build the pulse graph neural network model architecture, select graph convolutional layers and pulse propagation mechanism, define the number of nodes and activation functions of input layer, hidden layer and output layer, determine the connection weights of the network, and set hyperparameters. The module uses the drug compound structure data and corresponding in vivo biological performance data collected by the data collection module as training samples for training.

[0012] The dynamic recognition module is used to input new molecules into a trained pulse diagram neural network model for recognition. The model's prediction output shows the in vivo performance of the new molecules, including drug distribution, metabolism, and effects. Based on the prediction results output by the model, candidate drug compounds that meet the preset threshold requirements are screened out, and the candidate drug compounds are evaluated and the model evaluation value is output.

[0013] The requirement setting module is used to customize the requirement threshold. After the threshold is set, the configuration is applied to the dynamic recognition module.

[0014] A data repository is used to organize candidate drug compounds that meet the required thresholds into a compound library, and to attach associated prediction data and model evaluation values ​​to each candidate compound;

[0015] The verification and acquisition module is used to collect experimental data obtained from compound data experiments in the data repository and to acquire actual in vivo performance data.

[0016] The judgment unit is used to compare the data predicted by the model with the results shown in the actual verification data to determine whether there is an error.

[0017] The error identification module is triggered when the judgment unit determines that there is an error. It calculates the error between the model prediction and the actual result and traces the source based on the input data, model structure and parameter settings.

[0018] The error adjustment module is used to adjust the hyperparameters, network structure, or input data features of the model building module according to the source tracing results of the error identification module.

[0019] Furthermore, the data collection module collects drug compound structural data attributes including: molecular formula, functional groups and three-dimensional conformation, and in vivo performance data attributes including: distribution status, metabolic pathway, excretion mode, biological activity, therapeutic effect and toxicity coefficient.

[0020] Furthermore, the data collection module is interconnected with a preprocessing module via a wireless network. The preprocessing module is used to standardize the drug structure and biological performance data collected by the data collection module, convert the compound structure information into a graph representation that matches the neural network processing, use the nodes and edges of the molecular graph to represent the atoms and chemical bonds in the compound structure molecule, and normalize the biological performance data.

[0021] Furthermore, the pulse graph neural network model constructed by the model building module converts the new drug molecule structure into a graph structure, which is then fed into the pulse graph neural network. The network's multi-layer transmission mechanism is used to perform deep learning on the input features and extract structural features.

[0022] Furthermore, the operating logic of the pulse graph neural network model in the dynamic recognition module is as follows:

[0023] ;

[0024] In the formula, This represents the drug performance results predicted by the model. θ represents the feedforward function of the pulse graph neural network, indicating the model's processing of the input data X, where θ represents the model parameters. Represents the total number of nodes in the network. Represents the node The weight, The activation function, through non-linear mapping, enables the model to learn complex features. Representative node The set of neighboring nodes, Representative node and neighboring nodes The connections between them are used to represent the elements of the adjacency matrix. Representative node In the The hidden state of a layer represents the feature representation of that node. Representative node The bias term.

[0025] Furthermore, the dynamic identification module, based on the drug performance results predicted by the model, filters out candidate drug compounds that meet preset threshold requirements. The specific process is as follows:

[0026] K = {X | Y > T};

[0027] In the formula, K represents the set of molecules that meet the conditions, Y represents the drug performance predicted by the model, T represents the preset demand threshold, and X represents the data currently input into the model.

[0028] Furthermore, the judgment unit has sub-modules deployed at its lower level, including a definition module, a threshold module, and a configuration module. The definition module and the threshold module are interconnected via a wireless network, and the threshold module and the configuration module are interconnected via a wireless network.

[0029] The definition module is used to define the standards and basis for judging data attributes, specify the format, range and data type of data, set the judgment rules for each data attribute, and build a standardized evaluation framework;

[0030] The threshold module is used to establish standard thresholds for several judgment data based on the standardized evaluation framework built by the threshold module, and to set different priorities or weights according to different attributes.

