PPARG on-line molecular activity prediction system

Abstract

The invention discloses a PPARG online molecular activity prediction system, and relates to the technical field of activity prediction. The system comprises a data input and preprocessing module, a prediction efficiency evaluation module, a prediction efficiency regulation and control module, a prediction precision evaluation module and a prediction precision regulation and control module. Through data input and preprocessing, prediction efficiency and precision evaluation and regulation and control modules, full-process automation of PPARG molecular activity prediction is achieved, finally efficient and accurate prediction of single molecule second-level response and batch task minute-level completion is achieved, the PPARG molecular activity prediction precision can meet research and development requirements, and the research and development requirements are met. Therefore, the on-line molecular activity prediction efficiency and prediction precision of the PPARG are improved, and the problems of low on-line molecular activity prediction efficiency and prediction precision of the PPARG in the prior art are solved.

Description

Technical Field

[0001] This invention relates to the field of activity prediction technology, and in particular to the PPARG online molecular activity prediction system. Background Technology

[0002] First, users submit the SMILES string or structure file of the compound to be tested through a web interface. The backend system (such as a cheminformatics toolkit based on RDKit or Open Babel) standardizes it. Then, molecular descriptor calculation and feature engineering are performed to automatically extract hundreds to thousands of quantitative molecular descriptors from the preprocessed structure, including physicochemical properties (such as LogP, molecular weight), topological structure (such as connectivity index), electronic distribution (such as electrostatic potential, HOMO energy level of the highest occupied molecular orbital), etc. High-dimensional features are learned from the molecular graph using a deep learning-based graph neural network, and the chemical structure is transformed into a machine-readable numerical vector. Then, a pre-trained machine learning or deep learning model is invoked, based on random forest or support vector machine, and trained on a large dataset containing biological activity data of known PPARG (Peroxisome Proliferator-Activated Receptor Gamma) agonists / antagonists. The model takes the obtained feature vector as input, and outputs a predicted activity value or classification probability (such as agonist / non-agonist) through nonlinear mapping. At the same time, similarity-based methods (such as k-nearest neighbor algorithm) are also used in parallel to evaluate the structural similarity between the molecule and known active molecules, as auxiliary evidence for the prediction.

[0003] After obtaining preliminary prediction results, molecular docking simulation is performed to provide mechanistic explanations. Molecules are docked into the crystal structure of the PPARG ligand binding domain. By searching the conformational space and evaluating the results using scoring functions, information such as QSAR prediction scores and docking conformations is summarized and dynamically presented to the user through the front-end interface. This completes a full and efficient online PPARG molecular activity prediction process, from macroscopic property prediction to microscopic binding mechanism explanation.

[0004] However, in the process of implementing the inventive technical solution in the embodiments of this application, it was found that the above-mentioned technology has at least the following technical problems: On the one hand, traditional experimental methods rely on the synthesis of physical molecules and in vitro / in vivo activity verification processes. This process is cumbersome and requires a lot of human and material resources, as well as a long time cost, making it unsuitable for the actual needs of large-scale candidate molecule screening in biomedical research and development, and making it difficult to improve the efficiency of PPARG molecule activity prediction. On the other hand, single models used for predicting the activity of PPAR molecules have technical limitations. Traditional machine learning models (such as KNN and SVM) are unable to fully capture the intrinsic relationship between the structural and topological features of PPARG-related molecules and their activities. Deep learning models have significant shortcomings in fitting molecular structural features and mining local similarities. Furthermore, existing technologies lack technical mechanisms that can effectively integrate the advantages of different types of models. As a result, the predictive performance of PPARG molecular activity is limited by the adaptability of a single algorithm, and it is impossible to simultaneously take into account the comprehensiveness of molecular feature extraction and the stability of prediction results. Consequently, the accuracy of PPARG molecular activity prediction cannot meet the research and development needs, leading to the problem of low efficiency and low accuracy in online PPARG molecular activity prediction. Summary of the Invention

[0005] This invention provides an online PPARG molecular activity prediction system, which can achieve PPARG molecular activity prediction accuracy that meets research and development needs, thereby improving the efficiency and accuracy of online PPARG molecular activity prediction. The technical solution provided by this application is as follows: According to a first aspect of this application, a PPARG online molecular activity prediction system is provided. This system includes: a data input and preprocessing module, a prediction efficiency evaluation module, a prediction efficiency control module, a prediction accuracy evaluation module, and a prediction accuracy control module. The data input and preprocessing module supports manual input of single-molecule SMILES sequences and batch uploading of multi-molecule Excel files, and performs preprocessing to obtain molecular feature vectors. The prediction efficiency evaluation module loads a pre-trained model based on the molecular feature vectors to perform PPARG molecular activity prediction, obtains prediction efficiency parameters, and obtains a molecular activity prediction efficiency evaluation value based on the prediction efficiency parameters to reflect the high efficiency performance of PPARG online molecular activity prediction. The prediction efficiency control module determines whether to perform activity prediction based on the molecular activity prediction efficiency evaluation value. The system dynamically adjusts efficiency to achieve a balance between resource utilization efficiency and online molecular activity prediction response speed. If the adjustment is successful, the online molecular activity prediction result is output after adjustment; otherwise, the online molecular activity prediction result is output directly. The prediction accuracy evaluation module acquires the prediction accuracy parameters during the output of the online molecular activity prediction result and obtains the molecular activity prediction accuracy evaluation value based on the prediction accuracy parameters to reflect the accuracy of the model in adapting to drug development for PPARG online molecular activity prediction. The prediction accuracy control module determines whether to perform dynamic adjustment of activity prediction accuracy based on the molecular activity prediction accuracy evaluation value to ensure the reliability of molecular activity prediction accuracy. If successful, the prediction result visualization step is performed after adjustment; otherwise, the prediction result visualization step is performed directly, forming a fully automated closed loop of input-prediction-output.

[0006] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. By quantifying the overall processing efficiency through the prediction efficiency evaluation module and combining it with the dual adaptive control mechanism of the prediction efficiency control module, a precise balance between system resource utilization and prediction response speed is achieved. For scenarios with varying numbers of parallel threads in batch molecule preprocessing, the threshold for the number of task fragments is dynamically adjusted to avoid resource waste caused by idle threads or scheduling overhead caused by excessive fragmentation, ensuring maximum parallel processing efficiency. Facing different concurrent prediction request volumes, the batch prediction task completion time interval is adaptively controlled. In high-concurrency scenarios, the interval is shortened to increase throughput, while in low-concurrency scenarios, the interval is maintained to ensure service stability. This effectively supports high-performance indicators such as second-level response for single molecules and minute-level completion for thousands of molecules. This dynamic adaptation mechanism can flexibly cope with varying task scales and request pressures, significantly improving the amount of molecules processed per unit time and solving the pain points of rigid resource configuration and efficiency degradation in high-concurrency scenarios in traditional prediction systems.

