State inference system and state inference method

Abstract

To provide a state inference system and a state inference method, which can improve the efficiency of bioproduction and research and development using a biological sample.SOLUTION: A state inference system 1a is provided with: a measurement unit 2 for measuring characteristic values of a plurality of culture media and at least one target culture medium; and a state inference unit 10a for inferring a culture state of a culture from characteristic values of the target culture medium on the basis of a feature amount extraction model that characterizes a distribution of characteristic values of the plurality of culture media. In the state inference system 1a, since it can be inferred in advance whether a desired culture state can be obtained in bioproduction and research and development using a target culture medium, a time required for making a proper decision on the target culture medium can be shortened. As a result, the efficiency of bioproduction or research and development using a culture medium can be improved.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a state estimation system and a state estimation method. [Background technology]

[0002] Raw materials of biological origin, such as natural media, are widely used as media, samples, and reagents for cell culture, but it is known that raw materials of biological origin have lot and batch differences. This is a direct cause of variation in the quality of cell culture, and in the processes of bioproduction, research and development, and testing, the phenomenon that the results of the process differ greatly even when the process is carried out under exactly the same conditions often occurs.

[0003] Analyzing the condition of a medium before using the medium to culture cells is important for improving the efficiency of biological production and research and development using the medium. For example, Patent Document 1 discloses a method for analyzing the components of a medium using spectroscopic analysis. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Special Publication No. 2013-544353 Summary of the Invention [Problem to be solved by the invention]

[0005] However, simply measuring the components of a medium using spectroscopic analysis makes it difficult to determine in advance from the measurement results whether or not the desired biological production or research and development results can ultimately be obtained using a biological sample such as a medium, and it is necessary to actually carry out biological production or research and development using the biological sample to the end to confirm whether the biological sample will ultimately produce the desired results. Therefore, since it takes time to determine whether a biological sample can produce the desired results in biological production and research and development using a biological sample, it is desired to shorten the time required for this determination and increase the efficiency of biological production and research and development using biological samples.

[0006] Therefore, an object of the present invention is to provide a state estimation system and a state estimation method that can improve the efficiency of biological production and research and development using biological samples. [Means for solving the problem]

[0007] The state estimation system of the present invention is provided with a culture medium, a culture cultivated in the culture medium, a culture medium extract extracted from the culture medium, and a culture extract extracted from the culture, which are used as a biological sample and a target biological sample, and includes a measurement unit that measures characteristics of a plurality of the biological samples at a certain point in time and at least one target biological sample of the same type as the biological samples at the certain point in time, a database that stores a feature extraction model that characterizes a distribution of input data including at least characteristic values ​​of the plurality of biological samples acquired via the measurement unit, and a state estimation unit that estimates a culture state of the culture from the input data of the target biological sample based on the feature extraction model, or estimates a culture time for the culture to reach a predetermined culture state.

[0008] Furthermore, the state estimation method of the present invention includes a measurement step in which a biological sample and a target biological sample are any of a culture medium, a culture cultured in the culture medium, a culture medium extract extracted from the culture medium, and a culture extract extracted from the culture, and a measurement unit is used to measure characteristics of a plurality of the biological samples at a certain time point and at least one of the target biological samples of the same type as the biological samples at the certain time point; a storage step in which a feature extraction model is stored in a database, the feature extraction model characterizing a distribution of input data including at least characteristic values ​​of the plurality of biological samples acquired via the measurement unit; and a state estimation step in which a state estimation unit estimates a culture state of the culture from the input data of the target biological sample based on the feature extraction model, or a culture time for the culture to reach a predetermined culture state is estimated by the state estimation unit. Effect of the Invention

[0009] According to the present invention, it is possible to estimate in advance whether a desired culture state can be obtained in biological production and research and development using a target biological sample, and the estimated result can be used as a guide to shorten the time required to properly judge the target biological sample, thereby improving the efficiency of biological production and research and development using the biological sample. [Brief description of the drawings]

[0010] [Figure 1] FIG. 1 is a block diagram showing a configuration of a state estimation system. [Diagram 2] FIG. 2 is a schematic diagram for explaining the timing of measuring the characteristics of a culture medium and the timing of measuring the culture state of a culture. [Diagram 3] 1 is a graph showing an example of a mass spectrum obtained from a mass spectrometer and an example of a measurement vector generated from the mass spectrum. [Figure 4] 1 is a table showing the culture results when Escherichia coli was cultured using each sample. [Diagram 5] 13 is a table showing the culture results when Escherichia coli is cultured using other samples and the state estimation results predicted by the state estimation system. [Figure 6] 13 is a graph showing the relationship between actual culture results and state estimation results. [Figure 7] 1 is a graph showing the measurement vector of an additive. [Figure 8] 1 is a table showing the dot products between the measurement vectors and the principal component vectors of additives. [Figure 9] 1 is a table showing the extent of improvement provided by each additive. [Figure 10] FIG. 13 is a schematic diagram for explaining another embodiment regarding the timing of measuring the characteristics of the culture medium. [Figure 11] FIG. 13 is a schematic diagram for explaining another embodiment regarding the timing of measuring the culture state of a culture. [Figure 12] FIG. 13 is a block diagram showing a configuration (2) of another state estimation system. [Figure 13] FIG. 1 is a schematic diagram for explaining an example in which culture A is induced to differentiate into culture B, culture C, or non-target cells through culture. [Figure 14] FIG. 13 is a block diagram showing a configuration of a state estimation system according to a fourth embodiment. [Figure 15] FIG. 1 is a schematic diagram for explaining a method for selecting culture A in a sorting section. [Figure 16] 16A is a schematic diagram illustrating an example in which culture A is induced to differentiate into culture E or non-target cells through culture, and 16B is a schematic diagram illustrating an example in which culture A is induced to differentiate into culture D or non-target cells through culture, and culture D is induced to differentiate into culture E or non-target cells through culture. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will now be described in detail with reference to the accompanying drawings. In the following description, the same components are designated by the same reference numerals, and duplicated description will be omitted.

