Cell culture state prediction method, neural network model training method and device

The prediction of cell culture status in bioreactors through neural network models is solved, and the prediction difficulties in the prior art are achieved, and more efficient biopharmaceutical production process and product consistency are achieved.

CN118969097APending Publication Date: 2024-11-15WUXI BIOLOGICS (SHANGHAI) CO LTD
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410994240.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-23
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The prior art is difficult to accurately predict the next or final state of cell culture in a bioreactor, resulting in a large amount of manpower and material resources required to conduct repeated experiments in the production process of biopharmaceuticals to optimize the culture benefits.

Method used

Using a neural network model, by obtaining the data of the bioreactor in the culture stage, the model is trained to predict the state of the future culture stage, considering historical states and operational actions, and decomposing them into multiple sub-problems to accurately predict the future state.

Benefits of technology

It greatly shortens the development and optimization time of biopharmaceutical manufacturing process, saves manpower and material resources, ensures product quality and consistency, and improves prediction accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118969097A_ABST
    Figure CN118969097A_ABST
Patent Text Reader

Abstract

The invention relates to a cell culture state prediction method and a neural network model training method and device. A method for predicting the status of a cell culture in a bioreactor, comprising: obtaining culture data for the bioreactor for each culture stage of a series of culture stages; and providing the acquired culture data to the trained neural network model to obtain a state of the bioreactor in a future culture phase following the series of culture phases. A loss function of the neural network model is determined based on losses generated by a set of sub-problems corresponding to training data of a series of training stages from a training data set, each sub-problem in the set of sub-problems is configured to predict a state of a respective one of the series of cultivation stages using cultivation data from each cultivation stage preceding the respective one of the series of cultivation stages of the training data set.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present disclosure generally relates to the field of biological cell culture technology, and more specifically, to a method and apparatus for predicting the state of cell culture in a bioreactor, a method and apparatus for training a neural network model for predicting the state of cell culture in a bioreactor, and related equipment, non-transitory storage media, and computer program products. Background Art

[0002] At present, the main production method of biological drugs is cell culture in bioreactors. Since there are many state parameters of cell culture and the changes are nonlinear, and the cell culture cycle is long, it is difficult to predict the next state or final state of cell culture based on the current state or initial state of cell culture and the operation of the bioreactor. Usually, biopharmaceutical companies need to invest a lot of manpower and material resources to conduct repeated experimental attempts on specific biological drugs, and use the final experimental results to select the operating methods and culture processes that can maximize the culture benefits. Summary of the invention

[0003] A brief overview of the disclosure is given below in order to provide a basic understanding of some aspects of the disclosure. However, it should be understood that this overview is not an exhaustive overview of the disclosure. It is not intended to identify the key or important parts of the disclosure, nor is it intended to limit the scope of the disclosure. Its purpose is simply to give some concepts of the disclosure in a simplified form as a prelude to a more detailed description given later.

[0004] According to a first aspect of the present disclosure, a method for predicting the state of cell culture in a bioreactor is provided. The method comprises: acquiring culture data of each culture stage in a series of culture stages of the bioreactor; and providing the acquired culture data to a trained neural network model to obtain the state of the bioreactor in a future culture stage after the series of culture stages. The loss function of the neural network model is determined based on the loss generated by a set of subproblems, the set of subproblems corresponding to the culture data of the series of culture stages from a training data set, and each subproblem in the set of subproblems is configured to predict the state of a corresponding culture stage using the culture data of each culture stage before the corresponding culture stage in the series of culture stages from the training data set.

[0005] According to a second aspect of the present disclosure, a method for training a neural network model for predicting the state of cell culture in a bioreactor is provided. The method comprises: obtaining a training data set, the training data set comprising culture data of each culture stage in each culture stage series in a plurality of culture stage series of a plurality of bioreactors; and training a neural network model using the training data set. The loss function of the neural network model is determined based on the loss generated by a set of sub-problems, the set of sub-problems corresponding to the culture data of the culture stage series from the training data set, and each sub-problem in the set of sub-problems is configured to predict the state of a corresponding culture stage using the culture data of each culture stage before the corresponding culture stage in the culture stage series from the training data set.

[0006] According to a third aspect of the present disclosure, a device for predicting the state of cell culture in a bioreactor is provided. The device includes: an acquisition unit configured to acquire culture data of each culture stage in a series of culture stages of the bioreactor; and a prediction unit configured to provide the culture data acquired by the acquisition unit to a trained neural network model to obtain the state of the bioreactor in a future culture stage after the series of culture stages. The loss function of the neural network model is determined based on the loss generated by a set of subproblems, the set of subproblems corresponding to the culture data of the series of culture stages from a training data set, and each subproblem in the set of subproblems is configured to predict the state of a corresponding culture stage using the culture data of each culture stage before a corresponding culture stage in the series of culture stages from the training data set.

[0007] According to a fourth aspect of the present disclosure, there is provided a device for training a neural network model for predicting the state of cell culture in a bioreactor. The device comprises: an acquisition unit configured to acquire a training data set, the training data set comprising culture data of each culture stage in each culture stage series in a plurality of culture stage series of a plurality of bioreactors; and a training unit configured to train the neural network model using the training data set acquired by the acquisition unit. The loss function of the neural network model is determined based on the loss generated by a set of sub-problems, the set of sub-problems corresponding to the culture data of the culture stage series from the training data set, and each sub-problem in the set of sub-problems is configured to predict the state of a corresponding culture stage using the culture data of each culture stage before a corresponding culture stage in the culture stage series from the training data set.

[0008] According to a fifth aspect of the present disclosure, a computing device is provided, comprising: one or more processors; and a memory storing computer-executable instructions, wherein when the computer-executable instructions are executed by the one or more processors, the one or more processors execute the method according to the first aspect or the second aspect of the present disclosure.

[0009] According to a sixth aspect of the present disclosure, a non-volatile storage medium having computer executable instructions stored thereon is provided, wherein the computer executable instructions, when executed by one or more processors, cause the one or more processors to execute the method according to the first aspect or the second aspect of the present disclosure.

[0010] According to a seventh aspect of the present disclosure, a computer program product is provided, wherein the computer program product comprises instructions, and when the instructions are executed by a processor, the method according to the first aspect or the second aspect of the present disclosure is implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The foregoing and other features and advantages of the present disclosure will become apparent from the following description of the embodiments of the present disclosure in conjunction with the accompanying drawings. The accompanying drawings are incorporated herein and form a part of the specification, and are further used to explain the principles of the present disclosure and enable those skilled in the art to make and use the present disclosure. Among them:

[0012] Figure 1 A flow chart showing a method for predicting the status of a cell culture in a bioreactor according to some embodiments of the present disclosure;

[0013] Figure 2 A flow chart showing a non-limiting example process in which a method for predicting the state of a cell culture in a bioreactor according to some embodiments of the present disclosure is applied;

[0014] Figure 3A shows a schematic structural block diagram of a neural network model for predicting the state of cell culture in a bioreactor according to some embodiments of the present disclosure, Figure 3B and Figure 3C Schematic structural block diagrams of generative pre-trained transformer modules of neural network models according to some embodiments of the present disclosure are respectively shown;

[0015] Figure 4 A flowchart showing a method for training a neural network model for predicting the state of cell culture in a bioreactor according to some embodiments of the present disclosure;

[0016] Figure 5A and Figure 5B A graph showing a cell culture state predicted by the method for predicting the state of cell culture in a bioreactor according to some embodiments of the present disclosure and an actual cell culture state;

[0017] Figure 6 A schematic block diagram of an apparatus for predicting the state of cell culture in a bioreactor according to some embodiments of the present disclosure is shown;

[0018] Figure 7 A schematic block diagram showing an apparatus for training a neural network model for predicting the state of cell culture in a bioreactor according to some embodiments of the present disclosure;

[0019] Figure 8 A schematic block diagram of a computing device according to some embodiments of the present disclosure is shown;

[0020] Fig. 9 A schematic block diagram of a computer system is shown on which embodiments of the present disclosure may be implemented.

