Control variable optimization method, bioresource production method, and bioresource production system

Abstract

The present invention provides a control variable optimization method that can determine more optimal culture conditions by means of a predictive model using machine learning, as well as a bioresource production method and a bioresource production system using the same. A bioresource production system S according to another aspect of the present invention comprises a culture system B that carries out bioresource production and a control variable optimization system A that carries out optimization of control variables acquired from the culture system, wherein the control variable optimization system separates control variables into initial variables and fluctuating variables, creates a predictive model in a form that adapts to each of the initial variables and the fluctuating variables, and combines the two to optimize the control variables.

Description

Control variable optimization method, bioresource production method, and bioresource production system

[0001] The present invention relates to a control variable optimization method, a bioresource production method, and a bioresource production system.

[0002] Bioproduction systems play an important role in fermentation technology, which uses microorganisms to produce useful proteins and antibiotics; cell culture technology, which cultivates animal and plant cells to produce pharmaceuticals and other products; and algae culture technology, which cultivates microalgae and produces useful organic matter through photosynthesis from atmospheric carbon dioxide.

[0003] Known technologies related to biological production systems include, for example, mathematical models (Non-Patent Document 1 below), experimental design (Non-Patent Document 2 below), and dimension reduction (Non-Patent Document 3 below).

[0004] In the mathematical model shown in Non-Patent Document 1 below, a dynamic mathematical model is constructed consisting of antibody production amount, viable cell number, apoptotic cell number, dead cell number, culture volume, and pH. The parameters of this mathematical model are estimated based on the results of culture tests with different pH values, and the optimal control value of pH is finally predicted using the mathematical model with estimated parameters.

[0005] Furthermore, the experimental design method shown in Non-Patent Document 2 below discloses a technique in which several medium components and fermentation parameters are optimized using a greedy method or the like, important factors are screened, and then the results are optimized using an experimental design or the like.

[0006] In addition, the dimension reduction described in Non-Patent Document 3 below combines Boruta, a dimension reduction method, with variable importance (VIP) in PLS, and applies this to near-infrared spectral data, and predicts lactate and glucose concentrations from the dimension-reduced near-infrared spectral data using a machine learning model.

[0007] Optimization of a pH-shift control strategy for producing monoclonal antibodies in Chinese hamster overary cell cultures using a pH-dependent dynamic model, T. Hogiri et al. , J. Biosci. Bioeng. , 2018, 125(2) Optimization of heterologous production of Bacillus ligniniphilus L1 laccase in Escherichia coli through statistical design of experiments, I. Mutanda et al. , Microbiol. Res. 2023, 274Transfer learning and wavelength selection method in NIR spectroscopy to predict glucose and lactate Concentrations in culture media using VIP-Boruta, H. Kaneko et al. , Anal. Sci. Adv. , 2021, 2, 470-479.

[0008] However, mathematical models cannot fully represent the complexities of real-world culture systems, making it difficult to accurately predict the behavior of a culture system using a mathematical model. Additionally, using only one machine learning model makes it difficult to optimize control variables with different qualities.

[0009] Furthermore, while greedy optimization methods are effective when each control variable is independent and does not interact with each other, there is a problem in that the control variables in cultivation often interact with each other, and therefore greedy optimization methods are likely to only obtain local solutions.

[0010] Furthermore, experimental design can optimize even when there are interactions, but when there are many types of control variables or when trying to consider changes to each control variable from moment to moment, the number of combinations becomes enormous, and even experimental design cannot fully optimize (optimization requires a huge number of experiments, but a characteristic of bioproduction is that there is a limit to the number of cultures that can be performed at one time, and each culture can be expensive and take a long time). Also, when there are many control variables, it becomes difficult for humans to fully understand and interpret the experimental results.

[0011] For the reasons mentioned above, experimental design cannot adequately handle optimization problems with many control variables, and currently the number of control variables is often limited, or constant control is used.

[0012] In view of the above, an object of the present invention is to provide a control variable optimization method that can determine more optimal culture conditions using a predictive model that uses machine learning, as well as a bioresource production method and a bioresource production system that use the same.

[0013] A control variable optimization method according to one aspect of the present invention that solves the above-mentioned problems is characterized in that it separates control variables into initial variables and variable variables, creates prediction models that are adapted to each of the initial variables and the variable variables, and optimizes the control variables by combining these two.

[0014] Furthermore, in this respect, although not limited thereto, it is preferable to calculate the priority of the initial variables and variable variables from the gradient of the control variable with respect to the objective variable, and determine the order of optimization of the initial variables and variable variables.

