CN115985404A - Method and device for monitoring and automatically controlling a bioreactor - Google Patents

Abstract

The present invention provides a method and apparatus for monitoring and automating the control of a bioreactor. Specifically, the method and the device for predicting the indexes of the cell culture fluid in the bioreactor and automatically controlling the bioreactor based on the prediction are included. The prediction comprises the following steps: acquiring online spectral data of a plurality of different time points of cell culture solution indexes and offline target values of the indexes, which are measured by sampling the cell culture solution; constructing a text convolution model by the online spectrum data and the offline target value; and inputting the real-time measured spectral data of the index of the cell culture solution to be predicted into the text convolution model so as to obtain the predicted value of the index.

Description

Method and device for monitoring and automatically controlling a bioreactor

Technical Field

The invention relates to the field of automated control of bioreactors, and in particular to a method and a device for monitoring and automatically controlling a bioreactor.

Background

A bioreactor is a device system which utilizes the function of enzyme or organism (such as cell) in vitro and carries out biochemical reaction, and is a biological function simulator which is widely applied to the production and development of biological medicine, including but not limited to the production of antibody and antigen and other products by using bioreactor to culture cell in suspension. Control of the reaction process becomes critical for the reactants and reaction process types that have been selected. The control of the bioreactor greatly affects the yield, quality and input cost of the biological medicine. In order to maintain the reaction in an optimal environment, various operation parameters of the reaction are monitored and controlled at any time.

The cell culture solution is an important component in the cell culture process of the bioreactor, and the cell culture is widely applied to the production and research and development of biological medicines. The indexes of the cell culture solution greatly determine the regulation and control operation of the bioreactor in the production process of the biological medicine. In the case of culturing cells in a bioreactor and performing production, the problems of the biological drug production company are that the controllable parameters of the bioreactor during the biological cell culture period are more (such as temperature, stirring, aeration, and the addition amount of various components of a filling material), the culture cycle state (components of a culture solution, viable cell density, temperature, pH, and the like) is more, the change is nonlinear, the culture cycle is longer, and how to detect the indexes of the cell culture solution in different time periods and adjust the culture process in each stage in the culture cycle according to the feedback of the detection result to obtain the best benefit.

Specifically, for example, in the development of biopharmaceuticals, a large number of biological reaction experiments are required, and the most reliable cell strain culture operation process is selected according to the experimental results, so that the long-term stable production process, high yield and drug quality are ensured. The biological reaction experiment needs to explore the influence of a large range of experimental parameters (such as sugar, lactic acid, oxygen, amino acid and the like) on cell strains at every moment, so that in each stage of cell production, the experiment parameters of the concentration of a cell culture solution are as close as possible and are kept at the level which is favorable for the growth and the secretion of cells to produce target substances is one of the core processes which need a great deal of experiment cost and time investment in the whole research and development of biological medicines.

In the present stage, many researchers have tried a feedback adjustment method, which uses historical data to predict the components of the cell culture solution by a statistical analysis method or a multivariate fitting mathematical modeling method, and adjusts the experimental parameters in advance according to the prediction result. Such a method obviously has certain drawbacks. First of all it is subject to the accuracy of the prediction, and when the prediction itself does not reach sufficient accuracy, adjustments made based on such predictions will also be problematic. Second, the adjustment behavior, including the timing of adjustments, the experimental parameters adjusted, and the amount of adjustments made based on predictions are currently heavily dependent on the individual judgment of the operator.

Therefore, there is a need for a method for monitoring and automatically controlling a bioreactor, which automatically controls the bioreactor based on a more accurate prediction of bioreactor index, so as to avoid the above-mentioned drawbacks.

Disclosure of Invention

The present inventors predicted indices in bioreactor cell culture fluids using machine learning models, and based on such predictions, adopted control strategies in advance, and adjusted the parameters by various means.

The present invention is based on the above-mentioned findings, and therefore, one aspect of the present invention relates to a method for predicting an index of a cell culture fluid in a bioreactor, comprising the steps of:

1) Acquiring online spectral data of a plurality of different time points of cell culture solution indexes and offline target values of the indexes, which are measured by sampling the cell culture solution;

2) Constructing a text convolution model by the online spectrum data and the offline target value;

3) And inputting the real-time measured spectral data of the index of the cell culture solution to be predicted into the text convolution model so as to obtain the predicted value of the index.

In one embodiment, said building a text convolution model comprises: setting a word vector mapping layer and performing feature dimension reduction; and after taking the mean value on the spatial dimension, obtaining the predicted value of the index by using a full-connection structure.

In one embodiment, after the word vector mapping layer is set and feature dimensionality reduction is performed, neighbor feature sampling is further performed.

In one embodiment, the online spectral data is raman spectral signal data.

In one embodiment, the indicator comprises one or more of: pH, carbon dioxide partial pressure, sodium ion concentration, potassium ion concentration, glucose concentration, lactic acid concentration, ammonium ion concentration, glutamine concentration, glutamic acid concentration, target protein concentration, lactate dehydrogenase concentration, terbium ion concentration, phosphate ion concentration, osmotic pressure, viable cell density, cell viability, viable cell average diameter, and amino acid concentration.

Another aspect of the invention relates to a method of controlling a bioreactor, the method comprising

1) Obtaining real-time measured spectral data of indexes of a cell culture solution in a bioreactor in real time;

2) By adopting the method, the predicted value of the index is obtained by measuring the spectral data obtained in the step 1) in real time;

3) Obtaining a state value according to the predicted value of the index;

4) And carrying out real-time automatic reverse control on the cell culture in the bioreactor through the Internet of things according to the state value.

In one embodiment, the deriving the state value from the predicted value of the indicator comprises one or more of:

1) Carbon dioxide state value: determining a carbon dioxide state value according to the predicted value of the partial pressure of the carbon dioxide;

2) Nutrient status value: determining a nutrient status value from the live cell density prediction value;

3) Osmotic pressure state value: determining an osmotic pressure state value according to the osmotic pressure predicted value;

4) Glucose concentration state value: determining a glucose concentration state value according to the glucose predicted value;

5) Temperature state value: determining a temperature state value according to the predicted value of the density value of the living cells;

6) Aeration stirring state value: and determining the aeration stirring state value according to the live cell density value predicted value.