[0012] (1) First embodiment (1-1) Configuration of state estimation system FIG. 1 is a block diagram showing the configuration of a state estimation system 1a according to this embodiment. In the first embodiment, a culture medium is shown as an example of a biological sample. The first embodiment is an example in which a mass spectrometer is applied as a measurement unit 2 described later. The state estimation system 1a according to this embodiment estimates the state of a culture medium used in a cell culture experiment and the culture state of a cultured product based on a feature extraction model that specifies feature amounts, and includes the measurement unit 2, a calculation processing unit 3a, a database 4, and a notification unit 5.

[0013] In this embodiment, the "culture state" refers to, for example, the state of the yield of the culture per unit volume of the medium used in the culture process, or the state of the quality of the culture. The yield of the culture also includes the yield of the culture extract extracted from the cultured culture and the yield of the components in the medium secreted by the cultured culture. For example, the yield of the culture extract is the yield of useful proteins (e.g., enzymes) extracted from E. coli when the culture is E. coli. For example, the yield of the components in the medium secreted by the cultured culture is the yield of biopharmaceuticals (e.g., antibodies) extracted from the medium when the culture is CHO cells. In this embodiment, the state of the culture state is mainly described as the state of the yield of the culture per unit volume of the medium used in the culture process.

[0014] The quality of a culture indicates the functionality of the culture. For example, when the culture is cells, the quality index indicates the degree of therapeutic effect (function) of the cultured cells against the target disease, or the differentiation state of the cells. The quality of a culture as the culture state of a biological sample also includes the quality of a culture extract extracted from the cultured culture and the quality of the components in the medium secreted by the cultured culture. For example, when the culture is E. coli, the quality of a culture extract is an index indicating the reaction efficiency (functionality) of a useful protein (e.g., an enzyme) extracted from E. coli. The quality of a component in the medium secreted by the culture is an index indicating the degree of the target function of a biopharmaceutical (e.g., an antibody) extracted from the medium.

[0015] The measurement unit 2 according to this embodiment is a mass spectrometer, and obtains a mass spectrum by performing mass analysis on the culture medium. In this embodiment, a mass spectrometer is used as the measurement unit, but the present invention is not limited to this. For example, an impedance sensor, a microscope device, a Raman spectrometer, a chromatography device, a digital PCR measurement device, a nuclear magnetic resonance device, an antibody quantification kit, or a nucleic acid sequence determination device may be used as the measurement unit. Examples of using these impedance sensors, microscope devices, etc. as the measurement unit will be described later.

[0016] Here, the mass spectrometer serving as the measurement unit 2 generates ions by irradiating a culture medium containing a mixture of a matrix, a solvent, etc. with laser light, detects the ions with an ion detector, and generates a mass spectrum based on the detection signal from the ion detector.

[0017] The mass spectrometer may be, for example, a mass spectrometer using an ion source such as EI (Electron Ionization), CI (Chemical Ionization), FAB (Fast Atom Bombardment), MALDI (Matrix Assisted Laser Desorption Ionization), or ESI (Electro Spray Ionization).The mass spectrometer may be, for example, a mass spectrometer using a mass separator (analyzer) such as a magnetic sector type, quadrupole type, ion trap type, or time-of-flight type, or a mass spectrometer using an ion detector such as an electron multiplier or a microchannel plate.

[0018] In this specification, the term "culture medium" refers to a medium for culturing a culture, which contains nutrients used to grow cells, microorganisms, or viruses (hereinafter simply referred to as cultures) and provides a growth environment for the culture. The medium according to this embodiment may be a known medium used to cultivate various cultures, and is appropriately determined depending on the type of culture. For example, the medium may be one containing inorganic salts, carbohydrates, amino acids, vitamins, proteins, peptides, fatty acids, lipids, serum, etc. in a buffer solution.