[0015] Furthermore, in this respect, although not limited thereto, it is preferable to optimize the control variables by repeatedly using two models adapted to the initial variables and the variable variables, respectively, and recursively performing inverse analysis until the optimal control values ​​of the initial variables and the variable variables converge.

[0016] Furthermore, in this respect, although not limited thereto, it is preferable that the control variables are optimized by assigning medium temperature and pH to variable variables, and medium components and inoculation amount to initial variables, and combining two models that are adapted to the initial variables and variable variables, respectively.

[0017] Furthermore, in this respect, although not limited to this, it is preferable to perform dimensionality reduction by extracting coefficients that minimize the approximation error by projecting the huge number of explanatory variables used in each model onto a Gompertz curve.

[0018] Furthermore, in this respect, although not limited thereto, it is preferable that when the optimal control values ​​of the initial variables and the variable variables are calculated, the distribution of the predicted values ​​of the control variables and the relative position of the optimal control values ​​with respect to the learning data are presented.

[0019] In addition, in this respect, although not limited thereto, it is preferable to separate the variable variables into time variables so that all the variable variables can be treated in the same manner as the initial variables.

[0020] In addition, a bioresource production method according to another aspect of the present invention separates control variables into initial variables and variable variables, creates predictive models that are adapted to each of the initial variables and variable variables, and combines these two to optimize the control variables, thereby performing cultivation using the optimized control variables.

[0021] In addition, a bioresource production system according to another aspect of the present invention comprises a culture system for bioresource production and a control variable optimization system for optimizing control variables obtained from the culture system, in which the control variable optimization system separates the control variables into initial variables and variable variables, creates predictive models that are adapted to each of the initial variables and the variable variables, and optimizes the control variables by combining these two.

[0022] As described above, the present invention provides a control variable optimization method that can determine more optimal culture conditions using a predictive model that uses machine learning, as well as a bioresource production method and a bioresource production system that use the same.

[0023] 1 is a functional block diagram of the entire bioresource production system according to an embodiment. FIG. 2 is a diagram showing the flow of data processing in the bioresource production method according to an embodiment. FIG. 3 is a diagram showing an image relating to determining the priority of control variables in the bioresource production method according to an embodiment. FIG. 4 is a diagram showing an image of performing dimension reduction in the bioresource production method according to an embodiment. FIG. 5 is a diagram showing an image of determining the validity of an optimal value in the bioresource production method according to an embodiment. FIG. 6 is a diagram showing an outline of the results of actual cultivation in the bioresource production method according to an embodiment. FIG. 7 is a diagram showing an outline of the application of the Gompertz curve when actual cultivation is performed in the bioresource production method according to an embodiment. FIG. 8 is a flowchart showing the calculation of optimal control values ​​in the bioresource production method according to an embodiment. FIG. 9 is a diagram showing functional blocks during real-time control of the entire bioresource production system according to an embodiment. FIG. 10 is a flowchart showing the calculation of optimal control values ​​during real-time control of the bioresource production system according to an embodiment.

[0024] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. However, the present invention can be embodied in many different forms and is not limited to the specific descriptions given in the following embodiments and examples.

[0025] First, FIG. 1 shows the overall functional blocks of a biological resource production system S according to this embodiment (hereinafter referred to as "this system").

[0026] As shown in the figure, this system S comprises a culture system B and a control variable optimization system A that adjusts the control variables for this culture system B. As shown in the figure, this system S cultivates, i.e., produces, a bioresource using the culture system B, while the control variables used in the culture system B are adjusted by the control variable optimization system A.

[0027] As described above, the culture system B in the present system S is a system capable of culturing bioresources. Specifically, the culture system B of the present system S includes at least an incubator B1, a culture control device B2 that controls the incubator B1, various sensors B3, and a sensor control device B4 that controls the sensors.

[0028] The incubator B1 in the present system S is a device including a culture vessel for culturing a bioresource. By introducing a specific target bioresource into the incubator B1, the bioresource can be cultivated and propagated, i.e., produced. The specific structure of the incubator B1 is not limited as long as it has the above-mentioned functions, but may include, for example, a culture vessel such as a jar fermenter, a temperature control device for controlling the temperature of the jar fermenter, a rotation device for rotating the culture vessel, an acid or base supply device for adjusting pH, an oxygen supply device for supplying oxygen, carbon dioxide, etc., a carbon dioxide supply device, etc., and, in some cases, an observation device for observing the inside of the culture vessel. As described above, the culture vessel of the incubator B1 of the present system S is not only introduced with the bioresource to be propagated, but also with a culture solution, medium, etc., as needed.