In one embodiment, the state value is derived by:

1) Multi-value averaging is carried out on the predicted values of the partial pressure of the carbon dioxide to obtain a carbon dioxide state value;

2) Determining the amount of nutrient consumed by the cells according to the average value of the predicted live cell density value, and taking the amount as a nutrient state value;

3) Carrying out multi-value averaging on the predicted osmotic pressure value to obtain an osmotic pressure state value;

4) Multi-value averaging is carried out on the glucose predicted value to obtain a glucose state value;

5) Multivalued averaging the predicted value of the living cell density to obtain a temperature state value;

6) And averaging the predicted live cell density values in multiple values to obtain a ventilation stirring state value.

In one embodiment, the automatic counter-control comprises one or more of the following control strategies:

1) The state value is a carbon dioxide state value, when the carbon dioxide state value is larger than the dead zone control upper limit set by the carbon dioxide partial pressure control module, air is introduced, and when the carbon dioxide state value is smaller than the dead zone control lower limit set by the carbon dioxide partial pressure control module, carbon dioxide is introduced;

2) The state value is a nutrient state value, the nutrient state value is fed back to the feeding device, and corresponding nutrients are fed by using a metering mode;

3) The state value is an osmotic pressure state value, and when the osmotic pressure state value is smaller than an osmotic pressure set value, nutrient substance feeding and/or glucose feeding are carried out;

4) The state value is a glucose concentration state value, and when the glucose concentration state value is smaller than a glucose concentration set value, the glucose feeding is controlled through a proportional-integral control formula;

5) The state value is a temperature state value, and the temperature is controlled through a linear module according to the temperature state value;

6) And the state value is a ventilation stirring state value, and ventilation stirring is controlled through a linear module according to the ventilation stirring state value.

Yet another aspect of the invention relates to an artificial intelligence internet of things device, wherein the device comprises the following equipment: a spectral sensor for monitoring the bioreactor, a GPU edge computing device for implementing the method of the invention, and a bioreactor controller for counter-control;

the real-time measured spectrum data obtained by the spectrum sensor is input into the GPU edge computing equipment, and the bioreactor controller is controlled by the computing result obtained by the GPU edge computing equipment.

Drawings

FIG. 1 is a flow of model building prediction and control illustrated by an embodiment of the present invention.

Fig. 2 is an AIOT smart internet of things device used in an embodiment of the present invention. In the device, a Raman spectrum file in a Raman controller is read through an SMB protocol, an AI algorithm is integrated, and feedback is executed on a Bioreactivor controller through an OPC protocol.

FIG. 3 is the feeding situation of the TextCNN Raman model automation control strategy compared with the traditional manual bioreactor operation strategy.

FIG. 4 is an additional sugar feed in addition to the conventional feed for a culture environment control strategy versus a conventional bioreactor operating strategy.

Detailed Description

Hereinafter, embodiments of the present invention will be described.

As previously described, in the present invention, machine learning models can be used to predict indicators in bioreactor cell culture fluids, and based on such predictions, control strategies can be taken in advance, with early intervention and adjustment of experimental parameters by various means.

The machine learning model can be a plurality of mathematical algorithms such as Partial Least Squares (PLS), cubic derivation (Cubist), random Forests (RF), support Vector Machines (SVM), exponential Smoothing (ES), inverse variance (RV) and the like, and the model can learn the capability of predicting the indexes such as glucose concentration, metal ion concentration, amino acid concentration and the like in the cell culture solution through the training of a large amount of detection data (training set) on the model, so that the machine learning combined model can accurately predict all the indexes of the cell culture solution.

In some cases, predictions can also be made using non-machine learning models, such as fuzzy mathematical rules, chemometric models, multivariate analysis methods of statistical analytics (MVA), principal Component Analysis (PCA), orthogonal Partial Least Squares (OPLS), multivariate regression, canonical correlation, factorial analysis, cluster analysis, graphical methods, multivariate fits, and the like. However, these methods are too simple to analyze or model, depend on a previously built-in judgment rule and a multi-parameter model for judgment, lack self-learning of the model, or cannot adapt to the diversity of biological drugs, the diversity of culture environments, and the diversity of operation methods but only aim at a single behavior, and thus cannot obtain an ideal prediction effect when various factors interact with each other under certain conditions.

Therefore, the present invention preferably predicts the indicator in the bioreactor cell culture fluid by a machine learning model. Among them, the text convolution-TextCNN model is more preferable.

Text convolution is the application of Convolutional Neural Networks (CNN) in text classification. Convolutional neural networks are a deep learning method, and are commonly used in image processing. When used for text processing, a specialized text convolution-TextCNN model is derived. Yoon Kim proposed TextCNN in the paper (2014 EMNLP) volumetric Neural Networks for Session Classification. The core idea of convolutional neural networks is to capture local features, which for text are sliding windows consisting of words, similar to N-grams. The convolutional neural network has the advantage that N-gram features can be automatically combined and screened to obtain semantic information of different abstract levels. The convolutional neural network CNN is applied to a text classification task, and key information in a sentence is extracted by using a plurality of convolutional kernels with different sizes, so that local correlation can be better captured.

Such properties make it well applicable in the context of the present invention. Since the relationship between the respective indices of the culture medium in the present invention is often correlated with each other, for example, an increase in the partial pressure of carbon dioxide may lead to a decrease in the pH value and an increase in the overall pressure of the system. Such parametric properties have similarities to natural language processing with convolutional neural networks.

The present invention uses spectral data as detection data. In one embodiment, the invention uses a raman spectrometer to detect raman spectrum signals in a cell culture solution as detection data (training set) in real time during cell culture, and uses a machine learning model to analyze the raman spectrum data in real time and establish a relation model with each index of the culture solution detected at the corresponding time under line.

