Biological medicine active component screening method and system based on artificial intelligence

Abstract

The invention discloses a biological medicine active component screening method and system based on artificial intelligence, and relates to the technical field of biological medicine manufacturing. The method comprises the following steps: acquiring component characteristic parameters and performance parameters of a biological medicine formula to be optimized; constructing and acquiring a performance prediction model, and acquiring a performance prediction confidence coefficient based on the component characteristic parameters and the performance parameters; on the basis of the component features, obtaining an optimization feasible region in combination with the performance prediction confidence; performing iterative optimization in the optimization feasible region by adopting a particle swarm algorithm to obtain an optimized biological medicine formula, and obtaining optimized performance parameters of the optimized biological medicine formula based on the performance prediction model; and according to the optimized biological medicine formula and the optimized performance parameters, generating a biological medicine active component screening result. The screening efficiency and precision of the active ingredients of the biological medicine are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of biopharmaceutical manufacturing technology, specifically to a method and system for screening biopharmaceutical active ingredients based on artificial intelligence. Background Technology

[0002] With the rapid development of artificial intelligence technology, methods such as deep learning and machine learning have been gradually applied to the biopharmaceutical field. Existing technologies typically construct a single performance prediction model, directly linking component characteristics with efficacy parameters, and using optimization algorithms to search for potential high-performance formulation combinations.

[0003] However, existing technologies still have significant limitations: on the one hand, single prediction models often lack quantitative assessment of the reliability of their results, leading to blind optimization in unverified high-risk areas, which may result in ineffective or even unstable formulations; on the other hand, conventional optimization algorithms are prone to getting stuck in local optima or having low computational efficiency when performing global searches in a vast component space, resulting in insufficient targeting and reliability of the screening process, ultimately leading to low efficiency and accuracy in screening biopharmaceutical active ingredients. Summary of the Invention

[0004] This invention provides a method and system for screening biopharmaceutical active ingredients based on artificial intelligence, aiming to solve the technical problems of low efficiency and accuracy in the screening of biopharmaceutical active ingredients in the prior art.

[0005] In view of the above problems, the present invention provides a method and system for screening biopharmaceutical active ingredients based on artificial intelligence.

[0006] In a first aspect, the present invention provides a method for screening biopharmaceutical active ingredients based on artificial intelligence, comprising: Obtain the component characteristic parameters and performance parameters of the biopharmaceutical formulation to be optimized; Construct a performance prediction model and obtain the performance prediction confidence level based on the component feature parameters and the performance parameters; Based on the component characteristics and the performance prediction confidence, the optimization feasible region is obtained; Within the optimized feasible region, the particle swarm optimization algorithm is used for iterative optimization to obtain the optimized biopharmaceutical formulation, and the optimized performance parameters of the optimized biopharmaceutical formulation are obtained based on the performance prediction model. Based on the optimized biopharmaceutical formulation and the optimized performance parameters, the screening results of biopharmaceutical active ingredients are generated.

[0007] Secondly, the present invention provides an artificial intelligence-based biopharmaceutical active ingredient screening system, comprising: The parameter acquisition module is used to acquire the component characteristic parameters and performance parameters of the biopharmaceutical formulation to be optimized. The confidence assessment module is used to construct a performance prediction model and obtain the performance prediction confidence based on the component feature parameters and the performance parameters. The feasible region determination module is used to obtain an optimized feasible region based on the component features and the performance prediction confidence. The iterative optimization module is used to perform iterative optimization using a particle swarm optimization algorithm within the optimization feasible region to obtain an optimized biopharmaceutical formulation, and to obtain optimized performance parameters of the optimized biopharmaceutical formulation based on the performance prediction model. The results generation module is used to generate screening results for biopharmaceutical active ingredients based on the optimized biopharmaceutical formulation and the optimized performance parameters.

[0008] One or more technical solutions provided in this invention have at least the following technical effects or advantages: This invention provides an artificial intelligence-based method and system for screening biopharmaceutical active ingredients. By constructing a performance prediction model with confidence assessment, the reliability of performance prediction is improved, providing a precise basis for optimization direction. By combining confidence with dynamic adjustment of the feasible region, screening efficiency and formulation coverage are balanced, avoiding the omission of high-quality solutions. By optimizing the iteration rules and stagnation expansion strategy of the particle swarm optimization algorithm, local optima are effectively broken, ultimately efficiently obtaining optimized formulations with both high activity and high stability, effectively improving the efficiency and accuracy of biopharmaceutical active ingredient screening. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 A schematic flowchart of an artificial intelligence-based biopharmaceutical active ingredient screening method provided in an embodiment of the present invention; Figure 2 A schematic diagram of the structure of an artificial intelligence-based biopharmaceutical active ingredient screening system provided in an embodiment of the present invention; The components represented by each number in the attached diagram are explained below: The module includes a parameter acquisition module 11, a confidence assessment module 12, a feasible region determination module 13, an iterative optimization module 14, and a result generation module 15. Detailed Implementation

[0011] This invention provides an artificial intelligence-based method and system for screening biopharmaceutical active ingredients, which addresses the technical problems of low efficiency and accuracy in screening biopharmaceutical active ingredients in existing technologies.

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0013] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0014] Example 1, as Figure 1 As shown, this invention provides a method for screening biopharmaceutical active ingredients based on artificial intelligence, the method comprising: S100: Obtain the component characteristic parameters and performance parameters of the biopharmaceutical formulation to be optimized.

[0015] In this embodiment of the invention, the component characteristic parameters and performance parameters of the biopharmaceutical formulation to be optimized are obtained. During the development of biopharmaceutical formulations, formulation data and performance test results are often scattered across various media, generally exhibiting fragmentation and inconsistent formats. Historical experimental records are mostly handwritten notes or non-standardized electronic documents, data in scientific literature is often primarily textual descriptions, standardized databases require targeted screening of suitable information, and high-throughput experimental platforms generate massive amounts of unorganized raw test data. Directly using unstructured data can lead to biases in the correlation analysis between components and performance, thus affecting the accuracy of formulation optimization. Therefore, it is necessary to collect and transform data using standardized methods to build a solid foundation of data support for biopharmaceutical formulation optimization.

[0016] Step S100 in the method provided in this embodiment of the invention includes: Obtain the component characteristic parameters of the biopharmaceutical formulation to be optimized, wherein the component characteristic parameters include component characteristics and component ratio parameters; Obtain the performance parameters of the biopharmaceutical formulation to be optimized, wherein the performance parameters include activity parameters and stability parameters.

[0017] First, the component characteristic parameters of the biopharmaceutical formulation to be optimized are obtained. These parameters include component characteristics and component ratio parameters. Component characteristics refer to the chemical or biological names of the active ingredients and key excipients in the formulation, used to uniquely identify the substance. Examples include sodium hyaluronate, polydeoxyribonucleotide (PDRN), glycine, and vitamin B12. Component ratio parameters refer to the relative content or concentration of each component in the formulation, usually expressed as a mass percentage, molar concentration, or volume ratio, defining the compositional structure of the formulation. Raw material composition information of the target biopharmaceutical formulation is collected selectively from historical experimental records, biopharmaceutical research literature, biopharmaceutical standardization databases, or high-throughput experimental platforms. The core identification information of the active ingredients and key excipients in the formulation is extracted, and the corresponding content or concentration data of each component is collected simultaneously. For example, the biopharmaceutical formulation to be optimized is an injectable gel, and the basic formulation F1 is collected from laboratory records. The component characteristics are recorded as: [sodium hyaluronate, polydeoxyribonucleotide (PDRN), glycine, PBS buffer]. The composition ratio parameters (mass percentage) are recorded as follows: [Sodium hyaluronate: 4%, Polydeoxyribonucleotide (PDRN): 2%, Glycine: 0.5%, PBS buffer: 93.5%].

[0018] Secondly, the performance parameters of the biopharmaceutical formulation to be optimized are obtained, including activity parameters and stability parameters. Performance parameters refer to the set of core indicators that measure the actual application effect and usage period of a biopharmaceutical formulation, and are a key basis for evaluating the quality of the formulation. They mainly include activity parameters and stability parameters. Activity parameters are quantitative indicators reflecting the ability of a biopharmaceutical to exert its core efficacy, and are directly related to the therapeutic or application effect of the drug. They are commonly found in pharmacodynamic test results, such as the collagen production rate in biologics. Stability parameters are indicators reflecting the ability of a biopharmaceutical to maintain its physicochemical properties and efficacy during storage, transportation, and use. They are mostly derived from physical and long-term stability test data, such as the in vitro degradation time of injectable gels. For example, for the basic formulation F1 of an injectable gel, the performance parameters of this formulation were collected from corresponding experimental test reports and scientific literature in the field of biopharmaceuticals. The activity parameters are recorded as: [Collagen production rate: 32%, Skin moisturizing improvement rate: 45%]. Stability parameters were recorded as follows: [In vitro degradation time: 80 h, shelf life under refrigeration (2-8 ℃): 18 months, osmolarity: 300 mOsmol / kg, pH value: 6.5].