[0021] Note that in the embodiments described below, sometimes the same reference numerals are used in common between different drawings to represent the same parts or parts with the same functions, and their repeated descriptions are omitted. In some cases, similar numbers and letters are used to represent similar items, so once an item is defined in one drawing, it does not need to be further discussed in subsequent drawings.

[0022] For ease of understanding, the position, size, range, etc. of each structure shown in the drawings and the like may not represent the actual position, size, range, etc. Therefore, the present disclosure is not limited to the position, size, range, etc. disclosed in the drawings and the like. DETAILED DESCRIPTION

[0023] Various exemplary embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings. It should be noted that unless otherwise specifically stated, the relative arrangement of components and steps, numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present disclosure.

[0024] The following description of at least one exemplary embodiment is in fact merely illustrative and is in no way intended to limit the present disclosure and its application or use. That is, the structures and methods herein are shown in an exemplary manner to illustrate different embodiments of the structures and methods in the present disclosure. However, those skilled in the art will appreciate that they merely illustrate exemplary ways of the present disclosure that can be implemented, rather than exhaustive ways. In addition, the drawings need not be drawn to scale, and some features may be enlarged to illustrate the details of specific components.

[0025] In addition, technologies, methods, and equipment known to ordinary technicians in the relevant art may not be discussed in detail, but where appropriate, the technologies, methods, and equipment should be considered as part of the specification.

[0026] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0027] The development of biological drugs requires a large number of biological reaction experiments, and the most reliable cell culture operation process is screened out through the experimental results, so as to ensure the stability, high yield and high quality of the long-term production process of biological drugs. Biological reaction experiments need to study the effects of a wide range of experimental control parameters (such as sugar supplementation, feed supplementation, cooling, ventilation, etc.) on cells and various parameters of culture medium (such as culture medium composition, live cell density, temperature, pH, etc.) at specific times.

[0028] It is very important to predict the correct culture trajectory to help adopt the correct bioreactor operation and culture process at each stage of the culture cycle to achieve the best culture yield. For example, cell culture state prediction methods can include methods based on mechanism models and methods based on traditional machine learning models.

[0029] Mechanistic models often only need to refer to a small number of parameters in several biological reaction experiments. However, the cell system is very complex, and the mathematical fitting of a small number of parameters may not be able to fully describe the material changes in the complex cell system. In addition, the mechanism model requires that the action on the time scale (the operation of the bioreactor) is relatively uniform. Once there is a change, the predictiveness of the mechanism model will deteriorate. In addition, the versatility and room for improvement of the mechanism model are limited. It is impossible to effectively use the existing massive data to improve its own performance, and it is impossible to consider the complexity of the action.

[0030] Traditional machine learning models or their hybrid models can use various parameters of complex cell systems. For example, the state data of each day of the previous rounds of cell culture experiments can be used for model training, so that the trained model can predict the state of the cells on each day of the new round of culture experiments. Specifically, when predicting the state of day t+1, the state and rate of change of day t can be used to predict the state of day t+1. Similarly, when predicting the state of day t+2, the state and rate of change of day t+1 can be used to predict the state of day t+2. However, in reality, considering only one day may not be enough to complete accurate predictions. Traditional machine learning models also do not take into account the effects of bioreactor operations (such as feeding, sugar supplementation, ventilation, cooling, pH adjustment, etc.). Moreover, when using traditional machine learning models, it takes manpower and material resources to generate more than four or five rounds of training data for the same cell sample before using this model to predict the current round.

[0031] Therefore, neither mechanistic models nor traditional machine learning models are well suited to the complex application environment of the biopharmaceutical industry.

[0032] To this end, the present disclosure provides a method for predicting the state of cell culture in a bioreactor, providing the known culture data of each culture stage in the culture stage series (in this article, the culture stage series refers to a series of culture stages along the culture trajectory of the bioreactor) of the bioreactor to a trained neural network model to predict the state of the bioreactor in the future culture stage. In particular, the neural network model of the present disclosure performs conditional state modeling on the problem of predicting the state of cell culture in the bioreactor to decompose the problem into multiple sub-problems of "next stage state prediction", so that the state of the future culture stage can be accurately predicted based on the culture data of all historical culture stages. In addition, the culture data can include state-related data and action-related data, so that the model can not only consider the historical culture state of the bioreactor but also the operation experienced by the bioreactor when predicting. The method of the present disclosure is applicable to bioreactors of various scales (e.g., small scale, medium scale and large scale) or small scale to large scale, and can be widely used in the biopharmaceutical industry, shortening the time of process development and optimization of biopharmaceutical manufacturing, greatly saving manpower and material resources used to explore the culture process, reducing the development and manufacturing costs of biopharmaceuticals, and helping to ensure the high quality and consistency of the final product.

[0033] In this document, the same or similar characters may be used to represent the same or similar variables. Therefore, once a variable is defined in a certain embodiment, it does not need to be repeatedly described in subsequent embodiments.

[0034] Below, the method for predicting the state of cell culture in a bioreactor according to various embodiments of the present disclosure will be described in detail in conjunction with the accompanying drawings. It is understood that the actual method may also include other steps, but in order to avoid blurring the main points of the present disclosure, these other steps are not discussed herein and are not shown in the accompanying drawings.

[0035] Figure 1 FIG. 1 is a flow chart showing a method 100 for predicting the state of a cell culture in a bioreactor according to some embodiments of the present disclosure. Figure 1 As shown, the method 100 includes step S102 and step S104.

[0036] In step S102, the culture data of the bioreactor in each culture stage in the culture stage series is obtained.

[0037] Since in the actual cell culture experiment process, the format of the data related to cell culture recorded by the experimenter is often not standard, therefore, in some cases, the original experimental data will be cleaned and organized into a standard format before proceeding to the subsequent steps.

[0038] In some embodiments, the culture data of each culture stage includes first data, which may include raw state data indicating the state of the bioreactor in the culture stage and raw action data indicating the action performed on the bioreactor in the culture stage. For example, the raw state data may include the sampling time and the data obtained by sampling and measuring the bioreactor (including the cells and culture fluid therein) at the sampling time, including but not limited to pH, carbon dioxide partial pressure, oxygen partial pressure, sodium ion concentration, temperature, glucose concentration, lactic acid concentration, ammonium ion concentration, target product concentration (e.g., target protein concentration), osmotic pressure, viable cell density, viability, average cell diameter, working volume, or a combination of any of the foregoing.

[0039] It can be understood that when one or more states remain constant or change in a predetermined pattern throughout the cultivation process (for example, this may depend on the specific cultivation scheme, such as the temperature remains constant throughout the cultivation process in a constant temperature cultivation mode, and the sampling time can be set to the same time every morning during the entire cultivation process, etc.), the original state data indicating the one or more states may not need to be input into the neural network model.