The Raman spectrum is a vibration spectrum for detecting and identifying substance molecules by detecting a Raman spectrum generated by a sample to be detected aiming at the Raman scattering effect of exciting light, and can perform nondestructive analysis on chemical components and molecular structures, and the number, frequency shift, band intensity, shape and the like of Raman spectrum bands generated by the substance due to the Raman scattering effect are directly related to the vibration and rotation of molecules. In particular, under certain conditions, the intensity is linear with the concentration of the substance. Therefore, the detection of the structure, the components and the concentration of the substance can be realized. Compared with spectral analysis means such as infrared, near-infrared and ultraviolet fluorescence, the Raman spectrum has outstanding advantages, including: a wide detection range; no damage, rapidness and no pollution; remote testing technology; high detection sensitivity and the like.

Thus, with improvements in laser sampling and detector technology, the use of raman spectroscopy in polymer, pharmaceutical, bio-manufacturing and biomedical analysis has proliferated over the last three decades. Due to the advances in these technologies, raman spectroscopy has now become a practical analytical technique for use inside and outside laboratories. In the field of bioreactor pharmacy, raman spectroscopy is often used for on-line monitoring. Since the first report of the use of in situ raman measurements in bio-manufacturing, they have been used to provide online real-time predictions of several key process states such as glucose, lactate, glutamate, glutamine, ammonia, VCD, etc.

However, according to the principle characteristic of raman spectroscopy, a certain value or a certain number of values in some bands is very sensitive to the concentration change of components and has a very large influence on the prediction result, and the change of some floating point numbers is slight, so that the change needs to be expanded and mapped to a two-dimensional space structure to be convenient for extracting the change. And the affected bands may be fragmented and not fixed. Therefore, the convolutional neural network with fixed convolution kernel can not accurately capture specific features.

Based on the above reasons, the inventor has referred to the model structure of text convolution-TextCNN often used in text processing to extract features mapping neighborhood of different step sizes in feature maps using one-dimensional convolution of different kernel sizes, and has guided the feature relationship of original position neighbors in full join operation to accurately predict indexes of components by using the spatial association constraint of the intermediate feature extraction layer.

In the present invention, items such as concentration, pressure, density, temperature, and content in the bioreactor are collectively referred to as indices.

In the bioreactor of the present invention, the control strategy can be divided into a culture environment control strategy and a feeding strategy.

The culture environment control strategy refers to a strategy for regulating the culture environment, and includes but is not limited to an ambient gas strategy, a temperature control strategy, a pressure control strategy and the like.

The environmental gas strategy comprises but is not limited to a strategy for controlling the content and pressure of various gases which have influence on the growth of cells and the production and secretion of target products, such as oxygen, carbon dioxide, carbon monoxide, nitrogen, argon, ammonia and the like. For example, the control may be directed to the content or pressure of a certain gas therein. In the present invention, the pressure of a certain gas is referred to as partial pressure for the purpose of distinguishing from the pressure control strategy described below. For example, the pressure of oxygen may be referred to as the oxygen partial pressure or the oxygen partial pressure, and the pressure of carbon dioxide may be referred to as the carbon dioxide partial pressure.

Among ambient gas strategies, carbon dioxide partial pressure strategies are often very common and important. This is because carbon dioxide is a metabolite of cells during cell culture, is an essential component for cell growth, and is involved in maintaining the pH of the culture solution. During the cell culture process, the culture solution becomes acidic with the increase of the metabolic release amount of carbon dioxide, and the pH value of the culture solution is maintained by adding an alkaline solution or increasing aeration. On the contrary, the culture solution is alkaline, and carbon dioxide can be actively introduced to maintain the pH value. In the prior manual control, the adjustment of the partial pressure of carbon dioxide is adjusted according to the result of off-line sampling every day and the on-line pH value in the cell culture period, the adjustment result can only maintain the pH range, and the partial pressure of carbon dioxide can only be automatically adjusted along with the change of the cell culture process.

The temperature control strategy may be to maintain the system temperature at a certain temperature value, to increase/decrease the temperature to a certain temperature value at a fixed or varying rate, or to increase/decrease the temperature in various desired ways. Temperature control strategies are generally controlled according to the stage of cell growth, so long as growth of the cells and production and secretion of the desired product are favored. In actual production, the cell culture process is usually a heat generation process, so when the living cell density reaches a certain amount, actively reducing the culture temperature can effectively maintain the cell survival rate and improve the yield of the target protein. However, in the initial stage of cell culture, it may be necessary to raise the temperature to an initial temperature favorable for cell growth. In the prior manual control, a time period when the living cell density reaches a target is predicted according to an empirical formula, the sample density is increased in the time period, and when the offline detection data reaches a specified density, the temperature is manually reduced.

The feeding strategy means that the nutrient consumption rate of the cells during the cell culture process carried out in the bioreactor needs to be supplemented with the corresponding nutrients so that it is sufficient to provide the nutrients required for the growth of the cells and the production secretion of the target product, and not to inhibit the growth of the cells and the production secretion of the target product in return due to an excess of nutrients. These nutrients include, but are not limited to, glucose, terbium ions, glutamine, sodium ions, potassium ions, lactic acid, ammonium ions, glutamic acid, lactate dehydrogenase, phosphate ions, water, and the like. In the conventional manual control, because off-line detection data fed back in real time does not exist, the feeding strategy can only set the feeding times, feeding ratio and glucose concentration in the cell culture period according to historical data and experience of experimenters. After the operator calculates the feeding parameters according to an empirical formula, the operator manually supplements the feeding into the bioreactor by depending on the scale and the peristaltic pump.

As described above, on the basis of accurately predicting each index of the cell culture solution, the invention determines a plurality of state values of a plurality of indexes output by the model through the method of the Internet of things, and associates the plurality of state values with various control measures such as feeding, ventilation and the like to control the target index.

In this specification, the target index may be the same as or different from an index for acquiring the real-time measured spectral data, predicted value, and state value.

The correspondence between the predicted value and the state value of the index output by the partial prediction model and a specific control strategy thereof will be described below.