[0019] In this embodiment of the invention, by collecting information on the formulation raw materials and performance test data of biopharmaceuticals, the invention comprehensively covers component characteristic parameters, component ratio parameters, and performance parameters in both activity and stability dimensions. This provides a complete and accurate data source for subsequent performance prediction model training, avoiding directional deviations caused by data gaps during subsequent optimization. At the same time, the standardized data format provides a unified evaluation benchmark for comparing component adjustments and performance changes between different optimization schemes, effectively reducing the trial-and-error cost of biopharmaceutical formulation optimization, improving R&D efficiency, and laying a solid data foundation for subsequent performance prediction model construction, optimization feasible domain determination, and particle swarm optimization iterative optimization.

[0020] S200: Construct a performance prediction model and obtain the performance prediction confidence level based on the component feature parameters and the performance parameters.

[0021] In this embodiment of the invention, a performance prediction model is constructed, and the performance prediction confidence level is obtained based on the component characteristic parameters and the performance parameters. There is a complex nonlinear relationship between the components and performance of biopharmaceutical formulations. Constructing only a single prediction model cannot accurately adapt to performance indicators of two different attributes: activity and stability. Furthermore, without quantifying the confidence level of the prediction results, subsequent adjustments to the feasible region for optimization will be blind, thus affecting the overall optimization efficiency and accuracy. Therefore, it is necessary to construct a dedicated performance prediction model and establish a corresponding confidence level evaluation system to provide accurate and reliable model support for formulation optimization.

[0022] Step S200 in the method provided in this embodiment of the invention includes: The construction of the performance prediction model includes: Obtain a set of sample biopharmaceutical formulations that have the same main components as the biopharmaceutical formulation to be optimized, and obtain the sample characteristic parameter set, sample activity parameter set, and sample stability parameter set of the sample biopharmaceutical formulation set; Construct a performance prediction model, wherein the performance prediction model includes an activity prediction branch and a stability prediction branch; Using the sample feature parameter set as input, and the sample activity parameter set and the sample stability parameter set as supervision, the activity prediction branch and the stability prediction branch are trained respectively until the model converges.

[0023] First, a sample set of biopharmaceutical formulations with the same main components as the formulation to be optimized is obtained, along with its sample feature parameter set, sample activity parameter set, and sample stability parameter set. The sample set of biopharmaceutical formulations refers to a collection of historical formulations with the same main components as the formulation to be optimized, which have been experimentally verified; it serves as the foundational data source for model training. The sample feature parameter set is the collection of component characteristics and component ratio parameters corresponding to all samples in the sample set of biopharmaceutical formulations, serving as input data for model training. The sample activity parameter set is the collection of activity quantification indicators corresponding to all samples in the sample set of biopharmaceutical formulations, used to supervise the training of the model's activity prediction branch. The sample stability parameter set is the collection of stability quantification indicators corresponding to all samples in the sample set of biopharmaceutical formulations, used to supervise the training of the model's stability prediction branch. Formulations with the same main components as the formulation to be optimized are selected from historical experimental records, standardized databases, and scientific literature to form the sample set of biopharmaceutical formulations. The component characteristics and component ratio parameters of each sample are extracted from this set and integrated to form the sample feature parameter set; simultaneously, the experimentally verified activity parameters and stability parameters of each sample are extracted and organized into the sample activity parameter set and sample stability parameter set, respectively.

[0024] For example, 100 historical formulations with sodium hyaluronate and polydeoxyribonucleic acid (PDRN) as the main components were screened from a biopharmaceutical experimental database to form a sample biopharmaceutical formulation set. The sample characteristic parameter set consists of the component characteristics of each sample [sodium hyaluronate, polydeoxyribonucleic acid (PDRN), active ingredient, PBS buffer] and the corresponding component ratio parameters [mass percentage data such as sodium hyaluronate concentration 2%-6%, PDRN concentration 1%-3%]. Each sample has 4 component characteristics and 4 component ratio parameters, for a total of 8 characteristic dimensions. The sample activity parameter set consists of [collagen production rate 25%-40%, skin moisturizing improvement rate 35%-50%] and the sample stability parameter set consists of [in vitro degradation time 50h-120h, shelf life 12 months-24 months under refrigeration (2-8℃)].

[0025] Secondly, a performance prediction model is constructed, comprising an activity prediction branch and a stability prediction branch. The performance prediction model is an algorithm specifically designed to predict the performance of biopharmaceutical formulations, containing independent yet collaborative activity and stability prediction branches. The activity prediction branch is a submodule within the performance prediction model focused on outputting the predicted activity parameters of the biopharmaceutical formulation. The stability prediction branch is a submodule within the performance prediction model focused on outputting the predicted stability parameters of the biopharmaceutical formulation. A performance prediction model with two independent branches is constructed. The activity prediction branch predicts the efficacy of the biopharmaceutical, while the stability prediction branch predicts the retention of the biopharmaceutical's physicochemical properties. The two branches share input layer data but each has its own independent computational and output layers.

[0026] For example, a performance prediction model is constructed, consisting of two independent random forest sub-models: an activity prediction branch and a stability prediction branch, sharing an 8-dimensional sample feature parameter input data source. The activity prediction branch uses a random forest regression model, setting the number of decision trees to 100, the maximum depth of each decision tree to 8, the maximum number of features to 4, and the minimum number of sample splits to 5. The model output layer contains two prediction results, corresponding to collagen generation rate and skin hydration improvement rate, respectively. The stability prediction branch also uses a random forest regression model, with hyperparameters such as the number of decision trees, maximum depth, maximum number of features, and minimum number of sample splits consistent with the activity prediction branch. The output layer contains two prediction results, corresponding to in vitro degradation time and refrigerated shelf life, respectively. The two branches independently store training parameters, without interfering with each other, sharing only the input data preprocessing module. The activity prediction branch outputs predicted values ​​for collagen generation rate and skin hydration improvement rate, while the stability prediction branch outputs predicted values ​​for in vitro degradation time and refrigerated shelf life. Both branches share the sample feature parameter input layer.

[0027] Furthermore, using the sample feature parameter set as input, and the sample activity parameter set and the sample stability parameter set as supervision, the activity prediction branch and the stability prediction branch are trained separately until the model converges. Using the sample feature parameter set as the unified input data for the model, the sample activity parameter set as the supervision signal for the activity prediction branch, and the sample stability parameter set as the supervision signal for the stability prediction branch, the gradient descent algorithm is used to train the two branches simultaneously and independently. By iteratively adjusting the weight parameters of each branch, the error between the predicted value and the actual value is continuously calculated. When the error decreases below a preset threshold and there is no significant fluctuation in error over multiple iterations, the model is considered to have converged, and training is stopped.

[0028] For example, using the sample feature parameter set as input and the sample activity parameter set and sample stability parameter set as supervision signals, two random forest branches are trained separately. Hyperparameters are adjusted through cross-validation. When the prediction error of the activity prediction branch drops to 4.8% and the prediction error of the stability prediction branch drops to 5.2%, and after readjusting the number of decision trees, the error fluctuation of both branches is less than 0.3%, the model is considered to have converged, and training is complete.

[0029] The process of obtaining performance prediction confidence based on the component feature parameters and the performance parameters includes: The component characteristic parameters are input into the performance prediction model to obtain the predicted activity parameters and predicted stability parameters, which are used as the predicted performance parameters. Based on the predicted performance parameters and the performance parameters, a performance prediction confidence level is obtained, wherein the performance prediction confidence level is obtained by weighted calculation based on the activity prediction confidence level and the stability prediction confidence level.

[0030] First, the component characteristic parameters are input into the performance prediction model to obtain predicted activity parameters and predicted stability parameters, which serve as predicted performance parameters. Predicted performance parameters refer to the predicted activity parameters and predicted stability parameters output after the component characteristic parameters of the formulation to be optimized are input into the trained and converged performance prediction model. The complete component characteristic parameters of the biopharmaceutical formulation to be optimized, including all component characteristics and corresponding component ratio parameters, are input into the trained and converged performance prediction model. The model outputs predicted activity parameters through its activity prediction branch and predicted stability parameters through its stability prediction branch; the combination of these two parameters forms the predicted performance parameters.

[0031] For example, the feature data of formulation F1, consisting of eight dimensions including component characteristic parameters [sodium hyaluronate, polydeoxyribonucleotide (PDRN), glycine, PBS buffer] and component ratio parameters [sodium hyaluronate: 4%, polydeoxyribonucleotide (PDRN): 2%, glycine: 0.5%, PBS buffer: 93.5%], are input into a convergent bi-branch random forest performance prediction model. The model outputs predicted activity parameters [predicted collagen production rate: 31%, predicted skin hydration improvement rate: 43%] and predicted stability parameters [predicted in vitro degradation time: 78h, predicted shelf life under refrigeration (2-8℃): 17 months]. These two parameters together constitute the predicted performance parameters.

