Experimental condition search method and experimental system

Abstract

This experimental system comprises: an experimental device capable of simultaneously executing a plurality of experiments with dissimilar experimental conditions; and a search device for searching for experimental conditions to be used in the experimental device. The search device sets a first experimental condition using a first experiment planning method, controls the experimental device so that the experimental device executes a first experiment based on the first experimental condition, and sets a second experimental condition using a second experiment planning method on the basis of the result of the first experiment. Each of the first experiment planning method and the second experiment planning method includes a Bayesian optimization method using a first acquisition function, and a Bayesian optimization method using a second acquisition function different from the first acquisition function.

Description

Method for searching experimental conditions and experimental system

[0001] The present disclosure relates to a method for searching for experimental conditions and an experimental system.

[0002] When searching for optimal experimental conditions, there are many parameters to consider, and the number of experimental conditions resulting from the combination of these parameters is enormous. Therefore, searching for optimal experimental conditions essentially requires a large number of experiments. However, since conducting a large number of experiments requires a huge amount of time and cost, it is desirable to search for optimal experimental conditions with as few experiments as possible.

[0003] For example, Non-Patent Document 1 ("Multi-Information Source Bayesian Optimization of Culture Media for Cellular Agriculture") describes a method for searching for optimal experimental conditions (cell culture conditions) using Bayesian optimization when conducting experiments to screen cell culture conditions.

[0004] "Multi-Information Source Bayesian Optimization of Culture Media for Cellular Agriculture", May 2022 Biotechnology and Bioengineering 119 (2009), Zachary Cosenza, Raul Astudillo, Peter I. Frazier, Keith Baar, David E. Block, Internet <URL: http: / / www.researchgate.net / publication / 360546529>

[0005] Bayesian optimization is a method for searching for optimal input conditions that maximize (or minimize) output when the relationship between input and output is unknown. Therefore, by performing Bayesian optimization using the experimental conditions and experimental results as input and output, respectively, it is possible to search for optimal experimental conditions.

[0006] Bayesian optimization searches for optimal input conditions using an "acquisition function" that represents the likelihood of the next input condition. The acquisition function must be determined by the user. However, it is difficult for users to select an appropriate acquisition function to achieve optimal experimental conditions.

[0007] The present disclosure has been made to solve such problems, and its purpose is to provide a method for searching for optimal experimental conditions with a small number of experiments.

[0008] A search method according to the present disclosure is a search method for experimental conditions used in an experimental apparatus capable of executing multiple experiments with different experimental conditions, and includes the steps of setting first experimental conditions using a first experimental design, causing the experimental apparatus to execute a first experiment based on the first experimental conditions, and setting second experimental conditions using a second experimental design based on the results of the first experiment. The first experimental design includes at least two of an experimental design based on a Bayesian optimization method using a first acquisition function, an experimental design based on a Bayesian optimization method using a second acquisition function different from the first acquisition function, and an experimental design other than the Bayesian optimization method. The second experimental design includes an experimental design based on a Bayesian optimization method using an acquisition function that is the same as or different from the acquisition function used in the first experimental design.

[0009] The experimental system according to the present disclosure includes an experimental apparatus capable of executing a plurality of experiments each having different experimental conditions, and a search device that searches for experimental conditions to be used in the experimental apparatus. The search device sets first experimental conditions using a first experimental design, causes the experimental apparatus to execute a first experiment based on the first experimental conditions, and sets second experimental conditions using a second experimental design based on the results of the first experiment. The first experimental design includes at least two of an experimental design based on a Bayesian optimization method using a first acquisition function, an experimental design based on a Bayesian optimization method using a second acquisition function different from the first acquisition function, and an experimental design other than the Bayesian optimization method. The second experimental design includes an experimental design based on a Bayesian optimization method using an acquisition function that is the same as or different from the acquisition function used in the first experimental design.

[0010] According to the present disclosure, optimal experimental conditions can be found with a small number of experiments.

[0011] 1 is a diagram schematically showing an example of the overall configuration of an experimental system. FIG. 2 is a diagram showing an example of experimental data and the flow of batch Bayesian optimization. FIG. 3 is a diagram schematically showing an image of the next experimental conditions searched for by exploitation-focused batch Bayesian optimization. FIG. 4 is a diagram schematically showing an image of the next experimental conditions searched for by exploration-focused batch Bayesian optimization. FIG. 5 is a diagram schematically showing an image of the next experimental conditions searched for by hybrid batch Bayesian optimization that combines exploitation-focused and exploration-focused optimization. FIG. 6 is a diagram showing an example of the results of searching for optimal input conditions that maximize the output of an Ackley function by hybrid batch Bayesian optimization. FIG. 7 is a flowchart (part 1) showing an example of the processing procedure of a search device. FIG. 8 is a flowchart (part 2) showing an example of the processing procedure of a search device.