[0019] Here, the "cell" is not particularly limited as long as it can be cultured in a medium, and examples thereof include insect cells, animal cells, human cells, and plant cells, as well as cardiomyocytes, neural stem cells, mesenchymal stem cells, embryonic stem cells, hepatic cells, corneal stem cells, pancreatic islet cells, tumor cells, iPS cells, ES cells, and blood cells. Specifically, 293T cells are included. The "cell" also includes cell tissues that are aggregates of cells, such as artificial meat, retinal tissue, myocardial sheets, and multicellular biological tissues (e.g., fetuses, plants, etc.). The "microorganisms" include bacteria, fungi, microalgae, protozoa, yeast, and the like. The "bacteria" include Escherichia coli, archaea (e.g., thermophilus), Bacillus subtilis, cyanobacteria, and the like.

[0020] In this specification, each lot or batch of medium that may have differences in components due to lot differences, batch differences, etc. is referred to as one medium. When there are multiple such media, that is, media prepared from multiple different lots or multiple different batches, they are referred to as "multiple media."

[0021] The state estimation system 1a according to this embodiment performs mass analysis on a plurality of culture media using the measurement unit 2, and generates a feature extraction model based on the obtained mass analysis data. The mass analysis data indicates a mass spectrum measured when the culture media are mass analyzed using the measurement unit 2, and a measurement vector (described later) generated from the mass spectrum, and can be a characteristic value of the culture media. The state estimation system 1a then performs mass analysis on a culture medium with unknown components using the measurement unit 2, and can estimate the culture state of a culture medium with unknown components based on the resulting mass analysis data and the above-mentioned feature extraction model. A biological sample for which the culture state of a culture is estimated using the generated feature extraction model is referred to as a "target biological sample". In this embodiment, the case where a culture medium is applied as an example of a biological sample will be described below, and a culture medium for which the culture state of a culture medium is estimated using the generated feature extraction model is referred to as a "target culture medium".

[0022] The measurement unit 2 performs mass analysis on each of the multiple culture media and outputs the obtained mass spectrum to the calculation processing unit 3a and the database 4. When the mass spectrum (mass analysis data) of the multiple culture media is input from the measurement unit 2, the database 4 stores the mass spectrum of each culture medium in association with the date and time when the mass spectrum was measured. The database 4 also stores information on the characteristics of the culture medium (the characteristics include, for example, the composition (spec information) of the culture medium, the model number, the lot number, the date of manufacture, the date of arrival, the past usage history, etc.), and also stores information on parameters related to the culture conditions (hereinafter also referred to as culture parameters) such as the culture temperature, culture time, pH of the culture medium, and the stirring speed of the culture medium in association with the mass analysis data. The database 4 also stores various data such as the feature extraction model generated by the calculation processing unit 3a.

[0023] The calculation processing unit 3a includes a state estimation unit 10a, a judgment unit 12, and a measurement object optimization estimation unit 14. The state estimation unit 10a receives mass spectra of a plurality of culture media from the measurement unit 2 or the database 4.

[0024] The state estimation unit 10a according to the present embodiment generates a feature extraction model by, for example, performing multivariate analysis using mass spectrometry data of multiple culture media obtained at a certain point in time via the measurement unit 2 and the results of the culture state of the culture when the culture is actually cultured using the culture media (hereinafter, also referred to as the culture result). With such a feature extraction model, it is possible to analyze the components of the culture medium that affect the culture result when the culture is cultured using the culture medium.

[0025] In the case of supervised learning, it is desirable to apply a regression model or the like to the feature extraction model, which inputs mass spectrometry data of the target medium and outputs an estimate of the future culture state (hereinafter also referred to as state estimation result) that will ultimately be obtained when the culture of the culture is continued in the target medium. In addition, when generating a feature extraction model by supervised learning, it is desirable to learn the correspondence between mass spectrometry data of multiple media and the culture results when these media are used, and characterize the distribution of mass spectrometry data from the culture results.

[0026] As another example of supervised learning of feature extraction models, mass spectrometry data of a plurality of culture media, the culture results when these culture media are used, and correct answer labels indicating whether or not the culture results are the desired results (e.g., the accuracy rate indicating whether or not the culture results are the desired results) may be used, and when the mass spectrometry data of the culture media and the culture results are learned, the correct answer labels may be assigned and the model may be trained to characterize the distribution of the mass spectrometry data.

[0027] As the feature extraction model, various feature extraction models can be applied, such as, for example, Principal Component Analysis (PCA), Partial Least Squares (PLS), polynomial regression, Gaussian process regression, Random Forest Regression, etc. In addition, as the feature extraction model, for example, a machine learning model generated by machine learning the characteristic values ​​(mass spectrometry data) of the medium and the culture results may be applied.

[0028] As the machine learning model, for example, a neural network (for example, a CNN (Convolutional Neural Network) or a BNN (Bayesian Neural Network)) can be applied. In addition to the above-mentioned supervised learning, the feature extraction model may be unsupervised learning. Details of generating a feature extraction model by unsupervised learning will be described later, but when generating a feature extraction model by unsupervised learning, it is desirable to use mass spectrometry data of multiple culture media and perform learning to characterize the distribution of these multiple mass spectrometry data.

[0029] The determination unit 12 determines the state of the target medium based on the estimation result of the culture state of the culture obtained by the state estimation unit 10a (state estimation result). The determination unit 12 according to this embodiment sets discrimination values ​​capable of discriminating between good and bad medium in a stepwise manner in a state estimation space in which principal component vectors (features) are defined in a feature extraction model, projects the mass spectrometry data of the target medium onto the state estimation space, and estimates the state of the target medium using the discrimination values ​​as a guide (state determination process).