[0007] 2. By combining prediction accuracy assessment with dynamic adjustment, the prediction accuracy assessment module integrates multi-dimensional accuracy parameters to comprehensively measure the consistency between prediction results and experimental data, as well as the model's generalization ability, avoiding the one-sidedness of single-index assessment. For scenarios where the accuracy assessment value is below standard, adaptive adjustment of the number of shared instances is used. The number of loaded instances is dynamically adjusted based on the model prediction execution time, solving the prediction bias caused by single-instance resource overload and avoiding the waste of memory resources occupied by idle instances. Adaptive adjustment of the output latency threshold adapts to the model loading latency, preventing result transmission interruptions or service anomalies due to insufficient latency threshold, ensuring stable output of high-precision prediction results. This closed-loop system ensures the model's robustness in predicting molecules with different structural types (such as active cliff compounds and structurally different molecules), effectively improving the reliability of prediction results and providing a reliable scientific basis for candidate molecule screening in drug development.

[0008] 3. Based on the data input and preprocessing modules, it supports both manual input of single molecules and batch upload of multiple molecules. Utilizing the RDKit tool, it achieves fully automated data verification, standardization, and feature conversion. Combined with the automatic loading of pre-trained models and parameter configuration-free design, it significantly lowers the operational threshold for researchers. The prediction result visualization module provides real-time display of single molecules and batch result export to Excel, forming a fully automated closed loop of input, preprocessing, prediction, dual evaluation, regulation, and output. This completes the transformation from molecular data to accurate prediction results without manual intervention. This closed-loop process not only avoids the drawbacks of traditional experimental methods—long cycles, high costs, and low throughput—but also significantly shortens the R&D cycle for PPARG-related molecular activity evaluation through dual guarantees of efficiency and accuracy, improving candidate molecule screening efficiency. It provides efficient and accurate technical support for drug development by biopharmaceutical companies and research institutions, accelerating the new drug development process.

[0009] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0010] The accompanying drawings are provided for a better understanding of this solution and do not constitute a limitation of this application. Wherein: Figure 1 This is a schematic diagram of the structure of the PPARG online molecular activity prediction system provided in an embodiment of the present invention; Figure 2 This is a transaction timeline diagram of the PPARG online molecular activity prediction system provided in this embodiment of the invention. Detailed Implementation

[0011] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of this application, including various details to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0012] This invention provides an online molecular activity prediction system for PPARG, such as... Figure 1 The diagram shown illustrates the structure of the PPARG online molecular activity prediction system, which includes: a data input and preprocessing module, a prediction efficiency evaluation module, a prediction efficiency control module, a prediction accuracy evaluation module, and a prediction accuracy control module.

[0013] As the first module of this system, the data input and preprocessing module supports manual input of single-molecule SMILES sequences and batch uploading of multi-molecule Excel files, and performs preprocessing to obtain molecular feature vectors. The preprocessing process integrates the RDKit tool to automatically perform SMILES validity checks, null value detection, molecular structure standardization, and molecular feature vector transformation.

[0014] It should be understood that the data input and preprocessing module, as the basic data processing unit of the PPARG online molecular activity prediction system, primarily supports two convenient data input modes: First, a manual input mode for single-molecule SMILES sequences, where users can directly enter the SMILES sequence of a single molecule on the front-end interface. The system simultaneously provides format examples and input specification prompts to help users ensure that the entered data format is legal and valid. Second, a batch upload mode for multi-molecule Excel files, where users can download standardized templates (the templates include a mandatory "SMILES" field to ensure a unified data structure), fill in the templates, and then upload multi-molecule data, significantly improving the input efficiency of large-scale molecular data. After data reception, the module integrates the RDKit tool to initiate a fully automated preprocessing process: First, SMILES validity verification is performed, eliminating invalid sequences with incorrect formats or those that cannot be parsed; simultaneously, null value detection is performed, filtering out empty entries without valid SMILES information; then, molecular structure standardization processing is carried out to unify the expression form of molecular structures, avoiding the impact of structural characterization differences on subsequent prediction accuracy; finally, molecular feature vector conversion is completed, and the RDKit tool calculates and generates two types of core feature vectors: molecular fingerprints and molecular descriptors, providing standardized, high-quality structured input data for subsequent model predictions.

[0015] The second module of this system, the prediction efficiency evaluation module, is used to load a pre-trained model based on molecular feature vectors to perform PPARG molecular activity prediction in order to obtain prediction efficiency parameters. Based on the prediction efficiency parameters, a molecular activity prediction efficiency evaluation value is obtained to reflect the high efficiency performance of PPARG online molecular activity prediction.

[0016] It should be understood that prediction efficiency parameters include single-molecule data input time, batch task processing throughput, and average time per prediction call. Single-molecule data input time refers to the total time from when the user completes the input of a single-molecule SMILES sequence on the front-end interface and clicks the submit button, to when the data is received and the preprocessing process is triggered. Batch task processing throughput refers to the number of molecules that can be predicted within a preset unit of time. Average time per prediction call refers to the average computation time consumed from inputting a preprocessed feature vector until the Stacking ensemble model returns the prediction result.

[0017] By correcting the result of the single-molecule data input time normalization process using the input time impact factor, the input time impact component is obtained. The specific constraint expression for the input time impact component is as follows: ; In the formula, E represents the input time-consuming influence component; n1 represents the input time-consuming influence factor obtained from the predefined activity prediction database; S max S represents the critical upper limit of the input time taken to obtain the data from a predefined activity prediction database; min This represents the lower limit of the input time threshold for retrieving data from a predefined activity prediction database; S represents the input time for single-molecule data. There is a negative correlation between the input time for single-molecule data and the molecular activity prediction efficiency evaluation value. The longer the input time for single-molecule data, the worse the user experience and the efficiency of single-molecule validation, resulting in a lower molecular activity prediction efficiency evaluation value.