[0012] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the drawings. In the drawings, the same or corresponding parts are designated by the same reference numerals, and description thereof will not be repeated.

[0013] [Configuration of Experimental System] Figure 1 is a diagram showing a schematic example of the overall configuration of an experimental system 1 according to this embodiment. The experimental system 1 is a system that can search for experimental conditions that will optimize the experimental results by conducting an experiment in a situation where the relationship between the experimental conditions and the experimental results (such as experimental productivity or the quality of the product) is unknown. The experimental system 1 includes a search device 100 and an experimental device 200.

[0014] The experimental apparatus 200 is an apparatus that automatically performs scientific experiments in response to a control signal. Fig. 1 shows an example of an experiment in which the experimental apparatus 200 performs a cell culture experiment. The cell culture conditions used in an experiment using the experimental apparatus 200 are an example of "experimental conditions" in the present disclosure, and the cell culture volume measured in an experiment using the experimental apparatus 200 is an example of "experimental results" in the present disclosure.

[0015] The cells cultured in the experimental apparatus 200 are not particularly limited, and may be, for example, cells that have a useful function in medicine or cells that produce substances that have a useful function in various industrial fields. Furthermore, the objects cultured in the experimental apparatus 200 are not limited to cells, and may also be microorganisms or bacteria (such as E. coli). Furthermore, experiments performed in the experimental apparatus 200 may be experiments other than cell culture (such as synthetic chemistry experiments).

[0016] The experimental device 200 is configured to be able to simultaneously perform multiple cell culture experiments with different culture conditions in a single process. Specifically, the experimental device 200 includes a medium preparation device 210, a culture device 220, and a measurement device 230.

[0017] The culture medium preparation device 210 is configured to automatically introduce culture medium and cell lines to be cultured under the culture medium conditions (amounts of components in the culture medium, pH, etc.) specified by the searching device 100 into n wells 212 (n is an integer of 2 or more; in the example of FIG. 1 , n=12×8=96) arranged in a well plate 211. Note that introduction of the culture medium and cell lines into each well 212 is not necessarily limited to being performed automatically by the culture medium preparation device 210, but may also be performed manually by the user.

[0018] When the well plate 211 in which the culture medium and cell line have been placed in each well 212 is set in the culture device 220, the culture device 220 cultures the cells in each well 212 in the environment (temperature, time, etc.) instructed by the search device 100. This allows n cell culture experiments with different culture conditions (culture medium conditions) to be performed simultaneously in a single process.

[0019] The measuring device 230 measures the results of a cell culture experiment performed by the culture device 220. The measuring device 230 according to this embodiment measures the cell culture volume in each of the n wells 212 as a result of the cell culture experiment. The measuring device 230 transmits experimental data combining the culture conditions and cell culture volume of each well 212 to the searching device 100 as the experiment result.

[0020] 1, the searching device 100 includes a processor 101, a memory 102, an input / output interface (I / F) 103, and a communication I / F 104.

[0021] The processor 101 executes various programs to enable the searching device 100 to perform various processes. The memory 102 stores the programs executed by the processor 101 and various data required for executing the programs. The input / output I / F 103 is an interface that allows the processor 101 to communicate with the experimental device 200. The communication I / F 104 is an interface that allows the processor 101 to communicate with devices outside the experimental system 1 via a network.

[0022] The memory 102 stores a database 130 that stores the experimental data received from the experimental device 200. The structure of the database 130 will be described later.

[0023] The search device 100 further includes a display 110 and an input device 120. The display 110 displays the results of the arithmetic processing performed by the processor 101 to the user. The input device 120 (such as a mouse, keyboard, or touch sensor) accepts data input operations to the processor 101.

[0024] [Bayesian Optimization and Batch Bayesian Optimization] The search device 100 searches for optimal experimental conditions that maximize (or minimize) the experimental results of the experimental device 200 based on experimental data obtained using the experimental device 200.