[0031] The configuration module is used to apply the standard threshold in practice. It compares the current data to be judged with the corresponding standard threshold to determine whether the current data meets the requirements.

[0032] Furthermore, during the error identification module's source tracing process based on input data, model structure, and parameter settings:

[0033] The traceability items for input data are: the state of outliers, and the distribution status of training and validation data.

[0034] The traceability items of the model structure are: the training error and validation error gap state, the learning curve stable state, and the feature extraction state;

[0035] The traceability items for parameter settings are: error fluctuation status, convergence speed status, and optimization algorithm performance status during training.

[0036] Furthermore, the error adjustment module adjusts the input data, network structure, and hyperparameters of the model building module as follows:

[0037] The strategies for adjusting input data are: data cleaning and supplementation, data balancing, and dataset augmentation;

[0038] The strategies for adjusting the network structure are: increasing or decreasing the number of layers or neurons in the graph neural network, changing the network structure, and regularization;

[0039] The hyperparameter tuning strategies are: adjusting the learning rate, adjusting the training data batch size, and selecting an optimization algorithm.

[0040] Furthermore, the management module and the data collection module are interconnected via a wireless network; the data collection module and the model building module are interconnected via a wireless network; the dynamic identification module, the model building module, the requirement setting module, and the data storage module are interconnected via a wireless network; the verification and acquisition module, the data storage module, and the judgment unit are interconnected via a wireless network; the judgment unit and the error identification module are interconnected via a wireless network; the error identification module and the error adjustment module are interconnected via a wireless network; and the error adjustment module and the model building module are interconnected via a wireless network.

[0041] Compared with known prior art, the technical solution provided by this invention has the following beneficial effects:

[0042] 1. Through pulse calculations using biological neural networks, the characteristics and dynamic processes of biological systems can be better simulated, thereby improving the prediction accuracy of complex drug molecule behavior. By converting molecular structures into graphical representations, the connectivity and interactions between molecules can be considered more comprehensively, improving the understanding of the internal structure of molecules. In addition to the structure of the drug, the system can also comprehensively consider the drug's distribution, metabolism, excretion, and effects in the body, providing a comprehensive analytical perspective.

[0043] 2. By dynamically adjusting the distribution of input data during training, the pulse graph neural network has a stronger generalization ability, enabling it to better adapt to new data inputs. It also has self-learning capabilities, allowing it to adjust and optimize itself after receiving new data, thereby improving the model's adaptability to new molecular reactions and enhancing its robustness.

[0044] 3. By introducing in-depth analysis and combining statistical methods with the comparison of model output data, we can capture more detailed sources of error, providing a basis for subsequent parameter adjustments and model design. We can build a feedback mechanism to continuously adjust the model structure and parameters from the model's output, achieve dynamic optimization, improve the model's predictive ability, and ensure that the model always keeps pace with the latest research results through dynamic model adjustment and retraining based on new data, thereby improving the accuracy in practical applications. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0046] Figure 1 This is a schematic diagram of the overall framework of the present invention;

[0047] Figure 2 This is a schematic diagram of the framework of the judgment unit in this invention.

[0048] The labels in the diagram represent: 1. Management module; 2. Data collection module; 3. Model building module; 4. Dynamic identification module; 5. Requirement setting module; 6. Data storage module; 7. Verification and acquisition module; 8. Judgment unit; 81. Definition module; 82. Threshold module; 83. Configuration module; 9. Error identification module; 10. Error adjustment module; 11. Preprocessing module. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0050] The present invention will be further described below with reference to embodiments.

[0051] Example 1: This example describes a small molecule drug dynamic analysis system using a spiking neural network, such as... Figure 1 As shown, it includes:

[0052] Management module 1 is used to control the editing and sending of operation commands for each functional module and unit;

[0053] Data collection module 2 is used to collect known drug compound structural data and corresponding in vivo performance data. The attributes of the drug compound structural data collected by data collection module 2 include: molecular formula, functional groups and three-dimensional conformation, and the attributes of the in vivo performance data include: distribution status, metabolic pathway, excretion mode, biological activity, therapeutic effect and toxicity coefficient.