[0018] By correcting the throughput impact factor for the normalized throughput of batch task processing, the throughput impact component is obtained. The specific constraint expression for the throughput impact component is as follows: ; In the formula, A represents the throughput impact component; n² represents the throughput impact factor obtained from the predefined activity prediction database; Q max This represents the critical upper limit of throughput obtained from a predefined activity prediction database; Q min This represents the critical lower limit of throughput obtained from the predefined activity prediction database; Q represents the batch task processing throughput. There is a positive correlation between batch task processing throughput and the molecular activity prediction efficiency evaluation value. The higher the batch task processing throughput, the more molecules can be predicted per unit time, resulting in better batch screening efficiency and a higher molecular activity prediction efficiency evaluation value.

[0019] The average time-consuming influence component is obtained by correcting the result of normalizing the average time-consuming process for a single prediction call using the average time-consuming influence factor. The specific constraint expression for the average time-consuming influence component is as follows: ; In the formula, M represents the average time-consuming influence component; n3 represents the average time-consuming influence factor obtained from the predefined activity prediction database; T max T represents the upper limit of the average time taken from a predefined activity prediction database; minThis represents the lower limit of the average time required to obtain data from the predefined activity prediction database; T represents the average time required for a single prediction call. The average time required for a single prediction call is negatively correlated with the molecular activity prediction efficiency evaluation value. A longer average time for a single prediction call affects the core time-consuming aspects of single-molecule and batch tasks, resulting in a lower molecular activity prediction efficiency evaluation value.

[0020] The input time-consuming component, throughput-consuming component, and average time-consuming component are coupled to obtain a molecular activity prediction efficiency evaluation value. This value is used to measure the overall processing efficiency and batch processing capacity of PPARG's online molecular activity prediction process, including molecular data input (single-molecule SMILES input, multi-molecule Excel batch upload), automated preprocessing, model prediction, and result output. This efficiency and capacity are reflected by performance indicators such as second-level response time for single-molecule prediction and minute-level batch prediction for thousands of molecules. The specific limiting expression for the molecular activity prediction efficiency evaluation value is as follows: ; In the formula, B represents the molecular activity prediction efficiency assessment value.

[0021] It should be noted that the positive and negative correlations among single-molecule data input time, batch task processing throughput, and average time per prediction call need to be considered in conjunction with the actual workflow of PPARG online molecular activity prediction, including input-preprocessing-prediction-output and resource constraints. Specifically: there is a negative correlation between single-molecule data input time and batch task processing throughput. The longer the single-molecule data input time, the more insufficient the resources. In batch processing, the total preprocessing time for thousands of molecules will increase proportionally, and the number of molecules predicted per unit time will decrease, thus reducing the batch task processing throughput. There is also a negative correlation between batch task processing throughput and average time per prediction call. The longer the average time per prediction call, the more the model inference needs to process multiple feature vectors simultaneously (parallel inference). If the time per call is too long, it indicates low model inference efficiency (e.g., insufficient GPU memory leading to frequent data swapping). Resource competition during parallel inference will further prolong the total time, reducing the batch task processing throughput.

[0022] As the third module of this system, the prediction efficiency control module determines whether to perform dynamic adjustment of the activity prediction efficiency based on the molecular activity prediction efficiency evaluation value in order to achieve a balance between resource utilization efficiency and online molecular activity prediction response speed. If yes, the online molecular activity prediction result is output after adjustment; otherwise, the online molecular activity prediction result is output directly.

[0023] Furthermore, the specific steps for determining whether to perform dynamic adjustment of activity prediction efficiency based on the molecular activity prediction efficiency evaluation value are as follows: compare the molecular activity prediction efficiency evaluation value with the preset prediction efficiency reference value. If the molecular activity prediction efficiency evaluation value is greater than or equal to the prediction efficiency reference value, then dynamic adjustment of activity prediction efficiency will not be performed; otherwise, dynamic adjustment of activity prediction efficiency will be performed. Dynamic adjustment of activity prediction efficiency includes adaptive control of the threshold for the number of task segments and adaptive control of the interval between prediction task completion.

[0024] As further explained in detail, the specific process for adaptive adjustment of the task fragmentation number threshold is as follows: The system retrieves the number of parallel threads running in the PPARG online molecular activity prediction preprocessing service, comparing this number with a control value. By dynamically matching the number of slices to the thread processing capacity, efficient utilization of parallel processing resources is achieved.

[0025] If the number of parallel threads in batch molecular preprocessing is less than or equal to the lower limit of the reference number of parallel threads, the parallel thread offset is input into the preset mapping relationship model to obtain the task fragment number threshold reduction coefficient. The task fragment number threshold reduction coefficient is multiplied by the current batch prediction task fragment number threshold to obtain the target batch prediction task fragment number threshold, so as to avoid the waste of parallel processing resources caused by thread idleness. The task fragment mapping relationship model is configured such that the input and output are positively correlated. The positive correlation ratio is pre-calibrated based on the prediction efficiency offset, the number of parallel threads in batch molecular preprocessing, and the batch prediction task fragment number threshold to ensure that the target task fragment number is adapted to the thread processing capacity. The input is the number of parallel threads in batch molecular preprocessing, and the output is the target task fragment number threshold. The prediction efficiency offset represents the negative difference between the molecular activity prediction efficiency evaluation value and the preset prediction efficiency reference value. The parallel thread offset represents the negative difference between the number of parallel threads in batch molecular preprocessing and the lower limit of the reference number of parallel threads. When the number of threads is detected to be less than or equal to a reference lower limit, it indicates that the system's parallel processing capacity is excessive, and the current number of task shards may be too high, leading to resource waste such as thread idleness and frequent scheduling. At this point, the system initiates a "resource-saving" adjustment: calculating the "parallel thread offset" and obtaining a "task shard number threshold reduction coefficient" less than 1 through a preset positive correlation mapping model. Multiplying this coefficient by the current threshold yields a smaller target threshold that matches the current low-concurrency thread count. This dynamically reduces the upper limit of the number of subtasks into which batch tasks are split. This ensures that the number of task shards precisely corresponds to the currently available, limited number of parallel threads, effectively eliminating thread idleness, frequent scheduling, and context switching overhead caused by excessive task shards, ultimately reducing overall resource consumption and embodying the "on-demand allocation" principle in intelligent resource management.