[0025] The search device 100 searches for optimal experimental conditions using a Bayesian optimization method. Bayesian optimization is a method for searching for optimal input conditions that maximize (or minimize) an output when the relationship between the input and the output is unknown. More specifically, Bayesian optimization is a method for searching for explanatory variables xi that maximize (or minimize) the objective variable y when the objective variable y is the output and n-dimensional explanatory variables xi (i = 1, 2, ... n) are the input, assuming that the function f of y = f(xi) follows a Gaussian process, and searching for explanatory variables xi that maximize (or minimize) the objective variable y in a situation where the function f is unknown. Assuming that the unknown function f follows a Gaussian process allows for highly optimized various objective variables y with simpler processing than when assuming that the objective variable y and the explanatory variables xi follow other distributions. Note that Bayesian optimization itself is a well-known method, and therefore further detailed description will be omitted in this specification.

[0026] As described above, the experimental apparatus 200 according to this embodiment can simultaneously perform multiple (n) cell culture experiments with different culture conditions in a single process. In view of this, the searching apparatus 100 according to this embodiment uses "batch Bayesian optimization," a type of Bayesian optimization, to search for multiple (n) experimental conditions at once. While ordinary Bayesian optimization searches for one condition as the next experimental condition based on experimentally performed data, batch Bayesian optimization can search for multiple conditions as the next experimental condition based on experimentally performed data. Note that the experimentally performed data includes combinations of experimental conditions and experimental results obtained from one or more experiments.

[0027] 2 is a diagram showing an example of experimental data stored in the database 130 of the searching device 100 and the flow of batch Bayesian optimization. Fig. 2 shows an example in which, each time an experiment is completed, medium conditions (contents of components A, B, and C, and pH values) for n wells 212 are searched for as the next experimental conditions, and ultimately three experiments are performed.

[0028] In the first experiment, cells are cultured under initial experimental conditions in the experimental device 200. The initial experimental conditions are arbitrarily set by the user, for example, and stored in the database 130. In the first experiment, the cell culture volumes in n wells 212 are measured. Data from the first experiment (n combinations of medium conditions and culture volumes) is sent to the searching device 100 and stored in the database 130.

[0029] The processor 101 of the search device 100 performs batch Bayesian optimization for the first batch based on the first experimental data stored in the database 130 to search for n culture medium conditions that are optimal for the next experiment (second experiment), and transmits the searched n culture medium conditions to the experimental device 200 as the next experimental conditions.

[0030] In the second experiment, cells are cultured under the n medium conditions searched for by the first batch of batch Bayesian optimization. The data obtained from the second experiment is sent to the search device 100 and added to the database 130. At this point, the database 130 stores the data obtained from the first and second experiments (2n combinations of medium conditions and culture volumes).

[0031] The processor 101 performs a second batch of batch Bayesian optimization based on the first and second experimental data stored in the database 130 to search for n optimal culture medium conditions for the next experiment (third experiment), and transmits the searched n culture medium conditions to the experimental device 200 as the next experimental conditions.

[0032] In the third experiment, cells are cultured under the n medium conditions searched for by the second batch of batch Bayesian optimization. Data from the third experiment is sent to the search device 100 and added to the database 130. Therefore, at this point, data from the first to third experiments (3n combinations of medium conditions and culture volumes) are stored in the database 130.

[0033] The experiment and batch Bayesian optimization are repeated until an experiment termination condition is met. The experiment termination condition is set, for example, to be at least one of the following conditions: the latest experiment result has reached a target value; and the number of batch Bayesian optimization processes (the number of batches) has exceeded a reference value. In the example of Figure 2, the third experiment result is determined to have reached the target value, and the culture medium conditions for the third experiment are displayed on the display 110 as optimal culture conditions and presented to the user.

[0034] [Batch Bayesian Optimization Using Multiple Acquisition Functions] In the above-described batch Bayesian optimization, as in ordinary Bayesian optimization, an "acquisition function" is used, which is an index of the likelihood of the next input condition. Conventionally, numerous acquisition functions with different characteristics have been proposed, and users have had to select one from among many acquisition functions. However, if the selected acquisition function is not well suited to the target experiment, there is a concern that optimal experimental conditions cannot be found with a small number of experiments.

[0035] Figure 3 is a diagram that shows a schematic image of the next experimental conditions searched for by the exploitation-oriented batch Bayesian optimization. The exploitation-oriented approach means that the search focuses on input conditions (experimental conditions) that result in high output (experimental results). Note that Figure 3 and Figure 4, which will be described later, show an example in which four experimental conditions are searched for per batch as the next experimental conditions.