[0030] In this specification, a medium in which the mass of cultured matter per unit volume of medium finally obtained when culturing a culture using the medium is equal to or greater than a predetermined optimal setting value (hereinafter referred to as the good judgment value) is referred to as a "good medium," and a medium in which the mass is equal to or less than a predetermined setting value (hereinafter referred to as the bad judgment value) lower than the good judgment value is referred to as a "bad medium."

[0031] The determination unit 12 may also set a threshold for abnormality determination in a state estimation space in which the principal component vectors are defined in the feature extraction model (abnormality determination process). In this way, if the state estimation vector obtained by projecting the mass spectrometry data of the target medium onto the state estimation space does not exceed the threshold, the determination unit 12 can determine that the target medium is one that will give good culture results, and on the other hand, if the state estimation vector is less than the threshold, the determination unit 12 can determine that the target medium is one that will give poor culture results.

[0032] In this specification, a medium whose characteristics do not meet the required standards due to differences in medium lots or batches is referred to as an abnormal medium, and a medium whose characteristics meet the required standards is referred to as a normal medium.

[0033] The measurement target optimization estimation unit 14 estimates additives to be added to the target medium to obtain a desired culture state based on the state estimation result obtained by the state estimation unit 10a (measurement target optimization estimation process). This allows an operator to add an appropriate additive to a target medium that is estimated by the state estimation unit 10a not to obtain a desired culture state, thereby correcting the target medium to obtain a desired culture state.

[0034] When the measurement target optimization estimation unit 14 estimates what kind of additive should be added to the target culture medium in order to obtain the desired culture result, it outputs the obtained additive estimation result to the notification unit 5 described below. The additive estimation result obtained by the measurement target optimization estimation unit 14 is presented to an operator, etc. via the notification unit 5. This allows the operator, etc. to add the additive presented by the state estimation system 1a (estimated additive) to the target culture medium, thereby executing a culture process in the target culture medium with improved performance, and attempting to obtain the desired culture result.

[0035] The notification unit 5 may be, for example, a display device that displays the state estimation result, the determination result, etc. by the state estimation unit 10a on a display unit to allow the operator, etc. to visually recognize it, a transmission device that sends an email to the operator, etc. to notify the operator, etc. of the state estimation result, the determination result, etc., or a sound emission device that notifies the operator, etc. of the state estimation result, the determination result, etc. by sound. The operator, etc. may be an operator of the state estimation system 1a, an experimenter of the culture experiment, or an arbitrarily determined manager, etc.

[0036] In addition, the notification unit 5 can allow an operator, etc. to understand the contents of the feature extraction model generated by the calculation processing unit 3a, the measurement results (mass spectrum and measurement vector) in the measurement unit 2, and the results of estimated additives obtained by the measurement object optimization estimation unit 14 using a notification device, a transmission device, etc.

[0037] Next, the above-mentioned state estimation unit 10a, determination unit 12, and measurement object optimization estimation unit 14 will be described in order.

[0038] (1-2) State Estimation Unit Next, the state estimation process in the state estimation unit 10a will be described. Here, as an example, an example of generating a feature extraction model by PLS regression will be described. As described above, the culture state of the culture estimated by the state estimation unit 10a may be the quality (functionality) of the culture obtained in the culture process, but here, as an example, the case where the yield of the culture obtained in the culture process is used will be mainly described.

[0039] Here, FIG. 2 is a schematic diagram showing the timings for measuring the mass spectrometry data of a plurality of culture media and the culture state of the culture, which are acquired by the state estimation unit 10a when generating a feature extraction model. In FIG. 2, S indicates the start time of the culture process, and F indicates the end time when a predetermined culture period has elapsed in the culture process. Note that the "start time S of the culture process" indicates the time when the culture of the culture starts in the culture medium under culture conditions in which a predetermined culture parameter is determined, and the "end time F of the culture process" indicates the time when the culture of the culture in the culture medium is ended. In the state estimation system 1a according to this embodiment, as shown in FIG. 2, before the start of the culture process (before the start time S), mass spectrometry data of a plurality of culture media is acquired via the measurement unit 2 (Sa1). In addition, in the state estimation system 1a, after the end of the culture process (after the end time F), the yield of the culture after the culture is actually cultured in each culture medium is examined, and the culture state of the culture is measured (Fa1).

[0040] The state estimation unit 10a characterizes the relationship between the mass spectrometry data of the medium before the start of the culture process (before use for culture) obtained in this manner and the yield of the culture contained in the medium after the end of the culture process (after use for culture) (i.e., the culture result), and estimates the yield of the culture (culture state of the culture) that will be obtained in the future when the target medium is used in the culture process from the mass spectrometry data measured on the target medium before use for culture, whose components are unknown. This makes it possible to estimate whether the target medium is one that will provide a desired culture state before the culture process is started.

