System and method for evaluating stability of anti-generative drug

Abstract

The invention relates to the field of drug stability evaluation, in particular to an anti-generative drug stability evaluation system and method. Comprising a confrontation sample simulation module, a drug stability test module and an intelligent identification module. The system generates an anti-noise, self-adaptive and stable adversarial sample based on a drug molecular structure, and predicts drug stability under different conditions through environmental physical simulation and a digital twinborn model. And the intelligent identification module executes molecular fragment sensitivity analysis, predicts stability under extreme conditions and outputs an evaluation report. The system solves the problems of efficiency, precision and application range in the prior art, realizes qualitative change evaluation through innovative technology combination, provides a scientific basis for drug research and development, transportation and storage, and remarkably improves the drug stability evaluation level.

Description

Technical Field

[0001] This invention relates to the field of drug stability assessment, and in particular to a system and method for assessing the stability of antigenic drugs. Background Technology

[0002] In recent years, with the rapid development of the pharmaceutical industry and increasingly stringent regulatory requirements, drug stability assessment has become increasingly important in new drug development and quality control. Traditional drug stability assessment methods mainly rely on accelerated laboratory testing and long-term observation. While reliable, these methods have significant drawbacks, including long cycles, high costs, and difficulty in simulating complex environmental conditions. For example, a standard long-term stability test typically requires 12-36 months, which severely restricts the speed of new drug launches.

[0003] To overcome the limitations of traditional methods, the industry has begun to explore various computer-aided techniques for predicting drug stability. Among these, molecular dynamics simulations can predict drug molecule behavior to some extent, but they are computationally expensive and struggle to account for complex environmental factors. On the other hand, machine learning-based prediction models, while computationally efficient, often fail to accurately predict the stability of drugs under extreme conditions or novel drugs due to limitations in training data.

[0004] The closest existing technique is a hybrid approach combining quantum chemical calculations and artificial neural networks. This method obtains the fundamental properties of drug molecules through quantum chemical calculations and then uses neural network models to predict their stability under different environmental conditions. However, this approach still has the following problems: First, the complexity of quantum chemical calculations limits its application in large-scale screening; second, this method struggles to effectively simulate the behavior of drugs in complex and variable real-world storage environments; and finally, it cannot provide molecular-level insights into structure-stability relationships, which are crucial for drug formulation optimization.

[0005] The shortcomings of these existing technologies highlight the necessity of a novel, efficient, comprehensive, and intelligent drug stability assessment system. An ideal system should be able to rapidly and accurately predict drug stability under various conditions, while providing in-depth molecular-level analysis and the ability to handle extreme conditions and novel drugs. Summary of the Invention

[0006] The antigenic drug stability assessment system and method of the present invention are proposed to solve the above-mentioned technical problems. This system aims to achieve rapid, accurate, and comprehensive drug stability assessment, overcoming the limitations of existing technologies in terms of efficiency, accuracy, and applicability.

[0007] This invention proposes an antigenetic drug stability assessment system and method, comprising:

[0008] The adversarial example simulation module is used for:

[0009] Based on the molecular structure of the drug to be evaluated, adversarial samples with noise resistance, adaptability and stability are generated.

[0010] The adversarial sample is output to the drug stability testing module;

[0011] The drug stability testing module, which is communicatively connected to the adversarial example simulation module, is used for:

[0012] Receive adversarial samples sent by the adversarial sample simulation module;

[0013] Based on the adversarial examples, construct environmental physical simulation and digital twin models;

[0014] Generate drug stability prediction models under different storage conditions and time;

[0015] The intelligent identification module, which is communicatively connected to the drug stability testing module, is used for:

[0016] Receive the prediction model sent by the drug stability testing module;

[0017] Based on the prediction model, perform molecular fragment sensitivity analysis;

[0018] Generate drug stability prediction results under extreme conditions;

[0019] Output a drug stability assessment report.

[0020] Preferably, the adversarial example simulation module includes:

[0021] The adversarial network submodule is used for:

[0022] Construct an adversarial generative network that includes a generator and a discriminator;

[0023] Based on the molecular structure of the drug to be evaluated, the generator is trained to generate adversarial examples;

[0024] The quality of the generated adversarial examples is evaluated using the discriminator.

[0025] The sample optimization submodule, which is communicatively connected to the adversarial network submodule, is used for:

[0026] Receive adversarial samples generated by the adversarial network submodule;

[0027] The adversarial example is optimized based on a predefined loss function.

[0028] The optimized adversarial sample is output to the drug stability testing module.

[0029] Preferably, the loss function in the sample optimization submodule includes:

[0030] The adversarial loss term is used to measure the similarity between generated samples and real samples;

[0031] The key point variation loss term is used to evaluate the reasonableness of key structural variations in the generated samples;

[0032] The molecular structure similarity loss term is used to ensure the overall structural similarity between the generated sample and the original drug molecule.

[0033] Preferably, the drug stability testing module includes:

[0034] The environment simulation submodule is used for:

[0035] Construct a physical model of the drug storage environment that includes factors such as temperature, humidity, and light.

[0036] Real-time monitoring and adjustment of simulated environment parameters;

[0037] The digital twin submodule, which is communicatively connected to the environment simulation submodule, is used for:

[0038] Based on the data from the environmental simulation submodule, a digital twin model of the drug storage environment is constructed;

[0039] Simulate the long-term storage process of drugs in a virtual environment;

[0040] The multimodal fusion submodule, which is communicatively connected to the digital twin submodule, is used for:

[0041] Integrating physical environment data and digital twin simulation data;

[0042] A multimodal fusion deep learning algorithm is applied to generate a drug stability prediction model.