[0054] The data collection module 2 is wirelessly connected to the preprocessing module 11. The preprocessing module 11 is used to standardize the drug structure and biological performance data collected by the data collection module 2, converting the compound structure information into a graph representation that matches the neural network processing, using the nodes and edges of the molecular graph to represent the atoms and chemical bonds in the compound structure molecule, and normalizing the biological performance data. This process lays the foundation for subsequent analysis, enabling the data to be processed in a unified format.

[0055] Model building module 3 is used to build the pulse graph neural network model architecture, select graph convolutional layers and pulse propagation mechanism, define the number of nodes and activation functions of input layer, hidden layer and output layer, determine the connection weights of the network, and set hyperparameters. It is trained based on the drug compound structure data and corresponding biological in vivo performance data collected by data collection module 2 as training samples to ensure that the model can learn the mapping between the internal rules of drug molecules and biological performance.

[0056] The dynamic identification module 4 is used to input new molecules into a trained pulse diagram neural network model for identification. The model predicts the in vivo performance of the new molecules, including drug distribution, metabolism, and effects. Based on the prediction results output by the model, candidate drug compounds that meet the preset threshold requirements are screened out, and the candidate drug compounds are evaluated and the model evaluation value is output. Based on the prediction results output by the model, candidate drug compounds that meet the preset threshold requirements can be screened out and further evaluated. This module can efficiently identify potentially effective drugs and improve R&D efficiency.

[0057] The requirement setting module 5 is used to customize the requirement threshold. After the setting is completed, the configuration is applied to the dynamic identification module 4, allowing users to customize the requirement threshold and specify the screening criteria for candidate drugs.

[0058] Data repository 6 is used to organize candidate drug compounds that meet the required threshold into a compound library, attaching associated prediction data and model evaluation values ​​to each candidate compound to facilitate subsequent review and in-depth analysis, forming a structured database;

[0059] The verification and acquisition module 7 is used to collect experimental data obtained from compound data experiments in the data repository 6, and to obtain actual in vivo performance data. By collecting actual in vivo performance data obtained from experiments, empirical support is provided for the accuracy of the model.

[0060] Judgment unit 8 is used to compare the data predicted by the model with the results shown in the actual verification data to determine whether there is an error, and to provide data basis for model optimization.

[0061] Error identification module 9 is triggered when judgment unit 8 determines an error, calculates the error between the model prediction and the actual result, and traces the source based on input data, model structure, and parameter settings; during the process of error identification module 9 tracing the source based on input data, model structure, and parameter settings:

[0062] The traceability items for input data are: the state of outliers, and the distribution status of training and validation data.

[0063] The traceability items of the model structure are: the training error and validation error gap state, the learning curve stable state, and the feature extraction state;

[0064] The traceability items for parameter settings are: error fluctuation status, convergence speed status, and optimization algorithm performance status during training.

[0065] Error adjustment module 10 is used to adjust the hyperparameters, network structure, or input data features of model building module 3 according to the source tracing results of error identification module 9; during the adjustment process of input data, network structure, and hyperparameters of model building module 3, error adjustment module 10:

[0066] The input data adjustment strategy is as follows: data cleaning and supplementation, data balancing and dataset augmentation. The data is denoised, missing values ​​are filled in and outliers are removed to ensure data quality. More training samples are generated through data balancing and augmentation techniques to improve the model's ability to learn different data patterns.

[0067] The network structure adjustment strategy is as follows: increase or decrease the number of layers or neurons in the graph neural network, change the network structure and regularization. Specifically, based on the performance of the model, the number of layers or neurons in the graph neural network is increased or decreased appropriately to balance the complexity and training ability of the model. Different types of graph neural network architectures and different activation functions are tried to improve the model's representation ability. Regularization techniques are introduced to avoid overfitting and improve the model's generalization ability.