[0036] If only one acquisition function is selected and that one acquisition function acts as a "utilization-oriented" function for the target experiment, as shown in Figure 3, the search will be biased toward experimental conditions that are close to the experimental data, resulting in a concern that the search will fall into a local solution and will not be able to find the optimal solution.

[0037] 4 is a diagram showing an image of the next experimental conditions searched for by the search-oriented batch Bayesian optimization. The search-oriented type means that the search focuses on the experimental conditions for which there is no experimental data.

[0038] If only one acquisition function is selected, and that one acquisition function acts as a "search-oriented" function for the target experiment, as shown in Figure 4, there is a concern that the search will be biased toward experimental conditions far from the optimal solution, making it impossible to search for the optimal solution with a small number of experiments.

[0039] As described above, when only one acquisition function is selected, there is a concern that the search may be biased toward experimental conditions that are close to the experimental data (utilization-oriented type), or toward experimental conditions that are far from the optimal solution (exploration-oriented type). In consideration of such problems, the search device 100 according to this embodiment searches for the next experimental conditions by hybrid batch Bayesian optimization using multiple acquisition functions.

[0040] Fig. 5 is a diagram showing an example of the next experimental conditions searched for by hybrid batch Bayesian optimization, which combines exploitation-oriented and exploration-oriented methods. Fig. 5 shows an example in which a total of four experimental conditions are searched for per batch, two each by exploitation-oriented and exploration-oriented methods.

[0041] By searching for the next experimental conditions using hybrid batch Bayesian optimization, it is possible to avoid the risk of biased searches for experimental conditions that are close to the experimental data or biased searches for experimental conditions that are far from the optimal solution, as shown in Figure 5. This makes it possible to efficiently search for optimal experimental conditions by taking into account a good balance between "exploration" and "utilization." As a result, it is possible to appropriately search for optimal experimental conditions to be used in an experimental device 200 that can simultaneously perform multiple experiments with different experimental conditions, with a small number of experiments.

[0042] Whether an acquisition function is exploitation-oriented or exploration-oriented is determined relatively based on the results of a search using that acquisition function, and it is difficult to determine in advance before conducting the search. However, by using at least multiple acquisition functions, the risk of falling into either the exploitation-oriented or exploration-oriented type can be avoided compared to using a single acquisition function. As a result, optimal experimental conditions can be efficiently searched for by taking into account a good balance between "exploration" and "exploitation."

[0043] 6 shows an example of the results of a search for 10 optimal input conditions for maximizing the output of an Ackley function with two-dimensional inputs (X1, X2) and one-dimensional output (Y) using hybrid batch Bayesian optimization. In the example shown in Fig. 6, the upper confidence bound (UCB) is used as the first acquisition function, and the expected improvement (EI) is used as the second acquisition function. Of the 10 input conditions searched, five conditions are derived using the first acquisition function "UCB," and the remaining five conditions are derived using the second acquisition function "EI."

[0044] Note that both "UCB" and "EI" are representative acquisition functions. For example, UCB for input x is expressed as UCB(x) = μ(x) + k·σ(x). μ(x) is a function representing the expected value of the posterior distribution of a Gaussian process, σ(x) is a function representing the uncertainty (variance) of the expected value, and "k" is a hyperparameter determined by the user. In this way, UCB is expressed using the expected value and uncertainty (variance) of the posterior distribution of a Gaussian process, as well as hyperparameters. Like "UCB," "EI" is also expressed using the expected value and uncertainty of the posterior distribution of a Gaussian process, but does not include hyperparameters.

[0045] 6, circles (○) indicate experimental data used as input for batch Bayesian optimization, squares (□) indicate first experimental conditions searched for using the first acquisition function UCB, and triangles (△) indicate second experimental conditions searched for using the second acquisition function EI. From the results shown in Fig. 6, it can be seen that by using the first acquisition function UCB and the second acquisition function EI, input coordinates close to the experimental data and input coordinates far from the experimental data are searched for in a balanced manner in the two-dimensional coordinate system of inputs X1 and X2.

[0046] 7 is a flowchart showing an example of a processing procedure of the searching device 100. The processing shown in this flowchart is started when a start command is input to the input device 120 by the user, for example.