In the present invention, a state value refers to a value influenced by a control strategy. The state values in the present invention include, but are not limited to:

carbon dioxide state value: and averaging the component predicted values of a plurality of carbon dioxide partial pressures to obtain a carbon dioxide state value. The predicted values of the components of the plurality of partial pressures of carbon dioxide may be a plurality of predicted values obtained from spectral data of one sensor over a period of timeOr a plurality of predicted values obtained from spectral data of a plurality of sensors over a period of time. The period of time and the number of selected predicted values can be determined according to actual needs. In one embodiment, the period of time is 10 hours, 9 hours, 8 hours, 7 hours, 6 hours, 5 hours, 4 hours, 3 hours, 2 hours, 100 minutes, 1 hour, 30 minutes. In one embodiment, the number of predictors selected may be 10, 9, 8, 7, 6, 5, 4, 3, 2.

Carbon dioxide partial pressure strategy

And determining a carbon dioxide partial pressure state value through a predicted value of the carbon dioxide partial pressure. The carbon dioxide partial pressure state value is associated to the carbon dioxide partial pressure control module, the carbon dioxide partial pressure control module is carried out in a dead zone control mode, when the calculated mean value of the carbon dioxide partial pressure is larger than the upper limit of the dead zone control, the electromagnetic valve is started to introduce air to exchange redundant carbon dioxide gas out of the cell culture solution, the real-time value of the carbon dioxide partial pressure is actively reduced, when the calculated mean value of the carbon dioxide partial pressure is smaller than the lower limit of the dead zone control, the electromagnetic valve is started to introduce carbon dioxide, and the carbon dioxide partial pressure value of the cell culture solution is actively increased. The method of dead zone control is matched with gas exchange to effectively and automatically control the partial pressure of the carbon dioxide.

Status value of nutrient: determining the amount of the nutrient consumed by the cells over a period of time based on a mean of a plurality of predicted values of viable cell density over the period of time. Specifically, the average value of the predicted live cell density values in a period of time is firstly obtained, the amount of nutrient substances consumed by the cells in the period of time is calculated according to a formula, the amount of the nutrient substances consumed by the cells is used as a nutrient substance state value, and the nutrient substances are automatically supplemented by controlling a material supplementing device according to the nutrient substance state value. The formula may be an empirical formula or an online formula that varies in real time.

On the other hand, in the case where the nutrient to be supplemented is an amino acid nutrient solution, the average of predicted values of amino acid concentration over a period of time is determined, and the average of predicted values of amino acid concentration is used as a nutrient state value, and the nutrient is automatically supplemented by controlling a supplementing device using the nutrient state value.

In one embodiment, the period of time is 10 hours, 9 hours, 8 hours, 7 hours, 6 hours, 5 hours, 4 hours, 3 hours, 2 hours, 100 minutes, 1 hour, 30 minutes. In one embodiment, the feeding apparatus may comprise a peristaltic pump, with automatic feeding using a metering mode. The nutrient is a generic term for a variety of substances that can be used in cell culture for biological reactions. For different nutrients, the state value of the nutrient can be obtained from the predicted value of the viable cell density, and the nutrient is fed. The nutrient may be commercial formula such as Cytiva 7a/7b nutrient supplement liquid, gibico FM016 nutrient supplement liquid, or self-prepared nutrient liquid such as single amino acid nutrient liquid, polyamino acid nutrient liquid. There are two feeding strategies for nutrient status values, which are described separately below:

nutrient supplement strategy 1)

Nutrient feeding strategies based on the rate of nutrient metabolism by the number of living cells. Specifically, the amount of nutrients consumed by a single living cell in a unit time (e.g., 5 minutes, the same as the raman detection time) reactor is calculated according to the live cell density value predicted by the real-time measured spectral data, and the feeding device calculates the amount of nutrients consumed according to the formula as described above and feeds the corresponding nutrients. This operation can be repeated in the next feedback cycle until the incubation is complete. In this way, feedback automatic regulation of the feed is achieved.

Nutrient supplement strategy 2)

Under the condition that the nutrient is single amino acid nutrient solution, the concentration of the amino acid is predicted according to the real-time measured spectral data, and the corresponding amino acid nutrient solution is supplemented to a proper set value through a supplementing device.

In both of the above feeding strategies, a plurality of predicted values of viable cell density are selected over a period of time to determine the amount of nutrients consumed by the cells over the period of time, and the amount of consumed nutrients is fed back to the feeding apparatus for feeding. The period of time and the number of values selected may be determined according to actual needs. In one embodiment, the period of time is 10 hours, 9 hours, 8 hours, 7 hours, 6 hours, 5 hours, 4 hours, 3 hours, 2 hours, 100 minutes, 1 hour, 30 minutes. In one embodiment, the number of values selected may be 10, 9, 8, 7, 6, 5, 4, 3, 2. In one embodiment, the feeding apparatus may comprise a peristaltic pump. In one embodiment, the peristaltic pump utilizes a metering mode, calculates the amount of make-up and automatically supplements. The skilled person can select the desired one according to the actual need.

Osmotic pressure status value and glucose concentration status value: for industrial convenience, in production, as described above, a known formulation or a commercially available medium or nutrient solution is often used for feeding. This results in the possible need for additional glucose addition. At the same time, both the nutrient concentration and the glucose concentration in the bioreactor influence the osmotic pressure in a coupled manner, so the invention also concerns osmotic pressure and glucose concentration. In the present invention, a plurality of glucose concentration predicted values are selected over a period of time, and the average value thereof is used as a glucose concentration state value. Similarly, a plurality of predicted osmotic pressure values are selected over a period of time, and the average value thereof is taken as an osmotic pressure state value.

And averaging the multiple predicted osmotic pressure values to obtain an osmotic pressure state value. The plurality of predicted values of osmotic pressure and the plurality of predicted values of glucose concentration may be a plurality of predicted values obtained from data of one sensor over a period of time, or a plurality of predicted values obtained from data of a plurality of sensors over a period of time. The period of time and the number of the selected osmotic pressure predicted value and the glucose concentration predicted value can be determined according to actual needs. In one embodiment, the period of time is 10 hours, 9 hours, 8 hours, 7 hours, 6 hours, 5 hours, 4 hours, 3 hours, 2 hours, 100 minutes, 1 hour, 30 minutes. In one embodiment, the number of osmolarity predictors and glucose concentrations chosen may be 10, 9, 8, 7, 6, 5, 4, 3, 2.

