Method and system for developing basal medium

The method and system leverage culture index evaluation and AI to optimize basal medium composition, addressing inefficiencies in conventional methods by rapidly identifying optimal compositions for high cell viability and density.

JP7592177B2Active Publication Date: 2024-11-29SHENZHEN TAILI BIOTECHNOLOGY CO LTD
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
JP2023543259
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Priority Date
2020-11-26
Filing Date
2021-11-17
Publication Date
2024-11-29
Estimated Expiration
2041-11-17

AI Technical Summary

Technical Problem

Conventional methods for developing basal media are time-consuming and require multiple specialized fields, often failing to achieve an optimal composition due to the inability to include all components in each test.

Method used

A method and system utilizing culture index evaluation and artificial intelligence to optimize basal medium composition through regression prediction models, machine learning algorithms, and a sample composition database to predict and verify the most effective medium compositions efficiently.

Benefits of technology

Significantly reduces development time and costs by identifying optimal basal medium compositions quickly and accurately, ensuring high cell viability and density.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007592177000016
    Figure 0007592177000016
  • Figure 0007592177000017
    Figure 0007592177000017
  • Figure 0007592177000018
    Figure 0007592177000018
Patent Text Reader

Abstract

This application discloses a method and system for developing a basal medium and a basal medium composition, which includes the steps of: (1) determining a regression prediction model for predicting a basal medium composition for a selected culture index; (2) generating multiple candidate basal medium compositions by obtaining the addition ranges of each component of the basal medium and randomly selecting from the list; (3) predicting the culture index using the regression prediction model and recommending a basal medium composition; and (4) conducting cell culture experiments to verify the culture index of the recommended basal medium composition.
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] This application relates to a Chinese patent application filed with the State Intellectual Property Office of the People's Republic of China on November 26, 2020, bearing application number 2020113431805 and title "Method for developing a basal medium based on culture index evaluation." Wish No. 6,313,935, filed on May 23, 2003, the contents of which are incorporated herein by reference in their entirety.

[0002] The present application belongs to the field of biotechnology, and more specifically, relates to a method for developing a basal medium, and a method and system for developing a basal medium composition. [Background technology]

[0003] Serum-free, animal-derived component-free, chemically defined basal media consist of carbon sources, amino acids, vitamins, trace metal ions, lipids, buffers and other additives. Traditional methods for developing basal medium compositions are based on one or several typical media (e.g., DEME / F12), and then adding multiple different components, employing single-factor testing or DOE screening testing to find out the critical components, and then using multiple DOE experimental designs such as response surface to optimize the concentration of each component to obtain the optimal composition; or optimizing the composition by finding out the changes in each component during cell growth and its effect on the target product and quality based on cell metabolism analysis, genomics analysis and proteomics analysis.

[0004] Conventional techniques require multiple tests, cannot include all components in each test, are time-consuming, and require a grasp of many specialized theoretical knowledge in basic chemistry, biochemistry, molecular biology, cell biology, etc., and the resulting composition may not be optimal.

[0005] The conventional method for developing a medium composition is to base it on one or several typical media (e.g., DEME / F12), add several different components, employ one-factor testing or DOE screening testing to find out the critical components, and then use several DOE experimental designs, such as response surface, to optimize the concentration of each component to obtain the optimal composition; or optimize the composition by finding out the change status of each component during cell growth and its effect on the target product and quality based on cell metabolism analysis, genomics analysis and proteomics analysis. Summary of the Invention [Problem to be solved by the invention]

[0006] In response to the above deficiencies or needs for improvements in the prior art, the present application provides a method for developing a basal medium, a method and a system for developing a basal medium composition. [Means for solving the problem]

[0007] One aspect of the present application provides a method for developing a basal medium based on culture index evaluation, comprising the steps of: A step (1) of selecting and optimizing a regression model for the selected culture index using a basal medium composition database verified by experiments, and determining a regression prediction model for predicting a basal medium for the selected culture index; (2) obtaining the range of addition of each component of the basal medium and generating a plurality of candidate basal medium compositions by enumerating and randomly selecting the amount of each component to be added; (3) predicting the culture indexes of the candidate basal medium compositions obtained in step (2) by using the regression prediction model obtained in step (1), and screening one or more compositions from the candidate basal medium compositions as recommended basal medium compositions based on the prediction results; and Step (4) of conducting cell culture experiments using the recommended basal medium composition obtained in step (3), verifying the culture indexes of the recommended basal medium composition, and identifying the optimal basal medium composition based on the results.

[0008] Preferably, in the method for developing a basal medium based on a culture index evaluation, the regression prediction model for predicting a basal medium for the culture index in step (1) is determined according to the following method: (1-1)Collecting training composition: The step of searching within the addition range of each component to form a training composition includes, but is not limited to, three methods: random generation composition, DOE experimental design composition, and mixed formation composition; (1-2) Acquisition of culture indicator data: Conduct a cell culture experiment to obtain culture index data for the training composition obtained in step (1-1); (1-3) Training sample data set composition: A training sample data group is organized by using the amount of each component of the training composition obtained in step (1-1) or its normalized value as an input matrix and the culture index data obtained in step (1-2) as an output matrix; (1-4) For the multiple candidate regression prediction models, training and testing are performed under the same conditions using the training sample data group obtained in step (1-3), and a superior regression prediction model is selected as a model for predicting the basal medium for the culture index.

[0009] The present application provides a method and system for developing a basal medium composition based on artificial intelligence, the purpose of which is to apply a machine learning algorithm to the complex composition optimization process, establish a sample composition database with good quality and sufficient quantity, select an appropriate machine learning algorithm and optimization algorithm, and recommend a basal medium composition that is most likely to have good culture effect within a short period of time, thereby lowering the hurdle for composition development and solving the technical problems of slow development speed and high development costs caused by the complexity of existing basal medium components.

[0010] An aspect of the present application further provides a method for developing a basal medium composition based on artificial intelligence, comprising: (1) constructing a sample composition database: obtaining candidate basal medium composition components, determining a search space for the addition ratio for each component, and searching within the search space for each component to form a basal medium sample composition; collecting the basal medium sample compositions to construct a sample composition database; (2) Acquiring a sample composition culture database: performing experimental verification on the basal medium sample compositions stored in the sample composition database obtained in step (1) according to the development purpose to obtain the culture effect of each basal medium sample composition, and collecting the basal medium sample composition data associated with the culture effect as a sample composition culture database; (3) Using the sample composition culture database obtained in step (2), a machine learning model is trained for the development goal to obtain a basal medium composition culture effect prediction model; (4) employing the basal medium composition culture effect prediction model obtained in step (3) within a search space of the addition ratio of each component in the basal medium composition to be optimized, to perform a culture effect regression prediction against the development target, and selecting a superior one from the recommended basal medium compositions based on the predicted culture effect.