[0047] The processor 101 of the searching device 100 acquires the experimental data from the database 130 (step S10).

[0048] Next, the processor 101 sets the experimental data acquired in step S10 as input data for batch Bayesian optimization (step S11).

[0049] Next, the processor 101 sets the next n experimental conditions by performing batch Bayesian optimization based on the input data set in step S11 (step S21). Specifically, the processor 101 sets m experimental conditions by performing batch Bayesian optimization using the first acquisition function (UCB) with the input data set in step S11. "m" is an integer between 1 and (n-1), for example, m = n / 2. Furthermore, the processor 101 sets (n-m) experimental conditions by performing batch Bayesian optimization using the second acquisition function (EI) with the input data set in step S11. Then, the processor 101 sets the next n experimental conditions as a combination of the m experimental conditions derived using the first acquisition function (UCB) and the (n-m) experimental conditions derived using the second acquisition function (EI). Specifically, the processor 101 generates control signals for causing the experimental apparatus 200 to execute cell culture experiments based on each of the n next experimental conditions.

[0050] Next, the processor 101 outputs the control signal generated in step S21 to the experimental device 200, thereby controlling the experimental device 200 to perform a cell culture experiment under the next n experimental conditions (step S24).

[0051] Next, the processor 101 acquires the experimental data including the next experimental result from the experimental device 200 (step S25).

[0052] Next, the processor 101 adds the experimental data acquired in step S25 to the database 130 (step S26).

[0053] Next, the processor 101 determines whether the experiment termination condition is satisfied (step S27). The experiment termination condition is set, for example, as described above, to be a condition that at least one of the following conditions is satisfied: the latest experiment result has reached the target value; and the number of batches in the batch Bayesian optimization has exceeded a reference value.

[0054] If the experiment end condition is not met (NO in step S27), the processor 101 returns the process to step S10 and repeats the batch processing of steps S10 to S26 until the experiment end condition is met.

[0055] If the experiment termination condition is met (YES in step S27), the processor 101 displays the most recent experimental conditions among the experimental conditions stored in the database 130 on the display 110 as the optimal experimental conditions to be presented to the user (step S30). Note that the experimental conditions that produced the best experimental results among the most recent experimental conditions may be set as the optimal experimental conditions.

[0056] As described above, the search device 100 according to this embodiment sets the next experimental conditions as a combination of m experimental conditions searched for by batch Bayesian optimization using the first acquisition function UCB and (n-m) experimental conditions searched for by batch Bayesian optimization using the second acquisition function EI. Therefore, compared to when all of the next experimental conditions are searched for by batch Bayesian optimization using only one acquisition function (e.g., only UCB or only EI), it is possible to avoid the risk of the next experimental conditions being biased toward experimental conditions that are close to the experimental data or toward experimental conditions that are far from the optimal solution. As a result, the experimental conditions to be used in the experimental device 200 can be appropriately searched for with a small number of experiments.

[0057] [Variant 1] In the flowchart of Figure 7 described above, assuming that batch Bayesian optimization using the first acquisition function and batch Bayesian optimization using the second acquisition function are performed within one batch process, a combination of the experimental conditions searched for by batch Bayesian optimization using the first acquisition function and the experimental conditions searched for by batch Bayesian optimization using the second acquisition function is set as the next experimental condition (see step S21 in Figure 7).

[0058] Alternatively, the batch Bayesian optimization using the first acquisition function and the batch Bayesian optimization using the second acquisition function may be performed in separate batch processes.

[0059] Fig. 8 is a flowchart showing an example of a processing procedure of the searching device 100 according to the present modified example 1. The flowchart of Fig. 8 is obtained by adding step S20 to the flowchart of Fig. 7 described above, and further by changing step S21 of the flowchart of Fig. 7 described above to steps S22 and S23. The other steps in Fig. 8 (steps having the same numbers as the steps shown in Fig. 7 described above) have already been described, and therefore detailed description thereof will not be repeated here.

[0060] After the experimental data is set in step S11, the processor 101 determines whether the previous experimental conditions are conditions searched for by batch Bayesian optimization using the first acquisition function (UCB) (step S20).

[0061] If the previous experimental conditions are not conditions searched for by batch Bayesian optimization using the first acquisition function (UCB) (NO in step S20), the processor 101 sets the next n first experimental conditions by batch Bayesian optimization using the first acquisition function (UCB) (step S22). Thereafter, the processor 101 executes the processes from step S24 onward.