[0043] Preferably, the multimodal fusion submodule further includes:

[0044] Near-infrared spectroscopy analysis unit, used for:

[0045] Collect near-infrared spectral data of drug samples;

[0046] The near-infrared spectral data is fused with environmental simulation data and digital twin data;

[0047] Optimize the drug stability prediction model based on the fused multimodal data.

[0048] Preferably, the intelligent recognition module includes:

[0049] The molecular fragment analysis submodule is used for:

[0050] Based on the drug stability prediction model, sensitive fragments in drug molecules are identified.

[0051] Calculate the local sensitivity index for each molecular fragment;

[0052] Output a molecular fragment sensitivity analysis report;

[0053] The extreme condition extrapolation submodule, which is communicatively connected to the molecular fragment analysis submodule, is used for:

[0054] Receive the sensitivity analysis report of the molecular fragment;

[0055] Based on recurrent neural networks, the stability changes of drugs under extreme environmental conditions are simulated;

[0056] Generate drug stability prediction results under extreme conditions;

[0057] The evaluation report generation submodule, which is communicatively connected to the extreme condition extrapolation submodule, is used for:

[0058] A comprehensive analysis of drug stability prediction results under both routine and extreme conditions was conducted.

[0059] Generate an assessment report that includes a stability score, risk factors, and improvement recommendations.

[0060] Preferably, the extreme condition deduction submodule further includes:

[0061] The mass balance calculation unit is used for:

[0062] Based on the principles of chemical reaction kinetics, calculate the changes in drug components under extreme conditions;

[0063] Output the forward and reverse equilibrium factors of the mass balance equation;

[0064] The reaction rate analysis unit, which is communicatively connected to the substance balance calculation unit, is used for:

[0065] Based on the aforementioned forward and reverse equilibrium factors, the rates of each chemical reaction under extreme conditions are calculated;

[0066] Predict the degradation trend and rate of drugs under extreme conditions.

[0067] As a preferred option, it also includes:

[0068] The data intelligence interface module, which is communicatively connected to the intelligent recognition module, is used for:

[0069] Provides multi-platform access interfaces for cloud, mobile, and desktop;

[0070] To enable the visualization of drug stability assessment results;

[0071] It supports interactive user queries and analysis.

[0072] As a preferred option, it also includes:

[0073] The model optimization module, which is communicatively connected to the drug stability testing module and the intelligent recognition module, is used for:

[0074] Collect actual drug stability test data;

[0075] Compare the deviations between the predicted results and the actual results;

[0076] Based on deviation analysis, the parameters of the prediction model are adaptively adjusted.

[0077] Regularly update the various prediction models in the system.

[0078] An antigenic drug stability assessment method, using the aforementioned system, includes the following steps:

[0079] S1: Obtain molecular structure data of the drug to be evaluated;

[0080] S2: Based on the molecular structure data, generate adversarial samples with noise resistance, adaptability and stability using an adversarial generative network;

[0081] S3: Construct a physical model and digital twin model of the drug storage environment that includes factors such as temperature, humidity, and light.

[0082] S4: Input the adversarial examples into the physical model and digital twin model to simulate the state changes of the drug under different storage conditions and time.

[0083] S5: Collect near-infrared spectral data of drug samples and perform multimodal fusion with environmental simulation data and digital twin data;

[0084] S6: Apply a multimodal fusion deep learning algorithm to generate a drug stability prediction model based on the fused data;

[0085] S7: Using the aforementioned prediction model, perform molecular fragment sensitivity analysis to identify sensitive fragments in drug molecules;

[0086] S8: Based on the principles of recurrent neural networks and chemical reaction kinetics, simulate the stability changes of drugs under extreme environmental conditions;

[0087] S9: Analyze the drug stability prediction results under both normal and extreme conditions to generate an assessment report;

[0088] S10: Collect actual drug stability test data, compare the deviation between the predicted results and the actual results, and adaptively adjust the parameters of the prediction model.

[0089] This invention innovatively integrates advanced methods such as adversarial sample generation, environmental physics simulation, digital twin technology, and intelligent recognition to construct a powerful drug stability assessment platform. This comprehensive approach not only significantly improves the efficiency and accuracy of the assessment but also expands the system's applicability and analytical depth.

[0090] From a macro perspective, the system of this invention significantly shortens the drug stability assessment cycle from months or even years using traditional methods to just a few days. This high efficiency directly accelerates the new drug development process, enabling pharmaceutical companies to bring innovative drugs to market faster and benefit patients. At the same time, the system's high accuracy and comprehensiveness significantly reduce the risks and costs of drug development, improving the innovation efficiency of the entire pharmaceutical industry.

[0091] At the microscopic level, the system of this invention, through adversarial sample generation technology, can simulate minute structural changes in drug molecules, thereby comprehensively assessing the stability risks of drugs. This method can not only predict the stability of known drugs but also provide reliable assessments for novel drugs, filling a gap in existing technologies for handling novel molecular structures.

[0092] Furthermore, the system of this invention achieves accurate simulation of complex and variable storage environments through the combination of environmental physical simulation and digital twin technology. This solves the problem that traditional methods struggle to simulate long-term and extreme environmental conditions, making it possible to assess the stability of drugs throughout their entire life cycle. For example, the system can accurately predict changes in drug stability during transportation across multiple climate zones, which is of great significance for global pharmaceutical supply chain management.