[0040] In addition, the raw action data may include, for example, the operation of the bioreactor during the culture stage at the corresponding sampling time, including but not limited to the amount of sugar / feed supplement (for example, the mass of the supplemented glucose solution, the mass of (one or more) culture medium components (such as CB7a (Cell Boost7a), CB7b (Cell Boost 7b) and other commercially available general culture medium components or self-developed culture medium components, etc.), temperature adjustment amount (usually cooling amount), pH adjustment amount, ventilation amount (for example, air or oxygen), or any combination of the foregoing. It can be understood that the temperature adjustment amount, pH adjustment amount, and ventilation amount can be respectively inferred from the temperature, pH, and oxygen partial pressure obtained by sampling and measuring during the corresponding culture stage. For example, if the temperature of culture stage i is T i The temperature of the cultivation stage i+1 is T i+1 ≠T i , it can be determined that a temperature adjustment operation was performed in the culture stage i+1 and the temperature adjustment amount was ΔT = (T i+1 -T i ), where a negative ΔT indicates a temperature drop and a positive ΔT indicates a temperature rise. Therefore, when the original state data includes one or more of temperature, pH, and oxygen partial pressure, the original motion data may not necessarily include the corresponding one or more of temperature adjustment amount, pH adjustment amount, and ventilation amount.

[0041] In addition to providing the raw state data to the neural network model, the present disclosure can also provide the raw action data to the neural network model so that the model can consider the impact of the operation of the bioreactor when predicting. The present disclosure can be applied to any culture mode, such as but not limited to batch fed-batch culture, one-time fed-batch culture, continuous flow culture, etc. Different culture modes determine different operations on the bioreactor, and these operations can be fed into the neural network model via action data.

[0042] In addition, in addition to providing the original data to the neural network model, the present disclosure can also provide the processed data obtained by feature engineering the original data to the neural network model, so that the model obtains more valuable and easier to process inputs, thereby improving the performance and efficiency of the model. For example, these feature engineering can help identify and extract the most useful features for the task, thereby improving the performance and generalization ability of the model, and can also reduce the dimension of the data to reduce the complexity and computational cost of model training. In addition, data processed by feature engineering may be more interpretable and relevant, so that the model converges faster and speeds up model training. Data processed by feature engineering is usually more stable and consistent, which helps to reduce the risk of overfitting the model and improve the robustness of the model.

[0043] Specifically, in some embodiments, the culture data of each culture stage also includes second data, which can be calculated by at least one of the original state data and the original action data. Exemplarily, the second data may include derived action data, derived state data, state change data and time-related data. Among them, the derived action data may refer to data related to the action performed on the bioreactor in the culture stage that cannot or need not be directly measured but can be derived from the first data, for example, including but not limited to the temperature adjustment amount calculated according to the temperature, the pH adjustment amount calculated according to the pH, the ventilation amount calculated according to the oxygen partial pressure, the sugar / feed concentration calculated according to the sugar / feed amount and the working volume (for example, glucose feed concentration, (one or more) culture medium component feed concentration (such as, CB7a feed concentration, CB7b feed concentration, etc.)) or any combination of the foregoing items. The derived state data may refer to data related to the state performed on the bioreactor in the culture stage that cannot or need not be directly measured but can be derived from the first data. The state change data may indicate the changes in various states of the bioreactor in the culture stage, etc. The time-related data may indicate the time statistics of the bioreactor in each culture stage.

[0044] As a non-limiting example combination of the first data and the second data, the original state data may include sampling time, temperature, live cell density, and working volume; the original action data may include the mass of the supplemented glucose solution and the mass of the supplemented culture medium components; the derived action data may include the glucose feed concentration calculated by the mass of the supplemented glucose solution and the working volume, the culture medium component feed concentration calculated by the mass of the supplemented culture medium components and the working volume, the distance cooling time calculated by the sampling time and the temperature, and the cooling amplitude calculated by the temperature; the derived state data may include the cell growth rate calculated by the sampling time and the live cell density; the state change data may include the change between adjacent culture stages calculated by each state in the original state data; the time-related data may include the cumulative time of the tank calculated by the sampling time. It can be understood that the first data and the second data cited in this article are exemplary and non-limiting, and those skilled in the art can obtain any other suitable data about cell culture in addition or in place according to the biological reaction experiment.

[0045] It is understandable that, in the case where one or more of the above-described states remain constant or change in a predetermined pattern throughout the culture process, because the original state data indicating the one or more states are known, even if the original state data indicating the one or more states are not input into the neural network model, the second data can still be calculated based on the original state data indicating the one or more states, and the calculated second data can be input into the neural network model. For example, the culture scheme has been determined to be a 25-degree constant temperature culture, and the original state data of the temperature is known to be 25 degrees and therefore does not need to be measured and input into the neural network model, while the temperature reduction amplitude can be calculated as 0 based on this and can be input into the neural network model as needed.

[0046] Continue to refer Figure 1 , at step S104 , the acquired culture data is provided to the trained neural network model to obtain the state of the bioreactor in a future culture stage after the series of culture stages.

[0047] For example, the trained neural network model can be configured to predict the state of the M+1th training stage based on the training data of the 1st to M training stages (M is a positive integer). The M+1th training stage is a future training stage along the training trajectory of the 1st to M training stages. If you want to further predict the state of the M+2th training stage, you can provide the training data of the 1st to M+1 training stages to the model, wherein the training data of the M+1th training stage includes the state of the M+1th training stage predicted by the model based on the training data of the 1st to M training stages and data of the action to be performed in the M+1th training stage, which can be received via user input, as the first data and optionally the second data calculated according to the first data as described above. If it is desired to further predict the state of the M+3th culture stage, the culture data of the 1st to M+2 culture stages can be provided to the model, wherein the culture data of the M+1th culture stage includes the state of the M+1th culture stage predicted by the model based on the culture data of the 1st to M culture stages and, for example, data of an action to be performed in the M+1th culture stage that can be received via user input as first data and, optionally, second data calculated from the first data as described above, and the culture data of the M+2th culture stage includes the state of the M+2th culture stage predicted by the model based on the culture data of the 1st to M+1 culture stages and, for example, data of an action to be performed in the M+2th culture stage that can be received via user input as first data and, optionally, second data calculated from the first data as described above. By analogy, the model can not only predict the state of one or more future culture stages after the series of culture stages that have occurred, but also predict the final culture state when the culture is completed (for example, at the end of the entire culture cycle), and can even predict the state after the culture is completed (for example, after the end of the entire culture cycle, at this time, the bioreactor is usually no longer operated, i.e., there is no original action data).

[0048] Therefore, in some embodiments, method 100 may further include providing the acquired culture data, together with the culture data of the future culture stage determined based on the predicted state of the future culture stage and the action to be performed in the future culture stage, to the trained neural network model to obtain the state of the bioreactor in the second future culture stage after the future culture stage. For example, the predicted state of the future culture stage and the data of the action to be performed in the future culture stage are equivalent to the first data of the future culture stage, and additionally, the second data of the future culture stage can be calculated based on the first data of the future culture stage as described above, so that the culture data of the future culture stage including both the first data and the second data are provided to the trained neural network model.