[0032] Secondly, based on the predicted performance parameters and the actual performance parameters, a performance prediction confidence score is obtained. This score is calculated by weighting the activity prediction confidence score and the stability prediction confidence score. The performance prediction confidence score is a core indicator for quantifying the degree of matching between predicted and actual performance parameters, reflecting the reliability of the model's prediction results. Its value ranges from [0,1], with a higher reliability indicating a value closer to 1. The activity prediction confidence score is calculated based on the deviation between the predicted and actual activity parameters, specifically evaluating the reliability of the activity prediction branch's output. The stability prediction confidence score is calculated based on the deviation between the predicted and actual stability parameters, specifically evaluating the reliability of the stability prediction branch's output. The deviation rates between the predicted and actual activity parameters and the predicted and actual stability parameters are calculated separately, and the activity prediction confidence score and stability prediction confidence score are obtained by reverse conversion based on these deviation rates. According to the current application scenarios and actual needs of biopharmaceutical formulations, weights are set for the two types of confidence scores (the sum of the weights is 1), and the final performance prediction confidence score is calculated using a weighted summation formula.

[0033] For example, based on the actual performance parameters of formulation F1, the calculated confidence level for activity prediction is 0.92, and the confidence level for stability prediction is 0.90. Since this formulation is used in skin repair applications, where activity is prioritized over stability, the weighting for activity prediction confidence level is set to 0.6, and the weighting for stability prediction confidence level is set to 0.4. The weighted calculation is: Performance Prediction Confidence Level = 0.92 × 0.6 + 0.90 × 0.4 = 0.912.

[0034] In this embodiment of the invention, a performance prediction model with dual branches of activity and stability is constructed and trained. This clarifies the model's hyperparameter settings and structural division, accurately capturing the complex nonlinear relationships between component characteristic parameters and different types of performance parameters, thus addressing the pain points of insufficient prediction accuracy and structural ambiguity inherent in single models. Furthermore, by subdividing the acquisition of prediction parameters and weighting the confidence scores, not only are accurate prediction performance parameters obtained, but the reliability of the prediction results is also quantified based on formulation requirements, avoiding the decision-making risks of blindly relying on model predictions. The trained performance prediction model possesses advantages such as high training efficiency and strong anti-interference capabilities, enabling rapid output of formulation performance prediction results. The performance prediction confidence score provides a core quantitative basis for the subsequent dynamic adjustment of the feasible region, enhancing the scientific rigor and specificity of biopharmaceutical formulation optimization, and laying a solid model foundation and evaluation standard for subsequent efficient iterative optimization using particle swarm optimization algorithms.

[0035] S300: Based on the component characteristics and the performance prediction confidence, obtain the optimization feasible region.

[0036] In this embodiment of the invention, an optimization feasible region is obtained based on the component characteristics and the performance prediction confidence level. The optimization feasible region is the boundary of the iterative optimization range of a biopharmaceutical formulation, and its rationality directly determines the efficiency and accuracy of the formulation optimization. Boundaries determined solely based on component characteristics, without considering the reliability of performance prediction results, suffer from a fixed limitation: if the boundary is too wide, the optimization range will be too large, increasing the iteration cost of the particle swarm optimization algorithm and even leading to invalid searches; if the boundary is too narrow, potential high-quality formulations may be missed. Performance prediction confidence level reflects the reliability of the model's prediction results and is highly correlated with the adaptability of the formulation optimization range. Therefore, it is necessary to first determine the initial boundary based on component characteristics, then dynamically adjust it based on the performance prediction confidence level, and use normalization to satisfy the total constraints of the formulation, thus constructing a scientifically reasonable optimization feasible region to provide accurate range support for subsequent efficient optimization.

[0037] Step S300 in the method provided in this embodiment of the invention includes: Based on the component characteristics, the initial independent feasible boundaries of each component are obtained; The initial independent feasible boundary is adjusted based on the performance prediction confidence level. Obtain the adjusted feasible boundaries of each component to form the optimized feasible region.

[0038] First, based on the aforementioned component characteristics, initial independent feasible boundaries are obtained for each component. These initial independent feasible boundaries refer to the content ranges individually set for each component in the formulation, based solely on its characteristics and general standards, safety regulations, and process limitations within the biopharmaceutical industry. They serve as the foundational benchmark for subsequent boundary adjustments. The physicochemical properties, biosafety data, and core functions of each component in the biopharmaceutical formulation to be optimized are analyzed. Combined with raw material usage standards, production process limitations, and clinical application safety thresholds within the biopharmaceutical industry, independent content ranges are defined for each component. The upper limit of the range must satisfy the component's solubility and stability limitations, while the lower limit must ensure the component can exert its basic efficacy, ultimately forming the set of initial independent feasible boundaries for each component.

[0039] For example, continuing with the basic formulation F1 of the injectable gel, its components are characterized as [sodium hyaluronate, polydeoxyribonucleic acid (PDRN), glycine, and PBS buffer]. Combining biopharmaceutical industry standards and the characteristics of each component, the initial independent feasible boundaries of each component are determined as follows: Sodium hyaluronate: 2%-6%, the lower limit ensures the basic filling and shaping effect, and the upper limit adapts to its solubility limitation in PBS buffer; Polydeoxyribonucleic acid (PDRN): 1%-3%, the lower limit meets the basic concentration requirements for cell regeneration, and the upper limit meets the biosafety threshold; Glycine: 0.2%-1.0%, the lower limit achieves the basic moisturizing and antioxidant effects, and the upper limit avoids excessive concentration leading to abnormal gel viscosity; PBS buffer: 89%-96.8%, derived by working backward from the boundary ranges of other components to ensure the solvent requirements of the formulation system.

[0040] Secondly, the initial independent feasible boundary is adjusted based on the performance prediction confidence level.

[0041] The adjustment of the initial independent feasible boundary based on the performance prediction confidence includes: Obtain a first confidence threshold and a second confidence threshold, wherein the second confidence threshold is greater than the first confidence threshold; If the performance prediction confidence is less than or equal to the first confidence threshold, then each initial independent feasible boundary is expanded outward by a first proportion to amplify the optimized feasible region. If the performance prediction confidence is greater than or equal to the second confidence threshold, then each initial independent feasible boundary is shrunk inward by a second proportion to narrow the optimized feasible region. If the performance prediction confidence level is between the first confidence threshold and the second confidence threshold, then a third ratio is determined based on linear interpolation, and each initial independent feasible boundary is linearly scaled according to this third ratio. Obtain the adjusted boundaries of each component, and normalize the upper bounds of all components so that their sum satisfies the sum constraint, thus forming the optimized feasible region.

[0042] First, a first confidence threshold and a second confidence threshold are obtained, wherein the second confidence threshold is greater than the first confidence threshold. The first confidence threshold is the low threshold critical point for performance prediction confidence; below this value, the prediction result has low reliability, and the optimization range needs to be expanded. The second confidence threshold is the high threshold critical point for performance prediction confidence; above this value, the prediction result has high reliability, and the optimization range can be narrowed to improve efficiency. Based on industry experience in biopharmaceutical formulation optimization and model training accuracy, the first and second confidence thresholds are preset, with the second confidence threshold being greater than the first confidence threshold. For example, based on model prediction experience for this type of injectable gel, the first confidence threshold is preset to 0.8, and the second confidence threshold is preset to 0.95.

[0043] Secondly, if the performance prediction confidence level is less than or equal to the first confidence threshold, then each initial independent feasible boundary is expanded outward by a first proportion to amplify the optimized feasible domain. The first proportion refers to the fixed proportion by which the boundary is expanded outward when the confidence level is less than or equal to the first threshold. If the performance prediction confidence level of the formulation to be optimized is less than or equal to the first confidence threshold, it indicates that the prediction result has low reliability. The initial independent feasible boundaries of each component are expanded outward by the first proportion to broaden the search range and cover potential high-quality formulations. For example, if the first proportion (expansion proportion) is set to 20%, and the performance prediction confidence level of formulation F1 is 0.75 ≤ 0.8, then the initial boundaries of each component are expanded outward by 20%. For example, the initial boundary of sodium hyaluronate (2%-6%) is expanded to 1.6%-7.2%, and the initial boundary of PDRN (1%-3%) is expanded to 0.8%-3.6%.