[0062] On the other hand, if the previous experimental conditions are conditions searched for by batch Bayesian optimization using the first acquisition function (UCB) (YES in step S20), the processor 101 sets n second experimental conditions by batch Bayesian optimization using the second acquisition function (EI) (step S23). Thereafter, the processor 101 executes the processes from step S24 onward.

[0063] By doing so, in this modification 1, an experiment using n experimental conditions searched for using the first acquisition function and an experiment using n experimental conditions searched for using the second acquisition function are alternately executed, which makes it possible to avoid the risk that the next experimental conditions will be biased toward experimental conditions that are close to the experimental data or that are far from the optimal solution, compared to when only one acquisition function for the next experimental conditions is used to search for them.

[0064] 8 shows an example in which the first and second acquisition functions are alternately used to set the next experimental conditions, but the first and second acquisition functions may be used irregularly to set the next experimental conditions instead of alternately. For example, the first acquisition function, the second acquisition function, the second acquisition function, the first acquisition function, and so on may be used in this order to set the next experimental conditions.

[0065] [Modification 2] The experimental design used in the above-described embodiment and modification 1 is not limited to the batch Bayesian optimization method. For example, all or part of the experimental design used in the above-described embodiment and modification 1 may be a normal Bayesian optimization method.

[0066] Furthermore, part of the experimental design used in the above-described embodiment and modification 1 may be an experimental design other than Bayesian optimization. For example, part of the experimental design used in the above-described embodiment and modification 1 may be an experimental design (grid search, random search, etc.) that comprehensively sets experimental conditions without using data that has already been tested.

[0067] Furthermore, when using a Bayesian optimization method that uses an acquisition function with a hyperparameter such as UCB, for example, it is possible to use in combination a Bayesian optimization method that uses UCB with hyperparameter k set to "k1" and a Bayesian optimization method that uses UCB with hyperparameter k set to "k2" different from k1.

[0068] [Aspects] It will be understood by those skilled in the art that the above-described embodiments and their modifications are specific examples of the following aspects.

[0069] (Section 1) A search method according to the present disclosure is a search method for experimental conditions used in an experimental apparatus capable of executing multiple experiments with different experimental conditions, the search method including the steps of setting first experimental conditions using a first experimental design, controlling the experimental apparatus to execute a first experiment based on the first experimental conditions, and setting second experimental conditions using a second experimental design based on the results of the first experiment. The first experimental design includes at least two of an experimental design based on a Bayesian optimization method using a first acquisition function, an experimental design based on a Bayesian optimization method using a second acquisition function different from the first acquisition function, and an experimental design other than the Bayesian optimization method. The second experimental design includes an experimental design based on a Bayesian optimization method using an acquisition function that is the same as or different from the acquisition function used in the first experimental design.

[0070] In the search method described in paragraph 1, first experimental conditions are set using at least two of an experimental design using a Bayesian optimization method with a first acquisition function, an experimental design using a Bayesian optimization method with a second acquisition function different from the first acquisition function, and an experimental design other than the Bayesian optimization method, and a first experiment is performed based on the set one experimental condition. Therefore, compared to when the first experimental conditions are set using a single experimental design, it is possible to avoid the risk of the first experimental conditions being searched for in a bias toward experimental conditions close to the experimental data or experimental conditions far from the optimal solution. As a result, it is possible to search for optimal experimental conditions with a small number of experiments.

[0071] (Section 2) In the search method described in Section 1, the first acquisition function is UCB (Upper Confidence Bound) and the second acquisition function is EI (Expected Improvement).

[0072] In the search method described in Section 2, experimental conditions set by Bayesian optimization using UCB and experimental conditions set by batch Bayesian optimization using EI can be searched for as the first experimental conditions.

[0073] (Clause 3) In the search method described in clause 1, the experimental design method different from the Bayesian optimization method is an experimental design method that comprehensively sets experimental conditions without using data that has already been tested.

[0074] In the search method described in paragraph 3, experimental conditions set by Bayesian optimization using experimental data and second experimental conditions set comprehensively without using experimental data can be searched for as first experimental conditions.

[0075] (Item 4) In the method for searching for experimental conditions according to any one of items 1 to 3, the experimental device is configured so that cells can be cultured in multiple wells in a single process.