[0093] The intelligent identification module of this invention, particularly its molecular fragment sensitivity analysis function, provides unprecedented molecular-level insights for drug development. This capability not only helps in understanding the intrinsic mechanisms of drug stability but also directly guides the structural optimization and formulation design of drug molecules. For example, by identifying key molecular fragments that affect stability, researchers can specifically improve drug structures to enhance their stability.

[0094] It is worth noting that the synergistic effect between the modules of this invention generates comprehensive advantages that transcend individual technologies. The combination of adversarial example generation and environmental simulation enables the system to assess drug stability in multiple dimensions; the combination of digital twin technology and intelligent recognition achieves comprehensive analysis from the macroscopic environment to the microscopic molecules. This multi-level, multi-angle approach not only improves the comprehensiveness and accuracy of the assessment but also provides rich decision support information for drug development.

[0095] The system of this invention also exhibits excellent adaptability and scalability. Through continuous learning and self-optimization, the system can continuously improve its prediction accuracy and adapt to new drugs and new environmental conditions. This adaptive capability ensures that the system maintains its long-term practical value in the rapidly evolving pharmaceutical industry.

[0096] In summary, the antigenic drug stability assessment system and method of this invention not only solves the efficiency, accuracy, and applicability issues of existing technologies, but also achieves a qualitative leap in drug stability assessment through innovative technology combinations and synergistic effects. It provides the pharmaceutical industry with a powerful tool that can accelerate new drug development, optimize quality control, reduce R&D risks, and ultimately drive the entire industry towards greater efficiency and accuracy, making a significant contribution to human health. Attached Figure Description

[0097] Figure 1 This is a system overall logic block diagram of the present invention;

[0098] Figure 2 This is a logic block diagram of the adversarial sample simulation module of the present invention;

[0099] Figure 3 This is a logic block diagram of the drug stability testing module of the present invention;

[0100] Figure 4 This is a logic block diagram of the intelligent recognition module of the present invention;

[0101] Figure 5 This is a logic block diagram of the model optimization module of the present invention. Detailed Implementation

[0102] Please refer to the attached document. Figure 1-5 This invention provides a system and method for assessing the stability of antigenic drugs, aiming to address the problems of insufficient accuracy in drug stability assessment and difficulty in simulating the effects of long-term and extreme conditions in existing technologies. The invention will be described in detail below with reference to specific embodiments.

[0103] The adversarial generative drug stability assessment system of the present invention includes an adversarial sample simulation module 1, a drug stability testing module 2, and an intelligent identification module 3. These three core modules realize a complete process from sample generation to stability assessment through data interaction and collaborative work.

[0104] The adversarial sample simulation module 1 is used to generate adversarial samples with noise resistance, adaptability, and stability based on the molecular structure of the drug to be evaluated. This module employs advanced adversarial generative network technology, enabling the generation of high-quality simulated samples, significantly improving the accuracy and reliability of subsequent stability assessments. Preferably, the adversarial sample simulation module 1 can generate samples based on the SMILES representation or 3D structural data of the drug molecule. For example, for a common antihypertensive drug such as amlodipine, its SMILES representation is "Clc1ccccc1C2C(=C( / N / C(=C2 / C(=O)OC)COCCN)c3ccccc3)C(=O)OCC", and the adversarial sample simulation module 1 will generate multiple samples with minor structural variations based on this structural representation.

[0105] The drug stability testing module 2 is communicatively connected to the adversarial example simulation module 1, and is used to receive adversarial examples and construct environmental physical simulations and digital twin models based on these examples. By combining actual physical environment simulation and virtual digital twin technology, this module can more comprehensively simulate the long-term behavior of drugs under various storage conditions. For example, the drug stability testing module 2 can simulate the stability changes of drugs under various environments such as 25°C / 60% relative humidity (normal conditions), 40°C / 75% relative humidity (accelerated testing conditions), and -20°C (freezing conditions). Through this method, the system of the present invention can obtain reliable predictions of the long-term stability of drugs in a short time.

[0106] The intelligent identification module 3 communicates with the drug stability testing module 2. Its core function is to perform molecular fragment sensitivity analysis and generate stability prediction results under extreme conditions. This module uses advanced machine learning algorithms, such as deep neural networks and support vector machines, to perform refined analysis of drug molecule structures. For example, for the aforementioned amlodipine molecule, the intelligent identification module 3 may identify the dihydropyridine ring as a key structure affecting its stability and predict that under extreme conditions such as pH values ​​below 3 or above 9, this structure may undergo hydrolysis, leading to drug inactivation.

[0107] The system of this invention also includes a data intelligence interface module (not shown in the figure), which provides multi-platform access interfaces for the cloud, mobile devices, and desktops, greatly improving the system's ease of use and versatility. For example, researchers can monitor the progress of drug stability trials in real time through a mobile application, while quality control personnel can conduct in-depth analysis of stability data through a desktop client.

[0108] In a preferred embodiment, the adversarial example simulation module 1 includes an adversarial network submodule 11 and a sample optimization submodule 12. The adversarial network submodule 11 constructs an adversarial generative network containing a generator and a discriminator. Both the generator and the discriminator employ a deep convolutional neural network structure, where the generator's task is to generate realistic drug molecule structure samples, and the discriminator is responsible for distinguishing whether these samples are real or generated.

[0109] The sample optimization submodule 12 receives adversarial samples generated by the adversarial network submodule 11 and optimizes these samples based on a predefined loss function. The design of the loss function is an innovation of this invention; it includes an adversarial loss term, a keypoint mutation loss term, and a molecular structure similarity loss term. The mathematical expressions of these three loss terms are as follows:

[0110] 1. Countermeasures against losses:

[0111] ,

[0112] in, Indicates the discriminator, Represents a generator. These are real samples. It is random noise input.