[0011] Another aspect of the present application further provides an artificial intelligence based development system for a basal medium composition, a sample composition generation module for searching within a search space of the addition ratio of each component in the basal medium composition to be optimized to form a basal medium sample composition and constructing a sample composition database; a sample composition culture database in which each basal medium sample composition in the sample composition database and associated culture effect data are stored; a regression model training module for selecting a regression model, employing the basal medium sample compositions and the associated culture effect data stored in the sample composition culture database to perform regression model training, and obtaining and storing a basal medium composition culture effect prediction model; and a composition recommendation module for applying the basal medium composition culture effect prediction model stored in the regression model training module to the basal medium composition culture effect prediction in the search space, and selecting an excellent one from the recommended basal medium compositions. [Brief description of the drawings]

[0013] [Figure 1] FIG. 1 is a schematic diagram showing the flow of a method for developing a basal medium based on culture index evaluation provided in the present application. [Diagram 2] 1 is a flow chart of a method for developing a basal medium based on culture index evaluation provided in the examples of the present application. [Diagram 3] 1 shows cell density plots of batch cultures of three optimal basal medium compositions obtained in the examples of this application. [Figure 4] 1 is a flow chart of the artificial intelligence-based basal medium composition development method provided in the present application. [Diagram 5] This is an evaluation of the accuracy of predicting maximum cell density within 7 days using various machine learning models against a sample composition database using 15-fold cross-validation. [Figure 6] This is an evaluation of the accuracy of predicting cell density on day 5 using various machine learning models against a sample composition database in a 15-fold cross-validation. [Figure 7] 1 is a flow chart of the method for developing the basal medium composition of the examples of the present application. [Figure 8] FIG. 1 shows data comparison between predicted and actual culture results of the basal medium composition recommended for maximum cell density within 7 days in the examples of the present application. [Figure 9] 1 shows cell growth curves of basal medium compositions recommended for maximum cell density within 7 days in the examples of the present application. [Figure 10] FIG. 1 shows data comparison between predicted and actual culture results of the basal medium composition recommended for cell density on day 5 in the examples of the present application. [Figure 11] 1 shows cell growth curves of recommended basal medium compositions for cell density on day 5 in the examples of the present application. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0014] In order to make the purpose, technical solution and advantages of the present application clearer, the present application will be described in more detail with reference to the following examples. It should be understood that the specific examples described in this specification are merely used to interpret the present application, and do not limit the present application. In addition, the technical features of each embodiment of the present application described below can be combined with each other as long as they are not mutually inconsistent.

[0015] The method for developing a basal medium based on culture index evaluation provided in the present application includes steps (1) to (4), as shown in FIG.

[0016] (1) For a selected culture index, a regression model is selected and optimized using a database of basal medium compositions verified through experiments, and a regression prediction model for predicting the basal medium for the culture index is determined.

[0017] The regression prediction model for predicting the basal medium for the culture index is determined according to the following method.

[0018] (1-1)Collecting training composition: The step of searching within the addition range of each component to form a training composition includes, but is not limited to, three methods: randomly generated composition, DOE experimental design composition, and mixed formation composition.

[0019] The randomly generated composition, that is, for each component in the basal medium composition, random values ​​are taken within the range of addition to form a basal medium sample composition.

[0020] The DOE experimental design composition is: Step S1: clustering the minimum addition ratio of each component in the basal medium to obtain multiple addition levels, and classifying each component in the basal medium into functional categories including amino acids, trace metal ions, vitamins, lipids, buffers, etc. according to their functions; Step S2, in which the different doping levels and function categories obtained in step S1 are combined to form DOE experimental factors, and space-filling DOE experimental designs, including ball packing, Latin hypercube, homogeneous and lowest potential methods, are adopted to form basic sample compositions, and it is preferable to use the Latin hypercube as the space-filling DOE experimental design to design the composition.

[0021] The mixed composition is obtained by screening and combining existing basal medium sample compositions to obtain an updated basal medium sample composition, preferably by screening and combining existing basal medium sample compositions according to the following method, verifying the culture effect of the existing basal medium sample compositions, and selecting compositions with relatively high cell viability, relatively high cell density, or relatively high protein expression and mixing them in pairs or adopting three or more compositions and mixing them randomly or in a preset ratio to prepare a new composition;

[0022] (1-2) Acquisition of culture indicator data: Conduct cell culture experiments and obtain culture index data for the training composition obtained in step (1-1).

[0023] (1-3) Training sample data set composition: The amounts of each component of the training composition obtained in step (1-1) added or their normalized values ​​are used as an input matrix, and the culture index obtained in step (1-2) is used as an output matrix to organize a training sample data group.

[0024] In order to ensure the versatility of the data, as many culture indices as possible should be covered when obtaining the culture indices of the training composition, including but not limited to cell viability, cell density, protein expression level, glucose, lactate, and ammonia. Specifically, each set of experimental data is input data (x1, x2, . . ., x n ) and output data (y1, y2, . . . , y m ), where x iis the i-th component of the basal medium composition, which is characterized during model training, validation and testing. y1 represents cell viability, y2 represents cell density, y3 represents protein expression level, ..., y m represents the mth output indicator, which includes but is not limited to parameters such as glucose, lactate, and ammonia. The input matrix of the regression model is denoted by the X matrix, where x ij represents the j-th component of the i-th composition. The output matrix is ​​denoted by the Y matrix. Here, y ij represents the jth output value of the ith composition.

[0025]

number

[0026] Preferably, the components in the composition are optimized by feature selection of the model.

[0027] (1-4) For multiple candidate regression prediction models, training and testing are performed under the same conditions using the training sample data group obtained in step (1-3), and a superior regression prediction model is selected as a model for predicting the basal medium for the culture index.

[0028] The candidate regression prediction models include, but are not limited to, SVR (Support Vector Regression) model, ElasticNet model, Xgboost model, Gradient Boosting Regression model, Logostic Regression model, regression model by multi-layer neural network, regression model by convolutional neural network, and / or regression model by recurrent neural network. Different regression models are implemented in different ways, i.e., the basic mathematical principles are not the same, so different regression models are adopted to predict the basal medium output, and different regression models cannot predict the prediction performance for the basal medium output. A regression prediction model selection step is required to perform optimization and selection. Different regression models are adopted to predict the basal medium output, and the prediction results obtained by various regression prediction models for different culture indicators are different, so the performance of different regression models is compared to select the one with the best performance.