[0029] Furthermore, in this system S, "bioresources" refers to resources produced from living organisms, and more specifically refers to, but is not limited to, plants, microorganisms, cells, algae, bacteria, viruses, etc. themselves, or useful products produced by their activities.

[0030] Specifically, this system S can be applied to microbial manufacturing, cell-based pharmaceutical production, algae manufacturing, and agriculture. For example, in "microbial manufacturing," the process of producing a substance using microorganisms involves initial variables such as medium components, inoculation volume, metabolic switching time, feed start time, initial culture volume, autoclave time and temperature, feed solution, feed solution concentration, and pH adjuster, as well as variable variables such as temp, pH, Agit, DO, aeration volume, feed rate, temperature dead zone, pressure, and oxygen concentration. Therefore, the microbial manufacturing process can be controlled using the initial variables and variable variables.

[0031] Furthermore, in "cellular pharmaceutical production," for example, in antibody (protein) production using CHO cells (animal cells), there are initial variables such as medium components, seeding amount, metabolic switching time, feed start time, culture medium volume at the start, autoclave time and temperature, feed solution, feed solution concentration, pH adjuster, etc., and variable variables such as temp, pH, Agit, DO, aeration volume, feed rate, temperature dead zone, pressure, oxygen concentration, etc. Therefore, the cellular pharmaceutical production process can be controlled using the initial variables and variable variables.

[0032] Furthermore, in the process of producing substances using algae in "algae production," there are initial variables such as medium components, inoculation amount, metabolic switching time, feed start time, culture solution volume at the start, autoclave time and temperature, feed solution, feed solution concentration, pH adjuster, and variable variables such as temp, pH, Agit, DO, aeration volume, feed rate, temperature dead zone, pressure, oxygen concentration, CO2 supply amount, and light intensity. Therefore, the process of algae production can be controlled using the initial variables and variable variables.

[0033] In "agriculture," for example, in organic and natural farming, there are initial variables such as chemical items (pH, EC, nitrogen ion concentration, phosphate ion concentration, potassium ion concentration), physical items (water content, maximum water holding capacity), and biological items (total bacterial count, total carbon amount, total nitrogen amount, total phosphorus amount, total potassium amount, CN ratio, CP ratio), and variable variables such as the amount of carbon input and nitrogen input from organic materials each year. Therefore, in organic and natural farming, control is possible using the initial variables and variable variables.

[0034] In addition, in the present system S, various sensors B3 are connected to the culture vessel. Here, the term "sensor" refers to a device used to monitor the conditions within and near the culture vessel. Examples of such sensors include, but are not limited to, culture condition sensors that monitor culture conditions, such as a pH sensor for detecting pH, a temperature sensor for detecting temperature, a dissolved oxygen sensor for detecting dissolved oxygen concentration (DO), and an oxygen sensor for detecting oxygen concentration. Furthermore, culture status sensors that monitor culture conditions, such as a carbon dioxide sensor for detecting carbon dioxide concentration, an infrared sensor for detecting infrared rays, an RGB sensor for detecting color, and an electric potential sensor for detecting electric potential, can be used. Using these various sensors makes it possible to detect the culture environment, determine the production status of bioresources, and so forth. Furthermore, after these sensors are detected, the culture conditions for bioresource production can be adjusted by controlling the culture control device B2, etc.

[0035] Furthermore, at least some of the various sensors B3 of the present system S are connected to a culture control device B2. Specifically, the culture control device B2 is connected to culture condition detection sensors, such as a pH sensor, a temperature sensor for detecting temperature, a dissolved oxygen sensor for detecting dissolved oxygen concentration (DO), and an oxygen sensor for detecting oxygen concentration, and the culture control device B2 can change the culture conditions detected by these sensors, such as pH, temperature, dissolved oxygen concentration, and oxygen concentration. This allows the present system S to achieve and maintain optimal culture conditions. As described below, data measured by some of the various sensors B3 is collected by the culture control device B2, and sensor data information (such as jar number, version number, and serial number) is added to the data before being sent to the control variable optimization system A.

[0036] Furthermore, at least some of the various sensors B3 of the present system S are connected to a sensor control device B4. Specifically, the sensor control device B4 is connected to culture status detection sensors that detect the culture status, such as a carbon dioxide sensor for detecting carbon dioxide concentration, an infrared sensor for detecting infrared rays, an RGB sensor for detecting color, and a potential sensor for detecting potential. By connecting to the above culture status detection sensors, the sensor control device B4 can perform the necessary control of these sensors and transmit acquired data to devices outside the culture vessel. As a result, not only can the culture status be confirmed, but the data can also be used as data for machine learning, as described below. As described below, data measured by some of the various sensors B3 is collected by the sensor control device B4, and sensor data information (such as jar number, version number, and serial number) is added and sent to the control variable optimization system A.