[0026] Adaptive control of the task fragment number threshold also includes: If the number of parallel threads for batch molecular preprocessing is within the reference range for the number of parallel threads, then the adaptive adjustment of the task shard number threshold will not be performed, and the current threshold for the number of batch prediction task shards will be maintained. The reference range for the number of parallel threads includes the open interval formed by the lower limit and the upper limit of the reference range for the number of parallel threads. This avoids system jitter and performance overhead caused by unnecessary and frequent configuration adjustments. It ensures the configuration stability and operational continuity of the system when handling normal loads, and ensures that the preprocessing service can provide a continuous and reliable data stream to the backend "model prediction and result encapsulation" module, which is an important guarantee for achieving "efficient response" and "stable operation" throughout the entire process.

[0027] If the number of parallel threads for batch molecular preprocessing is greater than or equal to the reference upper limit for the number of parallel threads, the parallel thread correction amount is input into the preset task partitioning mapping model to obtain the task partitioning number threshold gain coefficient. This task partitioning number threshold gain coefficient is then multiplied by the current batch prediction task partitioning number threshold to obtain the target batch prediction task partitioning number threshold, thereby improving the parallel processing efficiency of batch molecular preprocessing. Here, the parallel thread correction amount represents the positive difference between the number of parallel threads for batch molecular preprocessing and the reference upper limit for the number of parallel threads. By increasing the number of task partitions, the massive batch molecular dataset can be divided into more and finer-grained subtasks. This is like adding more workstations to a busy production line, enabling the system to fully utilize the processing power of all high-concurrency threads, allowing each thread to efficiently process tasks, thus significantly improving the parallel processing efficiency and overall throughput of batch molecular preprocessing. This achieves the performance vision of "thousand-level molecular batch prediction completed in minutes" and lays a solid technical foundation for supporting larger-scale predictions in the future.

[0028] In this embodiment, by increasing the number of partitions, the parallel processing capability of multithreading is fully released, allowing more effective threads to participate in subtask processing simultaneously. This significantly improves the parallel processing efficiency and overall throughput of batch molecular preprocessing, solving the pain point of insufficient utilization of thread resource redundancy in the traditional fixed partitioning mode. The entire control process responds in stages to different thread count scenarios, avoiding resource waste while maximizing parallel processing efficiency. It achieves dynamic optimal matching between task partitioning and thread capabilities, providing core support for the system to efficiently handle batch molecular preprocessing tasks of different scales.

[0029] As further explained in detail, the specific process for adaptive adjustment of the prediction task completion interval is as follows: Get the number of concurrent predicted requests currently pending and compare it with the reference value for the number of requests.

[0030] If the number of concurrent prediction requests is less than or equal to the reference value, the adaptive adjustment of the prediction task completion interval will not be executed, and the current batch prediction task completion interval will be maintained. This avoids introducing unnecessary adjustment delays when the system pressure is low. It ensures that tasks can be submitted and processed as quickly and uninterruptedly as possible, thereby guaranteeing low-latency response and stable and smooth operation of the system under normal circumstances, providing users with a good interactive experience.

[0031] If the number of concurrent prediction requests exceeds the reference value, the prediction request offset is input into a preset time interval mapping model to obtain a time interval adjustment factor. The current batch prediction task completion time interval is multiplied by the time interval adjustment factor and then rounded down to the nearest integer as the target batch prediction task completion time interval. This improves the throughput and response timeliness of the entire molecular activity prediction process. The time interval mapping model is configured such that the input and output are negatively correlated. The negative correlation coefficient is pre-calibrated based on the prediction efficiency offset, the number of concurrent prediction requests, and the batch prediction task completion time interval to ensure the target batch prediction task completion time interval matches the concurrent request handling capacity. The prediction request offset represents the positive difference between the number of concurrent prediction requests and the reference value. By significantly shortening the task completion interval, the system speeds up task delivery to the backend processing module, allowing CPU, memory, and other computing resources to be deployed earlier and more intensively, reducing resource idle waiting time and greatly improving the throughput efficiency of the entire data pipeline from request reception and preprocessing to model prediction. Faster task delivery means that the waiting time of a single task in the queue is shortened, and the overall completion time of the entire batch task is advanced. This allows users to obtain prediction results faster, achieving the goal of "high-efficiency response capability", especially in high-concurrency scenarios, thereby ensuring the continuous stability and service reliability of the system under high load pressure.

[0032] In this embodiment, dynamic optimization of system throughput and response timeliness is achieved by precisely matching the concurrent request volume with the task time interval configuration. By acquiring the current number of concurrent prediction requests to be processed in real time and comparing it with a preset request reference value, when the number of concurrent prediction requests is less than or equal to the reference value, the current batch prediction task completion time interval is maintained, avoiding unnecessary system overhead caused by adjustments and ensuring service stability and processing efficiency in low-concurrency scenarios. When the number of concurrent prediction requests exceeds the reference value, the prediction request offset, representing the degree of request excess, is input into a negatively correlated time interval mapping model to generate a time interval down-adjustment factor. This factor is then multiplied and rounded down to obtain the target time interval adapted to high-concurrency scenarios. By shortening the start interval of continuous batch prediction tasks, the system's idle resources are fully utilized to improve task scheduling density. At the same time, by leveraging the proportional coefficient based on prediction efficiency offset, number of concurrent requests, and current time interval calibration, the system ensures that the target interval is neither too short, leading to increased resource contention, nor too long, affecting throughput. Ultimately, this achieves a significant improvement in throughput and precise guarantee of response timeliness throughout the entire molecular activity prediction process in high-concurrency scenarios. It effectively solves the pain points of low-concurrency resource idleness and high-concurrency response lag in the traditional fixed time interval mode, allowing the system to flexibly adapt to varying request pressures and maintain efficient and stable operation under different load scenarios.

[0033] As the fourth module of this system, the prediction efficiency control module determines whether to perform dynamic adjustment of the activity prediction efficiency based on the molecular activity prediction efficiency evaluation value in order to achieve a balance between resource utilization efficiency and online molecular activity prediction response speed. If yes, the online molecular activity prediction result is output after adjustment; otherwise, the online molecular activity prediction result is output directly.