[0049] For example, refer to Figure 2 , Figure 21 shows a flow chart of a non-limiting example process 200 in which the method 100 for predicting the state of a cell culture in a bioreactor according to some embodiments of the present disclosure is applied. Figure 2 As shown, the original state data and original action data of each training stage from the 1st to the Mth training stage are obtained (step S202). Determine that the future training stage to be predicted is the M+1th training stage (step S204) (M, l are positive integers). Initialize t to 1 (step S206). Feature engineering is performed on the acquired state and action data (step S208). The acquired state and action data and their feature engineering results are provided to the trained neural network model (step S210). The state of the M+tth training stage is predicted by the trained neural network model (step S212). Determine whether t is equal to l (step S214). If t is not equal to l (step S214 "No"), the predicted state of the M+tth training stage and the data of the action to be performed in the M+tth training stage are obtained (step S216), and t=t+1 (step S218), and then return to step S208. If t is equal to 1 (“Yes” in step S214 ), the predicted state of the M+1th culture stage is output (step S220 ).

[0050] In particular, the neural network model used in step S104 can perform conditional state modeling for the problem of predicting the state of cell culture in the bioreactor, so as to decompose the problem into multiple sub-problems of "predicting the state of the next stage". For example, if the 1st to 3rd culture stages have occurred but the 4th culture stage has not yet occurred, the problem of "predicting the state of the 4th culture stage" can be decomposed into three sub-problems of "predicting the state of the 2nd culture stage", "predicting the state of the 3rd culture stage", and "predicting the state of the 4th culture stage". More specifically, these sub-problems can be implemented as "predicting the state of the 2nd culture stage based on the culture data of the 1st culture stage", "predicting the state of the 3rd culture stage based on the culture data of the 1st to 2nd culture stages", and "predicting the state of the 4th culture stage based on the culture data of the 1st to 3rd culture stages".

[0051] Therefore, the loss function of the neural network model can be determined based on the loss generated by a set of sub-problems, the set of sub-problems corresponding to the cultivation data of a series of cultivation stages from a training data set, and each sub-problem in the set of sub-problems is configured to predict the state of a corresponding cultivation stage using the cultivation data of each cultivation stage before a corresponding cultivation stage in the series of cultivation stages from the training data set.

[0052] For non-limiting illustration purposes, it is assumed that the training data set has a training stage series C with M training stages, and its training data is {c1, c2, ..., c M}, where c tis the training data of the t-th training stage (M, t are positive integers), for example, including original state data, original action data and its feature engineering results. The sub-problem set for the training stage series C can be in, is the t-th sub-problem, which means predicting the state of the t-th cultivation stage using the cultivation data of the 1st to t-1st cultivation stages before the t-th cultivation stage of the cultivation stage series C. Thus, based on the true state s of the t-th cultivation stage t (as training labels) and predicted states The loss generated by the t-th sub-problem can be determined. Any suitable loss function can be used to characterize the loss generated by the sub-problem, including but not limited to the mean square error, etc., which is not particularly limited here. Then, the loss function of the entire neural network model can be determined based on the losses generated by each sub-problem. For example, the loss function of the neural network model can be determined as

[0053] In order to improve the training accuracy of the neural network model, each batch of training samples of the training data set may include the cultivation data of multiple cultivation stage series. In some embodiments, the loss function of the neural network model may be determined based on the loss generated by the first plurality of sub-problem sets, each sub-problem set in the first plurality of sub-problem sets corresponds to the cultivation data of a corresponding one of the multiple cultivation stage series from the training data set, and each sub-problem in each sub-problem set in the first plurality of sub-problem sets is configured to predict the state of the corresponding one of the cultivation stages using the cultivation data of each cultivation stage before the corresponding one of the cultivation stage series from the training data set.

[0054] For non-limiting illustration purposes, assume that a batch of training samples of a training dataset D includes a series of P training stages {C1, C2, ..., C P}, where C i represents the i-th cultivation stage series, each cultivation stage series C i It includes M training stages, and its training data is {c i1 ,c i2 ,...,c iM}, where c it Series C for the cultivation stage i The training data of the t-th training stage (M, t, P, i are positive integers), for example, including original state data, original action data and its feature engineering results. The first plurality of sub-problem sets include Among them, for the cultivation stage series C i The set of sub-problems can be in, For the cultivation stage series C i The t-th subproblem in the subproblem set means using the training phase series C i The state of the t-th culture stage is predicted based on the culture data of the 1st to t-1st culture stages before the t-th culture stage. i The real status of the t-th training stage s it (as training labels) and predicted states The loss generated by the t-th subproblem can be determined. Any suitable loss function can be used to characterize the loss generated by the subproblem, including but not limited to the mean square error, etc., which is not particularly limited here. Then, the loss function of the entire neural network model can be determined based on the loss generated by each subproblem of each subproblem set. For example, the loss function of the neural network model can be determined as

[0055] The state of cell culture generally includes multiple states, for example, the viable cell density, the average cell diameter, the target protein concentration, the glucose concentration, the lactate concentration, etc., as described above. The states in the acquired culture data may be some or all of the states of all cell cultures. The state of interest may be a subset or a full set of the states in the acquired culture data.

[0056] In order to optimize the prediction performance of the state of interest, a loss function can be set based on the state of interest. In some embodiments, the loss function of the neural network model can be determined based on the loss generated by a second plurality of sub-problem sets, each of the second plurality of sub-problem sets corresponds to cultivation data related to a corresponding one of the one or more states from a cultivation stage series of the training data set, and each sub-problem in each of the second plurality of sub-problem sets is configured to predict the corresponding one of the states of the corresponding one of the cultivation stages using the cultivation data related to the corresponding one of the states from each cultivation stage before the corresponding one of the cultivation stages in the cultivation stage series of the training data set.

[0057] For non-limiting illustration purposes, assume that the culture data c of the t-th culture stage in the M culture stages of the culture stage series C of the training data set D is t ={c t1 ,c t2 ,...,c tS}, where c tj It is c t The training data related to the jth state in the S states (M, t, S, j are positive integers). The second set of multiple sub-problems includes The sub-problem set for the j-th state of the training phase series C can be in, is the t-th sub-problem in the set of sub-problems for the cultivation stage series C, which means predicting the j-th state of the t-th cultivation stage using the cultivation data related to the j-th state in the 1st to t-1st cultivation stages before the t-th cultivation stage in the cultivation stage series C. Thus, based on the true state s of the t-th cultivation stage in the cultivation stage series C, tj (as training labels) and predicted states The loss generated by the t-th subproblem can be determined. Any suitable loss function can be used to characterize the loss generated by the subproblem, including but not limited to the mean square error, etc., which is not particularly limited here. Then, the loss function of the entire neural network model can be determined based on the loss generated by each subproblem of each subproblem set. For example, the loss function of the neural network model can be determined as

[0058] In addition, in combination with the above-mentioned embodiment, for example, the loss function of the neural network model can also be determined as

[0059] The present disclosure specifically designs the loss function of the neural network model to be determined based on the losses generated by a set of sub-problems, rather than being determined solely based on the deviation between the predicted state of a single cultivation stage and its measured actual state. This enables the neural network model trained using the specially designed loss function to comprehensively consider all historical data, thereby being able to accurately predict the state of future cultivation stages.

[0060] In addition, in some embodiments, before providing the acquired culture data to the trained neural network model, the trained neural network model may also use culture data of other batches of cell samples of the bioreactor in the past for small sample learning. Since the trained neural network model may be a model with good generality trained based on the training data set, using culture data of other batches of cell samples of the bioreactor to be predicted in the past to make the trained "general" neural network model perform small sample learning again can make the model more "specialized" for the cell samples of the bioreactor to be predicted, thereby improving the accuracy of predicting the state of cell culture in the bioreactor.