[0029] The step of training and testing using a training sample data group under the same conditions is preferably performed by dividing the training sample data group into training, validation, and testing in a ratio of 8:1:1, and conducting a crossover experiment, in which validation is used to avoid overfitting of data, and testing is used to evaluate the model generalization ability and prediction ability, and the prediction ability can be evaluated by adopting MSE (mean square error), RMSE (root mean square error), and / or R-SQUARED to screen the regression prediction model according to the prediction goal.

[0030] (2) Obtaining the range of addition of each component of the basal medium, enumerating the amount of each component to be added and randomly selecting the amount to be added to generate a plurality of candidate basal medium compositions; the step of enumerating the amount of each component to be added specifically involves forming a large amount of composition data by taking as many values ​​as possible within the range of addition of each component and taking the average value.

[0031] The present application avoids overfitting of the model and combines measures such as enumerating the amounts of components to be added, thereby avoiding the optimization trap in which a locally optimal solution based on experimental experience becomes the globally optimal solution, thereby efficiently obtaining the globally optimal basal medium composition.

[0032] Specifically, the process includes the following steps:

[0033] (2-1) for each component of the basal medium, taking the same number of points within the range of possible values ​​for each component, to form a sequence of possible values ​​for the component, Preferably, the same number of points form an arithmetic or geometric progression, which has better uniformity and covers a better range of possible values ​​respectively; (2-2) randomly sorting the sequence of possible component values ​​obtained in step (2-1) to obtain a sequence of possible component values ​​after rearrangement; (2-3) forming a sequence of possible component values ​​by using all the sequence of possible component values ​​obtained after the rearrangement in step (2-2) as rows or columns, and obtaining a candidate basal medium composition by setting the columns or rows of the sequence of possible component values ​​as possible values ​​for each component of the composition; (2-4) Repeating steps (2-1) to (2-3) to obtain more candidate basal medium compositions, preferably the number of the candidate basal medium compositions is 1,000 to 1,000,000.

[0034] (3) predicting the culture indexes of the candidate basal medium compositions obtained in step (2) using the regression prediction model obtained in step (1), and screening one or more recommended basal medium compositions from the candidate basal medium compositions based on the prediction results.

[0035] (4) Conduct cell culture experiments using the recommended basal medium composition obtained in step (3), verify the culture index of the recommended basal medium composition, and identify the optimal basal medium composition based on the results.

[0036] Regression analysis is a predictive modeling technique that studies the relationship between a dependent variable (target) and an independent variable (predictor). By using regression analysis, a model is first constructed between each component of the medium (independent variable) and protein expression level, cell density, cell viability, etc. (dependent variable) using existing basal medium composition experimental data, and then the model is used to predict the output indicators of a large number of basal medium compositions, that is, the indicators such as protein expression level, viability, and density produced after using the basal medium composition. A large amount of composition data can be formed according to the range of possible values ​​of each component in the composition. While the artificially optimized composition data has a relatively narrow range of possible values, each component in the automatically generated composition data has a wide range of possible values. Then, from the many prediction results, a basal medium composition that matches one or more indicators is selected as a candidate basal medium composition. Finally, the basal medium composition is verified using the candidate basal medium cultured cells. The candidate basal medium composition that matches the screening requirements in the verification experiment is determined as the optimal basal medium composition, which significantly reduces the number of experiments, saves labor and costs, and shortens development time.

[0037] The system for developing a basal medium based on culture index evaluation provided in the present application includes a regression model selection module, a candidate basal medium composition generation module, and a basal medium composition recommendation module; The regression model selection module is used to select and optimize a regression model based on a basal medium composition database verified by experiments, determine a regression prediction model for predicting a basal medium for the culture index, and submit the regression prediction model to the basal medium composition recommendation module; The candidate basal medium composition generating module is used to obtain the addition range of each component of the basal medium, enumerate and randomly select the addition amount of each component to generate a plurality of candidate basal medium compositions, and submit them to the basal medium composition recommendation module, the candidate basal medium composition generating module includes an enumeration submodule, a sorting submodule, and a combination submodule; The enumeration submodule is used to take the same number of points within the range of possible values ​​for each component of the basal medium, form a sequence of possible values ​​for the component, and submit the sequence to the sorting submodule; the sorting submodule is used for randomly sorting the possible value sequence of the components to obtain a sorted possible value sequence of the components, and submitting the sorted possible value sequence of the components to the combination submodule; The combination submodule is used to obtain a candidate basal medium composition by forming a sequence of possible component values ​​as rows or columns of the sequence of possible component values ​​of all the components after the rearrangement, and setting the columns or rows of the sequence of possible component values ​​as possible values ​​of each component of the composition; The basal medium composition recommendation module is used to predict the culture index by employing the regression prediction model for the candidate basal medium compositions, and screen one or more recommended basal medium compositions from the candidate basal medium compositions based on the prediction result.

[0038] The following is an example.

[0039] The method for developing a basal medium based on culture index evaluation includes steps (1) to (4) as shown in FIG.

[0040] (1) For a selected culture index, a regression model is selected and optimized using a database of basal medium compositions verified through experiments, and a regression prediction model for predicting the basal medium for the culture index is determined.

[0041] The regression prediction model for predicting the basal medium for the culture index is determined according to the following method.

[0042] (1-1)Collecting training composition: In this embodiment, computer AI technology is used to compile experimental data generated by cell culture for randomly generated compositions, compositions based on DOE experimental designs, and a large number of compositions formed by mixing two or more types of media in a certain ratio for existing compositions, to form a training composition set.

[0043] The step of searching within the addition range of each component to form a training composition includes, but is not limited to, three methods: randomly generated composition, DOE experimental design composition, and mixed formation composition.

[0044] A specific method for generating the random composition is to search the materials to find out the components and component concentration ranges contained in the medium composition, and then each component randomly takes a value within the concentration range, so that each component takes a random value to form a composition.

[0045] The DOE experimental design comprises the steps of: Step S1: Clustering is performed on the minimum addition ratio of each component in the basal medium to obtain multiple addition levels, and each component in the basal medium is classified into functional categories including amino acids, trace metal ions, vitamins, lipids, buffers, etc. according to their functions. Specifically, Except for minor, constant components in the composition (e.g. glucose), all other components are divided into five broad categories according to other substances such as amino acids, trace metal ions, vitamins, lipids, and buffering agents. Step S2: The different doping levels and functional categories obtained in step S1 are combined to form DOE experimental factors, and a space-filling DOE experimental design is adopted to form a basic sample composition; specifically, In each large category, the maximum addition value of each component is set to 100%, and the minimum value divided by the maximum value is set as the minimum addition % in the composition. The components with the closest minimum addition % are selected to form a new category, and in this way, nine large categories, i.e., nine factors, are formed on the basis of the five large categories, and 90 compositions are designed by adopting space-filling Latin hypercube DOE experiments.