[0034] It should be understood that the specific steps to obtain the molecular activity prediction accuracy assessment value are as follows: The prediction accuracy parameters include the prediction efficiency evaluation value for the control group, the molecular activity prediction cycle, and the molecular activity prediction rate. The prediction efficiency evaluation value for the control group refers to the value obtained again if dynamic adjustment of the activity prediction efficiency has been performed; otherwise, the current molecular activity prediction efficiency evaluation value is used. The molecular activity prediction cycle refers to the total time from when the user initiates a PPARG molecular activity prediction request (single molecule submission / batch file upload) to when the molecular activity prediction processing is completed and the final prediction result is returned to the user. The single molecule prediction cycle refers to the total time from when the user inputs and submits a single SMILES sequence, through front-end verification, back-end reception, preprocessing, model prediction, and result display. The batch prediction cycle refers to the total time from when the user uploads an Excel file containing multiple molecules, through file upload, format parsing, batch preprocessing, batch model prediction, result file generation, and download notification. The molecular activity prediction rate refers to the number of PPARG molecules whose activity prediction can be completed within a preset unit of time.

[0035] The predictive efficiency influence component is obtained by correcting the results of the analysis of the proportion of the predicted efficiency evaluation value and the critical value of the predicted efficiency by the predictive efficiency influence factor. The specific constraint expression for is as follows: ; In the formula, F represents; v1 represents the value obtained from the predefined activity prediction database; D0 represents the value obtained from the predefined activity prediction database; and D represents the value obtained from the predefined activity prediction database. There is a positive correlation between the evaluation value of the prediction efficiency of the control and the evaluation value of the molecular activity prediction accuracy. The larger the evaluation value of the prediction efficiency of the control, the higher the efficiency of the entire molecular activity prediction process, and the higher the evaluation value of the molecular activity prediction accuracy.

[0036] By correcting the result of normalizing the molecular activity prediction period using the prediction period influence factor, the prediction period influence component is obtained. The specific constraint expression for is: ; In the formula, G represents the prediction cycle influence component; v2 represents the prediction cycle influence factor obtained from the predefined activity prediction database; G maxThis represents the critical upper limit of the prediction period obtained from a predefined activity prediction database; G min This represents the critical lower limit of the prediction period obtained from the predefined activity prediction database; G represents the molecular activity prediction period. There is a negative correlation between the molecular activity prediction period and the molecular activity prediction accuracy assessment value. The longer the molecular activity prediction period, the more resource competition leads to data preprocessing and model inference queuing, resulting in incomplete data transmission during model inference and a lower molecular activity prediction accuracy assessment value.

[0037] The predicted rate influence component is obtained by correcting the result of normalizing the molecular activity predicted rate using the predicted rate influence factor. The specific constraint expression for is as follows: ; In the formula, Q represents the predicted rate influence component; v3 represents the predicted rate influence factor obtained from a predefined activity prediction database; R max R represents the critical upper limit of the prediction rate obtained from a predefined activity prediction database; min This represents the critical lower limit of the prediction rate obtained from the predefined activity prediction database; R represents the molecular activity prediction rate. There is a positive correlation between the molecular activity prediction rate and the molecular activity prediction accuracy assessment value; the faster the molecular activity prediction rate, the faster the processing speed and throughput of molecules, and the higher the molecular activity prediction accuracy assessment value.

[0038] The prediction efficiency, prediction period, and prediction rate components are coupled to obtain a molecular activity prediction accuracy evaluation value, which is used to measure the consistency between the prediction results (such as activity classification, activity value, binding energy, etc.) output by PPARG online molecular activity prediction and experimental results. The specific constraint expression is as follows: ; In the formula, L represents the molecular activity prediction accuracy assessment value.

[0039] It should be noted that, based on the workflow logic (request initiation, data processing, model prediction, result output) and parameter definitions of the PPARG online molecular activity prediction system, the positive and negative correlations among the control prediction efficiency evaluation value, molecular activity prediction cycle, and molecular activity prediction rate are as follows: The control prediction efficiency evaluation value is negatively correlated with the molecular activity prediction cycle; a higher value significantly improves the efficiency of data analysis, preprocessing, and model inference, resulting in a shorter molecular activity prediction cycle. The control prediction efficiency evaluation value is positively correlated with the molecular activity prediction rate; a higher value results in faster molecule processing and throughput, leading to a faster molecular activity prediction rate. The molecular activity prediction cycle is negatively correlated with the molecular activity prediction rate; a longer cycle results in a longer total time for the entire PPARG online molecular activity prediction process and a slower molecular activity prediction rate.

[0040] As the fifth module of this system, the prediction accuracy control module determines whether to dynamically adjust the activity prediction accuracy based on the molecular activity prediction accuracy assessment value to ensure the reliability of the molecular activity prediction accuracy. If so, the prediction result visualization step is performed after adjustment; otherwise, the prediction result visualization step is performed directly, forming a fully automated closed loop of input-prediction-output. The prediction result visualization step displays the single-molecule prediction results in real time on the front-end interface, and batch prediction results are exported in Excel format with a one-click export function.

[0041] Furthermore, the specific process for determining whether to perform dynamic adjustment of activity prediction accuracy is as follows: compare the molecular activity prediction accuracy assessment value with the preset prediction accuracy reference value. If the molecular activity prediction accuracy assessment value is greater than or equal to the prediction accuracy reference value, then dynamic adjustment of activity prediction accuracy is not performed; otherwise, dynamic adjustment of activity prediction accuracy is performed. Dynamic adjustment of activity prediction accuracy includes adaptive control of the number of shared instances and adaptive control of the output delay threshold.

[0042] As further explained in detail, the specific steps for adaptive adjustment of the number of shared instances are as follows: Obtain the single-time model prediction execution time of the current Stacking ensemble model and compare it with the execution time reference interval, which represents the closed interval formed by the lower limit of the execution time reference and the upper limit of the execution time reference.

[0043] If the execution time of a single model prediction is greater than or equal to the upper limit of the execution time reference, the execution time offset is input into the preset instance number mapping relationship model to obtain the shared instance number adjustment. The current model loading shared instance number is coupled with the shared instance number adjustment to obtain the target model loading shared instance number to reduce the time extension caused by single instance resource overload. The instance number mapping relationship model is configured such that the input and output are positively correlated. The input represents the execution time of a single model prediction, and the output represents the target model loading shared instance number. The positive correlation ratio coefficient of the instance number mapping relationship model is pre-calibrated based on the prediction accuracy offset, the execution time of a single model prediction, and the model loading shared instance number to ensure that the target model loading shared instance number is adapted to the execution time of a single model prediction. The prediction accuracy offset represents the negative difference between the molecular activity prediction accuracy evaluation value and the preset prediction accuracy reference value, and the execution time offset represents the positive difference between the execution time of a single model prediction and the upper limit of the execution time reference. By adding parallel model instances, the system can distribute the influx of prediction requests across multiple instances, significantly reducing the CPU / memory load on a single instance and avoiding processing delays caused by resource contention.