[0068] The hyperparameter tuning strategy is as follows: adjust the learning rate, adjust the training data batch size, and select an optimization algorithm. Use learning rate decay or learning rate scheduler to dynamically adjust the learning rate, select an appropriate batch size to balance training efficiency and stability, try different optimization algorithms, and select the optimizer that is most suitable for the current model and data.

[0069] The management module 1 and the data collection module 2 are connected via a wireless network. The data collection module 2 and the model building module 3 are connected via a wireless network. The dynamic identification module 4, the model building module 3, the requirement setting module 5, and the data storage module 6 are connected via a wireless network. The verification and acquisition module 7, the data storage module 6, and the judgment unit 8 are connected via a wireless network. The judgment unit 8 and the error identification module 9 are connected via a wireless network. The error identification module 9 and the error adjustment module 10 are connected via a wireless network. The error adjustment module 10 and the model building module 3 are connected via a wireless network.

[0070] Compared with existing technologies, the dynamic analysis system using spiking neural network technology, especially in dynamic identification and error source identification and adjustment, provides an effective solution to problems such as insufficient processing of structural features, poor model generalization ability, incomplete error analysis, and lack of feedback mechanism in existing technologies. It can not only improve the understanding and prediction accuracy of small molecule drugs, but also accelerate the drug screening and development process, and improve overall efficiency and accuracy.

[0071] Example 2: At other levels, this example also provides the operating logic of a pulse graph neural network model, specifically as follows:

[0072] ;

[0073] In the formula, This represents the drug performance results predicted by the model. θ represents the feedforward function of the pulse graph neural network, indicating the model's processing of the input data X, where θ represents the model parameters. Represents the total number of nodes in the network. Represents the node The weight, The activation function, through non-linear mapping, enables the model to learn complex features. Representative node The set of neighboring nodes, Representative node and neighboring nodes The connections between them are used to represent the elements of the adjacency matrix. Representative node In the The hidden state of a layer represents the feature representation of that node. Representative node The bias term.

[0074] In this example, the pulse graph neural network model converts the new drug molecule structure into a graph structure so that it can be processed by the pulse graph neural network. The calculation formula of the model is then used to... By utilizing the multi-layer transmission mechanism of the network to perform deep learning on the input features, structural features are extracted. These features will help the model predict the biological performance of molecules, and the calculated results represent the model's predicted performance for new molecules.

[0075] The dynamic identification module 4, based on the drug performance results predicted by the model, filters out candidate drug compounds that meet the preset threshold requirements. The specific process is as follows:

[0076] K = {X | Y > T};

[0077] In the formula, K represents the set of molecules that meet the conditions, Y represents the drug performance predicted by the model, T represents the preset demand threshold, and X represents the data currently input into the model.

[0078] The model prediction result Y is compared with the preset demand threshold T to identify candidate molecules that meet the requirements. The demand threshold T can be set according to the actual research objectives, such as the lowest effective concentration or the expected bioactivity value. New molecules that meet the condition Y>T are included in the candidate drug set. These molecules are considered to have potential drug activity. Finally, the screened candidate compounds are compiled into a compound library and accompanied by the corresponding prediction data and model evaluation values ​​to facilitate subsequent experimental verification and in-depth analysis.

[0079] Example 3: In this example, as Figure 2As shown, the judgment unit 8 has sub-modules deployed below it. These sub-modules include: a definition module 81, a threshold module 82, and a configuration module 83. The definition module 81 and the threshold module 82 are interconnected via a wireless network, and the threshold module 82 and the configuration module 83 are interconnected via a wireless network.

[0080] Module 81 defines the standards and criteria for judging data attributes, specifies the format, range and data type of data, sets judgment rules for each data attribute, and builds a standardized evaluation framework.

[0081] Threshold module 82 is used to formulate standard thresholds for several judgment data based on the standardized evaluation framework constructed by threshold module 82, and set different priorities or weights according to different attributes;

[0082] Configuration module 83 is used to apply the standard threshold in practice. It compares the current data to be judged with the corresponding standard threshold to determine whether the current data meets the requirements.