[0037] The system S also includes a control variable optimization system A connected to the culture system B. The control variable optimization system A can optimize the control variables contained therein based on the outputs from the culture condition detection sensor and the culture status detection sensor. Details of this will be described later.

[0038] In addition, in this system S, the culture system B and the control variable optimization system A are connected so as to enable the exchange of data, but they may be connected directly by wiring such as a cable, or may be connected via a LAN (Local Area Network) or WAN (Wide Area Network). In the example shown in this figure, the culture control device B2 and the sensor control device B4 are connected to the LAN via Wi-Fi or the like, and further connected to the control variable optimization system A via the WAN. A specific example of the connection between the LAN, WAN, etc. and various devices is a communication device such as a modem.

[0039] Furthermore, the present system S may further include a control computer C. By providing the control computer C and connecting it to the culture system B, the control variable optimization system A, etc., it becomes possible to control these as a whole.

[0040] In this system S, the control computer C can be literally implemented, and the control variable optimization system A can be implemented using a so-called computer. The computer is a device that can perform predetermined processing based on input data and output data, and is preferably configured to include, for example, components of a general information processing device, such as a central processing unit (CPU), a non-volatile recording medium such as a hard disk or flash memory, a volatile recording medium such as memory, input devices such as a keyboard or mouse, a communication device such as a modem, and a bus connecting these. The computer also preferably includes a display device that displays the results of data processing.

[0041] Furthermore, as described above, the control variable optimization system A of the present system S can optimize the control variables contained therein based on the outputs from the culture condition detection sensor and the culture status detection sensor, and this can be achieved by executing a so-called computer program. That is, according to the present system S, a control variable optimization method can be realized by having a computer execute a program for optimizing the control variables (a control variable optimization program (hereinafter referred to as the "program")), and the present system S can realize an optimized bioresource production method by using the optimized control variables.

[0042] This diagram shows functional blocks for explaining the flow (processing flow) of control optimization realized by this program. In other words, the control variable optimization system A in this system S functions as each block by executing this program.

[0043] First, as is clear from the above description, values ​​acquired by various sensors B3 in culture system B are uploaded and accumulated as various data via a LAN, WAN, etc. to a cloud DB in control variable optimization system A. The cloud DB can be realized by recording data in a predetermined area of ​​a non-volatile recording medium such as a computer hard disk.

[0044] Furthermore, data that has been uploaded to the cloud database and is to be processed is collected (data collection), converted to a standardized format, and prepared for subsequent processing. This converted data is then written back to the database.

[0045] Furthermore, it is preferable to perform preprocessing on the data acquired by the various sensors B3 not only to unify the above-mentioned formats but also to perform specific processing efficiently. For example, since the data acquired by the various sensors B3 contains noise, it is preferable to smooth the data by, for example, performing a moving average process to remove the noise, from the perspective of further improving the optimization accuracy. It is also preferable to perform processing to convert the measurement values ​​of each sensor into different numerical values. For example, conversion to Lab values ​​is preferable for RGB sensors, and absorbance conversion is preferable for near-infrared sensors. The preprocessed data is then written back to the DB.

[0046] Then, in the control variable optimization system A of this system S, machine learning is performed using the data that has undergone the above preprocessing, a prediction model is constructed, and a process (control variable optimization process) is performed to calculate optimal control values. The specific contents will be clear from the description below, but the data obtained by the various sensors B3 above are part of the control variables, i.e., variable variables, and the control variables are optimized including other variables (initial variables).

[0047] Furthermore, the data (control variables) calculated by the above-described control variable optimization process are subjected to a process of determining whether they are appropriate as optimal values ​​(validity determination process). The validity is determined based on their relative position with respect to the learning data and whether the optimal control value is in a high distribution of the objective variable. In this process, if the data (control variables) are determined to be appropriate, they are written back to the cloud DB; if they are determined to be inappropriate, they are returned to the above-described data collection step or control variable optimization process step and processed again.

[0048] Then, data (control variables) that are judged to be valid in the above process are sent to the culture system B, more specifically to the culture control device B2, and the culture conditions of the incubator B1 are adjusted by the culture control device B2 to match the above control variables, thereby optimizing the production of bioresources.