[0061] The following describes a non-limiting example structure of a neural network model that can be used in the present disclosure with reference to the accompanying drawings. It should be understood that the actual model may have other additional components, but in order to avoid obscuring the main points of the present disclosure, these other components are not discussed herein and are not shown in the accompanying drawings. It should also be understood that the structure of an untrained neural network model and the structure of a trained neural network model will not change due to training, and therefore, when describing the structure of a neural network model herein, the trained neural network model and the untrained neural network model may not be distinguished.

[0062] refer to Figure 3A In some embodiments, the neural network model 300 includes: an input module 302 configured to receive input data; a plurality of generative pre-trained transformer (GPT) modules 3041, ..., 304 connected to the output end of the input module 302 and connected in series with each other; r (In this article, it can also be collectively referred to as a generative pre-trained transformer module 304, where r is a positive integer); and an output module 306 connected to the output end of the plurality of generative pre-trained transformer modules, configured to output a prediction result. In some examples, r can be set between 6 and 24, for example, set to 12. Through these generative pre-trained transformer modules, long-range context information contained in different stages of the cell culture process can be learned.

[0063] In some embodiments, the input module 302 is configured to receive the first matrix and the second matrix, and add the first matrix and the second matrix element by element after they are respectively converted into the feature dimensions of the neural network model 300 to obtain a third matrix, and input the third matrix to the plurality of generative pre-trained transformer modules 3041, ..., 304 r (Specifically, generative pre-trained transformer module 3041). The first matrix and the second matrix are obtained based on the culture data of each culture stage in the culture stage series of the bioreactor. Each row of the first matrix corresponds to a corresponding culture stage in the culture stage series, and each column of the row corresponds to a corresponding piece of culture data of the corresponding culture stage. The second matrix indicates the culture stage position code.

[0064] For example, the first matrix can be determined based on the culture data other than the sampling time, and the second matrix can be determined based on the sampling time (equivalent to using the sampling time to represent the culture stage). In other examples, the first matrix can also be determined based on all the culture data, and the second matrix can be determined based on the sequence number of the culture stage. Thus, the third matrix can contain the fusion result of the culture stage and the corresponding state / action, which can achieve better feature expression than directly inputting the culture data into the GPT module.

[0065] Compared with the traditional neural network model mainly used in natural language processing (NLP), the neural network model of the present invention targets continuous data rather than discrete data. Therefore, compared with the NLP model, the neural network model of the present invention does not adopt the conversion process from token to embedding, but can adjust the input matrix through a linear conversion layer to conform to the feature dimension of the neural network model. Specifically, in some embodiments, the input module 302 may include a fully connected layer configured to implement the above-mentioned conversion of the first matrix and the second matrix. In some embodiments, the feature dimension of the neural network model 300 can be set between 128 and 2048, for example, it can be set to 1024.

[0066] For non-limiting illustration purposes, assuming that the culture stage series includes three culture stages, and the culture data of each culture stage includes sampling time t, pH, viable cell density VCD, and sodium ion concentration Na, then the first matrix can be determined as The kth row of the first matrix contains the pH value of the kth culture stage. k , viable cell density VCD k 、Sodium ion concentration Na k (k=1, 2, 3). The second matrix can be determined as Assuming that the feature dimension of the neural network model 300 is 1024, the first matrix and the second matrix can be added element by element to obtain a third matrix after being converted to a dimension of 1024 by the corresponding fully connected layer of the input module 302. For example, a 3×3 first matrix can be multiplied by a 3×1024 first conversion matrix to become 3×1024, and a 3×1 second matrix can be multiplied by a 1×1024 second conversion matrix to become 3×1024.

[0067] Of course, in other embodiments, the first matrix including the sampling time may also be constructed, for example There is no need to construct a second matrix, and then directly convert the first matrix into a feature dimension having the neural network model 300 and input it into the generative pre-trained transformer module 3041.

[0068] Further, refer to Figure 3B and Figure 3C In some embodiments, each generative pre-trained transformer module 304 of the trained neural network model 300 may include a multi-head self-attention layer 3042, a first normalization layer 3044, and a fully connected layer 3046 connected in series, and further includes a multi-head self-attention layer 3042 upstream (such as Figure 3B) or connected in series in the fully connected layer 3046 (as shown Figure 3C The second normalization layer 3048 downstream of the multi-head self-attention layer 3042 is shown in FIG. 1 , wherein the input and output of the multi-head self-attention layer 3042 are residually connected to each other, and the input and output of the fully connected layer 3046 are residually connected to each other. For example, the number of attention heads included in the multi-head self-attention layer 3042 can be set between 4 and 8. The multi-head self-attention layer 3042 can use the self-attention mechanism to refine the high-order features contained in the input data.

[0069] The feature dimension of the neural network model 300, the number of generative pre-trained transformer modules 304, and the number of attention heads included in the multi-head self-attention layer 3042 can be specifically set according to the amount of data that the neural network model 300 needs to process. In some embodiments, the feature dimension of the neural network model 300 can be 1024, the number of generative pre-trained transformer modules 304 can be 12, and the number of attention heads in the multi-head self-attention layer 3042 can be 4.

[0070] In some embodiments, the output module 306 may include a plurality of generative pre-trained transformer modules 3041, ..., 304 r (Specifically, the generative pre-trained transformer module 304 r ) is converted into a fully connected layer having a dimension of the number of columns of the first matrix. Following the above non-limiting embodiment, the prediction result finally output by the output module 306 can be This includes the predicted pH for stage 4 Viable cell density Sodium ion concentration

[0071] The neural network model 300 can perform multi-state regression tasks, and its architecture has unique advantages and application potential in processing continuous input features and predicting multiple culture states for cells.

[0072] refer to Figure 4 , which shows a flow chart of a method 400 for training a neural network model for predicting the state of cell culture in a bioreactor according to some embodiments of the present disclosure. It can be understood that the actual method 400 may also include other steps, but in order to avoid obscuring the main points of the present disclosure, these other steps are not discussed herein and are not shown in the drawings.

[0073] like Figure 4As shown, method 400 includes: at step S402, obtaining a training data set, the training data set including culture data of each culture stage in each culture stage series in multiple culture stage series of multiple bioreactors; at step S404, training a neural network model using the training data set, wherein a loss function of the neural network model is determined based on a loss generated by a set of sub-problems, the set of sub-problems corresponding to the culture data of the culture stage series from the training data set, and each sub-problem in the set of sub-problems is configured to predict the state of a corresponding culture stage in the culture stage series from the training data set using the culture data of each culture stage before the corresponding culture stage.

[0074] In some embodiments, the loss function of the neural network model is determined based on the loss generated by a first plurality of sub-problem sets, each sub-problem set in the first plurality of sub-problem sets corresponds to cultivation data of a corresponding one of a plurality of cultivation stage series from a training data set, and each sub-problem in each sub-problem set in the first plurality of sub-problem sets is configured to predict the state of a corresponding one of the cultivation stages in the training data set using the cultivation data of each cultivation stage before the corresponding one of the cultivation stage series.