[0044] Then, if the performance prediction confidence level is greater than or equal to the second confidence threshold, each initial independent feasible boundary is shrunk inward by a second proportion to narrow the optimization feasible region. The second proportion refers to the fixed proportion by which the boundary is shrunk inward when the confidence level is higher than or equal to the second threshold. If the performance prediction confidence level is greater than or equal to the second confidence threshold, it indicates that the prediction result is highly reliable. The initial independent feasible boundaries of each component are shrunk inward by the second proportion to narrow the search range and improve optimization efficiency. For example, if the second proportion (shrinkage proportion) is set to -10%, and the performance prediction confidence level of formulation F1 is 0.96 ≥ 0.95, then the initial boundaries of each component are shrunk inward by 10%. For example, the initial boundary of sodium hyaluronate (2%-6%) is shrunk to 2.2%-5.4%, and the initial boundary of PDRN (1%-3%) is shrunk to 1.1%-2.7%.

[0045] Furthermore, if the performance prediction confidence level is between the first confidence threshold and the second confidence threshold, a third ratio is determined based on linear interpolation, and each initial independent feasible boundary is linearly scaled according to this third ratio. Linear interpolation refers to a mathematical method that calculates the boundary scaling ratio corresponding to the current confidence level using the ratio values ​​corresponding to the two thresholds when the performance prediction confidence level is between two thresholds. The third ratio refers to the dynamic scaling ratio obtained through linear interpolation, suitable for scenarios where the confidence level is between the two thresholds. If the performance prediction confidence level is between two thresholds, the third ratio is calculated using the confidence level as a variable and the preset benchmark ratios corresponding to the first and second thresholds as endpoints, and the initial independent feasible boundaries are dynamically scaled according to this ratio.

[0046] For example, the confidence level of formulation F1 is 0.912, which is between 0.8 and 0.95. A first threshold corresponds to a baseline proportion of 20%, a second threshold corresponds to a baseline proportion of -10%, and a third proportion is calculated to be 2.8% through linear interpolation. The initial boundaries are adjusted according to this proportion. For example, the calculated initial boundaries for each component after adjustment are: sodium hyaluronate 1.94%-6.17%, PDRN 0.97%-3.09%, glycine 0.19%-1.03%, and PBS buffer 88.71%-96.90%.

[0047] Based on this, the adjusted boundaries of each component are obtained, and the upper bounds of all component boundaries are normalized to ensure that their sum satisfies the total constraint, thus forming the optimized feasible region. Normalization refers to the process of adjusting the upper bounds of each component boundary proportionally to ensure that their sum satisfies the preset constraint, in cases where the sum of the adjusted upper bounds of each component boundary may deviate from the total formulation constraint. The total constraint is a fixed condition that the sum of the content percentages of all components in a biopharmaceutical formulation must satisfy; for mass percentage formulations, the total constraint is usually 100%. The optimized feasible region refers to the set of value ranges for the content of each component after initial setting, dynamic adjustment, and normalization; it is the search interval for subsequent formulation optimization. The adjusted feasible boundaries of each component are summarized, and the sum of the upper bounds of all component boundaries is calculated. If the sum does not meet the preset total constraint, the upper bounds of each component boundary are normalized, the total constraint is allocated according to the original upper bound percentage of each component, the lower bound of each component remains unchanged, and only the upper bound is adjusted until the sum satisfies the constraint. The feasible boundaries of all components after normalization are integrated to form the final optimized feasible region.

[0048] For example, for the basic formulation F1 of the injectable gel, the following operations are performed: The adjusted upper boundaries of each component are summarized: Sodium hyaluronate 6.17%, PDRN 3.09%, glycine 1.03%, PBS buffer 96.90%, with a total upper boundary of 107.19%, exceeding the 100% mass percentage constraint. Normalization is then performed: Using 100% as the total constraint, the upper boundaries are adjusted according to the original percentage of each component. The adjustment formula is: New upper boundary = Original upper boundary × (Total constraint / Original upper boundary sum). The calculations yield: New upper boundary for sodium hyaluronate = 6.17% × (100% / 107.19%), new upper boundary for PDRN = 3.09% × (100% / 107.19%), new upper boundary for glycine = 1.03% × (100% / 107.19%), and new upper boundary for PBS buffer = 96.90% × (100% / 107.19%).

[0049] Finally, the adjusted feasible boundaries of each component are obtained to form the optimized feasible domain. For example, by integrating the normalized boundaries, the final optimized feasible domain is: sodium hyaluronate 1.94%-5.76%, polydeoxyribonucleotide (PDRN) 0.97%-2.88%, glycine 0.19%-0.96%, and PBS buffer 88.71%-90.40%.

[0050] In this embodiment of the invention, initial independent feasible boundaries for each component are first determined based on component characteristics, establishing a basic benchmark for the optimization range. Then, combined with performance prediction confidence, threshold judgments are used to expand, shrink, or linearly scale the boundaries, adapting to different prediction reliability scenarios and solving the problem of poor adaptability of traditional fixed boundaries. Simultaneously, normalization ensures that the upper bounds of each component boundary satisfy the summation constraint, guaranteeing the physical rationality of the feasible region. The final constructed optimized feasible region avoids missing high-quality formulations due to overly narrow boundaries when prediction reliability is low, and also solves the pain point of low optimization efficiency caused by overly wide boundaries when prediction reliability is high. It achieves dynamic matching between the optimization range and prediction accuracy, providing scientific range support for the efficient and accurate iterative optimization of subsequent particle swarm optimization algorithms, and improving the targeting and practicality of biopharmaceutical formulation optimization.

[0051] S400: Within the optimized feasible region, the particle swarm optimization algorithm is used for iterative optimization to obtain the optimized biopharmaceutical formulation, and based on the performance prediction model, the optimized performance parameters of the optimized biopharmaceutical formulation are obtained.

[0052] In this embodiment of the invention, within the optimized feasible region, a particle swarm optimization algorithm is used for iterative optimization to obtain an optimized biopharmaceutical formulation. Based on the performance prediction model, the optimized performance parameters of the biopharmaceutical formulation are obtained. Biopharmaceutical formulation optimization is essentially a collaborative optimization of multi-dimensional parameters. Traditional trial-and-error methods or single optimization algorithms struggle to quickly find the optimal solution that balances activity and stability within a complex feasible region. The optimized feasible region clearly defines the reasonable value ranges for each component, while the particle swarm optimization algorithm possesses swarm intelligence search characteristics, enabling efficient traversal of potential formulations within the feasible region through information sharing and collaboration among particles. Simultaneously, by combining fitness parameters with activity and stability weights, the actual application requirements of the formulation can be accurately matched, preventing the optimization direction from deviating from the core objective. Without a systematic iterative strategy, it is easy to get trapped in local optima, resulting in the optimized formulation failing to achieve optimal overall performance. Therefore, using a particle swarm optimization algorithm within the optimized feasible region, through the process of generating an initial particle swarm, predicting performance, calculating fitness, and iteratively optimizing, can efficiently obtain biopharmaceutical formulations with superior overall performance.

[0053] Step S400 in the method provided in this embodiment of the invention includes: Multiple particles are randomly generated within the optimized feasible region to obtain an initial particle swarm, wherein each initial particle represents an initial biopharmaceutical formulation. The initial biopharmaceutical formulation is input into the performance prediction model to obtain the predicted initial activity parameters and the predicted initial stability parameters. Based on the biopharmaceutical formulation to be optimized, activity weights and stability weights are obtained, and fitness parameters are obtained by combining the predicted initial activity parameters and the predicted initial stability parameters. The initial particle swarm is used for iterative optimization to obtain the particle with the largest predicted initial performance parameter, which is then used as the optimized biopharmaceutical formulation, and the optimized performance parameter is obtained.

[0054] First, multiple particles are randomly generated within the feasible region to obtain an initial particle swarm, where each initial particle represents an initial biopharmaceutical formulation. A particle is the basic search unit in the particle swarm optimization algorithm; each particle corresponds to a complete biopharmaceutical formulation, and its dimension corresponds to the number of components in the formulation. Each dimension value of the particle represents the content percentage of the corresponding component. The initial particle swarm is a collection of multiple randomly generated particles within the feasible region, serving as the initial search sample library for iterative optimization. Based on the boundary range of each component within the feasible region, the dimensions of the particles and the constraints on the values ​​of each dimension are determined. The particle swarm size is set, i.e., the number of initial formulations. A uniform random sampling method is used to randomly assign values ​​to each dimension of each particle within the feasible boundaries of each component, ensuring that the total component percentage of each particle satisfies the sum constraint. Invalid particles that exceed the feasible region or do not satisfy the sum constraint are filtered out, ultimately forming the initial particle swarm.