[0113] 2. Keypoint variation loss term:

[0114] ,

[0115] in, Indicates the first Feature extraction function for key points, That is the corresponding weight.

[0116] 3. Molecular structure similarity loss term:

[0117] ,

[0118] in, and These represent the sets of atoms in the generated molecule and the original molecule, respectively. This is actually calculating the Tanimoto similarity.

[0119] The total loss function is:

[0120] ,

[0121] in, , γ and γ are weighting coefficients used to balance the various loss terms. In practical applications, these coefficients can be adjusted according to the specific drug type and evaluation requirements. For example, for structurally complex macromolecular drugs, it may be necessary to increase γ. The value is used to ensure that the generated samples maintain sufficient structural similarity.

[0122] Through this carefully designed loss function, the system of the present invention can generate adversarial samples that retain the key characteristics of the original drug while possessing sufficient diversity, thereby providing a rich and reliable data foundation for subsequent stability assessment.

[0123] In practical applications, the system of this invention can effectively predict the stability of various drugs under different conditions. For example, for heat-sensitive protein drugs, the system may predict that at temperatures above 37°C, the protein may undergo significant conformational changes within 48 hours, thereby losing its activity. This prediction can help pharmaceutical companies optimize storage conditions and extend the shelf life of drugs.

[0124] In summary, the adversarial generative drug stability assessment system of this invention achieves comprehensive, accurate, and efficient assessment of drug stability through innovative adversarial sample generation technology, multimodal data fusion methods, and intelligent recognition algorithms. This not only accelerates the new drug development process but also improves the quality control level of existing drugs, which is of great significance for promoting technological progress in the pharmaceutical industry.

[0125] In a preferred embodiment of the present invention, such as Figure 2 As shown, the drug stability testing module 2 includes an environmental simulation submodule 21, a digital twin submodule 22, and a multimodal fusion submodule 23. This modular design enables the system of the present invention to comprehensively simulate and analyze the stability changes of drugs under various complex environments.

[0126] The environmental simulation submodule 21 is responsible for constructing a physical model of the drug storage environment, including factors such as temperature, humidity, and light. This submodule employs advanced sensor technology and a precision control system to accurately simulate various storage conditions. For example, when simulating tropical climate conditions, the environmental simulation submodule 21 can set the temperature to 30°C ± 2°C, control the relative humidity at 75% ± 5%, and simulate a 12-hour / day light cycle. This accurate environmental simulation provides a reliable experimental basis for drug stability assessment.

[0127] The digital twin submodule 22 is closely connected to the environmental simulation submodule 21, constructing a digital twin model of the drug storage environment based on actual environmental data. This virtual environment simulation technology allows the system to simulate the long-term storage process of drugs in a short time, greatly improving evaluation efficiency. For example, for new drugs requiring 36 months of long-term stability data, the system of this invention can generate reliable prediction results within a few weeks using digital twin technology, significantly shortening the drug development cycle.

[0128] The multimodal fusion submodule 23 is a key innovation of the system of this invention. This submodule is responsible for fusing physical environment data and digital twin simulation data, and applying a multimodal fusion deep learning algorithm to generate a drug stability prediction model. The multimodal fusion algorithm used here can be represented as follows:

[0129] ,

[0130] in, It is the feature representation after fusion. Indicates the first Data for each modality (such as temperature, humidity, light intensity, etc.). That is the corresponding weight matrix. It is a non-linear activation function, such as ReLU or Sigmoid.

[0131] By employing this method, the system of this invention can fully utilize data from diverse sources to generate more comprehensive and accurate stability prediction models. For example, when evaluating a photosensitizing drug, the system may discover an interaction between temperature and light intensity, where the drug's sensitivity to light increases significantly under high-temperature conditions. This complex interaction might be overlooked in traditional single-modal analysis.

[0132] In another embodiment of the invention, the multimodal fusion submodule 23 further includes a near-infrared spectroscopy analysis unit (not shown in the figure). This unit is responsible for acquiring near-infrared spectral data of the drug sample and fusing this data with environmental simulation data and digital twin data to further optimize the drug stability prediction model. Near-infrared spectroscopy technology can acquire drug molecular structure information non-destructively and rapidly, and is particularly suitable for monitoring minute changes in drugs during storage.

[0133] For example, for a common lipid-lowering drug like statins, a near-infrared spectroscopy unit might detect changes in the carbonyl peak within the wavenumber range of 1650–1800 cm⁻¹, which could indicate oxidative degradation of the drug. By correlating this spectral change with environmental condition data, the system can more accurately predict drug stability trends.

[0134] The intelligent recognition module 3 is one of the core components of the system of this invention. For example... Figure 3 As shown, this module includes a molecular fragment analysis submodule 31, an extreme condition simulation submodule 32, and an evaluation report generation submodule 33. This structural design enables the system to conduct in-depth analysis of drug stability at the molecular level and, based on this, perform stability simulations under extreme conditions.

[0135] The molecular fragment analysis submodule 31 identifies sensitive fragments in drug molecules based on a drug stability prediction model. This submodule employs graph neural network (GNN) technology to represent the drug molecule as a graph structure, where atoms are nodes and chemical bonds are edges. Through message passing and aggregation operations on the graph, the system can learn the contribution of different molecular fragments to drug stability. Mathematically, this process can be represented as:

[0136] ,

[0137] Represents a node In the Layer feature representation, It is a node The neighborhood group, and It is a learnable parameter matrix. It is a non-linear activation function.