[0044] The adaptive adjustment of the number of shared instances also includes: If the execution time of a single model prediction is less than or equal to the lower limit of the execution time reference, the execution time correction is input into the preset instance number mapping model to obtain the shared instance number reduction. The difference between the current model-loaded shared instance number and the shared instance number reduction is then calculated to obtain the target model-loaded shared instance number. This reduces the time extension caused by single instance resource overload, thereby reducing the number of instances and avoiding idle instances consuming system memory resources. The execution time correction represents the negative difference between the execution time of a single model prediction and the lower limit of the execution time reference. Deep learning model instances (such as GCN) typically consume a large amount of memory. Reducing the number of unnecessary instances can immediately free up valuable memory resources and alleviate system memory pressure. The freed memory and computing resources can be used by other modules of the system (such as preprocessing or feature calculation), thereby improving the overall resource utilization efficiency and stability of the entire PPARG prediction system and avoiding unnecessary resource waste.

[0045] In this embodiment, by dynamically matching the number of shared instances loaded in the model with the execution time of a single model prediction, the system achieves coordinated optimization of resource utilization efficiency, prediction accuracy, and response speed. The core technology has significant and comprehensive effects: By acquiring the execution time of a single prediction of the Stacking integrated model in real time and comparing it with a preset execution time reference range, the resource requirements of different time consumption scenarios are accurately adapted. When the execution time of a single model prediction is greater than or equal to the upper limit of the execution time reference, it indicates that the single instance resource is overloaded. At this time, the execution time offset is input into the positively correlated instance number mapping relationship model to obtain the increase in the number of shared instances. By increasing the number of shared instances loaded in the model to distribute the computational pressure, the time extension caused by single instance overload is effectively reduced. At the same time, based on the prediction accuracy offset, the single execution time, and the proportional coefficient of the current number of instances, it is ensured that the number of new instances is accurately matched with the execution time and accuracy requirements, avoiding inefficiencies due to insufficient number of instances. Bottlenecks or excessive instance counts can lead to resource contention. When the execution time of a single model prediction is less than or equal to the lower limit of the execution time reference, it indicates that the current number of instances exceeds the actual demand. The execution time correction is input into the instance number mapping relationship model to obtain the shared instance number reduction amount. By reducing the number of instances, idle instances are avoided from occupying system memory resources, thus improving the overall resource utilization. The entire adjustment process not only solves the pain points of low efficiency in high-time-consuming scenarios and resource waste in low-time-consuming scenarios under the traditional fixed instance number mode, but also ensures the stability and accuracy of the Stacking integrated model prediction process through dynamic adaptation of instance number, execution time, and accuracy, achieving a balance between resource utilization, prediction efficiency, and prediction accuracy.

[0046] As a further explanation, the specific process for adaptive adjustment of the output delay threshold is as follows: Obtain the actual model loading time of the Stacking integrated model and compare the model loading time with the reference value.

[0047] If the model loading latency is less than or equal to the reference loading latency value, then adaptive adjustment of the output latency threshold will not be performed, and the current model prediction output latency threshold will be maintained. Under optimal system conditions, this avoids introducing unnecessary additional waiting time. It ensures that in most scenarios, user requests receive the fastest possible response, thereby optimizing the user experience and reflecting the system's high-efficiency design goals.

[0048] If the model loading latency is greater than the loading latency reference value, the loading latency offset is input into the preset latency threshold mapping model to obtain the latency threshold adjustment factor. The current model prediction result output latency threshold and the latency threshold adjustment factor are multiplied and rounded up to obtain the target model prediction result output latency threshold. This is to avoid prediction result transmission interruption or service abnormality due to insufficient output latency threshold. Here, the loading latency offset represents the positive difference between the model loading latency and the loading latency reference value. The latency threshold mapping model is configured such that the input and output are positively correlated. The input represents the loading latency offset, and the output represents the target model prediction result output latency threshold. The positive correlation coefficient is pre-calibrated based on the prediction accuracy offset, the actual model loading latency, and the model prediction result output latency threshold to ensure that the target model prediction result output latency threshold is adapted to the model loading latency. By dynamically increasing the output latency threshold, a flexible buffer time is reserved for the model loading process, fundamentally avoiding interruptions in the transmission of prediction results and service connection failures caused by loading timeouts. This ensures that the prediction task can be executed completely and successfully returned to the front end. This strategy also reduces the probability of client anomalies triggered by timeouts and duplicate user requests, significantly improving the stability and professionalism of the system's external services, while ensuring the integrity and traceability of prediction data in scientific research scenarios.

[0049] In this embodiment, by dynamically adapting the model loading time and the prediction result output latency threshold, a dual guarantee of prediction service stability and user experience is achieved. By acquiring the actual loading time of the Stacking integrated model in real time and comparing it with a preset loading time reference value, when the model loading time is less than or equal to the reference value, the current model prediction result output latency threshold is maintained, avoiding system fluctuations caused by unnecessary threshold adjustments and ensuring the efficiency of prediction result transmission under normal loading scenarios. When the model loading time exceeds the reference value, the loading latency offset, representing the degree of excessive loading latency, is input into a positively correlated latency threshold mapping model to generate a latency threshold adjustment factor. This factor is then multiplied and rounded up to obtain the target output latency threshold. This design can accurately adapt to the increase in overall process time caused by model loading latency, effectively avoiding delays caused by output latency. This addresses issues such as interrupted prediction result transmission, data loss, or service anomalies caused by insufficient latency thresholds. It also leverages a pre-calibrated proportional coefficient based on prediction accuracy offset, actual model loading time, and the current output latency threshold. This ensures that the target threshold is neither insufficiently increased, posing a risk of service anomalies, nor excessively increased, leading to prolonged user wait times and impacting the user experience. Ultimately, it reliably guarantees complete prediction result output even in scenarios with model loading delays. This solves the pain point of traditional fixed output latency thresholds failing to adapt to fluctuations in model loading time, allowing the system to maintain reliable prediction result transmission capabilities under different loading states, providing users with a stable and consistent user experience.