[0055] For example, for the basic formulation F1 of an injectable gel, its optimized feasible region is: sodium hyaluronate 1.94%-5.76%, polydeoxyribonucleotide (PDRN) 0.97%-2.88%, glycine 0.19%-0.96%, and PBS buffer 88.71%-90.40%. The following steps are performed: Determine particle dimension: The formulation contains four components, and the particle dimension is set to 4, corresponding to the mass percentages of sodium hyaluronate, PDRN, glycine, and PBS buffer, respectively. Set particle swarm size: Select 50 particles as the initial particle swarm, i.e., 50 initial formulations. Randomly generate particles: Through uniform random sampling, assign values ​​to each particle within the feasible region of each component. For example, the values ​​of an initial particle P1 are: sodium hyaluronate 4.2%, PDRN 2.3%, glycine 0.7%, and PBS buffer 92.8%, totaling 100%, and all component values ​​are within the feasible region. Screening for effective particles: Three groups of invalid particles whose sum deviates from 100% (error > 0.1%) are removed, resulting in 47 groups of effective particles that form the initial particle swarm.

[0056] Next, the initial biopharmaceutical formulation is input into the performance prediction model to obtain predicted initial activity parameters and predicted initial stability parameters. The initial biopharmaceutical formulation refers to the specific formulation corresponding to each particle in the initial particle swarm, i.e., the combination of the content proportions of each component. The predicted initial activity parameter is the predicted value of the activity quantification index output by inputting the initial biopharmaceutical formulation into the activity prediction branch of the performance prediction model. The predicted initial stability parameter is the predicted value of the stability quantification index output by inputting the initial biopharmaceutical formulation into the stability prediction branch of the performance prediction model. Each effective particle in the initial particle swarm is traversed, and the values ​​of each dimension of the particle are extracted, i.e., the content proportions of each component, forming standardized initial biopharmaceutical formulation data. Each set of formulation data is then input into the trained and converged bi-branch random forest performance prediction model. The predicted initial activity parameter of the formulation is output through the activity prediction branch, and the predicted initial stability parameter is output through the stability prediction branch, and these parameters are associated with and stored with the corresponding initial formulation.

[0057] For example, the following operations are performed on the above initial particle swarm: Extract initial formulation data: Taking initial particle P1 as an example, its corresponding initial biopharmaceutical formulation is [sodium hyaluronate: 4.2%, PDRN: 2.3%, glycine: 0.7%, PBS buffer: 92.8%]. Model prediction: Input the formulation data into a two-branch random forest performance prediction model. The activity prediction branch outputs the predicted initial activity parameters as [collagen production rate: 34%, skin moisturizing improvement rate: 47%]; the stability prediction branch outputs the predicted initial stability parameters as [in vitro degradation time: 85h, shelf life under refrigeration (2-8℃): 19 months]. Association and storage: Particle P1, the corresponding formulation data, the predicted initial activity parameters, and the predicted initial stability parameters are bound and stored. The performance prediction of all 47 initial formulations is completed according to the same process.

[0058] Further, based on the biopharmaceutical formulation to be optimized, activity weights and stability weights are obtained. Combined with the predicted initial activity parameters and the predicted initial stability parameters, a fitness parameter is obtained. The activity weight is a coefficient set according to the application scenario of the formulation, used to measure the importance of the activity parameter in the overall performance evaluation, with a value range of [0,1]. The stability weight is a coefficient used to measure the importance of the stability parameter in the overall performance evaluation, with a value range of [0,1], and its sum with the activity weight is 1. The fitness parameter is a comprehensive performance evaluation index calculated based on the activity weight, stability weight, and the corresponding predicted parameters. It is the core basis for selecting the optimal particle in the particle swarm optimization algorithm; a larger value indicates better overall formulation performance. Based on the application scenario and core requirements of the biopharmaceutical formulation to be optimized, activity weights and stability weights (the sum of which is 1) are set. The predicted initial activity parameters of each initial formulation were normalized, and the parameter values ​​were converted to the [0,1] interval. The overall activity score was calculated. The predicted initial stability parameters were also normalized, and the overall stability score was calculated. The fitness parameter was calculated using a weighted summation formula: Fitness parameter = Overall activity score × Activity weight + Overall stability score × Stability weight.

[0059] For example, for the base formulation F1 of the injectable gel, the following operations are performed: Weighting: This formulation is used for skin repair, where activity is prioritized over stability. The previously set activity weight is 0.6, and the stability weight is 0.4 (totaling 1). Parameter Normalization: Activity parameter normalization: Based on the predicted collagen production rate (range 28%-38%) and skin hydration improvement rate (range 40%-52%) of all formulations in the particle swarm, the min-max normalization formula is used to calculate: Normalized value = (actual value - minimum value) / (maximum value - minimum value). For example, the collagen production rate of particle P1 at 34% is normalized to (34-28) / (38-28) = 0.6, and the skin hydration improvement rate at 47% is normalized to (47-40) / (52-40) ≈ 0.583. The overall activity score is (0.6 + 0.583) / 2 ≈ 0.5915. Stability parameter normalization: Based on the predicted in vitro degradation time (range 70h-95h) and refrigerated shelf life (range 16 months-22 months) in the particle swarm, the normalized in vitro degradation time of particle P1 at 85h is (85-70) / (95-70)=0.6, and the normalized refrigerated shelf life at 19 months is (19-16) / (22-16)=0.5. The overall stability score is (0.6+0.5) / 2=0.55. Fitness parameter calculation: Fitness parameter of particle P1 = 0.5915×0.6+0.55×0.4≈0.575. Fitness parameters for all 47 initial formulations were calculated using the same method.

[0060] Finally, the initial particle swarm is used for iterative optimization to obtain the particle with the largest predicted initial performance parameter, which is then used as the optimized biopharmaceutical formulation, and the optimized performance parameter is obtained.

[0061] The iterative optimization using the initial particle swarm includes: Initialize particle positions within the optimized feasible region and randomly initialize particle velocities; Based on the performance prediction confidence level, the optimization stagnation discrimination round threshold is dynamically set, wherein the lower the performance prediction confidence level, the smaller the optimization stagnation discrimination round threshold is; During the iterative optimization process, monitor the improvement of the global optimal fitness parameter of the particle swarm. If the number of consecutive iterations reaches the optimization stagnation discrimination round threshold, and the improvement of the global optimal fitness parameter is less than the preset improvement threshold, or the predicted initial stability parameter is less than the stability threshold, or the predicted initial activity parameter is less than the activity threshold, then optimization stagnation is determined. The optimization stagnation discrimination round threshold is obtained based on the performance prediction confidence, and the stability threshold and the activity threshold are obtained based on the component characteristics of the biopharmaceutical formulation. When optimization stalls, an extended search is performed. Based on the particle swarm distribution at the time of optimization stall, the feasible region is expanded, and some particles are re-initialized within the expanded feasible region for further search. The expanded range is obtained based on the activity prediction confidence and stability prediction confidence.

[0062] The iterative optimization using the initial particle swarm also includes: When updating particle velocities, their inertial weights are adjusted based on the performance prediction confidence level, which is inversely proportional to the inertial weights.

[0063] First, the particle positions are initialized within the optimized feasible region, and the particle velocities are randomly initialized. The values ​​of each particle in the initial particle swarm are used as the initial positions. Based on the boundary width of each component in the optimized feasible region, the range of particle velocity values ​​is set, typically ±5% to ±15% of the feasible region width of the corresponding component. Uniform random sampling is used to assign initial velocities to each dimension of each particle to ensure that the velocity direction and magnitude are reasonable. For example, for the initial particle swarm of injectable gel, the following operation is performed: Particle position initialization: The current values ​​of 47 effective particles are used as the initial positions, and the initial position of particle P1 is [4.2%, 2.3%, 0.7%, 92.8%]. Particle velocity initialization: The velocity range is set to ±10% of the feasible domain width of each component. For example, the feasible domain width of sodium hyaluronate is 5.76%-1.94%=3.82%, and its velocity range is ±3.82%×10%≈±0.38%; the feasible domain width of PDRN is 2.88%-0.97%=1.91%, and its velocity range is ±0.19%. Initial velocities are randomly assigned to particle P1: [+0.25%, +0.12%, +0.05%, -0.42%].

[0064] Secondly, based on the performance prediction confidence level, a threshold for the optimization stagnation discrimination round is dynamically set. The lower the performance prediction confidence level, the smaller the optimization stagnation discrimination round threshold. The optimization stagnation discrimination round threshold is the critical number of iterations required to determine if optimization has stalled. If the optimization improvement requirements are not met after this number of consecutive iterations, it is considered stalled. The threshold is negatively correlated with the performance prediction confidence level; the lower the confidence level, the smaller the threshold, avoiding invalid iterations in low-reliability prediction scenarios. A threshold baseline range is set to establish a mapping relationship between the performance prediction confidence level and the threshold: the higher the confidence level, the closer the threshold is to the baseline upper limit; the lower the confidence level, the closer the threshold is to the baseline lower limit. A linear mapping formula is used to calculate the specific threshold: Threshold = Baseline Lower Limit + (Confidence Level - Minimum Confidence Level) × (Baseline Upper Limit - Baseline Lower Limit) / (Maximum Confidence Level - Minimum Confidence Level), where the confidence level ranges from [0,1].