[0083] Ensuring the accuracy and consistency of data attributes provides a standardized framework for data analysis; setting reasonable standards based on drug development goals and scientific evidence provides clear judgment criteria for data evaluation; screening and classifying candidate drugs according to set standard thresholds, and dynamically feeding back and optimizing thresholds improve the accuracy and efficiency of the system. This ensures that all steps in data analysis follow scientific logic and the actual needs of drug development, ultimately helping to accurately screen drug molecules that meet the standards.

[0084] Working principle: In specific implementation, the present invention sends control commands through management module 1, collects raw data of drug compound structure through data collection module 2, preprocesses raw data through preprocessing module 11, constructs a spiking neural network model through model building module 3, trains the spiking neural network model using known drug compound structure data and its corresponding in vivo performance data to capture the relationship between drug compound structure and data on distribution, metabolism, excretion and effects in vivo, sets demand thresholds through demand setting module 5, and dynamically identifies new molecules using the trained model to screen out candidate drug compounds that meet the demand thresholds, so as to build a compound library through data storage module 6.

[0085] The verification acquisition module 7 acquires the actual verification data of the new molecule, the error identification module 9 evaluates the results based on the analysis, analyzes whether there is an error, identifies the source of the error, and the error adjustment module 10 adjusts the parameters of the pulse graph neural network model corresponding to the source of the error.

[0086] Through the collaborative work of various modules, dynamic analysis of small molecule drugs can be performed efficiently, improving efficiency and accuracy in drug development and screening. This structured system enables more precise drug development and provides solid data support for drug validation before clinical application.

[0087] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dynamic analysis system for small molecule drugs using a spiking neural network, characterized in that, include: The management module (1) is used to control the editing and sending of operation commands for each functional module and unit; The data collection module (2) is used to collect known drug compound structure data and corresponding in vivo performance data. The model building module (3) is used to build the pulse graph neural network model architecture, select the graph convolutional layer and pulse transmission mechanism, define the number of nodes and activation functions of the input layer, hidden layer and output layer, determine the connection weights of the network, and set hyperparameters. The drug compound structure data and corresponding biological in vivo performance data collected by the data collection module (2) are used as training samples for training. The dynamic recognition module (4) is used to input new molecules into the trained pulse diagram neural network model for recognition. The model predicts the in vivo performance of the new molecules, including drug distribution, metabolism and effects. Based on the prediction results output by the model, candidate drug compounds that meet the preset threshold requirements are selected, and the candidate drug compounds are evaluated and the model evaluation value is output. The requirement setting module (5) is used to customize the requirement threshold. After the setting is completed, the configuration is applied to the dynamic recognition module (4). The data repository (6) is used to organize candidate drug compounds that meet the demand threshold into a compound library, and attach associated prediction data and model evaluation values ​​to each candidate compound; The verification acquisition module (7) is used to collect experimental data obtained from the compound data experimental verification in the data storage repository (6) and obtain actual in vivo performance data. The judgment unit (8) is used to compare the data predicted by the model with the results shown in the actual verification data to determine whether there is an error; The error identification module (9) is triggered when the judgment unit (8) determines that there is an error, calculates the error between the model prediction and the actual result, and traces the source based on the input data, model structure and parameter settings. The error adjustment module (10) is used to adjust the hyperparameters, network structure or input data features of the model construction module (3) according to the source tracing results of the error identification module (9).

2. The small molecule drug dynamic analysis system based on a spiking neural network according to claim 1, characterized in that, The data collection module (2) collects the following attributes of drug compound structure: molecular formula, functional group and three-dimensional conformation. The attributes of biological manifestation include: distribution status, metabolic pathway, excretion mode, biological activity, therapeutic effect and toxicity coefficient.

3. The small molecule drug dynamic analysis system according to claim 1, characterized in that, The data collection module (2) is connected to the preprocessing module (11) via a wireless network. The preprocessing module (11) is used to standardize the drug structure and biological performance data collected by the data collection module (2), convert the compound structure information into a graph representation that matches the neural network processing, use the nodes and edges of the molecular graph to represent the atoms and chemical bonds in the compound structure molecule, and normalize the biological performance data.