[0075] In some embodiments, the loss function of the neural network model is determined based on the loss generated by a second plurality of sub-problem sets, each of the second plurality of sub-problem sets corresponds to cultivation data associated with a corresponding one of one or more states in a series of cultivation stages from a training data set, and each sub-problem in each of the second plurality of sub-problem sets is configured to predict a corresponding one of the states of the corresponding one of the cultivation stages using cultivation data associated with the corresponding one of the states in each cultivation stage before the corresponding one of the cultivation stages in the series of cultivation stages from the training data set.

[0076] In some embodiments, the culture data of each culture stage includes first data, the first data including raw state data of the state of the corresponding bioreactor in the culture stage and raw action data indicating the action performed on the corresponding bioreactor in the culture stage. In some embodiments, the culture data of each culture stage also includes second data, the second data including data calculated from at least one of the raw state data and the raw action data.

[0077] Various embodiments of method 400 are similar to the various embodiments related to the neural network model described above, and therefore reference may be made to the description of the various embodiments related to the neural network model described above, which will not be repeated here.

[0078] For illustrative purposes, in order to evaluate the predictive performance of the trained neural network model of the present disclosure, a five-fold cross-validation method can be used to validate the model. The validation method includes: dividing the training data set into five independent subsets, each subset is used as a validation set in turn, and the remaining four subsets are used for model training. In this way, each training data has the opportunity to be used as validation data, ensuring a comprehensive evaluation of all data in the training data set.

[0079] The evaluation results of some culture states predicted by the five-fold cross-validation model are shown in Table 1 below, where R 2 It is a statistic that measures the goodness of fit of a model. Its value range is between 0 and 1. The closer it is to 1, the better the model fits the data, and the closer it is to 0, the worse the model fits the data.

[0080] Table 1

[0081]

[0082] As shown in Table 1, it can be seen that the prediction results of the trained neural network model of the present disclosure have excellent prediction accuracy and robustness.

[0083] In addition, in order to further comprehensively evaluate the performance of the trained neural network model disclosed in the present invention and ensure its effectiveness and accuracy in practical applications, wet experiment verification can also be performed. The following is an example of 14-day batch feeding culture of a cell sample whose culture data is not included in the training data set. Among them, the specific action strategy of the batch feeding culture is known, that is, it is known what action will be performed on the bioreactor on which day.

[0084] On the one hand, during the cell culture process, the actual culture trajectory can be sampled and measured while the culture trajectory can be predicted in advance using the trained neural network model. Thus, the data point on day 1 includes the actual measured culture status, and the data points on day 2 and every subsequent day include the actual measured culture status and the predicted culture status based on the data of all the days before this day. Figure 5A The wet experiment evaluation results of the pre-one-step prediction tasks of pH, activity rate and target protein concentration are shown as examples. It can be seen that the trained neural network model of the present disclosure not only has accurate R prediction results for each state, but also has good 2, and can also make accurate predictions for key actions and time nodes in the cell culture process. For example, in the actual culture process, the pH value decreased in the early stage, and began to rise after the bioreactor was cooled on the 4th day. It can be seen that the trained neural network model of the present disclosure also accurately predicted in advance the trend of the pH value rising after the cooling on the 4th day. This accurate prediction of the inflection point also shows that the trained neural network model of the present disclosure has a deeper understanding of the action, which conforms to the biological laws, and increases the possibility of applying the trained neural network model of the present disclosure in more scenarios.

[0085] On the other hand, the trained neural network model of the present invention can also utilize the characteristics of autoregression to predict key growth indicators and product quality for each day from day x+1 to day 14 (cell harvest day) based on early culture data from day 1 to day x (x is a positive integer) and the action strategy to be taken during the culture cycle. Figure 5B The wet experiment evaluation results of the prediction task based on the known data of very few days (here 4 days) for lactate concentration, live cell density, viability and target protein concentration are shown as examples. It can be seen that the actual cell growth trajectory and the predicted cell growth trajectory of the complete culture cycle are consistent. Therefore, even if the known data is limited to the early stage of culture, the trained neural network model of the present disclosure can also make reliable predictions on the overall development trend and important turning points of cell culture states such as cell growth (live cell density, viability), metabolism (lactic acid concentration), and yield (target protein concentration).

[0086] The predictive capability of the trained neural network model disclosed herein can guide researchers and production personnel to timely adjust the culture conditions of the bioreactor and the action strategies to be executed, optimize the growth environment of cells, and thus improve the efficiency and output of cell culture.

[0087] In another aspect, the present disclosure also provides a device for predicting the state of cell culture in a bioreactor. Figure 6 , Figure 6 FIG. 6 is a schematic block diagram of an apparatus 600 for predicting the state of cell culture in a bioreactor according to some embodiments of the present disclosure. Figure 6As shown, the device 600 includes an acquisition unit 602 and a prediction unit 604. The acquisition unit 602 is configured to acquire the culture data of the bioreactor in each culture stage in the culture stage series. The prediction unit 604 is configured to provide the culture data acquired by the acquisition unit 602 to the trained neural network model to obtain the state of the bioreactor in the future culture stage after the culture stage series, wherein the loss function of the neural network model is determined based on the loss generated by the sub-problem set, the sub-problem set corresponds to the culture data of the culture stage series from the training data set, and each sub-problem in the sub-problem set is configured to predict the state of a corresponding culture stage using the culture data of each culture stage before the corresponding culture stage in the culture stage series from the training data set.

[0088] Various embodiments of the device 600 are similar to various embodiments of the aforementioned method 100 for predicting the state of cell culture in a bioreactor, and therefore reference may be made to the aforementioned description of various embodiments of the method 100 , which will not be repeated herein.

[0089] In another aspect, the present disclosure also provides a device for training a neural network model for predicting the state of cell culture in a bioreactor. Figure 7 , Figure 7 FIG. 7 is a schematic block diagram of an apparatus 700 for training a neural network model for predicting the state of cell culture in a bioreactor according to some embodiments of the present disclosure. Figure 7 As shown, the apparatus 700 includes an acquisition unit 702 and a training unit 704. The acquisition unit 702 is configured to acquire a training data set, the training data set including culture data of each culture stage in each culture stage series in a plurality of culture stage series of a plurality of bioreactors. The training unit 704 is configured to train a neural network model using the training data set acquired by the acquisition unit 702, wherein a loss function of the neural network model is determined based on a loss generated by a set of sub-problems, the set of sub-problems corresponding to the culture data of the culture stage series from the training data set, and each sub-problem in the set of sub-problems is configured to predict the state of a corresponding culture stage using the culture data of each culture stage before the corresponding culture stage in the culture stage series from the training data set.

[0090] The various embodiments of the device 700 are similar to the various embodiments of the method 400 for training a neural network model for predicting the state of cell culture in a bioreactor, and thus the description of the various embodiments of the method 400 may be referred to above and will not be repeated here.

[0091] The present disclosure also provides a computing device, which may include one or more processors and a memory storing computer executable instructions, wherein when the computer executable instructions are executed by the one or more processors, the one or more processors execute the method according to any of the aforementioned embodiments of the present disclosure. Figure 8 As shown, the device 800 includes (one or more) processors 802 and a memory 804 storing computer executable instructions, which, when executed by the (one or more) processors 802, cause the (one or more) processors 802 to perform the method according to any of the foregoing embodiments of the present disclosure. The (one or more) processors 802 may be, for example, a central processing unit (CPU) of the device 800. The (one or more) processors 802 may be any type of general-purpose processor, or may be a processor specifically designed to predict the state of cell culture in a bioreactor and / or train a neural network model for predicting the state of cell culture in a bioreactor, such as an application-specific integrated circuit ("ASIC"). The memory 804 may include various computer-readable media accessible by the (one or more) processors 802. In various embodiments, the memory 804 described herein may include volatile and non-volatile media, removable and non-removable media. For example, the memory 804 may include any combination of the following: random access memory ("RAM"), dynamic RAM ("DRAM"), static RAM ("SRAM"), read-only memory ("ROM"), flash memory, cache memory, and / or any other type of non-transitory computer-readable media. The memory 804 may store instructions that, when executed by the processor 802 , enable the processor 802 to perform the method according to any of the aforementioned embodiments of the present disclosure.