[0138] Using this method, the system of the present invention can accurately calculate the local sensitivity index of each molecular fragment. For example, for β-lactam antibiotics, the system may identify the β-lactam ring as the most sensitive structure, with a local sensitivity index as high as 0.8 (assuming an index range of 0-1). This analytical result can directly guide the optimization of drug formulations and the improvement of packaging design.

[0139] The extreme condition simulation submodule 32 is another innovation of this invention. This submodule is based on a recurrent neural network (RNN) to simulate the stability changes of drugs under extreme environmental conditions. The advantage of RNNs is their ability to process time-series data, making them very suitable for simulating the stability trends of drugs over time. Its core formula can be expressed as:

[0140] ,

[0141] ,

[0142] yes The hidden state at all times It is input. It is the output. and These are the weights and bias parameters.

[0143] In practical applications, the extreme conditions simulation submodule 32 can simulate the stability changes of drugs within a temperature range of -40°C to 60°C and a relative humidity range of 0% to 100%. Stability data under such extreme conditions is crucial for assessing the safety of drug transportation and storage.

[0144] The assessment report generation submodule 33 is responsible for comprehensively analyzing the drug stability prediction results under both routine and extreme conditions to generate a comprehensive assessment report. This report not only includes stability scores and risk factors but also provides specific improvement recommendations. For example, for a biological product requiring cold chain transportation, the report might recommend transportation and storage within a temperature range of 0-2°C and warn that even short-term temperature fluctuations could lead to a significant decrease in drug activity.

[0145] Through this comprehensive and in-depth analysis, the antigenetic drug stability assessment system of this invention provides pharmaceutical companies with a powerful tool that can significantly improve drug development efficiency, reduce quality risks, and ultimately benefit patient populations.

[0146] In another preferred embodiment of the present invention, the system further includes a data intelligence interface module 4. This module is designed to provide users with a convenient, intuitive, and powerful user interface, enabling drug developers, quality control personnel, and even regulatory agencies to easily access and utilize the system's functions.

[0147] The Data Intelligence Interface Module 4 provides multi-platform access interfaces, including cloud, mobile, and desktop. This multi-platform design greatly improves the system's usability and flexibility. For example, R&D personnel can use the desktop client to conduct in-depth analysis of drug stability data, quality control personnel can use the mobile application to monitor the stability of production batches in real time, and management can view the overall drug stability assessment report at any time through the cloud interface.

[0148] Preferably, the data intelligence interface module 4 adopts a responsive design, adapting to different screen sizes and resolutions. This ensures a consistently good user experience whether on a large-screen workstation or a small-screen mobile device. For example, on the mobile interface, the system automatically adjusts the size and layout of data visualization charts to ensure key information is clearly visible.

[0149] The data intelligence interface module 4 also enables the visualization of drug stability assessment results. Advanced data visualization technologies are employed, such as interactive heatmaps and 3D molecular structure views. For example, when demonstrating the stability of a drug under different temperature and humidity conditions, the system might use an interactive 3D surface plot, with the x-axis representing temperature, the y-axis representing humidity, and the z-axis representing the drug's stability index. Users can rotate and zoom this 3D graph to intuitively understand the impact of environmental factors on drug stability.

[0150] Furthermore, the data intelligence interface module 4 supports interactive user queries and analysis. Users can quickly obtain stability prediction results under specific conditions by setting various parameters, such as temperature range, humidity range, and light intensity. The system also provides advanced query functions, allowing users to perform complex data analysis using SQL-like language. For example, a user might query "Which batches of drugs experienced a stability reduction of more than 10% under 40°C / 75%RH conditions in the past 6 months?" The system will quickly return data that meets the criteria and display it intuitively in chart form.

[0151] In another embodiment of the invention, the system further includes a model optimization module 5. The main function of this module is to improve the predictive accuracy and adaptability of the system through continuous learning and optimization. The model optimization module 5 is closely connected with the drug stability testing module 2 and the intelligent recognition module 3, forming a closed-loop, self-improving system.

[0152] Model optimization module 5 is first responsible for collecting actual drug stability test data. This data may come from accelerated stability tests in the laboratory, intermediate testing in the production workshop, or the results of regular random sampling of drugs on the market. For example, for a newly developed oral solid dosage form, model optimization module 5 may collect its actual stability data after storage at 25°C / 60%RH for 3 months, 6 months, and 12 months.

[0153] Next, model optimization module 5 compares the deviations between the predicted and actual results. Several statistical metrics are used to quantify the prediction deviation, including root mean square error (RMSE), mean absolute error (MAE), and the R² value. Mathematically, these metrics can be expressed as:

[0154] ,

[0155] MAE ,

[0156] ,

[0157] These are actual observed values. It is a predicted value. It is the average of the observed values. It refers to the number of samples.

[0158] Based on these bias analysis results, model optimization module 5 adaptively adjusts the prediction model parameters. An online learning algorithm based on gradient descent is employed, the core idea of ​​which is to minimize the prediction error. The optimization process can be expressed as:

[0159] ,

[0160] in, yes Model parameters at time 10:00 It's the learning rate. It is a loss function. This represents the gradient.

[0161] Through this continuous optimization process, the system of this invention can continuously improve its predictive accuracy. For example, the system may discover that the influence weight of humidity needs to be appropriately increased for certain types of drugs. Alternatively, the system may learn the impact of certain seasonal factors on drug stability, thereby incorporating elements of time series analysis into the predictive model.