Osmotic pressure and glucose concentrationControl strategy

First, according to actual needs, a predetermined osmotic pressure set value and a glucose concentration set value are given. The osmotic pressure set point and the glucose concentration set point are predetermined, beneficial to the biological reaction and cell growth production osmotic pressure and glucose concentration, and can be determined according to actual needs. The person skilled in the art knows the osmotic pressure and the glucose concentration required for a particular biological reaction.

When the glucose concentration state value is larger than a preset glucose concentration set value, the output power of a control module of the material supplementing device is zero, and when the average value obtained by calculating the measured spectral data in real time is smaller than the glucose concentration set value, the control module controls the material supplementing device to supplement the glucose. When the osmotic pressure state value is larger than a preset osmotic pressure set value, the output power of a control module of the feeding device is zero, and when the mean value obtained by calculating the measured spectral data in real time is smaller than a glucose concentration set value, the control module controls the feeding device to feed nutrient substances or glucose. The output power of the control module of the feeding device operates according to control modes such as a proportion integration formula PID, linear control, logic negation, formula calculation and the like, and the proportion integration formula is optimized. In one embodiment, the feeding device may comprise a peristaltic pump.

Temperature state value and aeration stirring state value:

in the process of cell culture, the temperature control operation and the aeration stirring operation are required to stabilize the cell growth state and improve the yield of the target product. Therefore, under the setting of the linearization formula module, when the average value of the predicted values of the viable cell density reaches the set value, the temperature and the ventilation stirring state are automatically adjusted, and the adjustment can be temperature rise or temperature reduction.

In the present invention, a plurality of predicted viable cell density values are selected over a period of time, and the average value thereof is used as a temperature state value and an aeration-agitation state value. The plurality of live cell density prediction values may be a plurality of prediction values obtained from data of one sensor over a period of time, or may be a plurality of prediction values obtained from data of a plurality of sensors over a period of time. The period of time and the number of viable cell density predictors selected may be determined according to actual requirements. In one embodiment, the period of time is 10 hours, 9 hours, 8 hours, 7 hours, 6 hours, 5 hours, 4 hours, 3 hours, 2 hours, 100 minutes, 1 hour, 30 minutes. In one embodiment, the number of viable cell density predictors selected may be 10, 9, 8, 7, 6, 5, 4, 3, 2.

Temperature control strategy and aeration stirring strategy

First, according to actual needs, a preset living cell density set value is given. The living cell density set value is a predetermined living cell density value which is beneficial to biological reaction and cell growth production and can be determined according to actual requirements. One skilled in the art knows the viable cell density required for a particular biological response.

Then, when the temperature state value, that is, the average value of the predicted values of the density of the living cells selected within a period of time is greater than the maximum value of the set value of the density of the living cells in the experimental design, the linear parameter control module automatically executes the temperature control operation to adjust the temperature to the preset ideal reaction temperature. In addition, the temperature state value is associated with the X value of the linear parameter control module, and the Y value of the linear parameter control module is associated with the real-time value of the temperature module, so that the temperature is inversely controlled.

And when the aeration stirring state value, namely the average value of a plurality of live cell density predicted values selected within a period of time is larger than the maximum value of the live cell density set value of the experimental design, the control module adjusts the aeration stirring state. For example, if the aeration agitation state value is larger than the maximum value of the set value of the viable cell density of the experimental design, the aeration is increased, the agitation is started, or the agitation is increased. The real-time value of the oxygen ventilation is determined according to the difference between the ventilation agitation state value and the set value of the density of the living cells.

The aeration is performed by associating the stirred aeration control module to the real-time value of oxygen aeration based on the real-time oxygen aeration. When the aeration reaches the upper limit of the equipment and the oxygen demand for maintaining the cell growth cannot be continuously provided, the oxygen demand of the cells can be reduced only by adjusting the stirring rotating speed. When the ventilation volume reaches the upper limit value, the stirring speed is automatically increased; when the aeration reaches the lower limit value, the stirring rotation is automatically reduced to maintain the aeration within a proper range.

The cell culture control methods and systems of the present aspects can use any suitable bioreactor. For example, the bioreactor may include a fermentor, a stirred tank reactor, a wall-mounted bioreactor, a wave-type bioreactor, a disposable bioreactor, and the like.

The bioreactor can be made of a variety of different materials. For example, in some embodiments, the bioreactor can be made of metal (e.g., stainless steel). Metal bioreactors are typically designed to be reusable. Alternatively, the bioreactor may comprise a disposable bioreactor made of a rigid polymer or a flexible polymer membrane. For example, when made of a rigid polymer, the bioreactor wall can be freestanding. Alternatively, the bioreactor can be made of a flexible polymeric membrane or a shape-conforming material, which can be liquid impermeable and can have an internal hydrophilic surface.

The bioreactor may have any suitable volume. In particular, in one embodiment, the bioreactor has a volume suitable for small-scale laboratory production, pilot scale up, or actual large scale production.

The bioreactor may have various additional devices, such as, for example, stirring devices, baffles, bubblers, gas supplies, heat exchangers or thermocycler ports, etc., which allow for the culturing and propagation of biological cells.

In some embodiments, the cell is a eukaryotic cell (e.g., a mammalian cell) or a prokaryotic cell. The mammalian cell can be, for example, a human or rodent or bovine cell line or cell strain. Examples of such cells, cell lines or cell strains are e.g.mouse myeloma (NSO) -cell lines, chinese Hamster Ovary (CHO) -cell lines, HT1080, H9, hepG2, MCF7, MDBK Jurkat, NIH3T3, PC12, BHK (baby hamster kidney cells), VERO, SP2/0, YB2/0, Y0, C127, L cells, COS (e.g.COS 1 and COS 7), QC1-3, HEK-293, VERO, PER. C6, heLA, EBl, EB2, EB3, oncolytic or hybridoma cell lines. Preferably, the mammalian cell is a CHO cell line. In some embodiments, the cell is a CHO cell.