[0065] For example, with a performance prediction confidence level of 0.912 for formulation F1, the following operations are performed: Set a baseline range: the lower limit of the optimization stagnation threshold is 5 rounds, and the upper limit is 20 rounds. Calculate the threshold: Substituting into the formula, we get threshold = 5 + (0.912 - 0) × (20 - 5) / (1 - 0) ≈ 18.68, rounded to 19 rounds. That is, if the optimization improvement requirements are not met after 19 consecutive iterations, it is judged as optimization stagnation.

[0066] Further, the inertia weights are adjusted and the particle velocity and position are updated. When updating particle velocity, the inertia weight is adjusted based on the performance prediction confidence level, which is inversely proportional to the inertia weight. The inertia weight is a parameter controlling the influence of the particle's historical movement speed on its current movement. Its value ranges from 0.4 to 0.9, and it is inversely proportional to the performance prediction confidence level; the higher the confidence level, the smaller the weight, and the easier it is for the particle to adjust towards the optimal direction. Velocity update refers to adjusting the movement direction and speed of the next round based on the particle's past movement, its own optimal formula gap, and the group's optimal formula gap, making it easier for the particle to approach the high-quality formula. Position update refers to adjusting the proportion of each component corresponding to the particle according to the updated velocity, ensuring that the adjusted position remains within the optimization feasible region. Determining the inertia weight: It is directly set based on the performance prediction confidence level. A smaller value (e.g., 0.4~0.6) is used for high confidence levels, and a larger value (e.g., 0.7~0.9) is used for low confidence levels, ensuring the weight is within the range of 0.4~0.9. Update particle velocity: Retain some historical velocity: Multiply the particle's velocity from the previous round by an inertia weight to maintain a certain inertia. Add self-optimization tendency: Calculate the difference in component proportions between the particle's current recipe and its historical best recipe, multiply by a fixed coefficient of 2 and a random number between 0 and 1, giving the particle a tendency to move closer to its own optimal state. Add group optimization tendency: Calculate the difference in component proportions between the particle's current recipe and the global best recipe of the entire particle swarm, multiply by a fixed coefficient of 2 and a random number between 0 and 1, giving the particle a tendency to move closer to the group's optimal state. Add these three factors together to obtain the new velocity, ensuring that the velocity does not exceed a preset reasonable range. Update particle position: Add the new velocity of the corresponding dimension to the current proportion of each component of the particle to obtain the adjusted component proportion. If the adjusted proportion of a component exceeds the boundary of the feasible optimization region, then directly take that boundary value.

[0067] For example, with a performance prediction confidence level of 0.912 for formulation F1, the following operations are performed: Determine the inertia weight: Due to the high confidence level, the inertia weight is set to 0.44. Update particle velocity: Taking particle P1 as an example, the velocity in the previous round was [+0.25% (sodium hyaluronate), +0.12% (PDRN), +0.05% (glycine), -0.42% (PBS buffer)]; the component proportions of its own historical best formulation were [4.8%, 2.5%, 0.8%, 91.9%], and the current proportions are [4.2%, 2.3%, 0.7%, 92.8%], with a difference of [0.6%, 0.2%, 0.1%, -0.9%]; the component proportions of the population's global best formulation were [5.1%, 2.6%, 0.8%, 91.5%], with a difference of [0.9%, 0.3%, 0.1%, -1.3%] from the current proportions; random numbers are taken as 0.3 and 0.6. Retaining some historical rates (e.g., sodium hyaluronate: 0.44 × 0.25% ≈ 0.11%), and adding its own optimization tendency (0.6% × 2 × 0.3 ≈ 0.36%) and the group optimization tendency (0.9% × 2 × 0.6 ≈ 1.08%), the final new rate of sodium hyaluronate is approximately 1.55%. The other components are adjusted according to the same logic, and the overall new rates are [1.55%, 0.53%, 0.20%, -2.29%], all within a reasonable range. Updated particle positions: The current component percentages of particle P1 are adjusted by adding the new velocity. Sodium hyaluronate is adjusted from 4.2% to 5.75%, PDRN from 2.3% to 2.83%, glycine from 0.7% to 0.90%, and PBS buffer from 92.8% to 90.51%. This exceeds the upper limit of 88.71%~90.40%, so it is truncated to 90.40%. The final adjusted particle positions are [5.75%, 2.83%, 0.90%, 90.40%], which meets the feasible region constraint.

[0068] Next, during the iterative optimization process, the improvement of the global optimal fitness parameter of the particle swarm is monitored. The global optimal fitness parameter refers to the maximum fitness parameter of all particles during the iteration process, corresponding to the current optimal recipe. The improvement rate refers to the ratio of the difference between the current round's global optimal fitness and the previous round's value to the previous round's value. After each iteration, the fitness parameters of all particles are recalculated, and the historical optimal fitness and position of the particles, as well as the global optimal fitness and position of the particle swarm, are updated; the global optimal fitness parameter of each round is recorded, and the improvement rate of adjacent rounds is calculated to monitor the optimization progress in real time. For example, for the particle swarm iteration process, the following operations are performed: the global optimal fitness in round 10 is 0.68, and in round 11 it is 0.695, with an improvement rate of (0.695-0.68) / 0.68≈2.21%; the global optimal fitness in round 20 is 0.75, and in round 21 it is 0.752, with an improvement rate of ≈0.27%, and this rate of change is continuously monitored.

[0069] Furthermore, if the number of consecutive iterations reaches the optimization stagnation threshold, and the improvement of the global optimal fitness parameter is less than a preset improvement threshold, or the predicted initial stability parameter is less than a stability threshold, or the predicted initial activity parameter is less than an activity threshold, then optimization stagnation is determined. The optimization stagnation threshold is obtained based on the performance prediction confidence level, and the stability threshold and activity threshold are obtained based on the component characteristics of the biopharmaceutical formulation. The improvement threshold is a critical proportion for determining whether the fitness improvement is effective; below this proportion, it is considered ineffective. The activity threshold is a minimum requirement for activity parameters set based on the formulation component characteristics to ensure that the core efficacy of the formulation is met. The stability threshold is a minimum requirement for stability parameters set based on the formulation component characteristics to ensure that the formulation's usage period is met. Key thresholds are set as follows: the improvement threshold is typically set to 0.1%~1%; the activity threshold is set according to the core efficacy of the component (e.g., collagen production rate ≥30%); and the stability threshold is set according to the component stability requirements (e.g., in vitro degradation time ≥75h). Stagnation determination logic: When the number of consecutive iterations reaches the optimization stagnation determination round threshold, optimization stagnation is determined if any of the following conditions are met: the improvement of the global optimal fitness parameter is less than the improvement threshold; the predicted initial activity parameter is less than the activity threshold; or the predicted initial stability parameter is less than the stability threshold.

[0070] For example, for formulation F1, the following operations were performed: Key thresholds were set: improvement threshold = 0.1%, activity threshold = collagen production rate ≥ 30%, and stability threshold = in vitro degradation time ≥ 75 h. Iterative monitoring: In 19 consecutive iterations, the global optimal fitness in iterations 60-78 (out of 19) increased from 0.778 to 0.779, an increase of approximately 0.13%, slightly higher than the improvement threshold; however, the predicted initial stability parameter for iteration 78 was 74.5 h < 75 h, meeting the stagnation criteria, and was therefore determined to be an optimization stagnation.

[0071] Subsequently, when optimization stagnation is detected, an extended search is performed. Based on the particle swarm distribution at the time of optimization stagnation, the feasible region is expanded, and some particles are reinitialized within the expanded feasible region for further searching. The expanded range is obtained based on the activity prediction confidence and stability prediction confidence. Extended search refers to a search strategy that expands the feasible region and reinitializes some particles when optimization stagnates, used to break local optima. The expanded range is the feasible region expansion ratio determined based on the activity prediction confidence and stability prediction confidence; the lower the confidence, the larger the expansion ratio. Calculating the expansion ratio: Set a baseline expansion ratio (e.g., 5%~15%), expansion ratio = baseline ratio × (2 - activity prediction confidence - stability prediction confidence), ensuring that the lower the confidence, the more significant the expansion. Expanding the feasible region: Expand the feasible boundaries of each component outward according to the expansion ratio, keeping the lower bound expansion ratio consistent with the upper bound. Reinitializing particles: Within the expanded feasible region, randomly generate new particles accounting for 20%~30% of the original particle swarm size, replacing particles with lower fitness in the original particle swarm.

[0072] For example, the activity prediction confidence level of formulation F1 is 0.92, and the stability prediction confidence level is 0.90. The expansion ratio is calculated as follows: setting the baseline ratio at 10%, the expansion ratio = 10% × (2 - 0.92 - 0.90) = 10% × 0.18 = 1.8%. The feasible domain is expanded: the original feasible domain of sodium hyaluronate (1.94%-5.76%) is expanded to 1.90%-5.86% (both upper and lower bounds ±1.8%), and the remaining components are expanded in the same proportion. The particles are reinitialized: the original particle swarm of 47 particles is regenerated into 10 new particles (approximately 21%), replacing the bottom 10 particles in fitness ranking, forming a new particle swarm for continued iteration.