[0162] Model optimization module 5 is also responsible for regularly updating the various prediction models in the system. This includes adversarial example generation models, environmental simulation models, and molecular fragment sensitivity analysis models. The update frequency can be dynamically adjusted based on the rate of data accumulation and changes in model performance. Typically, the system may undergo a small-scale update monthly and a large-scale update quarterly.

[0163] It is worth noting that the model optimization module 5 takes a series of measures to ensure the stability and reliability of the system when updating the model. For example, before deploying a new model, the system runs a comparative experiment between the old and new models in an isolated test environment. Only when the new model significantly outperforms the old model in all metrics will it be officially deployed. In addition, the system retains the model version control function, allowing for a quick rollback to a previous stable version if necessary.

[0164] By introducing the data intelligence interface module 4 and the model optimization module 5, the adversarial generative drug stability assessment system of this invention is not only more functionally complete, but also possesses the ability to continuously learn and self-optimize. This enables the system to adapt to the ever-changing drug development environment and provide long-term technical support for pharmaceutical companies.

[0165] Finally, this invention also provides a corresponding method for evaluating the stability of antigenic drugs. This method includes the following steps:

[0166] S1: Obtain molecular structure data of the drug to be evaluated;

[0167] S2: Based on the molecular structure data, generate adversarial samples with noise resistance, adaptability and stability using an adversarial generative network;

[0168] S3: Construct a physical model and digital twin model of the drug storage environment that includes factors such as temperature, humidity, and light.

[0169] S4: Input the adversarial examples into the physical model and digital twin model to simulate the state changes of the drug under different storage conditions and time.

[0170] S5: Collect near-infrared spectral data of drug samples and perform multimodal fusion with environmental simulation data and digital twin data;

[0171] S6: Apply a multimodal fusion deep learning algorithm to generate a drug stability prediction model based on the fused data;

[0172] S7: Using the aforementioned prediction model, perform molecular fragment sensitivity analysis to identify sensitive fragments in drug molecules;

[0173] S8: Based on the principles of recurrent neural networks and chemical reaction kinetics, simulate the stability changes of drugs under extreme environmental conditions;

[0174] S9: Analyze the drug stability prediction results under both normal and extreme conditions to generate an assessment report;

[0175] S10: Collect actual drug stability test data, compare the deviation between the predicted results and the actual results, and adaptively adjust the parameters of the prediction model.

[0176] This method, through a systematic approach, enables end-to-end drug stability assessment, encompassing data acquisition, sample generation, environmental simulation, and predictive analysis. In particular, it introduces innovative technologies at several key stages, such as generative adversarial networks, digital twins, and multimodal fusion deep learning, significantly improving the accuracy and efficiency of the assessment.

[0177] For example, in step S2, the use of an adversarial generative network enables the system to generate a large number of high-quality simulated samples that retain the key characteristics of the original drug while introducing sufficient variability. This provides a rich and reliable data foundation for subsequent stability analysis. In practical applications, for a novel anticancer drug, the system may generate thousands of molecular samples with slight structural variations, which may have minor differences in side chain length, substituent positions, or stereoconfiguration.

[0178] In steps S3 and S4, the combination of the physical model and the digital twin model enables the system to simulate the long-term storage process of a drug in a short period of time. This method is particularly suitable for evaluating new drugs that require long-term stability data. For example, for a biologic that needs to be stored at 2-8°C, the system can simulate its stability trends over a 3-year storage period within a few weeks.

[0179] The multimodal fusion in steps S5 and S6 is another innovation of this method. By fusing near-infrared spectral data, environmental simulation data, and digital twin data, the system can capture the complex patterns of drug stability being affected by multiple factors. For example, for a photosensitive drug, the system may find that its stability is not only affected by light intensity but also interacts with temperature and humidity.

[0180] The molecular fragment sensitivity analysis in step S7 provides direct guidance for drug formulation optimization. For example, for a common antihypertensive drug such as losartan, the system might identify the tetrazolium ring as the most sensitive structure, whose stability is easily affected by pH changes. Based on this finding, pharmaceutical companies can consider adding pH buffers to the formulation or improving packaging design to enhance drug stability.

[0181] The extreme condition simulations and comprehensive analyses in steps S8 and S9 enable this method to fully assess the stability risks of a drug. This is crucial for ensuring the safety of drugs during transportation and storage. For example, for a drug intended for use in tropical regions, the system might simulate long-term storage at 40°C / 75%RH and provide specific shelf-life and storage recommendations.

[0182] Finally, the adaptive optimization in step S10 ensures that the system can continuously learn and improve. With the accumulation of more real-world data, the accuracy of the prediction model will gradually increase. For example, the system may discover that certain types of drugs degrade faster under high-temperature conditions than initially predicted, and adjust the relevant parameters accordingly.

[0183] In summary, the antigenic drug stability assessment system and method provided by this invention innovatively combines multiple advanced technologies to achieve a comprehensive, accurate, and efficient assessment of drug stability. This not only accelerates the new drug development process and reduces development risks but also improves the quality control level of existing drugs, ultimately benefiting a broad patient population. In the future, with the further development of artificial intelligence technology and the accumulation of more practical application data, this system is expected to play an increasingly important role in the pharmaceutical industry, driving the entire industry towards greater efficiency and precision.

[0184] To fully verify the superiority of the anti-inflammatory drug stability assessment system and method of the present invention, a common antihypertensive drug—amlodipine—was selected as the research object, and a series of simulation experiments were conducted. The following is a detailed description and analysis results of the examples and comparative cases.