[0041] In this way, estimating in advance what kind of culture state the culture will be when the target medium is used in the culture process before the culture process is started is important for quickly detecting the success or failure of the culture process in cases where the initial medium varies due to lot differences, such as in the production of biopharmaceuticals. In addition, by identifying in advance the cause of culture failure caused by the medium and taking measures such as correcting the medium so that the desired culture result can be obtained depending on the cause, the yield and quality can be stabilized. Furthermore, by using the state estimation result obtained by the state estimation system 1a, culture failure can be prevented in advance, and therefore significant economic damage caused by culture failure can be prevented.

[0042] In this embodiment, for example, a culture medium containing yeast extract is used to culture Escherichia coli as a culture medium. In such a culture of Escherichia coli, even if the same amount of yeast extract is added to the same medium, the growth performance of Escherichia coli may vary depending on the lot of the yeast extract.

[0043] In this case, the state estimation system 1a finds a correlation between the mass spectrometry data of each medium to which different yeast extracts have been added and the culture results when E. coli is cultured using the media to which these different yeast extracts have been added, and constructs a feature extraction model to clearly detect the difference in function between these media. By using such a feature extraction model, the state estimation system 1a can estimate the state of the yeast extract in the medium, which is related to the growth performance of E. coli, for a target medium whose components are unknown.

[0044] In this embodiment, for example, mass analysis is performed on multiple types of yeast extract aqueous solutions (culture media) from multiple different lots by the measurement unit 2, and a mass spectrum of the mass analysis for each yeast extract aqueous solution is obtained. 3A, 3B, and 3C in Figure 3 show examples of mass spectra obtained by mass analysis of different types of media (yeast extract aqueous solutions).

[0045] In Figures 3A, 3B, and 3C, the horizontal axis is m / z and the vertical axis is signal intensity. The m / z on the horizontal axis is a measurement value that can be read directly from the primary measurement results of a mass spectrometer, and is generally defined as "a dimensionless quantity obtained by dividing the mass of an ion by the unified atomic mass unit (= Dalton), and then further dividing it by the absolute value of the charge number of the ion." The peak intensity in the mass spectrum corresponds to the amount of ions with each m / z generated by ionization of the analyte (culture medium).

[0046] When the state estimation unit 10a according to this embodiment receives a mass spectrum by mass spectrometry from the measurement unit 2 or the database 4, it performs standardization and coarse-graining on the mass spectrum to generate a measurement vector with a reduced amount of information on signal intensity. Specifically, for standardization, each peak intensity is corrected so that the maximum peak intensity of the signal intensity of the mass spectrum becomes 100. For coarse-graining, the horizontal axis m / z of the mass spectrum is divided into a predetermined number of dimensions (30 dimensions in this case), the signal intensity is averaged for each dimension, and a measurement vector is obtained that shows the peak intensity for each dimension relative to the maximum peak intensity of 100. In this embodiment, the measurement vector obtained from the mass spectrum is mass spectrometry data and is used as a characteristic value of the target culture medium.

[0047] 3D, 3E, and 3F of FIG. 3 show examples of measurement vectors, where 3D shows a measurement vector generated from the mass spectrum of 3A, 3E shows a measurement vector generated from the mass spectrum of 3B, and 3F shows a measurement vector generated from the mass spectrum of 3C.

[0048] In addition, Figures 3D, 3E, and 3F show examples in which the horizontal axis m / z of a mass spectrum obtained by mass spectrometry is divided into a predetermined number of dimensions, the maximum peak intensity of the signal intensity for each dimension is defined as 100, and the maximum and minimum values ​​are subtracted within each dimension to determine the peak intensity for each dimension relative to the maximum peak intensity of 100.

[0049] The database 4 also stores the measurement vectors generated by the state estimation unit 10a. In addition, the database 4 stores, for example, as the culture results, the final culture results when the culture medium is actually used to culture Escherichia coli, as shown in FIG. 4. In FIG. 4, the culture media from which the culture results were obtained are represented as samples 0, 1, 2, ..., and the actual culture results are shown for each culture medium. In addition, "train" shown in FIG. 4 indicates that the attribute of the data is sample data for generating a feature extraction model.

[0050] In this case, the same E. coli is actually cultured using each medium, the E. coli is recovered from the culture solution after the E. coli culture, and the dry mass (hereinafter also simply referred to as mass or yield) (g / L) of the E. coli per unit volume of medium is measured to obtain the culture results. The culture results (actual results) obtained in this way when E. coli is actually cultured using each medium are stored in database 4 in correspondence with the measurement vector of each medium.

[0051] The state estimation unit 10a performs PLS regression using measurement vectors of multiple media (i.e., mass spectrometry data, which are characteristic values ​​of the media) as explanatory variables and the mass (yield) (g / L) of E. coli per unit volume of medium, which is the final culture result when E. coli (culture product) is cultured using each medium, as the objective variable. PLS regression is a method of selecting principal component vectors (basis) of the measurement vectors so as to maximize the covariance between the substance content in the measurement vectors and the E. coli yield.