[0049] Here, we will explain the details of the control variable optimization method (hereinafter referred to as "this method") specifically performed by the control variable optimization system A. This method is characterized by separating the control variables into initial variables and variable variables, creating a prediction model that is adapted to each of the initial variables and the variable variables, and optimizing the control variables by combining these two. Note that Figure 2 shows the data processing flow of this method.

[0050] As shown in the figure, this method preferably first uses data obtained by experimental design, specifically, the control variables varied in various ways by the present system S, and the data obtained as a result of varying these control variables.

[0051] As described above, in this method, the control variables are separated into initial variables and variable variables. Here, "control variables" refer to conditions (variables) necessary for optimizing the culture, "variable variables" refer to observable variables that change from moment to moment and can be controlled, and "initial variables" refer to observable variables that are set only once as the initial value for the culture and do not change thereafter. In this method, the optimization process is performed including these initial variables and variable variables.

[0052] Specific examples of variable variables include, but are not limited to, medium temperature, pH, Agit, dissolved oxygen concentration (DO), flow rate, feed rate, temperature dead zone, pressure, oxygen concentration, etc. Examples of initial variables include, but are not limited to, medium components, inoculation amount, metabolic switching time, feed start time, culture medium volume at the start, autoclave time and temperature, feed solution, feed solution concentration, pH adjuster, etc.

[0053] Furthermore, as described above, the control variables are separated into fluctuation variables and initial variables, but the explanatory variables of the prediction model include state variables in addition to the control variables, and the explanatory variables of the prediction model are separated into "fluctuation variables," "initial variables," and "state variables." However, "state variables" are observable but cannot be controlled during cultivation, and are not subject to optimization in the prediction model. Therefore, the optimization process described below is performed only on the fluctuation variables and initial variables among the explanatory variables.

[0054] Specific examples of state variables include, but are not limited to, exhaust CO2, exhaust O2, RGB, potential, visible light, near infrared, fluorescence, UV, viscosity, vibration, odor, respiratory quotient, ventilation volume, operating time of the heater / cooler pump, operating time of the pH adjuster pump, feed weight, kLA, glucose, enzyme sensor, etc.

[0055] In the above, after extracting the variable variables and initial variables from the control variables, this method determines which variable variables have higher priority. Figure 3 illustrates an example of how the priority of variable variables and initial variables is determined. Specifically, this figure shows a method for calculating the influence of a control variable on a response variable. Here, the "response variable" is a variable that serves as an index for determining the optimal value. Large fluctuations in the predicted value of this response variable are considered to be high priority, while small fluctuations are considered to be low priority. As shown in this figure, possible values ​​of the control variable within the set range of the control variable are generated, and this data is input into a machine learning model to calculate how the response variable will fluctuate depending on the value of the control variable. By calculating the value obtained by subtracting the minimum value from the maximum value of the calculated fluctuation (= the influence on the response variable (gradient of the control variable with respect to the response variable)), it is possible to estimate which variables have the greatest influence on control. Since it is better to optimize variables with greater influence first when performing sequential optimization, in this invention, variables with greater influence calculated by the above method are determined to have higher priority and are optimized first.

[0056] After the priority of the control variables is determined by the above process, a predetermined process is performed on the variable variables or initial variables that are determined to have the highest priority. However, the same process is performed on the variable variables and initial variables.

[0057] First, if it is determined that the priority of the initial variable is high, machine learning is performed based on a predictive model of the initial variable to calculate an optimal value for the initial variable. Then, machine learning is performed based on a predictive model of the variable variable to calculate an optimal value for the variable variable. On the other hand, if it is determined that the priority of the variable variable is high, machine learning is performed based on a predictive model of the variable variable to calculate an optimal value for the variable variable. Then, machine learning is performed based on the predictive model of the initial variable to calculate an optimal value for the initial variable. In other words, the key point of this method is that predictive models are created corresponding to each of the initial variables and the variable variables, and by combining them, the control variables can be optimized.

[0058] In machine learning based on a predictive model of variable variables, the predictive model uses a data format consisting of variable variables, initial variables, state variables, and objective variables as training data, with values ​​that change the variable variables, values ​​that fix the initial variables and state variables, and the objective variable as the resulting value. Machine learning is performed using the training data collected by this predictive model. After machine learning is completed, the control value of the variable variables that maximizes the objective variable is determined by inverse analysis of this predictive model.