[0046] The mixed composition is obtained by screening and combining existing basal medium sample compositions to obtain an updated basal medium sample composition. Specifically, the existing basal medium sample compositions are screened and combined according to the following method, the culture effect of the existing basal medium sample compositions is verified, and compositions with relatively high cell viability, relatively high cell density, or relatively high protein expression are selected and mixed two by two, or three or more compositions are adopted and mixed randomly or in a preset ratio to prepare a new composition.

[0047] (1-2) Acquisition of culture indicator data: Conduct cell culture experiments to obtain culture index data for the training composition obtained in step (1-1), specifically, In order to ensure the versatility of the data, as many culture indices as possible should be covered when obtaining the culture indices of the training composition, including cell viability, cell density, cell expression level, glucose, lactate, and ammonia, specifically: The large amount of training compositions formed by the three methods in the above invention content were subjected to cell batch culture, the culture vessel was a 50mL mini bioreactor, the culture volume was 10mL, the theoretical inoculation density was 0.5E+06cells / mL, the culture time was 7 days, sampling and counting were performed on the 3rd and 5th days, and the contents of biochemical parameters such as glucose and lactate were detected. On the 7th day, sampling and counting were performed to detect parameters such as glucose, lactate and protein expression, and then the batch culture was terminated.

[0048] (1-3) Training sample data set composition: The amounts of each component of the training composition obtained in step (1-1) added or their normalized values ​​are used as an input matrix, and the culture index obtained in step (1-2) is used as an output matrix to organize a training sample data group.

[0049] Each set of experimental data is input data (x1, x2, . . . , x n ) and output data (y1, y2, . . . , y m), where x i is the i-th component of the basal medium composition, which is characterized during model training, validation and testing. y1 represents cell viability, y2 represents cell density, y3 represents cell expression level, ..., y m represents the mth output indicator, which includes but is not limited to parameters such as glucose, lactate, and ammonia. The input matrix of the regression model is denoted by the X matrix, where x ij represents the j-th component of the i-th composition. The output matrix is ​​denoted by the Y matrix. Here, y ij represents the jth output value of the ith composition.

[0050]

number

[0051] Preprocessing of experimental data: Since the measurement units of each component of the basal medium composition and each value of the output index are different, if they are used directly, it will affect the efficiency of model training and validation, so normalization preprocessing is required for the input and output data.

[0052] In this embodiment, the maximum cell density y2 during a 7-day batch culture process is taken as an example. When specifically realized, it is also possible to model, verify, test and predict the values ​​of different output indicators at different stages.

[0053] The composition data of the basal medium is written into an Excel file, and each row represents information about the basal medium. During model training, testing or composition prediction, the data is read directly from the Excel file. For convenience of processing, the data in the Excel file can also be written into other files such as CSV files, database files, and the corresponding file formats can be formed.

[0054] Feature Selection: Each component of the basal medium composition has a different contribution to the output index. Therefore, in order to improve training efficiency and reduce prediction error, it is necessary to determine which components contribute most to the prediction of the output index. To determine which features contribute most to the prediction of the output value, various features are scored using methods such as correlation feature selection and mutual information feature selection, and the features with the highest scores are selected to perform a prediction experiment. The feature that corresponds to when the prediction result reaches the highest value is the feature with the highest predictive ability of the output value. In concrete terms, after preprocessing the experimental data, each feature is scored using methods such as correlation feature selection and mutual information feature selection. The higher the score of a certain component, the greater the predictive contribution to the output index, and the components are sorted in order of increasing score value to form a feature scoring sequence.

[0055] The results of the feature screening in this example show that the feature values ​​of the example, L-glutamic acid, L-alanine, L-tryptophan, manganese sulfate monohydrate, sodium selenite, cobalt chloride hexahydrate, pyridoxal hydrochloride, sodium pyruvate, HEPES (pH buffer), and sodium bicarbonate, contribute significantly to the maximum cell density in the 7-day batch culture process.

[0056] Select two or three regression models with good performance, and select the components with the highest score value (5, 10, 15, 20,..., all components) from the scoring sequence in different proportions to perform regression model training, verification, and test experiments. Select a 10-fold crossover experiment during the experiment, and evaluate the regression model performance based on the average value of the 10-fold crossover experiment, such as MSE and R2, as regression model indicators. The same applies to other experiments in this embodiment. Then, compare the performance when each model contains different proportions of component data, and select the component with the largest prediction contribution. The component with the largest prediction contribution is the component whose performance does not change when the component is increased in the model, and whose performance decreases when the component is decreased, and the corresponding component sequence is the component set with the best prediction performance.

[0057] (1-4) For the multiple candidate regression prediction models, training and testing are performed under the same conditions using the training sample data group obtained in step (1-3), and a superior regression prediction model is selected as a model for predicting the basal medium for the culture index.

[0058] The candidate regression prediction models include, but are not limited to, a Support Vector Regression (SVR) model, an ElasticNet model, an Xgboost model, a Gradient Boosting Regression model, a Logostic Regression model, a regression model based on a multi-layer neural network, a regression model based on a convolutional neural network, and a regression model based on a recurrent neural network.

[0059] Since the implementation forms of different regression models are different, that is, the basic mathematical principles are not the same, different regression models have different prediction performances for basal medium output. Optimization and selection are required through the regression prediction model selection step. Different regression models are adopted to predict basal medium output, and the prediction results obtained by various regression prediction models for different culture indicators are different, so the performances of the different regression models are compared and the one with the best performance is selected.

[0060] The step of training and testing using a training sample data group under the same conditions divides the training sample data group into training, validation, and testing in a ratio of 8:1:1, and performs cross-sectional experiments, in which validation is used to avoid overfitting of data, and testing is used to evaluate the model generalization ability and prediction ability. Depending on the prediction goal, the prediction ability can be evaluated by adopting MSE (mean square error), RMSE (root mean square error), and / or R-SQUARED to screen the regression prediction model.

[0061] Different regression algorithms can be used to predict the basal medium, but the prediction accuracy is different. To determine which regression algorithm has a higher prediction effect on the basal medium, different regression algorithms are used for training and validation to build a regression model. Different regression models are used to perform composition prediction experiments on the same test data, and the experimental results are compared to determine which regression model has the highest prediction efficiency for the basal medium.

[0062] In this embodiment, the mean square error (MSE), the root mean square error (RMSE), and the R-SQUARED index are adopted to perform comprehensive screening, and the regression model based on the convolutional neural network is screened as the model with the highest effect.