In some embodiments, the cell culture product (an expression or secretion of a cell (e.g., a recombinant therapeutic or diagnostic product)). Examples of the product of the cell culture include, but are not limited to, antibody molecules (e.g., monoclonal antibodies, bispecific antibodies), antibody mimetics (polypeptide molecules that specifically bind to an antigen but are structurally unrelated to an antibody (e.g., DARPin, affibody, adnectin, or IgNAR)), fusion proteins (e.g., fc fusion proteins, chimeric cytokines), other recombinant proteins (e.g., glycosylated proteins, enzymes, hormones), viral therapeutic agents (e.g., anti-cancer oncolytic viruses, viral vectors for gene therapy and viral immunotherapy), cellular therapeutic agents (e.g., pluripotent stem cells, mesenchymal stem cells, and adult stem cells), vaccines or lipid-encapsulated particles (e.g., exosomes, virus-like particles), RNA (e.g., siRNA), or DNA (e.g., plasmid DNA), antibiotics, or amino acids. In some embodiments, the devices, apparatuses, and methods can be used to produce biosimilar drugs.

The method of the present invention can be used in various bioreactor cell cultures without being limited to a process specifically used thereof, as long as it is advantageous for cell growth and secretion of a target substance. However, in various bioreactor cell culture processes, it is preferred to use the method of the invention in one or more of the following culture processes: conventional batch feed process (TFB), intensive batch feed process (IFB), concentrated batch feed process (CFB), continuous Perfusion process (Perfusion).

The Internet of Things (Internet of Things, ioT for short) is to collect any object or process needing monitoring, connection and interaction in real time and collect various required information such as sound, light, heat, electricity, mechanics, chemistry, biology and location through various devices and technologies such as various information sensors, radio frequency identification technologies, global positioning systems, infrared sensors and laser scanners, and to realize ubiquitous connection of objects and people through various possible network accesses, and to realize intelligent sensing, identification and management of objects and processes.

Examples

The method of the present invention is further specifically described below by way of examples.

Establishing accurate Raman spectrum model and verifying

First, raman spectral data and off-line target values are collected. Obtaining online spectrum data of different indexes changing according to time from a cell culture solution of a bioreactor by using a Raman spectrometer; and calculating the off-line target value of the index according to time, and sampling at regular time in the experimental process and obtaining the off-line target value of each batch by using the off-line detection equipment so as to correct and calculate the off-line target value.

Specifically, the present invention uses a monoclonal antibody-producing cell line (Chinese hamster ovary cells) in cell culture. The culture conditions of this cell line in a seed stage shaker (Kuhner) are 36.5 ℃, 110rpm,6% CO ₂ The concentration level is that Hyclone Actipro medium of Cytiva company is used, 3L and 250L reactors are used in the production and culture stage, the initial culture volumes are respectively 1.5L and 140L, the culture temperature is 36.5 ℃, the pH value is set to be 6.90+0.25, the dissolved oxygen saturation is 40%, and the initial inoculation density is 1.0 multiplied by 10 ⁶ cells/mL, and the feed medium was Hyclone Cell boost 7a/7b (10%/1%) from Cytiva.

The online spectrum data refers to full Raman shift spectrum data collected by a Raman spectrometer. The present invention uses a raman Rxn2 analyzer (Kaiser Optical Systems) equipped with an immersion Optical probe. The probes were mounted in a 3L bioreactor (Applikon) and immersed directly into the cell culture suspension. Raman spectra of the different bioreactors were recorded throughout the experiment. For a single recorded spectrum, 30 subsequent spectra were captured with a 10s exposure time and averaged, resulting in an acquisition interval of approximately 5min per bioreactor. The laser excitation wavelength was 785nm, providing a spectral coverage (Raman shift) of 100-3425 cm-1.

The off-line target value refers to a cell culture solution index obtained by real-time field sample detection. In the production and culture stage, the Raman equipment probe is put into the culture solution. Samples were taken 5 times per day by time and viable Cell density was measured using Vi-Cell XR from Beckman, glucose, lactate and target protein concentrations using a Cedex Bio analyzer from Roche, and amino acid concentrations using HPLC from Agilent.

Secondly, a machine learning model is established. The process of establishing the machine learning model comprises the following steps: data processing, model building, data parameter adjustment and model verification.

Data processing: and (3) reading spectral data, wherein each sample comprises a spectral file and an off-line target value, the spectral files are converted into spectral numerical values, and the detection time of the spectrum corresponds to the detection time of the off-line target value one by one to obtain the characteristic data (spectral values corresponding to different Raman shifts) and the target value of the model. And (3) data preprocessing, namely randomly splitting training data verification data and test data, standardizing and normalizing the characteristic values of the training data, and adapting the test data to a training data standardization rule for standardization.

Building a model: and (3) taking the spectral data with different sizes and lengths as input to construct various machine learning models. The machine learning model established in this embodiment includes partial least squares regression (PLS), xgboost, convolutional Neural Network (CNN), residual error network (Resnet), and text convolution (TextCNN).

Data parameter adjustment: and training the model by using the training set, and adjusting the hyper-parameters of the model according to the performance of the verification set. The specific parameters of each model are as follows:

PLS- -filtering with a Savitzky-Golay filter and then taking the second derivative, and inputting the derivative into a model, wherein n _ components is 5.

Xgboost- -first normalize the spectral data, PCA reduces the dimensions to 50,n _ u components

Set to 500.

A Convolution Neural Network (CNN) model-1-dimensional convolution superposition maximum pooling layer, and outputting results after connecting two full-connection layers after repeating for 4 times.

Residual neural network (Resnet) -not less than 2 layers of residual network layers, each of which contains one-dimensional convolutional layers, maximum pooling layers, and full-link layers.