[0059] Note that the prediction model in machine learning based on a prediction model of initial variables is similar to the above. Specifically, a data format consisting of variable variables, initial variables, state variables, and objective variables is used as training data, with values ​​that change the initial variables, values ​​that fix the variable variables and state variables, and the objective variable being the resulting value. Machine learning is performed using the training data collected by this prediction model. After machine learning is completed, the control value of the initial variables that maximizes the objective variable is determined by inverse analysis of this prediction model.

[0060] In this method, the training data for the prediction model used for initial variables contains state variables equal to (variable variables + number of sensors) x incubation time (e.g., an incubation time of 1,000 minutes would result in 1,000 variables per state variable). Therefore, creating a prediction model without dimension reduction is likely to result in poor prediction accuracy due to problems such as multicollinearity and the curse of dimensionality. Therefore, in this method, it is also useful to reduce the dimension by approximating the sensor time-series data to a Gompertz curve y = a × exp(-b × exp(-c × x)) + d. In the above example, approximating a certain state variable X to a Gompertz curve reduces the 1,000 variables to four variables: coefficients a, b, c, and d, which determine the Gompertz curve. It is preferable to use a logarithmic function, an exponential function, a Gompertz curve, or other functions based on biomedical knowledge. Figure 4 shows an example of this approach. In other words, this process reduces the dimension of the huge number of explanatory variables used in each model by projecting them onto a Gompertz curve and extracting coefficients that minimize the approximation error.

[0061] In the above description of this method, the variable variables and the initial variables have been treated as different variables, but in some cases, it is also possible to treat the variable variables as initial variables. Specifically, the variable variables can be divided into variables corresponding to the incubation time, such as the variable variable X0 at 0 min, the variable variable X10 at 10 min, the variable variable X20 at 20 min, ..., and the variable variable Xn at n min, and can therefore be treated as initial variables. In this case, it is possible to omit the determination of the priority of the control variables and the machine learning of the variable variables. In other words, it is also preferable to separate the variable variables into time-specific variables so that all variable variables can be treated equally to the initial variables.

[0062] Once the machine learning is complete, a determination is made as to whether the optimal value obtained as a result of this predictive model has converged. If convergence has occurred, the process proceeds to the next step. If convergence has not occurred, the process returns to the priority determination step, where machine learning is performed again to calculate the optimal value. Regarding "convergence," various determination methods can be used. For example, it is preferable to determine whether convergence has occurred by checking whether the difference between successively calculated solutions gradually decreases and is smaller than a predetermined amount. That is, by repeatedly using two models adapted to the initial variables and the variable variables, respectively, and recursively performing inverse analysis until the optimal control values ​​of the initial variables and the variable variables converge, more accurate results can be obtained by optimizing the control variables.

[0063] If the optimal value has converged, it is then determined whether this optimal value is appropriate as the final control value, and if it is not appropriate, the process returns to the experimental design method and increases the amount of collected data, continuing until the final control value is extracted in a way that prevents the optimal value from becoming a particular solution due to the amount and bias of the learning data.On the other hand, if the optimal value is determined to be appropriate as the optimal control value, the obtained result is confirmed and proposed as the optimal control value, and the process ends.

[0064] This method involves time-series control of variable variables and optimization of multiple initial variables, resulting in complex, multidimensional optimal conditions. Therefore, it is difficult for humans to determine whether the proposed control values ​​are likely to lead to good cultivation results. Therefore, by reducing the dimensions of the multidimensional control variables of the training data and optimal control values ​​and projecting the training data and optimal control values ​​into the reduced-dimensional space, it becomes possible to confirm the relative position of the optimal control values ​​and quantitatively visualize the influence of the control variables on the target variable. This is shown in Figure 5.

[0065] In this case, a prediction model is created using the learning data in the projected space as input, and contour lines of the predicted values ​​are calculated. By adding the contour lines to the above diagram, the distribution of the predicted values ​​can be visualized, making it possible to confirm whether the optimal control value is in a high distribution of the objective variable. The validity of the obtained optimal control value can then be determined by checking the relative position of this optimal control value and whether the optimal control value is in a high distribution of the objective variable. In other words, when the optimal control values ​​of the initial variables and the variable variables are calculated, it is preferable to present the distribution of the predicted values ​​of the control variables and the relative position of the optimal control value with respect to the learning data.

[0066] As described above, this embodiment can provide a control variable optimization method that can determine more optimal culture conditions using a prediction model using machine learning, as well as a bioresource production method and a bioresource production system that use the same.

[0067] To explain the effects of the present invention more specifically, in this embodiment, optimal culture conditions are proposed using a predictive model that uses machine learning, making it possible to improve the productivity of microorganisms by nearly 10% compared to culture conditions using conventional optimization methods.