[0063] (2) Obtain the range of each component of the basal medium, enumerate the amount of each component to be added, and randomly select a plurality of candidate basal medium compositions. The step of enumerating the amount of each component to be added is specifically to form a large amount of composition data by taking as many values ​​as possible and averaging them within the range of each component to be added. The specific method is as follows: First step: define an empty matrix to store the formed composition data, and name it PF. Second step: Form a sequence of length M based on the range of possible values ​​of each component of the basal medium composition. Any type of sequence can be formed according to the actual situation. This patent forms experimental data for arithmetic progression. Users can form an arithmetic progression of length m and common difference d. For example, the range of possible values ​​of xi is (x i_min ,x i_max ), and the arithmetic sequence {x 1_min ,x 1_min +d,x 1_min +i*d,......,x i_max}, where x i_min is the minimum value of the range of possible values, and x i_max is the maximum value in the range of possible values. By processing the range of possible values ​​of all components using the above method, we obtain the following arithmetic progression matrix.

number

number

[0064] By adjusting the length m of the arithmetic progression and the number of cycles k through the above steps, a large number of candidate basal medium compositions can be obtained.

[0065] (3) predicting the culture indices for the candidate basal medium compositions obtained in step (2) using the regression prediction model obtained in step (1), and screening one or more compositions from the candidate basal medium compositions as recommended basal medium compositions based on the prediction results.

[0066] in particular, First step: predict the composition of the resulting candidate basal medium using a regression model. Second step: From the prediction result, the condition {y1 ≧ y 1_optimal ∩y2≧y 2_optimal ∩y3…≧y m_optimal} are simultaneously selected as candidate compositions. 1_optimal is the lowest optimal value of y1, and y 2_optimal is the lowest optimal value of y2, and y m_optimal Yes m is the lowest optimal value.

[0067] For example, based on the range of each component in the basal medium composition, an arithmetic progression of length M is formed for each component, and if the basal medium contains five components, the ranges of their possible values ​​are shown in Table 1 below.

[0068] [Table 1]

[0069] An arithmetic progression is formed based on the range of possible values ​​of each component of the basal medium composition, and in this example, an arithmetic progression of length 5 (i.e., M=5) is formed, and the results are shown in Table 2 below. The experimenter can form an arithmetic progression of length M according to the actual situation.

[0070] [Table 2]

[0071] The data in the arithmetic progression matrix is ​​randomly sorted to form a new matrix, as shown in Table 3 below.

[0072] [Table 3]

[0073] The randomly sorted matrix is ​​transposed to obtain M sets of new compositions, where M is the length of the arithmetic progression, and in this example, the length of the arithmetic progression is 5, so 5 sets of new compositions are obtained, as shown in Table 4 below.

[0074] [Table 4]

[0075] The purpose of randomly sorting and transposing an arithmetic progression matrix is ​​to form different compositions, and more compositions can be formed by randomly sorting and transposing it several times during use. In this example, the compositions formed after randomly sorting and transposing it twice are shown in Table 5 below.

[0076] [Table 5]

[0077] The underlined parts in Table 5 are the compositions formed after the second random sort and transposition. After K random sorts and transpositions, K*M compositions can be formed, where M is the length of the arithmetic progression.

[0078] After obtaining a large amount of basal medium compositions by the above method, the optimal composition prediction model is used to predict each output index of all compositions. In this example, a total of 900,000 sets of basal medium compositions are formed, and the composition prediction model is used to predict the output index-cell density. After sorting the prediction results from top to bottom, the top 10 compositions are selected and recommended as the recommended basal medium compositions. The prediction results are shown in Table 6 below (only 10 component concentrations are listed in the table).

[0079] [Table 6]

[0080] (4) Conduct cell culture experiments using the recommended basal medium composition obtained in step (3) to verify the culture index of the recommended basal medium composition, and identify the optimal basal medium composition based on the results.

[0081] Specifically, a cell culture experiment is carried out using the candidate compositions, and one or more compositions with the best Y1, Y2, ..., Ym values ​​are selected from the experimental results as the optimal basal medium composition.

[0082] In this example, a medium was prepared according to the 10 optimal compositions recommended by AI, and cell culture experiments were carried out to verify the optimal compositions. The experimental results are shown in Table 7 below.

[0083] [Table 7]

[0084] The actual and predicted cell densities of the 10 optimal predicted compositions were compared, and it was found that the predicted values ​​of compositions numbered 3, 6, and 9 were close to the actual values, and the cell densities were very high, meeting the requirements. Therefore, these three medium compositions were selected as the optimal basal medium compositions, and the final optimal compositions and culture results are shown in Table 8 below.

[0085] [Table 8]

[0086] As can be easily understood by those skilled in the art, the above are only preferred embodiments of the present application, and do not limit the present application. Any amendments, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

[0087] The method for developing a basal medium composition based on artificial intelligence provided in the present application includes steps (1) to (4), as shown in FIG.

[0088] (1) Construction of a sample composition database: For each component in the basal medium composition to be developed, a search space for its addition rate is determined, and a basal medium sample composition is formed by searching within the search space of each component. The basal medium sample compositions are collected to construct a sample composition database, and the addition rate of the component is the ratio of the addition value to the maximum addition value of the component, the search space ranging from the minimum addition rate to 100%, and the minimum addition rate is the ratio of the minimum and maximum addition values ​​of the component.

[0089] The step of searching within the search space of each component to form a training composition includes, but is not limited to, four methods: randomly generated composition, DOE experimental design composition, mixture-formed composition, and historical AI recommended composition.

[0090] The randomly generated composition, ie, for each component in the basal medium composition, randomly takes values ​​within the search space to form the basal medium sample composition.

[0091] The DOE experimental design comprises the following steps: Step S1: clustering the minimum addition ratio of each component in the basal medium to obtain multiple addition levels, and classifying each component in the basal medium into functional categories including amino acids, trace metal ions, vitamins, lipids, buffers, etc. according to their functions; In step S2, it is preferable to combine the different doping levels and function categories obtained in step S1 to form DOE experimental factors, adopt space-filling DOE experimental designs including ball packing, Latin hypercube, homogeneous and lowest potential methods to form basic sample compositions, and use Latin hypercube as the space-filling DOE experimental design to design the compositions.

[0092] The mixed composition is, that is, an updated basal medium sample composition is obtained by screening and combining existing basal medium sample compositions. Preferably, the existing basal medium sample compositions are screened and combined according to the following method, the culture effect of the existing basal medium sample compositions is verified, and compositions with relatively high cell viability, relatively high cell density, or relatively high protein expression are selected and mixed two by two, or three or more types of compositions are adopted and mixed randomly or in a preset ratio to prepare a new composition.

[0093] The historical AI recommended compositions include basal medium compositions developed based on artificial intelligence according to the composition development method of the present application.