Text convolution (TextCNN) -when building a text convolution, the following further applies:

1) Setting a word vector mapping layer: and mapping the input one-dimensional floating point number vector into two dimensions of a specified length and a characteristic dimension. In the mapping, because each Raman displacement value only has one floating point number, and the size of a specific numerical value needs to be considered for feature extraction, a word vector matrix cannot be used, so that a full-connection structure is used for completing feature mapping to establish a fixed mapping relation between input and output, and a spatial position structure is also reserved to a certain extent while feature mapping is completed.

2) And (3) performing characteristic dimension reduction: the input floating point number vector is a feature vector of 3000xN (N is a feature mapping dimension) generated by directly mapping a spectrum number value with a length of 3000, so that the dimension generated after mapping needs to be compressed, and a large amount of operation consumption is prevented from being introduced in the initial operation stage. In the specific implementation, the feature dimension after the predefined mapping is AxN, the starting A is the length of the mapping feature vector, and N is the feature dimension. Original input vectors are directly mapped into length AxN vectors through a full connection structure, then reshape operation is used for generating mapping characteristics, and the characteristics are mapped while characteristic lengths are compressed.

3) Neighbor feature sampling: according to the requirement, after the word vector mapping layer is set and feature dimensionality reduction is carried out, neighbor feature sampling can be further carried out. Since different substances are known to correspond to one or several segments of feature regions in the spectrum, feature extraction and feature extraction in all input dimensions in the operation are not effective, and efficient feature extraction requires that feature operations can be performed between adjacent spectra. Because the full-connection operation as global feature integration disturbs the original spatial structure distribution in setting the word vector mapping layer and performing feature dimension reduction, spatial association needs to be introduced again in the subsequent structure. Specifically, with reference to a textCNN structure, one-dimensional convolution with different kernel sizes is used to extract features of neighborhoods with different step sizes in a mapping feature map, and the spatial association constraint of an intermediate feature extraction layer is used to guide the feature relationship of original position neighbors in full-connection operation.

4) And after the average value is taken on the spatial dimension in prediction, predicting the corresponding target numerical value by using the full-connection structure.

Firstly, the model is taken to be verified to obtain a verification result of the test set. Furthermore, to ensure that the model production is available, we finally selected prospective wet experiments to evaluate all model effects.

Specifically, the wet experimental procedure is described as follows: under the culture conditions the same as historical training data, 2 Raman sensors scan 2 bioreactors regularly to collect Raman spectrum data, 4 points are collected in one day, samples are collected at the same time to collect offline data, a complete culture period is collected, and 140 data are collected in total. And (5) counting the rmse of the off-line data and the on-line data. The results are shown in Table 1. Through comparison of multiple models of various parameters, textCNN is found to have obvious superiority in most indexes.

Based on the above results, the embodiment selects the TextCNN model with excellent prediction effect for deployment of edge devices, and also integrates the following devices for monitoring and counter-controlling the bioreactor: an artificial intelligence internet of things device (AIOT, see figure 2) edge computing gateway is formed by a Raman (Raman) sensor and a Bioreactor (Bioreactor) controller, raman spectrum files in the Raman industrial personal computer are read in real time through an SMB protocol, various indexes are predicted by combining with an artificial intelligence algorithm sinking in the edge computing gateway, state values are calculated by concentration, and then corresponding strategies are issued by the edge computing gateway to control the Bioreactor controller.

TABLE 1 model prediction values and off-line actual measurement values in cell culture media rmse

/>

TextCNN Raman model automation control strategy

Based on the comparison of the predicted effects in the previous step, the inventors have selected a TextCNN model based on raman spectral data. After the model is obtained, an automation control strategy based on the model is adopted. Specifically, a state value is generated by analyzing indexes predicted by a Raman spectrum in real time through a TextCNN model deployed on an AI platform, and the state value corresponds to the operation of an automatic control bioreactor, so that the bioreactor is controlled.

The specific environmental control strategy is described in the embodiments. In the carbon dioxide partial pressure strategy, the carbon dioxide partial pressure state value is assigned by a predicted value of the carbon dioxide partial pressure, for example, in this case, 5 points per 100 minutes. And when the calculated mean value of the carbon dioxide partial pressure is smaller than the lower limit of the dead zone control, the electromagnetic valve is started to introduce carbon dioxide, so that the carbon dioxide partial pressure value of the cell culture solution is actively increased. The dead zone control method is matched with gas exchange to effectively and automatically control the partial pressure of the carbon dioxide.

For the cooling strategy, the mean value obtained by calculating the live cell density parameter is associated with the X value of the linear parameter control module, and the Y value of the linear parameter module is associated with the real-time value of the temperature module. And setting the temperature parameters after cooling, and automatically executing cooling operation by the linear parameter control module when the calculated average value of the living cell density is greater than the maximum value of the experimental design.

There are additionally three feeding strategies.

Feeding strategy 1 relates to a feeding strategy of cells which is related to the rate of nutrient metabolism of viable cell numbers. Calculating the amount of nutrient consumed by the cells in 100 minutes according to the average value of the viable cell density parameter fed back every 100 minutes, feeding back the calculated value to a peristaltic pump, automatically supplementing the calculated supplement amount by using a metering mode through the peristaltic pump, and repeating the operation in the next feedback period until the culture is finished. Realize the feedback automatic adjustment of the feeding.

The feeding strategy 2 relates to the concentration of nutrients in the bioreactor, which can be reflected to the osmotic pressure concentration of the cell culture solution in a representation manner, the mean value of the analytic calculation is related to the real-time value of the osmotic pressure control module, the output power of the control module is related to the peristaltic pump through PID control, and the feeding operation is automatically operated in real time according to the mean value of the analytic data.

Feed strategy 3 involved glucose. In this embodiment, the mean value calculated from the glucose concentration parameter is correlated with the real-time value of the glucose control module, which operates in a PID control mode, and the peristaltic pump is then correlated to the output power of the control module. Firstly, a set value of the glucose concentration of the cell culture solution is given, when the feedback mean value is larger than the set value, the output power is zero, and when the feedback mean value is smaller than the set value, the output power operates according to PID. The glucose concentration in the bioreactor was automatically controlled back according to the above model.