[0185] Example 1: Using the Antigenic Drug Stability Assessment System of the Present Invention

[0186] In this embodiment, the stability of amlodipine was comprehensively evaluated using the system of the present invention. First, the system generated 1000 adversarial samples based on the molecular structure of amlodipine (SMILES: CCc1ncc([N+](=O)[O-])n1CCO). These samples, while maintaining the original molecular core structure, introduced minor changes in the side chains and substituent positions. Subsequently, the system simulated the behavior of these samples under different storage conditions using an environmental simulation submodule and a digital twin submodule.

[0187] The system employs a multimodal fusion deep learning algorithm, combining near-infrared spectral data, environmental simulation data, and digital twin data to generate a stability prediction model for amlodipine. Through molecular fragment sensitivity analysis, the system identifies the dihydropyridine ring as the key structure affecting the stability of amlodipine. Based on recurrent neural networks and chemical reaction kinetics principles, the system also simulates the stability changes of amlodipine under extreme environmental conditions.

[0188] Comparative Example 1: Traditional Accelerated Stability Testing Method

[0189] In this comparative example, a conventional accelerated stability test method was used. Amlodipine samples were placed under accelerated testing conditions of 40°C / 75%RH, and samples were taken for analysis periodically at 0, 1, 2, 3, and 6 months. High-performance liquid chromatography (HPLC) was used to determine the sample content and assess its stability.

[0190] Comparative Example 2: Simple Machine Learning Prediction Methods

[0191] In this comparative example, a prediction method based on a simple machine learning algorithm (random forest) is used. This method considers only three factors: temperature, humidity, and time, and trains the model based on historical data to predict the stability of amlodipine.

[0192] Test metrics and methods:

[0193] 1. Prediction accuracy: The root mean square error (RMSE) is used to assess the deviation between the predicted results and the actual results.

[0194] 2. Time efficiency: Record the time required from inputting the molecular structure to obtaining a complete stability assessment report.

[0195] 3. Extreme Condition Prediction Capability: Assess the accuracy of predictions under temperature and humidity conditions that exceed normal ranges.

[0196] 4. Sensitive Fragment Identification: Evaluate whether the method can accurately identify key molecular fragments that affect drug stability.

[0197] Test results:

[0198]

[0199] The results are analyzed as follows:

[0200] 1. Prediction Accuracy: The system of this invention (Example 1) exhibits excellent prediction accuracy, with an RMSE of only 0.15%. This means that the system's predictions are very close to the actual observations. In contrast, the simple machine learning method (Comparative Example 2) has an RMSE of 2.8%, indicating significantly lower prediction accuracy. While the traditional accelerated experiment method (Comparative Example 1) can provide accurate experimental data, it cannot make predictions and is therefore unsuitable for this metric.

[0201] 2. Time Efficiency: The system of this invention can complete a comprehensive stability assessment in just 48 hours, significantly better than the 6 months required by traditional accelerated testing methods. Even compared to the 72 hours required by simple machine learning methods, this system demonstrates a significant time advantage. This high efficiency is of great significance for accelerating the drug development process.

[0202] 3. Extreme Condition Prediction Capability: By combining adversarial example generation and digital twin technology, the system of this invention can accurately simulate drug behavior under extreme conditions. For example, the system successfully predicted the degradation trend of amlodipine under 50°C / 90%RH conditions, which is difficult to achieve with traditional methods. Simple machine learning methods have poor prediction capabilities under extreme conditions, possibly due to the lack of samples under extreme conditions in the training data.

[0203] 4. Sensitive Fragment Identification: The system of this invention successfully identified the dihydropyridine ring as a key structure affecting the stability of amlodipine. This finding is consistent with literature reports and confirms the accuracy of the system's molecular fragment sensitivity analysis function. Traditional accelerated experiments and simple machine learning methods cannot provide this level of molecular insight.

[0204] Based on the above test results, Example 1 is undoubtedly the best embodiment. In this embodiment, the system of the present invention not only demonstrates excellent prediction accuracy and time efficiency, but also can handle stability prediction under extreme conditions and provide molecular-level insights into structure-stability relationships.

[0205] Specifically, when predicting the content of amlodipine at 35°C / 65%RH for 12 months, the system gave a predicted value of 98.7%, while the actual measured value was 98.5%, with a prediction error of only 0.2%. This high-precision prediction can help pharmaceutical companies more accurately determine the shelf life of drugs and optimize quality control strategies.

[0206] Furthermore, the system's identification of dihydropyridine ring sensitivity provides important guidance for amlodipine formulation development. Based on this finding, researchers can consider adding antioxidants to the formulation or improving packaging design to enhance the drug's long-term stability.

[0207] In summary, these test results fully demonstrate the superiority of the system of this invention in the field of drug stability assessment. It not only provides more accurate and faster predictions but also handles complex extreme conditions and offers molecular-level insights. These capabilities have significant practical implications for accelerating new drug development, optimizing existing drug formulations, and improving drug storage and transportation strategies.

[0208] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An antigenic drug stability assessment system, characterized in that, include: The adversarial example simulation module is used for: Based on the molecular structure of the drug to be evaluated, adversarial samples with noise resistance, adaptability and stability are generated. The adversarial sample is output to the drug stability testing module; The drug stability testing module, which is communicatively connected to the adversarial example simulation module, is used for: Receive adversarial samples sent by the adversarial sample simulation module; Based on the adversarial examples, construct environmental physical simulation and digital twin models; Generate drug stability prediction models under different storage conditions and time; The intelligent identification module, which is communicatively connected to the drug stability testing module, is used for: Receive the prediction model sent by the drug stability testing module; Based on the prediction model, perform molecular fragment sensitivity analysis; Generate drug stability prediction results under extreme conditions; Output a drug stability assessment report.