[0094] The quality and number of basal medium sample compositions are important requirements and prerequisites for artificial intelligence to optimize the basal medium optimization effect. In order to train a more accurate model by machine learning, the number of basal medium sample compositions should reach more than 100, and the basal medium sample compositions should cover known and unknown different dimensions and different effects, but have basal medium sample compositions that are more distributed in the high-dimensional space area with good effect and in the high-dimensional space area with obvious changes. Therefore, in order to increase the number and select basal medium sample compositions more densely in the high-dimensional space area with good effect and in the high-dimensional space area with obvious changes, the present application gives more priority to the DOE experimental design composition, and increases the random generation of basal medium sample compositions and supports the method of mixed formation composition to increase the coverage of sample compositions with different dimensions and different effects. Experiments have shown that the sample composition database constructed by combining the above three methods has a better machine learning model training effect.

[0095] Preferably, the ratio of the number of randomly generated compositions to the number of DOE experimentally designed compositions in the sample composition database is 1-4:10, the randomly generated compositions can ensure the generalization ability of the method, and the DOE experimentally designed compositions can better improve the prediction accuracy.

[0096] The total number of samples in the sample composition database is more than 100, that is, the artificial intelligence based method for developing a basic medium composition provided in this application can be realized. Preferably, the total number of samples in the sample composition database is more than 1000, including 100-200 randomly generated compositions and 50-200 DOE experimental design compositions, thereby controlling the time and cost involved in medium preparation. The historical AI recommended compositions are always prepared and verified for culture effect during the continuous experiment process, so there is no need to increase the costs of medium preparation and verification of culture effect. The remaining amount is a mixed medium, which is mixed using a medium that has already been prepared, thereby greatly reducing the time and cost involved in medium preparation.

[0097] (2) Acquisition of a sample composition culture database: The step of performing experimental verification according to the optimization purpose for the basal medium sample compositions stored in the sample composition database obtained in step (1) to obtain the culture effect of each basal medium sample composition, collecting the basal medium sample composition data associated with the culture effect as a sample composition culture database, and performing experimental verification according to the optimization purpose to obtain the culture effect of each basal medium sample composition is specifically as follows: The basal medium sample composition is used to culture target cells, and the culture process involves sampling and detecting cell states over time, the cell states including cell viability, cell density, and / or biochemical indicators, the biochemical indicators being protein expression levels, glucose, lactate, ammonia, and / or glutamine content, and a cell viability curve for the culture time of the basal medium sample composition can be obtained by fitting to the cell viability, and a cell growth curve for the culture time of the basal medium sample composition can be obtained by fitting to the cell density, and the culture effect of the basal medium sample composition can be a cell growth curve, a cell viability curve, or a cell density, cell viability, or biochemical indicator at a specific time point for the culture time of the basal medium, and may be the above multiple indicators (correspondingly adopting a machine learning model of multi-objective optimization), or a comprehensive index of the above multiple indicators (for example, a weighted sum of a specific multiple indicators).

[0098] (3) Using the sample composition culture database obtained in step (2), a machine learning model is trained for the optimization goal to obtain a basal medium composition culture effect prediction model; The machine learning model includes, but is not limited to, a support vector machine regression model, a K-nearest neighbor model, XGBoost, ridge regression, LightGBM, random forest, GBDT, or a deep learning model, and the deep learning model includes, but is not limited to, a fully connected neural network, a convolutional neural network, or a recurrent neural network, among which, the support vector machine regression model has a better effect. Through 15-fold cross-validation, as shown in Figures 5 and 6, it can be seen that the support vector machine regression model has a better model representation, and the model is a continuously differentiable model, which also has great advantages in later optimal composition recommendation. The remaining machine learning models, such as the K-nearest neighbor model and the ridge regression model, have relatively low accuracy in the test set, and the tree-based machine learning models such as XGBoost, LightGBM, random forest, and GBDT are not continuously differentiable models. Deep learning models include fully connected neural networks, convolutional neural networks, recurrent neural networks, etc., where convolutional neural networks are suitable for translationally invariant data such as image processing, and recurrent neural networks are suitable for sequence-type data such as speech-to-text processing, and have no particular advantage in predicting optimal compositions. Deep learning models also require large amounts of data and are more expensive. Therefore, overall, the support vector machine regression model has the lowest mean square error and is the preferred model.

[0099] (4) Within the search space of the addition ratio of each component in the basal medium composition to be optimized, the basal medium composition culture effect prediction model obtained in step (3) is used to perform a culture effect regression prediction for the optimization target, and a superior one is selected from the recommended basal medium compositions based on the predicted culture effect. Specifically, Preferably, a global optimization algorithm or a heuristic algorithm is adopted to search for the basal medium composition in the search space to perform the culture effect regression prediction, and the heuristic algorithm includes, but is not limited to, a genetic algorithm, a greedy algorithm, an annealing algorithm, an ant swarm algorithm, a particle swarm algorithm, an artificial bee colony algorithm, an artificial fish swarm algorithm, a shuffled frog jumping algorithm, a fireworks algorithm, a bacterial foraging optimization algorithm, and a firefly algorithm, and the global optimization algorithm includes, but is not limited to, a Newton method, a quasi-Newton method, a conjugate gradient method, and a gradient descent method commonly used in deep learning, and common gradient descent methods include SGD, Momentum, Adagrad, RMSprop, Adam, Nadam, etc.

[0100] In terms of space complexity, most heuristic algorithms, such as group optimization algorithms such as genetic algorithms, ant swarm algorithms, particle swarm algorithms, artificial bee colony algorithms, artificial fish swarm algorithms, shuffled frog jumping algorithms, fireworks algorithms, bacterial foraging optimization algorithms, and firefly algorithms, need to occupy a large amount of cache and require a large amount of calculation. Meanwhile, it is not easy for greedy algorithms to achieve the overall optimum value in problems where multiple factors interact with each other, such as recommending a composition of various component contents. Meanwhile, annealing algorithms only use increments as an evaluation index, and are inferior to gradient descent, which uses gradients as an evaluation index. In addition, algorithms such as Newton's method, quasi-Newton's method, and conjugate gradient method use second-order gradient approximation, which is theoretically superior to the first-order gradient approximation of gradient descent, but due to the complex composition, such methods place a greater burden on computational power. Therefore, various variants of gradient descent are preferred.

[0101] The artificial intelligence-based basal medium composition development system provided in the present application comprises: a sample composition generation module for searching within a search space of the addition ratio of each component in the basal medium composition to be optimized to form a basal medium sample composition and constructing a sample composition database; a sample composition culture database in which each basal medium sample composition in the sample composition database and associated culture effect data are stored; a regression model training module for selecting a regression model, employing the basal medium sample compositions and the associated culture effect data stored in the sample composition culture database to perform regression model training, and obtaining and storing a basal medium composition culture effect prediction model; and a composition recommendation module for applying the basal medium composition culture effect prediction model stored in the regression model training module to the basal medium composition culture effect prediction in the search space, and selecting an excellent one from the recommended basal medium compositions.