[0092] The present disclosure also provides a non-volatile storage medium having computer executable instructions stored thereon, which, when executed by one or more processors, enables the one or more processors to execute the method according to any of the aforementioned embodiments of the present disclosure.

[0093] The present disclosure also provides a computer program product, which may include instructions, and when the instructions are executed by a processor, the method described in any of the foregoing embodiments of the present disclosure may be implemented. The instructions may be any set of instructions to be executed directly by one or more processors, such as machine code, or any set of instructions to be executed indirectly, such as a script. The instructions may be stored in an object code format for direct processing by one or more processors, or stored in any other computer language, including scripts or collections of independent source code modules that are interpreted on demand or compiled in advance.

[0094] Fig. 9A schematic block diagram of a computer system 900 on which an embodiment of the present disclosure may be implemented is shown. The computer system 900 includes a bus 902 or other communication mechanism for transmitting information, and a processing device 904 coupled to the bus 902 for processing information. The computer system 900 also includes a memory 906 coupled to the bus 902 for storing instructions to be executed by the processing device 904, which may be a random access memory (RAM) or other dynamic storage device. The memory 906 may also be used to store temporary variables or other intermediate information during the execution of the instructions to be executed by the processing device 904. The computer system 900 also includes a read-only memory (ROM) 908 or other static storage device coupled to the bus 902 for storing static information and instructions for the processing device 904. A storage device 910 such as a magnetic disk or optical disk is provided and coupled to the bus 902 for storing information and instructions. The computer system 900 may be coupled to an output device 912 via the bus 902 for providing output to a user, such as, but not limited to, a display (such as a cathode ray tube (CRT) or a liquid crystal display (LCD)), a speaker, etc. An input device 914 such as a keyboard, a mouse, a microphone, etc. is coupled to the bus 902 for transmitting information and command selections to the processing device 904. The computer system 900 may perform embodiments of the present disclosure. Consistent with certain implementations of the present disclosure, the results are provided by the computer system 900 in response to the processing device 904 executing one or more sequences of one or more instructions contained in the memory 906. Such instructions may be read into the memory 906 from another computer-readable medium such as the storage device 910. The execution of the sequence of instructions contained in the memory 906 causes the processing device 904 to perform the methods described herein. Alternatively, hard-wired circuitry may be used instead of software instructions or in combination with software instructions to implement the present teachings. Therefore, implementations of the present disclosure are not limited to any particular combination of hardware circuitry and software. In various embodiments, the computer system 900 can be connected to one or more other computer systems such as the computer system 900 across a network via a network interface 916 to form a networked system. The network can include a private network or a public network such as the Internet. In a networked system, one or more computer systems can store data and supply data to other computer systems. The term "computer-readable medium" as used herein refers to any medium that participates in providing instructions to the processing device 904 for execution. Such media can take many forms, including but not limited to non-volatile media, volatile media, and transmission media. Non-volatile media include, for example, optical disks or disks such as storage devices 910. Volatile media include dynamic memories such as memory 906. Transmission media include coaxial cables, copper wires, and optical fibers, including wiring that includes bus 902.Common forms of computer-readable media or computer program products include, for example, floppy disks, flexible disks, hard disks, tapes, or any other magnetic media, CD-ROMs, digital video disks (DVDs), Blu-ray disks, any other optical media, thumb drives, memory cards, RAMs, PROMs and EPROMs, flash EPROMs, any other memory chips or boxes, or any other tangible media from which a computer can read. Various forms of computer-readable media may be involved when carrying one or more sequences of one or more instructions to processing device 904 for execution. For example, the instructions may initially be carried on a disk of a remote computer. The remote computer may load the instructions into its dynamic memory and send the instructions over a telephone line using a modem. A modem local to computer system 900 may receive data over a telephone line and convert the data into an infrared signal using an infrared transmitter. An infrared detector coupled to bus 902 may receive the data carried in the infrared signal and place the data on bus 902. Bus 902 carries the data to memory 906, and processing device 904 retrieves the instructions from memory 906 and executes the instructions. Optionally, the instructions received by memory 906 may be stored on storage device 910 either before or after execution by processing device 904 .

[0095] According to various embodiments, instructions configured to be executed by a processing device to perform a method are stored on a computer readable medium. A computer readable medium may be a device that stores digital information. For example, a computer readable medium includes a compact disk read-only memory (CD-ROM) as known in the art for storing software. The computer readable medium is accessed by a processor adapted to execute the instructions configured to be executed.

[0096] One or more exemplary embodiments of the present disclosure are described above. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0097] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a server system. Of course, the present disclosure does not exclude that with the development of computer technology in the future, the computer that implements the functions of the above embodiments may be, for example, a personal computer, a laptop computer, a vehicle-mounted human-computer interaction device, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0098] Although one or more embodiments of the present disclosure provide method operation steps as described in the embodiments or flowcharts, more or fewer operation steps may be included based on conventional or non-creative means. The order of steps listed in the embodiments is only one way of executing the order of many steps and does not represent the only execution order. When the device or terminal product in practice is executed, it can be executed in sequence or in parallel according to the method shown in the embodiments or the drawings (for example, in a parallel processor or multi-threaded processing environment, or even a distributed data processing environment).

[0099] The terms "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, product or apparatus comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, product or apparatus. In the absence of further restrictions, it is not excluded that there are other identical or equivalent elements in the process, method, product or apparatus comprising the elements. For example, if the words "first", "second" and the like are used to indicate names, they do not indicate any particular order.

[0100] For the convenience of description, the above devices are described in various modules according to their functions. Of course, when implementing one or more embodiments of the present disclosure, the functions of each module can be implemented in the same or more software and / or hardware, or the module implementing the same function can be implemented by a combination of multiple sub-modules or sub-units, etc. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0101] The present disclosure is described with reference to the flowchart and / or block diagram of the method, device (system), and computer program product according to the embodiment of the present disclosure. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the process and / or box in the flowchart and / or block diagram can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device for implementing the function specified in one process or multiple processes in the flowchart and / or one box or multiple boxes in the block diagram.

[0102] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a particular manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction device that implements the functions specified in one or more flows of a flowchart and / or one or more blocks of a block diagram. These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operating steps are performed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more flows of a flowchart and / or one or more blocks of a block diagram.

[0103] Those skilled in the art will appreciate that one or more embodiments of the present disclosure may be in the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, one or more embodiments of the present disclosure may be in the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0104] One or more embodiments of the present disclosure may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. One or more embodiments of the present disclosure may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communications network. In a distributed computing environment, program modules may be located in local and remote computer storage media, including storage devices.