[0050] It is necessary to understand that, such as Figure 2 The diagram shows the transaction sequence of the PPARG online molecular activity prediction system provided in this embodiment of the invention. The specific logic is as follows: The user inputs a single molecule sample or uploads a batch of Excel files on the front-end interface (C-end) and initiates a prediction request. The front-end encapsulates the request data into JSON or a file stream and sends it to the PPARG system's FastAPI server (P-end) via an HTTP request. After receiving the front-end request, the PPARG system (P-end) performs different processing based on the request type (single-molecule prediction or batch prediction): For single-molecule input, it generates a molecular feature vector using RDKit and calls a stacking model for prediction. For batch files, it parses the Excel data, generates a molecular feature matrix, and calls a stacking model for prediction item by item or batch prediction. The PPARG system (P-end) processes the prediction results, integrates the base model output through a meta-model LR, generates the final activity prediction value and confidence level, and encapsulates the results returned by the PPARG system to the terminal (C-end) into a JSON or Excel file and returns it to the front-end according to the front-end's requirements. Prediction results and single-molecule prediction results are displayed in real time on the interface, while batch prediction results can be viewed by downloading an Excel file. Users ultimately obtain molecular activity prediction results and corresponding confidence levels.

[0051] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims can be performed in a different order than that shown in the embodiments and still achieve the desired results. Additionally, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are also possible or may be advantageous.

[0052] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device, equipment, and storage medium embodiments are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0053] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0054] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. The PPARG online molecular activity prediction system, characterized in that, It includes a data input and preprocessing module, a prediction efficiency evaluation module, a prediction efficiency control module, a prediction accuracy evaluation module, and a prediction accuracy control module. The data input and preprocessing module is used to support manual input of single-molecule SMILES sequences and batch uploading of multi-molecule Excel files, and to perform preprocessing to obtain molecular feature vectors. The prediction efficiency evaluation module is used to load a pre-trained model based on the molecular feature vector to perform PPARG molecular activity prediction in order to obtain prediction efficiency parameters. Based on the prediction efficiency parameters, a molecular activity prediction efficiency evaluation value is obtained to reflect the high efficiency performance of PPARG online molecular activity prediction. The prediction efficiency control module determines whether to perform dynamic adjustment of activity prediction efficiency based on the molecular activity prediction efficiency evaluation value in order to achieve a balance between resource utilization efficiency and online molecular activity prediction response speed. If yes, the online molecular activity prediction result is output after adjustment; otherwise, the online molecular activity prediction result is output directly. The prediction accuracy evaluation module acquires prediction accuracy parameters in the process of outputting online molecular activity prediction results, and obtains molecular activity prediction accuracy evaluation values ​​based on the prediction accuracy parameters to reflect the accuracy of the model in predicting the online molecular activity of PPARG in drug development. The prediction accuracy control module determines whether to perform dynamic adjustment of the activity prediction accuracy based on the molecular activity prediction accuracy evaluation value to ensure the reliability of the molecular activity prediction accuracy. If yes, the prediction result visualization step is performed after adjustment; otherwise, the prediction result visualization step is performed directly to form a fully automated closed loop of input-prediction-output.

2. The PPARG online molecular activity prediction system as described in claim 1, characterized in that, The prediction efficiency parameters include the time consumed for single-molecule data input, the throughput of batch task processing, and the average time consumed for a single prediction call. By correcting the result of the single-molecule data input time normalization process using the input time influence factor, the input time influence component is obtained; The throughput impact component is obtained by correcting the throughput normalization result of the throughput impact factor for batch task processing. The average time-consuming influence component is obtained by correcting the result of normalizing the average time-consuming process of a single prediction call by the average time-consuming influence factor. The input time-consuming component, throughput-consuming component, and average time-consuming component are coupled to obtain a molecular activity prediction efficiency evaluation value, which is used to measure the overall processing efficiency of molecular data input model prediction and result output in the PPARG online molecular activity prediction process.

3. The PPARG online molecular activity prediction system as described in claim 1, characterized in that, The specific steps for determining whether to perform dynamic adjustment of activity prediction efficiency based on the molecular activity prediction efficiency evaluation value are as follows: The molecular activity prediction efficiency evaluation value is compared with the preset prediction efficiency reference value. If the molecular activity prediction efficiency evaluation value is greater than or equal to the prediction efficiency reference value, the dynamic adjustment of activity prediction efficiency is not performed. Otherwise, the dynamic adjustment of activity prediction efficiency is performed. The dynamic adjustment of activity prediction efficiency includes adaptive control of the threshold of the number of task segments and adaptive control of the interval between prediction task completion.

4. The PPARG online molecular activity prediction system as described in claim 3, characterized in that, The specific process for adaptively adjusting the threshold for the number of task fragments is as follows: Obtain the number of parallel threads for batch molecular preprocessing that are running in the PPARG online molecular activity prediction preprocessing service process, and compare the obtained number of parallel threads for batch molecular preprocessing with the reference value for the number of parallel threads. If the number of parallel threads for batch molecular preprocessing is less than or equal to the lower limit of the reference number of parallel threads, the parallel thread offset is input into a preset mapping model to obtain a task fragment number threshold reduction coefficient. The task fragment number threshold reduction coefficient is multiplied by the current batch prediction task fragment number threshold to obtain the target batch prediction task fragment number threshold, so as to avoid the waste of parallel processing resources caused by thread idleness. The task fragment mapping model is configured such that the input and output are positively correlated. The proportional coefficient of the positive correlation is pre-calibrated based on the prediction efficiency offset, the number of parallel threads for batch molecular preprocessing, and the batch prediction task fragment number threshold to ensure that the target task fragment number is adapted to the thread processing capacity. The input is the number of parallel threads for batch molecular preprocessing, the output is the target task fragment number threshold, the prediction efficiency offset represents the degree of negative deviation between the molecular activity prediction efficiency evaluation value and the preset prediction efficiency reference value, and the parallel thread offset represents the degree of negative deviation between the number of parallel threads for batch molecular preprocessing and the lower limit of the reference number of parallel threads.

5. The PPARG online molecular activity prediction system as described in claim 4, characterized in that, The adaptive adjustment of the task fragment number threshold also includes: If the number of parallel threads for batch molecular preprocessing is within the reference range of the number of parallel threads, then the adaptive adjustment of the task fragment number threshold will not be performed, and the current batch prediction task fragment number threshold will be maintained. The reference range of the number of parallel threads includes the open interval formed by the lower limit of the reference range of the number of parallel threads and the upper limit of the reference range of the number of parallel threads. If the number of parallel threads in batch molecular preprocessing is greater than or equal to the reference upper limit of the number of parallel threads, the parallel thread correction amount is input into the preset task partitioning mapping relationship model to obtain the task partitioning number threshold gain coefficient. The task partitioning number threshold gain coefficient is multiplied by the current batch prediction task partitioning number threshold to obtain the target batch prediction task partitioning number threshold, so as to improve the parallel processing efficiency of batch molecular preprocessing. The parallel thread correction amount represents the degree of positive deviation between the number of parallel threads in batch molecular preprocessing and the reference upper limit of the number of parallel threads.