Finally, a schematic diagram comparing the culture environment control strategy of the TextCNN raman model automated control strategy with the conventional manual bioreactor operation strategy is obtained, which is shown in fig. 3. According to FIG. 3, the total amount of the traditional manual feeding is 84g (black column representation) in 3 days, the total amount of the real-time feeding of the automatic reverse control feeding is almost 60.59g (orange column representation) in every 20 minutes, and the yield of the target protein under the automatic control strategy of the TextCNN Raman model is higher than that under the operation strategy of the traditional manual bioreactor under the condition of reduced feeding amount.

TABLE 2 Total feed and target protein contrast for TextCNN Raman model automation control strategy and conventional manual bioreactor operating strategy

Fig. 4 (a) shows the results of glucose feeding compared to the culture environment control strategy of the TextCNN raman model automation control strategy. The reverse-controlled glucose concentration was compared to the manually-controlled glucose concentration over a 3 day period. The glucose controlled inversely followed the set value of 3.0g/L. Manually controlled glucose follows the 3-6g range of the experimental design. The total amount of sugar is manually controlled to be 6g only for 1 time within 3 days, and the total amount of sugar is automatically and repeatedly supplemented by a small amount of 3.11g, so that the sugar supplement amount is reduced while the real-time control is realized. The yield of the target protein is increased instead.

Claims (10)

1. A method of predicting an indicator of cell broth in a bioreactor comprising the steps of:

1) Acquiring online spectral data of a plurality of different time points of cell culture solution indexes and offline target values of the indexes, which are measured by sampling the cell culture solution;

2) Constructing a text convolution model by the online spectrum data and the offline target value;

3) And inputting the real-time measured spectral data of the index of the cell culture solution to be predicted into the text convolution model so as to obtain the predicted value of the index.

2. The method of claim 1, said building a text convolution model comprising:

setting a word vector mapping layer and performing feature dimension reduction; and

and after the average value is taken on the spatial dimension, obtaining the predicted value of the index by using a full-connection structure.

3. The method of claim 2, further performing neighbor feature sampling after setting a word vector mapping layer and performing feature dimensionality reduction.

4. The method of claim 1, wherein the online spectral data is raman spectral signal data.

5. The method of claim 1, the indicator comprising one or more of: pH, carbon dioxide partial pressure, sodium ion concentration, potassium ion concentration, glucose concentration, lactic acid concentration, ammonium ion concentration, glutamine concentration, glutamic acid concentration, target protein concentration, lactate dehydrogenase concentration, terbium ion concentration, phosphate ion concentration, osmotic pressure, viable cell density, cell viability, viable cell average diameter, and amino acid concentration.

6. A method of controlling a bioreactor, the method comprising

1) Obtaining real-time measured spectral data of indexes of a cell culture solution in a bioreactor in real time;

2) Obtaining a predicted value of the indicator from the real-time measured spectral data obtained in step 1) using the method of claim 1;

3) Obtaining a state value according to the predicted value of the index;

4) And carrying out real-time automatic reverse control on the cell culture in the bioreactor through the Internet of things according to the state value.

7. The method of claim 6, wherein the deriving the state value from the predicted value of the indicator comprises one or more of:

1) Carbon dioxide state value: determining a carbon dioxide state value according to the predicted value of the partial pressure of the carbon dioxide;

2) Nutrient status value: determining a nutrient status value from the live cell density prediction value;

3) Osmotic pressure state value: determining an osmotic pressure state value according to the osmotic pressure predicted value;

4) Glucose concentration state value: determining a glucose concentration state value according to the glucose predicted value;

5) Temperature state value: determining a temperature state value according to the predicted value of the density value of the living cells;

6) Aeration stirring state value: and determining the aeration stirring state value according to the predicted value of the viable cell density value.

8. The method of claim 7, wherein the state value is derived by:

1) Multi-value averaging is carried out on the predicted values of the partial pressure of the carbon dioxide to obtain a carbon dioxide state value;

2) Determining the amount of nutrients consumed by the cells according to the average value of the predicted values of the density of the living cells, and taking the amount as the state value of the nutrients

3) Carrying out multi-value averaging on the predicted osmotic pressure value to obtain an osmotic pressure state value;

4) Multi-value averaging is carried out on the glucose predicted value to obtain a glucose state value;

5) Carrying out multi-value averaging on the predicted live cell density value to obtain a temperature state value;

6) And carrying out multi-value averaging on the predicted values of the density of the living cells to obtain values of the aeration stirring state.

9. The method of claim 6, wherein the automatic counter-control comprises one or more of the following control strategies:

1) The state value is a carbon dioxide state value, when the carbon dioxide state value is larger than the dead zone control upper limit set by the carbon dioxide partial pressure control module, air is introduced, and when the carbon dioxide state value is smaller than the dead zone control lower limit set by the carbon dioxide partial pressure control module, carbon dioxide is introduced;

2) The state value is a nutrient state value, the nutrient state value is fed back to the feeding device, and corresponding nutrients are fed by using a metering mode;

3) The state value is an osmotic pressure state value, and when the osmotic pressure state value is smaller than an osmotic pressure set value, nutrient substance feeding and/or glucose feeding are carried out;

4) The state value is a glucose concentration state value, and when the glucose concentration state value is smaller than a glucose concentration set value, the glucose feeding is controlled through a proportional-integral control formula;

5) The state value is a temperature state value, and the temperature is controlled through a linear module according to the temperature state value;

6) And the state value is a ventilation stirring state value, and ventilation stirring is controlled through a linear module according to the ventilation stirring state value.

10. An artificial intelligence thing networking device, its characterized in that, the device includes following equipment: a spectral sensor for monitoring a bioreactor, a GPU edge computing device for implementing the method of claim 6, and a bioreactor controller for counter-control;

the real-time measured spectrum data obtained by the spectrum sensor is input into the GPU edge computing equipment, and the bioreactor controller is controlled by the computing result obtained by the GPU edge computing equipment.

Images