[0102] The following is an example.

[0103] The method for developing a basal medium composition based on artificial intelligence includes the following steps (1) to (4) using a small amount of compositional components as an example.

[0104] (1) Construction of a sample composition database: L-tryptophan, L-cysteine, L-glycine, L-alanine, manganese sulfate monohydrate, cobalt chloride hexahydrate, pyridoxal hydrochloride, ethanolamine, sodium bicarbonate, poloxamer 188. Determine the search space of the addition rate of each component, and search within the search space for each component to form a basal medium sample composition. Collect the basal medium sample compositions to construct a sample composition database, where the addition rate of the component is the ratio of the addition value to the maximum addition value of the component, and the search space is from the minimum addition rate to 100%, and the minimum addition rate is the ratio of the minimum addition value to the maximum addition value of the component.

[0105] The step of searching within the search space of each component to form a training composition includes four methods: randomly generated composition, DOE experimental design composition, mixture formation composition, and historical AI recommended composition.

[0106] The randomly generated composition, ie, for each component in the basal medium composition, randomly takes values ​​within the search space to form the basal medium sample composition.

[0107] This embodiment adopts the above DOE experimental design composition, specifically, Except for the small amount of constant components in the composition (e.g. glucose), all other components are divided into five major categories according to other substances such as amino acids, trace metal ions, vitamins, lipids, buffers, etc. In each major category, the maximum addition value of each component is 100%, and the minimum addition % in the composition is determined by dividing the minimum value by the maximum value. Components with the closest minimum addition % are selected to form a new category, and nine major categories, i.e. nine factors, are formed on the basis of the five major categories, and the Latin hypercube method in space-filling DOE experimental design is adopted to design 90 compositions.

[0108] Specifically, the mixed composition used in this embodiment is prepared by verifying the culture effect of existing basal medium sample compositions, selecting compositions with relatively high cell viability, relatively high cell density, or relatively high protein expression, and mixing two of them together or using three or more types of compositions and mixing them in a random ratio to prepare a new composition.

[0109] The historical AI recommended compositions include basal medium compositions that have been optimized based on artificial intelligence according to the composition optimization method of the present application.

[0110] The sample composition database finally constructed in this embodiment contains 1,000 to 1,500 basal medium compositions, of which 90 DOE experimental design compositions, 200 random compositions, and 100 to 200 historical AI recommended compositions, and the remaining 500 to 700 mixed compositions.

[0111] (2) Acquisition of a sample composition culture database: The basal medium sample compositions stored in the sample composition database obtained in step (1) are subjected to experimental verification according to the optimization purpose to obtain the culture effect of each basal medium sample composition, and the basal medium sample composition data associated with the culture effect is collected as a sample composition culture database. In this embodiment, the step of performing experimental verification according to the optimization purpose to obtain the culture effect of each basal medium sample composition is specifically as follows: The batch culture method was adopted, and the batch culture method was 0.5 × 10 6 The cells were inoculated at a cell density of 1000000000000000 cells / mL, the culture volume was 10 mL, the culture vessel was a 50 mL mini bioreactor, the rocker rotation speed was 180 rpm, and the culture time was 7 days. During the culture, samples were taken on the 3rd, 5th, and 7th days, respectively, to count the cell density and detect biochemical parameters such as glucose, lactate, ammonia, glutamine, and protein expression levels. Glucose was replenished up to 4-5 g / L according to the glucose consumption status. Sampling was performed daily to obtain complete test data.

[0112] When AI builds a model, it can perform single-objective modeling or multi-objective modeling using only the cell viability or cell density or biochemical parameters of a certain sampling point, and can also employ cell viability plots or cell density plots to build a regression model.

[0113] In this example, data is obtained for the cell density at each sampling point, the maximum cell density within 7 days, and a cell growth plot plotted using the cell density data.

[0114] The content of each component in the composition, data on the medium preparation process, and related information such as the culture effect are recorded and stored in a culture database.

[0115] (3) The sample composition culture database obtained in step (2) is used to train a regression model to obtain a basal medium composition culture effect prediction model, specifically, Using the python language, the support vector machine regression model was loaded, and the RBF (Gaussian) kernel function was adopted to perform 15-fold cross-validation. The composition culture effect prediction model 1 was obtained for the maximum cell density within 7 days, with the mean square error being approximately 0.39 with two decimal places; the composition culture medium effect prediction model 2 was obtained for the cell density on the 5th day, with the mean square error being approximately 0.41 with two decimal places. It is expected that the model can perfectly predict the yield and quality of cultured cells under different contents of each component of the culture medium composition.

[0116] If the training result does not meet the criteria, repeat steps (1)-(2) to increase the amount of training sample data.

[0117] (4) Within the search space of the addition ratio of each component in the basal medium composition to be optimized, the basal medium composition culture effect prediction model obtained in step (3) is used to perform a culture effect regression prediction for the optimization target, and a superior one is selected from the recommended basal medium compositions based on the predicted culture effect. Based on the maximum cell density within the above 7 days and the cell density on the 5th day, the model information is dug up, and the most likely cell culture effect to appear at different contents of each component is numerically simulated by computer. The simulation calculation adopts the gradient descent method, and the gradient is the change effect on the cell culture effect caused by increasing each component by one unit at a certain content, and this gradient is divided into two types: positive and negative, and the positive gradient indicates that increasing the content of this component is favorable to the cell culture effect, and the negative gradient indicates that increasing the content of this component is harmful to the cell culture effect. Based on the above simulation results, the component content is gradually corrected (when the gradient is positive, the component content is increased, and when the gradient is negative, the component content is decreased, and the content value of the increased or decreased component is proportional to the gradient value), and the simulation calculation is further repeated, and the components are adjusted according to the gradient. The above process is repeated until it is found through numerical simulation that the gradient is close to 0 and the culture effect of the cell simulation cannot be improved by correcting the component content, and this medium composition is the optimal composition in the model simulation. Algorithmic forms of the gradient descent method include SGD, Momentum, Adagrad, RMSprop, Adam, etc.

[0118] By successfully constructing the above machine learning model, it is possible to successfully measure the effect of each component on the cell culture effect. If the machine learning model fails to converge during training or has low accuracy, poor generalization ability, or other problems occur, it is determined that the data is insufficient, and steps (1) to (3) are repeated to continue random generation, DOE composition design, or mixed composition to form more basal medium sample compositions, expand the data, and optimize the machine learning model.