6. The PPARG online molecular activity prediction system as described in claim 3, characterized in that, The specific process for adaptively adjusting the prediction task completion interval is as follows: Get the current number of concurrent predicted requests to be processed, and compare the number of concurrent predicted requests with the reference value of the number of requests; If the number of concurrent prediction requests is less than or equal to the reference value for the number of requests, the adaptive adjustment of the prediction task completion interval will not be performed, and the current batch prediction task completion interval will be maintained. If the number of concurrent prediction requests exceeds the reference value for the number of requests, the prediction request offset is input into a preset time interval mapping model to obtain a time interval adjustment factor. The current batch prediction task completion time interval is multiplied by the time interval adjustment factor and then rounded down to the nearest integer as the target batch prediction task completion time interval, thereby improving the throughput and response timeliness of the entire molecular activity prediction process. The time interval mapping model is configured such that the input and output quantities are negatively correlated. The proportional coefficient of the negative correlation is pre-calibrated based on the prediction efficiency offset, the number of concurrent prediction requests, and the batch prediction task completion time interval to ensure that the target batch prediction task completion time interval is adapted to the concurrent request processing capability. The prediction request offset represents the degree of positive deviation between the number of concurrent prediction requests and the reference value for the number of requests.

7. The PPARG online molecular activity prediction system as described in claim 1, characterized in that, The specific steps for obtaining the molecular activity prediction accuracy assessment value are as follows: The prediction accuracy parameters include the prediction efficiency evaluation value of the control group, the molecular activity prediction cycle, and the molecular activity prediction rate. The results of the analysis of the proportion of the predicted efficiency evaluation value and the predicted efficiency critical value of the control group were corrected by the predicted efficiency influencing factor to obtain the predicted efficiency influencing component. By correcting the result of normalizing the molecular activity prediction cycle by the prediction cycle influence factor, the prediction cycle influence component is obtained. The prediction rate influence component is obtained by correcting the result of normalizing the molecular activity prediction rate by the prediction rate influence factor. The prediction efficiency component, prediction period component, and prediction rate component are coupled to obtain a molecular activity prediction accuracy evaluation value, which is used to measure the consistency between the prediction results output by PPARG online molecular activity prediction and the experimental results. The specific process for determining whether to perform dynamic adjustment of activity prediction accuracy is as follows: The molecular activity prediction accuracy assessment value is compared with the preset prediction accuracy reference value. If the molecular activity prediction accuracy assessment value is greater than or equal to the prediction accuracy reference value, the dynamic adjustment of activity prediction accuracy is not performed. Otherwise, the dynamic adjustment of activity prediction accuracy is performed. The dynamic adjustment of activity prediction accuracy includes adaptive control of the number of shared instances and adaptive control of the output delay threshold.

8. The PPARG online molecular activity prediction system as described in claim 7, characterized in that, The specific steps for adaptive adjustment of the number of shared instances are as follows: Obtain the single-time model prediction execution time of the current Stacking integrated model and compare it with the execution time reference interval, which represents the closed interval formed by the lower limit of the execution time reference and the upper limit of the execution time reference; If the execution time of a single model prediction is greater than or equal to the upper limit of the execution time reference, the execution time offset is input into the preset instance number mapping relationship model to obtain the shared instance number adjustment. The current model loading shared instance number is coupled with the shared instance number adjustment to obtain the target model loading shared instance number to reduce the time extension caused by single instance resource overload. The instance number mapping relationship model is configured such that the input and output are positively correlated. The input represents the execution time of a single model prediction, and the output represents the target model loading shared instance number. The positive correlation ratio coefficient of the instance number mapping relationship model is pre-calibrated based on the prediction accuracy offset, the execution time of a single model prediction, and the model loading shared instance number to ensure that the target model loading shared instance number is adapted to the execution time of a single model prediction. The prediction accuracy offset represents the degree of negative deviation between the molecular activity prediction accuracy evaluation value and the preset prediction accuracy reference value. The execution time offset represents the degree of positive deviation between the execution time of a single model prediction and the upper limit of the execution time reference.

9. The PPARG online molecular activity prediction system as described in claim 8, characterized in that, The adaptive adjustment of the number of shared instances also includes: If the execution time of a single model prediction is less than or equal to the lower limit of the execution time reference, the execution time correction amount is input into the preset instance number mapping relationship model to obtain the shared instance number reduction amount. The difference between the current model loading shared instance number and the shared instance number reduction amount is processed to obtain the target model loading shared instance number. This reduces the time extension caused by the overload of single instance resources and reduces the number of instances to avoid idle instances occupying system memory resources. The execution time correction amount represents the degree of negative deviation between the execution time of a single model prediction and the lower limit of the execution time reference.

10. The PPARG online molecular activity prediction system as described in claim 7, characterized in that, The specific process for adaptive adjustment of the output delay threshold is as follows: Obtain the actual model loading time of the Stacking integrated model and compare the model loading time with the loading time reference value; If the model loading time is less than or equal to the loading time reference value, then the adaptive adjustment of the output time threshold will not be performed, and the current model prediction result output time threshold will be maintained. If the model loading latency is greater than the loading latency reference value, the loading latency offset is input to the preset latency threshold mapping model to obtain the latency threshold adjustment factor. The current model prediction result output latency threshold and the latency threshold adjustment factor are multiplied and rounded up to obtain the target model prediction result output latency threshold, so as to avoid prediction result transmission interruption or service abnormality due to insufficient output latency threshold. Here, the loading latency offset represents the degree of positive deviation between the model loading latency and the loading latency reference value. The latency threshold mapping model is configured such that the input and output are positively correlated. The input represents the loading latency offset, and the output represents the target model prediction result output latency threshold. The positive correlation coefficient is pre-calibrated based on the prediction accuracy offset, the actual model loading latency, and the model prediction result output latency threshold to ensure that the target model prediction result output latency threshold is adapted to the model loading latency.

Images