2. The antigenic drug stability assessment system according to claim 1, characterized in that, The adversarial example simulation module includes: The adversarial network submodule is used for: Construct an adversarial generative network that includes a generator and a discriminator; Based on the molecular structure of the drug to be evaluated, the generator is trained to generate adversarial examples; The quality of the generated adversarial examples is evaluated using the discriminator. The sample optimization submodule, which is communicatively connected to the adversarial network submodule, is used for: Receive adversarial samples generated by the adversarial network submodule; The adversarial example is optimized based on a predefined loss function. The optimized adversarial sample is output to the drug stability testing module.

3. The antigenic drug stability assessment system according to claim 2, characterized in that, The loss function in the sample optimization submodule includes: The adversarial loss term is used to measure the similarity between generated samples and real samples; The key point variation loss term is used to evaluate the reasonableness of key structural variations in the generated samples; The molecular structure similarity loss term is used to ensure the overall structural similarity between the generated sample and the original drug molecule.

4. The antigenic drug stability assessment system according to claim 1, characterized in that, The drug stability testing module includes: The environment simulation submodule is used for: Construct a physical model of the drug storage environment that includes factors such as temperature, humidity, and light. Real-time monitoring and adjustment of simulated environment parameters; The digital twin submodule, which is communicatively connected to the environment simulation submodule, is used for: Based on the data from the environmental simulation submodule, a digital twin model of the drug storage environment is constructed; Simulate the long-term storage process of drugs in a virtual environment; The multimodal fusion submodule, which is communicatively connected to the digital twin submodule, is used for: Integrating physical environment data and digital twin simulation data; A multimodal fusion deep learning algorithm is applied to generate a drug stability prediction model.

5. The antigenic drug stability assessment system according to claim 4, characterized in that, The multimodal fusion submodule also includes: Near-infrared spectroscopy analysis unit, used for: Collect near-infrared spectral data of drug samples; The near-infrared spectral data is fused with environmental simulation data and digital twin data; Optimize the drug stability prediction model based on the fused multimodal data.

6. The antigenic drug stability assessment system according to claim 1, characterized in that, The intelligent recognition module includes: The molecular fragment analysis submodule is used for: Based on the drug stability prediction model, sensitive fragments in drug molecules are identified. Calculate the local sensitivity index for each molecular fragment; Output a molecular fragment sensitivity analysis report; The extreme condition extrapolation submodule, which is communicatively connected to the molecular fragment analysis submodule, is used for: Receive the sensitivity analysis report of the molecular fragment; Based on recurrent neural networks, the stability changes of drugs under extreme environmental conditions are simulated; Generate drug stability prediction results under extreme conditions; The evaluation report generation submodule, which is communicatively connected to the extreme condition extrapolation submodule, is used for: A comprehensive analysis of drug stability prediction results under both routine and extreme conditions was conducted. Generate an assessment report that includes a stability score, risk factors, and improvement recommendations.

7. The antigenic drug stability assessment system according to claim 6, characterized in that, The extreme condition simulation submodule also includes: The mass balance calculation unit is used for: Based on the principles of chemical reaction kinetics, calculate the changes in drug components under extreme conditions; Output the forward and reverse equilibrium factors of the mass balance equation; The reaction rate analysis unit, which is communicatively connected to the substance balance calculation unit, is used for: Based on the aforementioned forward and reverse equilibrium factors, the rates of each chemical reaction under extreme conditions are calculated; Predict the degradation trend and rate of drugs under extreme conditions.

8. The antigenic drug stability assessment system according to claim 1, characterized in that, Also includes: The data intelligence interface module, which is communicatively connected to the intelligent recognition module, is used for: Provides multi-platform access interfaces for cloud, mobile, and desktop; To enable the visualization of drug stability assessment results; It supports interactive user queries and analysis.

9. The antigenic drug stability assessment system according to claim 1, characterized in that, Also includes: The model optimization module, which is communicatively connected to the drug stability testing module and the intelligent recognition module, is used for: Collect actual drug stability test data; Compare the deviations between the predicted results and the actual results; Based on deviation analysis, the parameters of the prediction model are adaptively adjusted. Regularly update the various prediction models in the system.

10. A method for assessing the stability of an antigenic drug, using the system described in any one of claims 1-9, characterized in that, Includes the following steps: S1: Obtain molecular structure data of the drug to be evaluated; S2: Based on the molecular structure data, generate adversarial samples with noise resistance, adaptability and stability using an adversarial generative network; S3: Construct a physical model and digital twin model of the drug storage environment that includes factors such as temperature, humidity, and light. S4: Input the adversarial examples into the physical model and digital twin model to simulate the state changes of the drug under different storage conditions and time. S5: Collect near-infrared spectral data of drug samples and perform multimodal fusion with environmental simulation data and digital twin data; S6: Apply a multimodal fusion deep learning algorithm to generate a drug stability prediction model based on the fused data; S7: Using the aforementioned prediction model, perform molecular fragment sensitivity analysis to identify sensitive fragments in drug molecules; S8: Based on the principles of recurrent neural networks and chemical reaction kinetics, simulate the stability changes of drugs under extreme environmental conditions; S9: Analyze the drug stability prediction results under both normal and extreme conditions to generate an assessment report; S10: Collect actual drug stability test data, compare the deviation between the predicted results and the actual results, and adaptively adjust the parameters of the prediction model.

Images