[0068] The results of applying this technology to the actual cultivation of a certain microorganism are shown in Figure 6. These results were obtained by creating a prediction model for the cultivation of a certain microorganism, calculating the optimal conditions for a certain objective variable, and culturing the microorganism according to those conditions.

[0069] Furthermore, there are a total of seven control variables in the main culture. Conventionally, the aim is to maximize the objective variable by selecting the value of each variable, whereas in this method, of the seven control variables, four are assigned as initial variables and three are assigned as variable variables.

[0070] The differences between the conventional and the present method in terms of the variable variables and initial variables are shown in (a) and (b), respectively. As can be seen from these figures, the variable variables in the present method change during the culture period, whereas in the conventional method they must be kept constant. Furthermore, the present method is able to extract optimal values ​​for the initial variables taking into account the changes in the variable variables, and it can be seen that values ​​completely different from those selected in the conventional method are selected.

[0071] Finally, the difference in productivity is shown in Figure 6(c). Compared to the conventional method (left bar graph), the objective variable value is higher when using this method (right bar graph), and it can be seen that productivity is nearly 10% higher.

[0072] Furthermore, in actual cultivation, the cultivation conditions constantly change depending on the weather, the inoculation, and the state of the medium. However, in this method, the variable variables are controlled in response to these changes, making it possible to cultivate with improved productivity maintained.

[0073] Furthermore, because culture conditions have traditionally been determined based on the experience of skilled engineers, it has been difficult to visualize the effect of control variables on the objective variable. In contrast, this method, as shown in Figure 5 above, performs dimensionality reduction, making it possible to plot a graph with the predicted value of the calculated objective variable on the vertical axis and the control variable on the horizontal axis, thereby visualizing the effect of the control variable on the objective variable, which has the advantage of being useful as information when constructing culture equipment.

[0074] Furthermore, as shown in Figure 4 above, projection onto the Gompertz curve enables dimensionality reduction and improves prediction accuracy by about 35% compared to when the Gompertz curve is not used. Figure 7 shows the results of applying this method to a certain microbial culture.

[0075] For a given objective variable of a certain microbial culture, predictive models were created without dimension reduction of the explanatory variables (three culture condition variables + one wavelength of the visible light sensor * 852 points) and with dimension reduction using Gompertz curve approximation (a) (three culture condition variables + four Gompertz curve variables). The predictive accuracy of each model (RMSE standardized by the objective variable) was compared, and a corresponding t-test was performed to determine whether there was a significant difference. The results are shown in Figure (b) (violin plot). In this figure, the width of the light color (gray) represents data density, the dashed line represents the average RMSE with dimension reduction, and the solid line represents the average RMSE with dimension reduction. Also, "****" in the figure indicates that the corresponding t-test yielded p<0.0001.

[0076] As shown in this figure, the prediction accuracy (RMSE) was significantly higher when dimension reduction was performed, with the average value being about 35% higher ((= RMSE without dimension reduction - RMSE with dimension reduction) / RMSE with dimension reduction).

[0077] An example of a flowchart for calculating the optimal control value is shown in Figure 8. In this calculation, learning data for variable variables and learning data for initial variables are created based on the converted data. In the learning data for variable variables, the variable variables are treated as control variables, and the initial variables and sensor data are treated as state variables. On the other hand, in the learning data for initial variables, the initial variables are treated as control variables, and the variable variables and sensor data are treated as state variables.

[0078] Here, each training dataset is converted into a dimension-reduced dataset by dimension reduction using PCA or approximation to a Gompertz curve, and feature selection using correlation coefficients, clustering, regularization, or the like.

[0079] Next, the dimension-reduced dataset is used as training data to create a predictive model using machine learning such as neural networks or random forests. Next, domain knowledge of the cultivation or calculation of the gradient of the objective variable with respect to the control variable using a model is used to identify control variables that have a large impact on the cultivation, and the priority of optimization of initial variables and variable variables is determined. Below, an example is explained using a case where the influence of variable variables is large.

[0080] Optimization begins by calculating the optimal control values ​​for the variable variables with the highest priority using the prediction model created first (i.e., inverse analysis). The data used for inverse analysis is one of the acquired data sets that was not used for learning and is used as batch data for simulation.