[0052] As a result, the state estimation unit 10a extracts highly relevant multidimensional (e.g., 30-dimensional) principal component vectors (supervised basis) to obtain a feature extraction model for the purpose of obtaining a high yield of E. coli per unit volume of culture medium and a good culture result. Based on the obtained principal component vector, the feature extraction model can identify the medium component that most contributed to the difference in the culture result.

[0053] Next, a case will be described in which the state of a yeast extract in a target medium whose components are unknown is estimated using a feature extraction model that characterizes the distribution of mass spectrometry data of the medium. In the state estimation system 1a, the measurement unit 2 performs mass analysis on the target medium whose components are unknown and whose state is to be estimated before the start of the culture process to obtain a mass spectrum. The state estimation unit 10a generates a measurement vector as mass spectrometry data from the mass spectrum of the target medium whose state is to be estimated. The state estimation unit 10a normalizes and projects the obtained measurement vector to the principal component vector defined in the state estimation space of the feature extraction model, thereby obtaining a state estimation vector (hereinafter also referred to as a state estimation result) for the target medium.

[0054] The value of the state estimation vector, which is the result of the state estimation, indicates the mass of E. coli per unit volume of culture medium predicted when culturing is performed using each yeast extract (i.e., the culture state of the culture after the culture process is completed). The more positive the state estimation vector is, the greater the yield of E. coli expected as a culture result.

[0055] In this way, the state estimation unit 10a can output a rough estimate of the yield of E. coli that is expected when E. coli is cultured in a target medium whose components are unknown in the future (hereinafter also referred to as the state estimation result).

[0056] The state estimation result obtained by the state estimation unit 10a is stored in the database 4 and is presented to the operator or the like via the notification unit 5. In this way, in the state estimation system 1a, the state estimation result output by the state estimation unit 10a can also be presented to the operator or the like via the notification unit 5 directly, without the state of the target culture medium being judged by the judgment unit 12 described below. Thus, in the state estimation system 1a, it is possible to have the operator or the like estimate whether or not the target culture medium will give good culture results, based on the state estimation result that estimates the culture state of the culture, before actually using the target culture medium for culture.

[0057] FIG. 5 is a table showing the actual culture results (experimental results) in which the same E. coli was actually cultured using five types of target media (yeast extract aqueous solutions) with different components, the E. coli was recovered from the culture solution after E. coli culture, and the dry mass (g / L) of E. coli per unit volume of medium was measured, and also showing the state estimation results (estimated yield results of E. coli per unit volume of medium contained in the target medium after being used for culture) obtained from the measurement vectors of the five types of target media using the feature extraction model generated in advance in the above-mentioned state estimation system 1a.

[0058] In Fig. 5, the target culture media with unknown components are denoted as samples 0, 1, 2, 3, and 4. Also, "test" in Fig. 5 indicates that the data attribute is sample data for analyzing the target culture media with unknown components using the generated feature extraction model.

[0059] As shown in FIG. 5, the state estimation results obtained by the state estimation system 1a are close to the actual culture results. Therefore, by presenting these state estimation results to an operator or the like, the operator or the like can predict the culture results that will be obtained when the target culture medium is used, without actually performing a culture process using the target culture medium.

[0060] Fig. 6 is a graph showing the relationship between the state estimation results shown in Fig. 5 and actual culture results (actual results) shown for reference. The horizontal axis of Fig. 6 shows the actual culture results (actual results), and the vertical axis shows the state estimation results. Fig. 6 also shows that, as a general trend, the state estimation results obtained by the state estimation system 1a are able to predict the actual culture results.

[0061] (1-3) Judgment section Next, there will be described the determination process in the determination unit 12. The determination unit 12 determines the state of the target culture medium based on the state estimation result obtained by the state estimation unit 10a.

[0062] The judgment processing in the judgment unit 12 includes a state judgment processing in which, without performing the abnormality judgment described below on the target medium before use for culture, it is simply judged in a step-by-step manner whether the state of the target medium is suitable for culturing a culture, and an abnormality judgment processing in which it is judged whether the target medium before use for culture is in an abnormal state unsuitable for culturing a culture.

[0063] When performing the state determination process, the determination unit 12 sets, prior to the state determination process, a discrimination value capable of discriminating between a good medium and a bad medium in a staged manner in a state estimation space in which a principal component vector (supervised basis) is defined as a feature in a feature extraction model. In addition, this discrimination value may be associated with a discrimination display that indicates the degree of goodness of the state of the medium in stages, for example, based on the culture result obtained when the medium is used for culturing a culture.

[0064] The discrimination value for the state estimation space may be arbitrarily set by an operator or the like, or the discrimination value may be automatically set by automatically dividing the principal component vectors defined in the state estimation space at predetermined intervals. The determination unit 12 determines the state of the target culture medium from the discrimination value and the value of the state estimation vector obtained by projecting the mass spectrometry data (measurement vector) of the target culture medium onto the state estimation space.

[0065] The determination unit 12 may set a threshold value for abnormality determination in the state estimation space of the feature extraction model based on a rule for the abnormality determination. In this case, the threshold value may be set in the state estimation space by an operator or the like, taking into consideration a desired culture result. For example, the threshold value is set so that if the threshold value is not exceeded, a good culture result equal to or greater than the reference value is obtained, and if the threshold value is exceeded, a poor culture result below the reference value is obtained.