4. The small molecule drug dynamic analysis system according to claim 1, characterized in that, The pulse graph neural network model constructed by the model building module (3) converts the new drug molecule structure into a graph structure, enters the pulse graph neural network, and uses the multi-layer transmission mechanism of the network to perform deep learning on the input features, extract structural features, and calculate the results to represent the model's predictive performance on the new molecule.

5. The small molecule drug dynamic analysis system according to claim 1, characterized in that, The operating logic of the pulse graph neural network model in the dynamic recognition module (4) is as follows: ； In the formula, This represents the drug performance results predicted by the model. θ represents the feedforward function of the pulse graph neural network, indicating the model's processing of the input data X, where θ represents the model parameters. Represents the total number of nodes in the network. Represents the node The weight, The activation function, through non-linear mapping, enables the model to learn complex features. Representative node The set of neighboring nodes, Representative node and neighboring nodes The connections between them are used as elements to represent the adjacency matrix. Representative node In the The hidden state of a layer represents the feature representation of that node. Representative node The bias term.

6. The small molecule drug dynamic analysis system according to claim 1, characterized in that, The dynamic identification module (4) filters out candidate drug compounds that meet the preset threshold requirements based on the drug performance results predicted by the model. The specific process is as follows: ； In the formula, K represents the set of molecules that meet the conditions, Y represents the drug performance predicted by the model, T represents the preset demand threshold, and X represents the data currently input into the model.

7. The small molecule drug dynamic analysis system according to claim 1, characterized in that, The judgment unit (8) has sub-modules deployed below it, including: a definition module (81), a threshold module (82), and a configuration module (83). The definition module (81) and the threshold module (82) are interconnected via a wireless network, and the threshold module (82) and the configuration module (83) are interconnected via a wireless network. The definition module (81) is used to define the standards and basis for judging data attributes, specify the format, range and data type of data, set judgment rules for each data attribute, and build a standardized evaluation framework. The threshold module (82) is used to formulate standard thresholds for several judgment data based on the standardized evaluation framework constructed by the threshold module (82), and set different priorities or weights according to different attributes; The configuration module (83) is used to apply the standard threshold in practice, and compare the current data to be judged with the corresponding standard threshold to determine whether the current data meets the requirements.

8. The small molecule drug dynamic analysis system according to claim 1, characterized in that, During the process of tracing the source of error based on input data, model structure, and parameter settings, the error identification module (9) is as follows: The traceability items for input data are: the state of outliers, and the distribution status of training and validation data. The traceability items of the model structure are: the training error and validation error gap state, the learning curve stable state, and the feature extraction state; The traceability items for parameter settings are: error fluctuation status, convergence speed status, and optimization algorithm performance status during training.

9. The small molecule drug dynamic analysis system according to claim 1, characterized in that, During the process of adjusting the input data, network structure, and hyperparameters in the model building module (3), the error adjustment module (10) performs the following: The strategies for adjusting input data are: data cleaning and supplementation, data balancing, and dataset augmentation; The strategies for adjusting the network structure are: increasing or decreasing the number of layers or neurons in the graph neural network, changing the network structure, and regularization; The hyperparameter tuning strategies are: adjusting the learning rate, adjusting the training data batch size, and selecting an optimization algorithm.

10. The small molecule drug dynamic analysis system according to claim 1, characterized in that, The management module (1) and the data collection module (2) are connected via a wireless network. The data collection module (2) and the model building module (3) are connected via a wireless network. The dynamic identification module (4) and the model building module (3), the requirement setting module (5), and the data storage module (6) are connected via a wireless network. The verification acquisition module (7) and the data storage module (6) and the judgment unit (8) are connected via a wireless network. The judgment unit (8) and the error identification module (9) are connected via a wireless network. The error identification module (9) and the error adjustment module (10) are connected via a wireless network. The error adjustment module (10) and the model building module (3) are connected via a wireless network.

Images