[0081] Next, the optimal control value of the variable variable calculated as the state variable of the prediction model in which the initial variable is the control variable is reflected, and the optimal control value of the initial variable is calculated. Furthermore, the optimal control value of the variable variable calculated as the state variable of the prediction model in which the variable variable is the control variable is reflected, and the optimal control value of the variable variable is calculated again. Subsequently, optimization of the variable variable and the initial variable is performed recursively until the optimal control value converges, and the finally converged optimal control values ​​of the variable variable and the initial variable are obtained.

[0082] The multidimensional control variables of the training data and optimal control values ​​are reduced in dimension using PCA or other methods, and the training data and optimal control values ​​are projected onto the reduced-dimensional space to confirm the relative positions of the optimal control values. A prediction model is then created using the training data in the projected space as input, contour lines of the predicted values ​​are calculated, and the distribution of the predicted values ​​is calculated by adding the contour lines to the above diagram. The contour lines are then added to the above diagram to obtain a diagram depicting the relative positions of the optimal control values ​​and the distribution of the predicted values, and the appropriateness of the optimal control values ​​is determined.

[0083] Then, by having a person input the optimal control value that is judged to be appropriate into the culture system, specifically the culture device, before starting culture, culture can be achieved according to the optimal control value predicted by AI.

[0084] It is also possible to use this method to calculate the optimal control value while controlling the culture in real time. Figure 9 shows a functional block diagram for calculating the optimal control value during real-time control. In this block diagram, only the control variable optimization system A is different, and the rest is the same as in Figure 1 above.

[0085] In this example, the optimal control value of the variable variable is calculated by inputting the real-time data converted data into the constructed prediction model and performing reverse analysis. The calculated optimal control value of the variable variable is fed back to the culture control device B2 in real time, thereby real-time culture control by AI is realized.

[0086] FIG. 10 shows a flowchart for calculating the optimal control value in this case, i.e., during real-time control. In this case, previously acquired data is used. That is, before the start of culture, one of the acquired data that was not used for learning is used as data for back analysis to calculate the optimal control value of the initial variable. After this calculation, the optimal control value of the initial variable is calculated and input into the culture control device B2.

[0087] The calculated optimal control values ​​of the initial variables are reflected as state variables in the real-time converted data.

[0088] After the start of cultivation, the real-time converted data is used as input to calculate the optimal control values ​​of the variable variables in real time.

[0089] The calculated optimal control values ​​of the variable variables are then fed back to the culture control device B2 in real time, thereby realizing real-time culture control using AI.

[0090] INDUSTRIAL APPLICABILITY The present invention has industrial applicability as a control variable optimization method, a bioresource production method, and a bioresource production system.

[0091] S: Bioresource production system A: Control variable optimization system B: Cultivation system B1: Incubator B2: Cultivation control device B3: Various sensors B4: Sensor control device C: Control computer

Claims

1. A control variable optimization method characterized by separating control variables into initial variables and variable variables, creating a predictive model that is adaptive to each of the initial variables and the variable variables, and optimizing the control variables by combining the two.

2. The method for optimizing controlled variables according to claim 1, further comprising the steps of: calculating the priority of the initial variables and the variable variables from the gradient of the controlled variables with respect to the objective variable; and determining the order of optimization of the initial variables and the variable variables.

3. The control variable optimization method according to claim 1, characterized in that the control variables are optimized by repeatedly using two models adapted to the initial variables and the variable variables, respectively, and recursively performing inverse analysis until the optimal control values ​​of the initial variables and the variable variables converge.

4. The control variable optimization method according to claim 1, characterized in that the control variables are optimized by assigning medium temperature and pH to variable variables, and medium components and inoculation amount to initial variables, and combining two models that are adapted to the initial variables and variable variables, respectively.

5. A control variable optimization method according to claim 1, characterized in that the explanatory variables used in each model are projected onto a Gompertz curve to extract coefficients that minimize the approximation error, thereby reducing the dimensions.

6. The control variable optimization method according to claim 1, characterized in that, when the optimal control values ​​of the initial variables and variable variables are calculated, the distribution of predicted values ​​of the control variables and the relative position of the optimal control values ​​with respect to the learning data are presented.

7. A method for optimizing control variables according to claim 1, characterized in that all fluctuating variables are separated into time-dependent variables so that they can be treated equally to initial variables.

8. A method for producing bioresources, comprising cultivating and cultivating organisms based on the optimized control variables as claimed in claim 1.

9. A bioresource production system comprising: a cultivation system for bioresource production; and a control variable optimization system for optimizing control variables obtained from the cultivation system. The control variable optimization system separates the control variables into initial variables and variable variables, creates a predictive model that is adapted to each of the initial variables and the variable variables, and optimizes the control variables by combining these two.

Images