[0073] Finally, the particle with the highest predicted initial performance parameters is selected as the optimized biopharmaceutical formulation, and the optimized performance parameters are obtained. A maximum number of iterations is set. If no optimization stagnation occurs during iteration or optimization progress resumes after an extended search, the iteration continues until the maximum number of iterations is reached, at which point the process terminates. If two consecutive extended searches still determine that optimization is stagnant, the process terminates early. After iteration terminates, the formulation corresponding to the globally optimal particle is extracted as the optimized biopharmaceutical formulation and input into the performance prediction model to obtain the optimized performance parameters.

[0074] For example, a maximum of 200 iterations is set. After one extended search, the recovery progress is optimized. When iterating to the 150th iteration, the global optimal fitness reaches 0.82, and the improvement in the subsequent 20 iterations is less than 0.1%, satisfying the termination condition. The optimized formulation is determined: the formulation corresponding to the globally optimal particle is [sodium hyaluronate: 5.6%, PDRN: 2.8%, glycine: 0.9%, PBS buffer: 90.7%]. Optimized performance parameters are obtained: inputting into the performance prediction model, the output is the activity parameters [collagen production rate: 37.5%, skin moisturizing improvement rate: 51%], and the stability parameters [in vitro degradation time: 95h, shelf life under refrigeration (2-8℃): 22 months].

[0075] In this embodiment of the invention, an efficient and accurate optimization process for biopharmaceutical formulations is achieved through a particle swarm optimization (PSO) iterative optimization procedure. Scientific initialization of particle position and velocity ensures the comprehensiveness of the search. Dynamic thresholds and inertia weight adjustments based on performance prediction confidence levels adapt to different prediction reliability scenarios, avoiding the inefficiency caused by fixed parameters. The optimization stagnation detection and extended search mechanism effectively breaks local optima, addressing the pain point of traditional PSO algorithms being prone to stagnation. The resulting optimized biopharmaceutical formulation strictly adheres to the feasible domain constraints of the components and achieves synergistic optimization of activity and stability through fitness parameters, with optimized performance parameters significantly superior to the initial formulation. The entire iterative process balances search efficiency and optimization accuracy, providing an efficient and feasible technical path for biopharmaceutical formulation development and reducing the cost of trial and error in research and development.

[0076] S500: Based on the optimized biopharmaceutical formulation and the optimized performance parameters, generate the screening results of biopharmaceutical active ingredients.

[0077] In this embodiment of the invention, the active ingredient screening results are generated based on the optimized biopharmaceutical formulation and the optimized performance parameters. First, the names and corresponding content percentages of the active ingredients in the optimized biopharmaceutical formulation are extracted. Then, their corresponding optimized performance parameters are correlated, and compared with the initial formulation or industry standards, to screen out the active ingredients that play a key role in performance improvement, clarify their content range and efficacy advantages, and finally form a concise screening result. For example, the active ingredients extracted from the formulation are polydeoxyribonucleotide (PDRN) and glycine, with corresponding proportions of 2.8% and 0.9%, respectively. Corresponding to their optimized performance: PDRN increases collagen production rate by 5.5% compared to the initial formulation, and glycine increases skin hydration rate by 6%, and both synergistically extend in vitro degradation time by 15 hours. Compared with industry standards, both meet and exceed the requirements. The final screening result is: the core active ingredients are: polydeoxyribonucleotide (PDRN) with an optimal proportion of 2.8%, whose core function is promoting cell regeneration; and glycine with an optimal proportion of 0.9%, whose core function is moisturizing and anti-oxidation. The two have good compatibility and can support the formulation to achieve the expected skin repair and filling effects.

[0078] Through the specific implementation methods described above, the embodiments of the present invention achieve the following technical effects: This invention provides an artificial intelligence-based method and system for screening biopharmaceutical active ingredients. By comprehensively collecting formulation and performance data, constructing a bi-branch performance prediction model and quantifying confidence levels, dynamically adjusting and optimizing the feasible region, and combining iterative optimization with an adaptive particle swarm optimization algorithm, the system ultimately and accurately screens key active ingredients and their optimal proportions. The entire process considers the synergistic optimization of activity and stability, effectively improving the efficiency and accuracy of biopharmaceutical active ingredient screening, and providing clear and practical technical support for the implementation of biopharmaceutical formulations.

[0079] Example 2, as Figure 2 As shown, this invention provides an artificial intelligence-based biopharmaceutical active ingredient screening system, the system comprising: The parameter acquisition module 11 is used to acquire the component characteristic parameters and performance parameters of the biopharmaceutical formulation to be optimized. The confidence assessment module 12 is used to construct a performance prediction model and obtain the performance prediction confidence based on the component feature parameters and the performance parameters. The feasible region determination module 13 is used to obtain an optimized feasible region based on the component features and the performance prediction confidence. The iterative optimization module 14 is used to perform iterative optimization using the particle swarm optimization algorithm within the optimization feasible region to obtain an optimized biopharmaceutical formulation, and to obtain the optimized performance parameters of the optimized biopharmaceutical formulation based on the performance prediction model. The result generation module 15 is used to generate screening results of biopharmaceutical active ingredients based on the optimized biopharmaceutical formulation and the optimized performance parameters.

[0080] In one embodiment, the parameter acquisition module 11 is further configured to: Obtain the component characteristic parameters of the biopharmaceutical formulation to be optimized, wherein the component characteristic parameters include component characteristics and component ratio parameters; Obtain the performance parameters of the biopharmaceutical formulation to be optimized, wherein the performance parameters include activity parameters and stability parameters.

[0081] In one embodiment, the confidence assessment module 12 is further configured to: The construction of the performance prediction model includes: Obtain a set of sample biopharmaceutical formulations that have the same main components as the biopharmaceutical formulation to be optimized, and obtain the sample characteristic parameter set, sample activity parameter set, and sample stability parameter set of the sample biopharmaceutical formulation set; Construct a performance prediction model, wherein the performance prediction model includes an activity prediction branch and a stability prediction branch; Using the sample feature parameter set as input, and the sample activity parameter set and the sample stability parameter set as supervision, the activity prediction branch and the stability prediction branch are trained respectively until the model converges.

[0082] The process of obtaining performance prediction confidence based on the component feature parameters and the performance parameters includes: The component characteristic parameters are input into the performance prediction model to obtain the predicted activity parameters and predicted stability parameters, which are used as the predicted performance parameters. Based on the predicted performance parameters and the performance parameters, a performance prediction confidence level is obtained, wherein the performance prediction confidence level is obtained by weighted calculation based on the activity prediction confidence level and the stability prediction confidence level.

[0083] In one embodiment, the feasible region determination module 13 is further configured to: Based on the component characteristics, the initial independent feasible boundaries of each component are obtained; The initial independent feasible boundary is adjusted based on the performance prediction confidence level. Obtain the adjusted feasible boundaries of each component to form the optimized feasible region.

[0084] The adjustment of the initial independent feasible boundary based on the performance prediction confidence includes: Obtain a first confidence threshold and a second confidence threshold, wherein the second confidence threshold is greater than the first confidence threshold; If the performance prediction confidence is less than or equal to the first confidence threshold, then each initial independent feasible boundary is expanded outward by a first proportion to amplify the optimized feasible region. If the performance prediction confidence is greater than or equal to the second confidence threshold, then each initial independent feasible boundary is shrunk inward by a second proportion to narrow the optimized feasible region. If the performance prediction confidence level is between the first confidence threshold and the second confidence threshold, then a third ratio is determined based on linear interpolation, and each initial independent feasible boundary is linearly scaled according to this third ratio. Obtain the adjusted boundaries of each component, and normalize the upper bounds of all components so that their sum satisfies the sum constraint, thus forming the optimized feasible region.

[0085] In one embodiment, the iterative optimization module 14 is further configured to: Multiple particles are randomly generated within the optimized feasible region to obtain an initial particle swarm, wherein each initial particle represents an initial biopharmaceutical formulation. The initial biopharmaceutical formulation is input into the performance prediction model to obtain the predicted initial activity parameters and the predicted initial stability parameters. Based on the biopharmaceutical formulation to be optimized, activity weights and stability weights are obtained, and fitness parameters are obtained by combining the predicted initial activity parameters and the predicted initial stability parameters. The initial particle swarm is used for iterative optimization to obtain the particle with the largest predicted initial performance parameter, which is then used as the optimized biopharmaceutical formulation, and the optimized performance parameter is obtained.