[0105] The same or similar parts between the various embodiments of the present disclosure can refer to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts refer to the partial description of the method embodiment. In the description of the present disclosure, the description of the reference terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure. In the present disclosure, the schematic representation of the above terms does not necessarily target the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine the different embodiments or examples described in the present disclosure and the features of the different embodiments or examples without contradiction.

[0106] In addition, when used in the present disclosure, the words "herein," "above," "below," "hereunder," "above," and words of similar meaning shall refer to the present disclosure as a whole rather than to any particular portion of the present disclosure. Furthermore, unless expressly stated otherwise or understood otherwise in the context of use, conditional language used herein, such as "may," "might," "for example," "such as," and the like, is generally intended to express that certain embodiments include, while other embodiments do not include, certain features, elements, and / or states. Thus, such conditional language is generally not intended to imply that one or more embodiments require features, elements, and / or states in any way, or whether such features, elements, and / or states are included or performed in any particular embodiment.

[0107] The above description is only an example of one or more embodiments of the present disclosure, and is not intended to limit one or more embodiments of the present disclosure. For those skilled in the art, one or more embodiments of the present disclosure may have various changes and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present disclosure shall be included in the scope of the claims.

Claims

1. A method for predicting the state of a cell culture in a bioreactor, comprising: Acquiring cultivation data of the bioreactor at each cultivation stage in a series of cultivation stages; as well as providing the acquired cultivation data to the trained neural network model to obtain the state of the bioreactor in a future cultivation phase following the series of cultivation phases, The loss function of the neural network model is determined based on the loss generated by a set of sub-problems, the set of sub-problems corresponding to the cultivation data of a series of cultivation stages from a training data set, and each sub-problem in the set of sub-problems is configured to predict the state of a corresponding cultivation stage using the cultivation data of each cultivation stage before a corresponding cultivation stage in the series of cultivation stages from the training data set.

2. The method according to claim 1, wherein: The loss function of the neural network model is determined based on at least one of the losses generated by the first plurality of sub-problem sets and the losses generated by the second plurality of sub-problem sets, Each of the first plurality of sub-problem sets corresponds to the cultivation data of a corresponding one of the plurality of cultivation stage series from the training data set, and each sub-problem in each of the first plurality of sub-problem sets is configured to predict the state of the corresponding one of the cultivation stages in the training data set by using the cultivation data of each cultivation stage before the corresponding one of the cultivation stage series. Wherein, each sub-problem set in the second plurality of sub-problem sets corresponds to cultivation data related to a corresponding one of one or more states in the cultivation stage series from the training data set, and each sub-problem in each sub-problem set in the second plurality of sub-problem sets is configured to predict the corresponding one of the states of the corresponding one of the cultivation stages using the cultivation data related to the corresponding one of the states in each cultivation stage before the corresponding one of the cultivation stages in the cultivation stage series from the training data set.

3. The method according to claim 1, wherein: the cultivation data of each cultivation stage comprises first data, the first data comprising raw state data indicating a state of the bioreactor in the cultivation stage and raw action data indicating an action performed on the bioreactor in the cultivation stage; or The culture data of each culture stage includes first data and second data, wherein the first data includes raw status data indicating the state of the bioreactor in the culture stage and raw action data indicating the action performed on the bioreactor in the culture stage, and the second data is calculated by at least one of the raw status data and the raw action data.

4. The method according to claim 3, wherein: The second data includes derived action data, derived state data, state change data and time-related data.

5. The method according to claim 4, wherein: Raw state data include sampling time, temperature, working volume, and viable cell density; The original action data include the mass of supplemented glucose solution and the mass of supplemented culture medium components; The derived action data include the glucose feed concentration calculated by the mass and working volume of the supplemented glucose solution, the medium component feed concentration calculated by the mass and working volume of the supplemented medium components, the distance cooling time calculated by the sampling time and temperature, and the cooling amplitude calculated by the temperature; Derived state data include cell growth rate calculated from sampling time and viable cell density; The state change data includes the change amount between adjacent culture stages calculated by each state in the original state data; The time-related data includes the cumulative time of loading the tank calculated by the sampling time.

6. The method according to claim 1, wherein: The neural network model includes: an input module configured to receive input data; A plurality of generative pre-trained transformer modules connected to the output of the input module and connected in series with each other; and An output module connected to the output ends of the plurality of generative pre-trained transformer modules is configured to output prediction results.

7. The method according to claim 6, wherein: A first matrix and a second matrix are obtained based on the acquired culture data of each culture stage of the bioreactor in the culture stage series, wherein each row of the first matrix corresponds to a corresponding culture stage in the culture stage series, and each column of the row corresponds to a corresponding piece of culture data of the corresponding culture stage, and the second matrix indicates the culture stage position code, The input module is configured to receive the first matrix and the second matrix and add the first matrix and the second matrix element by element after they are respectively converted into the feature dimensions of the neural network model to obtain a third matrix, and input the third matrix into the multiple generative pre-trained transformer modules.

8. The method according to claim 7, wherein: The neural network model satisfies at least one of the following: The input module comprises a fully connected layer configured to implement the conversion; or The feature dimension is set between 128 and 2048; or The output module includes a fully connected layer configured to convert the outputs of the plurality of generative pre-trained transformer modules into a dimension having a number of columns of the first matrix.

9. The method according to claim 6, wherein: Each generative pre-trained transformer module includes a multi-head self-attention layer, a first normalization layer and a fully connected layer connected in series in sequence, and also includes a second normalization layer connected in series upstream of the multi-head self-attention layer or in series downstream of the fully connected layer, wherein the input and output ends of the multi-head self-attention layer are residually connected to each other, and the input and output ends of the fully connected layer are residually connected to each other.

10. The method according to claim 9, wherein: The number of generative pre-trained transformer modules included in the multiple generative pre-trained transformer modules is set between 6 and 24, and the number of attention heads included in the multi-head self-attention layer is set between 4 and 8.

11. The method according to claim 1, further comprising: The acquired culture data, together with the culture data of the future culture stage determined based on the predicted state of the future culture stage and the actions to be performed in the future culture stage, are provided to the trained neural network model to obtain the state of the bioreactor in a second future culture stage after the future culture stage.

12. The method according to claim 1, wherein: Before providing the acquired culture data to the trained neural network model, the trained neural network model performs small sample learning using culture data of other batches of the cell sample of the bioreactor in the past.

13. A method for training a neural network model for predicting the state of a cell culture in a bioreactor, comprising: Acquire a training data set, the training data set comprising culture data of each culture stage in each culture stage series in a plurality of culture stage series of a plurality of bioreactors; The neural network model is trained using the training data set, wherein a loss function of the neural network model is determined based on a loss generated by a set of subproblems, the set of subproblems corresponding to cultivation data of a series of cultivation stages from the training data set, and each subproblem in the set of subproblems is configured to predict a state of a corresponding cultivation stage in the series of cultivation stages from the training data set using cultivation data of each cultivation stage preceding the corresponding cultivation stage.

14. A computing device comprising: one or more processors; as well as A memory storing computer executable instructions which, when executed by the one or more processors, cause the one or more processors to perform a method according to any one of claims 1 to 13.

15. A non-transitory storage medium having stored thereon computer executable instructions which, when executed by one or more processors, cause the one or more processors to perform the method according to any one of claims 1 to 13.

16. A computer program product comprising instructions which, when executed by a processor, implement the method according to any one of claims 1 to 13.

Citation Information

Cited By

  • Prediction method and device for transfected cell screening, electronic equipment and storage medium

    CN119851773A