[0119] The recommended optimal composition obtained by the numerical simulation after the above machine learning is detailed, and the content of each component is clarified, and the medium can be prepared according to the composition to carry out a batch culture experiment. If the culture effect of the cell culture does not meet the experimental requirements, step (4) is repeated. After the model is constructed by the above machine learning, the recommended composition after adjusting the components is tested, and the basal medium composition is generated for each maximum cell density within 7 days, and the cells are inoculated into the medium and then batch culture is carried out. The cell viability remains stable, the 7-day mortality rate is 0%, the cell density is extremely high, the model prediction value is close to the actual value, and the predicted composition is reliable. The specific composition data is shown in the table below, and the comparison between the predicted value and the actual value data of the cultured cells is shown in Figure 8, and the cell growth curve is shown in Figure 9.

[0120] [Table 9]

[0121] For the cell density on the 5th day, a basal medium composition is prepared, and the cells are inoculated into the medium and then batch culture is performed. The cell viability remains stable, the cell death rate is 0% after 7 days, the cell density is extremely high, the model predicted value is close to the actual value, and the predicted composition is reliable. The specific composition data is shown in the table below. The comparison between the predicted value and the actual value data of the cultured cells is shown in Figure 10, and the cell growth curve is shown in Figure 11.

[0122] [Table 10]

[0123] The composition optimization cycle in this embodiment, including the construction of a sample composition culture database (more than 1,000 compositions) and machine learning model training, is approximately 5 months. When in use, the learning model is trained, compositions are recommended, and effects are verified. In just half a month, compositions can be developed for the culture effects already contained in the database, significantly shortening the development cycle of the basal medium and lowering the hurdle for developing a basal medium.

[0124] As can be easily understood by those skilled in the art, the above are only preferred embodiments of the present application, and do not limit the present application. Any amendments, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A method for developing a basal medium, comprising the steps of: A step of selecting and optimizing a regression model for the selected culture index using a basal medium composition database verified by experiments, and determining a regression prediction model for predicting a basal medium for the selected culture index; obtaining a range of addition of each component of the basal medium and generating a plurality of candidate basal medium compositions by enumerating and randomly selecting the amount of addition of each component; Employing the regression prediction model to predict the culture index for the obtained candidate basal medium compositions, and screening one or more compositions from the candidate basal medium compositions as recommended basal medium compositions based on the prediction result; A step of performing a cell culture experiment using the obtained recommended basal medium composition, verifying the culture index of the recommended basal medium composition, and determining an optimal basal medium composition based on the verified culture index; The step of obtaining the range of addition of each component of the basal medium and generating a plurality of candidate basal medium compositions by enumerating and randomly selecting the amount of each component to be added, obtaining an equal number of points within the range of possible values ​​for each component of the basal medium to form a sequence of possible values ​​for each component; Sorting all possible component value sequences to obtain a permuted possible component value sequence; constructing a sequence of possible component values ​​using the rearranged sequence of possible component values ​​as rows or columns, setting the columns or rows of the sequence of possible component values ​​as possible values ​​for each component, and obtaining a candidate basal medium composition.

2. The step of determining a regression prediction model for predicting a basal medium for the culture indicator comprises: Searching within the range of addition of each component of the basal medium to form a training composition; Conducting a cell culture experiment to obtain culture index data of the training composition; constructing a training sample data group using the amount of each component added in the training composition or its normalized value as an input matrix and the culture index data as an output matrix; The method for developing a basal medium according to claim 1, further comprising the steps of: training and testing a plurality of candidate regression prediction models under the same conditions using the training sample data group; and selecting one of the plurality of candidate regression prediction models that have been trained and tested as a regression prediction model for predicting the basal medium for the culture indicator.

3. The method for developing a basal medium according to claim 2, wherein the culture indicators include cell viability, cell density, protein expression level, and the contents of biochemical parameters of glucose, lactate, and ammonia after a certain culture period.

4. The step of constructing a training sample data group by using the amount of each component added in the training composition or a normalized value thereof as an input matrix and the culture index data as an output matrix, Input data (x 1 , x 2 , ..., x n ) and output data (y 1 , y 2 ,... ,y m ) is constructed as a set of experimental data, where x i is the i-th component of the basal medium composition, characterized during model training, validation and testing, and y 1 represents cell viability, and y 2 represents the cell density, and y 3 represents the protein expression level, and y m represents the mth output index, and the input matrix of the regression model is the X matrix, where x ij represents the jth component of the ith composition, and the output matrix is ​​the Y matrix, where y ij The method for developing a basal medium according to claim 2, further comprising the step of: where j represents the jth output value of the i-th composition. [0010]

5. The step of constructing a training sample data group by using the amount of each component added in the training composition or a normalized value thereof as an input matrix and the culture index data as an output matrix, The method for developing a basal medium according to claim 2, further comprising optimizing components in the training composition by feature selection of a regression model.

6. The method for developing a basal medium according to claim 2, wherein the plurality of candidate regression prediction models include a support vector regression model, an elastic network model, an Xgboost model, a gradient boosting regression model, a logistic regression model, a regression model based on a multilayer neural network, a regression model based on a convolutional neural network, and a regression model based on a recurrent neural network.

7. The method for developing a basal medium according to claim 1, wherein the number of candidate basal medium compositions is between 1,000 and 1,000,000.

8. 1. A basal medium development system comprising: A regression model selection module for selecting and optimizing a regression model based on a basal medium composition database verified by experiments for the selected culture index, and determining a regression prediction model for predicting the basal medium for the selected culture index; a candidate basal medium composition generation module for obtaining the range of addition of each component of the basal medium and generating a plurality of candidate basal medium compositions by enumerating and randomly selecting the amount of addition of each component; a basal medium composition recommendation module for predicting the culture index by employing the regression prediction model for the obtained candidate basal medium compositions, and screening one or more compositions from the candidate basal medium compositions as a recommended basal medium composition based on the prediction result; and an optimal basal medium composition determination module for performing a cell culture experiment using the obtained recommended basal medium composition, verifying the culture index of the recommended basal medium composition, and determining an optimal basal medium composition based on the verified culture index; The candidate basal medium composition generation module comprises: an enumeration submodule for obtaining an equal number of points within the range of possible values ​​for each component of the basal medium to form a sequence of possible values ​​for each component; a sorting submodule for sorting all the component possible value sequences to obtain a sorted component possible value sequence; A combination submodule for constructing a sequence of possible component values ​​using the rearranged sequence of possible component values ​​as rows or columns, setting the columns or rows of the sequence of possible component values ​​as possible values ​​for each component, and obtaining a candidate basal medium composition.

Citation Information

Patent Citations

  • Culture medium optimization method

    CN106096077A

  • Experiment design method for culture medium optimizing by use of artificial neural network method

    CN106503856A

  • Method for determining the optimal cell culture medium composition

    JP2014503220A