[0086] The iterative optimization using the initial particle swarm includes: Initialize particle positions within the optimized feasible region and randomly initialize particle velocities; Based on the performance prediction confidence level, the optimization stagnation discrimination round threshold is dynamically set, wherein the lower the performance prediction confidence level, the smaller the optimization stagnation discrimination round threshold is; During the iterative optimization process, monitor the improvement of the global optimal fitness parameter of the particle swarm. If the number of consecutive iterations reaches the optimization stagnation discrimination round threshold, and the improvement of the global optimal fitness parameter is less than the preset improvement threshold, or the predicted initial stability parameter is less than the stability threshold, or the predicted initial activity parameter is less than the activity threshold, then optimization stagnation is determined. The optimization stagnation discrimination round threshold is obtained based on the performance prediction confidence, and the stability threshold and the activity threshold are obtained based on the component characteristics of the biopharmaceutical formulation. When optimization stalls, an extended search is performed. Based on the particle swarm distribution at the time of optimization stall, the feasible region is expanded, and some particles are re-initialized within the expanded feasible region for further search. The expanded range is obtained based on the activity prediction confidence and stability prediction confidence.

[0087] The iterative optimization using the initial particle swarm also includes: When updating particle velocities, their inertial weights are adjusted based on the performance prediction confidence level, which is inversely proportional to the inertial weights.

[0088] It should be noted that the order of the above embodiments of the present invention 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 specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0089] 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 spirit and principles of the present invention should be included within the protection scope of the present invention.

[0090] This specification and accompanying drawings are merely illustrative examples of the invention and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its scope. Therefore, if such modifications and modifications fall within the scope of the invention and its equivalents, the invention is intended to include these modifications and modifications.

Claims

1. A method for screening biopharmaceutical active ingredients based on artificial intelligence, characterized in that, include: Obtain the component characteristic parameters and performance parameters of the biopharmaceutical formulation to be optimized; Construct a performance prediction model and obtain the performance prediction confidence level based on the component feature parameters and the performance parameters; Based on the component characteristics and the performance prediction confidence, the optimization feasible region is obtained; Within the optimized feasible region, the particle swarm optimization algorithm is used for iterative optimization to obtain the optimized biopharmaceutical formulation, and the optimized performance parameters of the optimized biopharmaceutical formulation are obtained based on the performance prediction model. Based on the optimized biopharmaceutical formulation and the optimized performance parameters, the screening results of biopharmaceutical active ingredients are generated.

2. The method for screening biopharmaceutical active ingredients based on artificial intelligence according to claim 1, characterized in that, Obtain the component characteristic parameters of the biopharmaceutical formulation to be optimized, including: Obtain the component characteristic parameters of the biopharmaceutical formulation to be optimized, wherein the component characteristic parameters include component characteristics and component ratio parameters; Obtain the performance parameters of the biopharmaceutical formulation to be optimized, wherein the performance parameters include activity parameters and stability parameters.

3. The method for screening biopharmaceutical active ingredients based on artificial intelligence according to claim 1, characterized in that, Construct a performance prediction model, including: Obtain a set of sample biopharmaceutical formulations that have the same main components as the biopharmaceutical formulation to be optimized, and obtain the sample characteristic parameter set, sample activity parameter set, and sample stability parameter set of the sample biopharmaceutical formulation set; Construct a performance prediction model, wherein the performance prediction model includes an activity prediction branch and a stability prediction branch; Using the sample feature parameter set as input, and the sample activity parameter set and the sample stability parameter set as supervision, the activity prediction branch and the stability prediction branch are trained respectively until the model converges.

4. The method for screening biopharmaceutical active ingredients based on artificial intelligence according to claim 1, characterized in that, Based on the component feature parameters and the performance parameters, the performance prediction confidence level is obtained, including: The component characteristic parameters are input into the performance prediction model to obtain the predicted activity parameters and predicted stability parameters, which are used as the predicted performance parameters. Based on the predicted performance parameters and the performance parameters, a performance prediction confidence level is obtained, wherein the performance prediction confidence level is obtained by weighted calculation based on the activity prediction confidence level and the stability prediction confidence level.

5. The method for screening biopharmaceutical active ingredients based on artificial intelligence according to claim 1, characterized in that, Based on the component characteristics and the performance prediction confidence, the optimization feasible region is obtained, including: Based on the component characteristics, the initial independent feasible boundaries of each component are obtained; The initial independent feasible boundary is adjusted based on the performance prediction confidence level. Obtain the adjusted feasible boundaries of each component to form the optimized feasible region.

6. The method for screening biopharmaceutical active ingredients based on artificial intelligence according to claim 5, characterized in that, Based on the performance prediction confidence level, the initial independent feasible boundary is adjusted, including: Obtain a first confidence threshold and a second confidence threshold, wherein the second confidence threshold is greater than the first confidence threshold; If the performance prediction confidence is less than or equal to the first confidence threshold, then each initial independent feasible boundary is expanded outward by a first proportion to amplify the optimized feasible region. If the performance prediction confidence is greater than or equal to the second confidence threshold, then each initial independent feasible boundary is shrunk inward by a second proportion to narrow the optimized feasible region. If the performance prediction confidence level is between the first confidence threshold and the second confidence threshold, then a third ratio is determined based on linear interpolation, and each initial independent feasible boundary is linearly scaled according to this third ratio. Obtain the adjusted boundaries of each component, and normalize the upper bounds of all components so that their sum satisfies the sum constraint, thus forming the optimized feasible region.

7. The method for screening biopharmaceutical active ingredients based on artificial intelligence according to claim 1, characterized in that, Within the optimized feasible region, based on the performance prediction model, predicted performance parameters of the optimized biopharmaceutical formulation are obtained. Iterative optimization is performed using a particle swarm optimization algorithm to obtain the optimized biopharmaceutical formulation, including: Multiple particles are randomly generated within the optimized feasible region to obtain an initial particle swarm, wherein each initial particle represents an initial biopharmaceutical formulation. The initial biopharmaceutical formulation is input into the performance prediction model to obtain the predicted initial activity parameters and the predicted initial stability parameters. Based on the biopharmaceutical formulation to be optimized, activity weights and stability weights are obtained, and fitness parameters are obtained by combining the predicted initial activity parameters and the predicted initial stability parameters. The initial particle swarm is used for iterative optimization to obtain the particle with the largest predicted initial performance parameter, which is then used as the optimized biopharmaceutical formulation, and the optimized performance parameter is obtained.

8. The method for screening biopharmaceutical active ingredients based on artificial intelligence according to claim 7, characterized in that, Iterative optimization using the initial particle swarm includes: Initialize particle positions within the optimized feasible region and randomly initialize particle velocities; Based on the performance prediction confidence level, the optimization stagnation discrimination round threshold is dynamically set, wherein the lower the performance prediction confidence level, the smaller the optimization stagnation discrimination round threshold is; During the iterative optimization process, monitor the improvement of the global optimal fitness parameter of the particle swarm. If the number of consecutive iterations reaches the optimization stagnation discrimination round threshold, and the improvement of the global optimal fitness parameter is less than the preset improvement threshold, or the predicted initial stability parameter is less than the stability threshold, or the predicted initial activity parameter is less than the activity threshold, then optimization stagnation is determined. The optimization stagnation discrimination round threshold is obtained based on the performance prediction confidence, and the stability threshold and the activity threshold are obtained based on the component characteristics of the biopharmaceutical formulation. When optimization stalls, an extended search is performed. Based on the particle swarm distribution at the time of optimization stall, the feasible region is expanded, and some particles are re-initialized within the expanded feasible region for further search. The expanded range is obtained based on the activity prediction confidence and stability prediction confidence.

9. The method for screening biopharmaceutical active ingredients based on artificial intelligence according to claim 7, characterized in that, Iterative optimization using the initial particle swarm also includes: When updating particle velocities, their inertial weights are adjusted based on the performance prediction confidence level, which is inversely proportional to the inertial weights.

10. An artificial intelligence-based biopharmaceutical active ingredient screening system, characterized in that, The system is used to implement the artificial intelligence-based biopharmaceutical active ingredient screening method according to any one of claims 1-9, the system comprising: The parameter acquisition module is used to acquire the component characteristic parameters and performance parameters of the biopharmaceutical formulation to be optimized. The confidence assessment module is used to construct a performance prediction model and obtain the performance prediction confidence based on the component feature parameters and the performance parameters. The feasible region determination module is used to obtain an optimized feasible region based on the component features and the performance prediction confidence. The iterative optimization module is used to perform iterative optimization using a particle swarm optimization algorithm within the optimization feasible region to obtain an optimized biopharmaceutical formulation, and to obtain optimized performance parameters of the optimized biopharmaceutical formulation based on the performance prediction model. The results generation module is used to generate screening results for biopharmaceutical active ingredients based on the optimized biopharmaceutical formulation and the optimized performance parameters.

Images