[0066] The determination unit 12 determines the presence or absence of an abnormality in the target medium based on the value of a state estimation vector obtained by projecting the mass spectrometry data of the target medium onto a state estimation space in which the principal component vectors are defined, and a threshold value. The above threshold value may be set to be divided into multiple stages to determine the degree of abnormality in the target medium.

[0067] The notification unit 5 receives the determination result of the target medium by the state determination process, and notifies an operator, etc. of the determination result of the state of the target medium before use for cultivation, which is determined based on the identification value. In addition, the notification unit 5 receives the determination result of the abnormality determination process, and notifies an operator, etc. of the abnormality or the degree of the abnormality.

[0068] For example, if an abnormality is detected, the notification unit 5 may notify the operator of information indicating the abnormality, including the characteristic values ​​of the target culture medium and an error code, and if no abnormality is detected, the notification unit 5 may notify the operator of the absence of an abnormality. Also, the operator may be notified only when there is a particular abnormality, or a display unit may be provided that lights up only when there is an abnormality, and a display may be provided only when there is an abnormality.

[0069] (1-4) Measurement object optimization estimation unit Next, the measurement object optimization estimation process in the measurement object optimization estimation unit 14 will be described. When the state estimation unit 10a obtains an estimation result (state estimation result) of the culture state of the culture based on the feature extraction model from the mass analysis data of the target medium before the start of the culture process, the measurement object optimization estimation unit 14 receives the state estimation result from the state estimation unit 10a. The measurement object optimization estimation unit 14 analyzes the state estimation result obtained by the state estimation unit 10a and estimates an additive that can be added to the target medium before use for culture to maximize the yield of the culture per unit volume of medium contained in the target medium after use for culture. The measurement object optimization estimation unit 14 corrects the components of the target medium used in the culture process by having an operator or the like add the estimated additive (also referred to as an estimated additive) to the target medium.

[0070] In this embodiment, an additive that maximizes the yield of the culture per unit volume of the medium contained in the target medium after use for culture is described as an example of an additive that corrects the components of the target medium, but the present invention is not limited to this. For example, the additive may be an additive that improves the quality of the culture contained in the target medium after use for culture (for example, brings the culture into a desired differentiation state, etc.).

[0071] In this manner, in this embodiment, the composition of the target medium can be changed by adding an appropriate type of additive to the target medium, and the yield of culture material (culture state) per unit volume of medium contained in the target medium after culture use can be improved before the target medium is actually used in a culture process.

[0072] Here, a case will be described in which the culture state is improved by adding one of three preselected additives to the target medium. In this case, the state estimation system 1a performs mass analysis on the three additives using the measurement unit 2 to obtain a mass spectrum, and the state estimation unit 10a obtains an additive measurement vector that reduces the amount of information in the mass spectrum.

[0073] Fig. 7 is a graph showing examples of additive measurement vectors for three types of additives, where Fig. 7A shows the additive measurement vector for a first additive, Fig. 7B shows the additive measurement vector for a second additive, and Fig. 7C shows the additive measurement vector for a third additive. The three types of additive measurement vectors obtained by state estimation unit 10a are stored in database 4.

[0074] The measurement target optimization estimation unit 14 calculates the relevance (similarity) between the additive measurement vector, which is mass spectrometry data on multiple types of additives, and the principal component vector defined in the state estimation space in the feature extraction model. The measurement target optimization estimation unit 14 estimates, from among the three types of additives, the additive of the additive measurement vector that is most highly related to the basis of the feature extraction model as the optimal additive to be added to the target medium. In this way, the culture results can be improved by adding the additive of the additive measurement vector that is most highly related to the basis of the feature extraction model to the target medium before use for culture.

[0075] Here, an additive whose additive measurement vector is highly correlated with the basis of the feature extraction model is an additive whose additive measurement vector has a high similarity to the principal component vector defined in the state estimation space in the feature extraction model. In this embodiment, the similarity between the principal component vector of the feature extraction model and the additive measurement vector is found by calculating the inner product between the principal component vector of the feature extraction model and the additive measurement vector.

[0076] In addition, when there are multiple principal component vectors in the feature extraction model, for example, the principal component vector with the largest eigenvalue may be selected, and the inner product between the selected principal component vector and the additive measurement vector may be calculated. Alternatively, the inner product with the additive measurement vector for each principal component vector may be calculated, and a sum value may be calculated by adding up the calculated inner products. In this case, the additive with the largest sum value among the sum values ​​calculated for each additive measurement vector is selected as the additive to be corrected.

[0077] FIG. 8 is a table summarizing an example of the inner products between the principal component vectors of the feature extraction model and three types of additive measurement vectors. In FIG. 8, the first additive is denoted as "additive0", the second additive is denoted as "additive1", and the third additive is denoted as "additive2". Here, the third additive "additive2" has the largest inner product with the principal component vectors of the feature extraction model and the additive measurement vector that is most related to the principal component vectors of the feature extraction model.