[0076] In the search method described in paragraph 4, for an experimental device configured to be able to culture cells in multiple wells in a single process, the experimental design for setting the culture conditions for some of the multiple wells can be different from the experimental design for setting the culture conditions for the remaining wells. Therefore, compared to when all culture conditions for multiple wells are set using a single experimental design, it is possible to avoid the risk of the search being biased toward culture conditions close to experimental data or toward culture conditions far from the optimal solution.

[0077] (5) In the method for searching for experimental conditions described in paragraph 1, the experimental apparatus is configured to automatically execute an experiment in response to a control signal. The step of setting first experimental conditions includes a step of generating a control signal for causing the experimental apparatus to execute the first experiment. The step of setting second experimental conditions includes a step of generating a control signal for causing the experimental apparatus to execute a second experiment based on the second experimental conditions.

[0078] (Section 6) An experimental system according to the present disclosure includes an experimental apparatus capable of executing multiple experiments with different experimental conditions, and a search device that searches for experimental conditions to be used in the experimental apparatus. The search device sets first experimental conditions using a first experimental design, controls the experimental apparatus to execute a first experiment based on the first experimental conditions, and sets second experimental conditions using a second experimental design based on the results of the first experiment. The first experimental design includes at least two of an experimental design based on a Bayesian optimization method using a first acquisition function, an experimental design based on a Bayesian optimization method using a second acquisition function different from the first acquisition function, and an experimental design different from the Bayesian optimization method. The second experimental design includes an experimental design based on a Bayesian optimization method using an acquisition function that is the same as or different from the acquisition function used in the first experimental design.

[0079] The experimental system described in paragraph 6 can achieve the same effects as the searching method described in paragraph 1.

[0080] The embodiments disclosed herein should be considered to be illustrative in all respects and not restrictive. The scope of the present invention is defined by the claims, not by the description of the above embodiments, and is intended to include all modifications within the meaning and scope of the claims.

[0081] 1 Experimental system, 100 Search device, 101 Processor, 102 Memory, 110 Display, 120 Input device, 130 Database, 200 Experimental device, 210 Culture medium preparation device, 211 Well plate, 212 Well, 220 Culture device, 230 Measurement device.

Claims

1. A method for searching experimental conditions used in an experimental apparatus capable of performing a plurality of experiments each having different experimental conditions, comprising the steps of: setting first experimental conditions using a first experimental design; controlling the experimental apparatus to perform a first experiment based on the first experimental conditions; and setting second experimental conditions using a second experimental design based on the results of the first experiment, wherein the first experimental design includes at least two of an experimental design by a Bayesian optimization method using a first acquisition function, an experimental design by a Bayesian optimization method using a second acquisition function different from the first acquisition function, and an experimental design other than the Bayesian optimization method, and the second experimental design includes an experimental design by a Bayesian optimization method using an acquisition function that is the same as or different from the acquisition function used in the first experimental design.

2. The method for searching experimental conditions according to claim 1, wherein the first acquisition function is UCB (Upper Confidence Bound), and the second acquisition function is EI (Expected Improvement).

3. A method for searching experimental conditions according to claim 1, wherein the experimental design method different from the Bayesian optimization method is an experimental design method that comprehensively sets experimental conditions without using any experimental data.

4. The method for searching experimental conditions according to claim 1, wherein the experimental apparatus is configured to enable cells to be cultured in multiple wells in a single process.

5. A method for searching experimental conditions as described in claim 1, wherein the experimental equipment is configured to automatically execute an experiment in response to a control signal, the step of setting the first experimental conditions includes a step of generating a control signal for causing the experimental equipment to execute the first experiment, and the step of setting the second experimental conditions includes a step of generating a control signal for causing the experimental equipment to execute a second experiment based on the second experimental conditions.

6. An experimental system comprising: an experimental apparatus capable of executing a plurality of experiments each having different experimental conditions; and a search device that searches for experimental conditions to be used in the experimental apparatus, wherein the search device sets first experimental conditions using a first experimental design method, controls the experimental apparatus to execute a first experiment based on the first experimental conditions, and sets second experimental conditions using a second experimental design method based on results of the first experiment, wherein the first experimental design method includes at least two of an experimental design by a Bayesian optimization method using a first acquisition function, an experimental design by a Bayesian optimization method using a second acquisition function different from the first acquisition function, and an experimental design other than the Bayesian optimization method, and wherein the second experimental design method includes an experimental design by a Bayesian optimization method using an acquisition function that is the same as or different from the acquisition function used in the first experimental design.