[0078] In this way, the measurement object optimization estimation unit 14 calculates the inner product between the principal component vector of the feature extraction model and the additive measurement vector as the similarity between the principal component vector of the feature extraction model and the additive measurement vector, and selects the additive with the largest inner product. The measurement object optimization estimation unit 14 outputs additive information such as the name of the selected additive as a correction candidate additive to the notification unit 5. The notification unit 5 notifies the content of the additive information received from the measurement object optimization estimation unit 14 to an operator, etc., by displaying it, etc.

[0079] Here, Figure 9 shows an example of the results of the yield of E. coli per unit volume of medium contained in each target medium after use for cultivation, in which three types of additives having the additive measurement vectors shown in Figure 7 were added to five types of target media (samples 0 to 4) shown in Figure 5, and E. coli was actually cultivated using each target medium whose composition was corrected with the additives.

[0080] The attributes in Figure 9 indicate that additives have been added to each target medium prepared as a sample to correct the components. For example, "corrected(additive2)" indicates that a third additive has been added to the sample, "corrected(additive1)" indicates that a second additive has been added to the sample, and "corrected(additive0)" indicates that a first additive has been added to the sample.

[0081] In this example, by adding a third additive (additive2) to the target medium, which has the additive measurement vector that is most strongly correlated with the principal component vector of the feature extraction model, the yield of E. coli per unit volume of medium contained in each target medium after cultivation use increases, and the improvement is greater than that of the other additives.

[0082] As described above, in the state estimation system 1a, even if the state estimation unit 10a estimates that the desired culture state cannot be obtained in the target culture medium, a correction method that can improve the culture state by adding an optimal additive to the target culture medium can be presented to an operator, etc. This allows the operator to add the optimal additive estimated by the measurement target optimization estimation unit 14 to the target culture medium, thereby executing a culture process in the target culture medium with improved performance, and obtaining a desired culture state.

[0083] In the above embodiment, the similarity between the principal component vector defined in the state estimation space in the feature extraction model and a plurality of types of additive measurement vectors is calculated, and additives of the additive measurement vectors having a high similarity to the principal component vector are specified by the measurement target optimization estimation unit 14. However, the present invention is not limited to this. For example, information on one or more additives may be associated in advance with the estimated result of the culture state that will be obtained by the state estimation unit 10a, and when the estimated result of the culture state is obtained from the mass spectrometry data of the target culture medium based on the feature extraction model, one or more additives associated in advance with the estimated result may be estimated as the optimal additives to be added to the target culture medium.

[0084] (1-5) Action and Effects In the above configuration, the state estimation system 1a includes a measurement unit 2 that measures mass spectrometry data of a plurality of culture media and at least one target culture medium as characteristic values ​​(input data), and a state estimation unit 10a that estimates the culture state of the culture from the characteristic values ​​of the target culture medium based on a feature extraction model that characterizes the distribution of the characteristic values ​​(input data) of the plurality of culture media. In this way, the state estimation system 1a can estimate in advance whether or not a desired culture state can be obtained in biological production and research and development using the target culture medium, thereby shortening the time required to properly determine the target culture medium. This can improve the efficiency of biological production and research and development using the culture medium.

[0085] (1-6) Unsupervised generation of feature extraction models In the above-described embodiment, a feature extraction model that characterizes the distribution of characteristic values ​​(mass spectrometry data) of multiple biological samples is described, in which a highly relevant multidimensional (e.g., 30-dimensional) principal component vector (supervised basis) is extracted to obtain a feature extraction model for the purpose of obtaining a large culture yield per unit volume of culture medium and a good culture condition.

[0086] Here, for example, a state estimation system 1a is described which performs PCA (principal component analysis) processing using mass spectrometry data of a plurality of culture media, extracts multidimensional principal component vectors (unsupervised basis) that characterize the distribution of the mass spectrometry data, and obtains a feature extraction model. Note that the measurement object optimization estimation process performed by the measurement object optimization estimation unit 14 using the state estimation results obtained based on the feature extraction model is the same as in the above-mentioned embodiment, and therefore a description thereof is omitted here.

[0087] In this case, the measurement unit 2 performs mass analysis on a plurality of culture media, and the obtained plurality of mass analysis data is used to perform, for example, multivariate analysis to construct a feature extraction model that can clearly detect differences between culture media. Specifically, when the state estimation unit 10a receives a mass spectrum from the mass analysis from the measurement unit 2 or the database 4, it standardizes and coarse-grains the mass spectrum to generate a measurement vector with a reduced amount of signal intensity information as mass analysis data. The state estimation unit 10a performs PCA processing on the measurement vectors of the plurality of culture media to extract principal component vectors (unsupervised basis) that characterize the distribution of the mass analysis data as features, thereby generating a feature extraction model.

[0088] The state estimation system 1a then obtains mass spectrometry data of a target medium whose components are unknown, and based on the mass spectrometry data of the target medium and the feature extraction model, it can estimate which medium the target medium whose components are unknown is most similar to when compared with past media.

[0089] In this way, even in the feature extraction model generated by PCA processing, a threshold for abnormality judgment can be set in the state estimation space of the feature extraction model based on a certain rule for abnormality judgment in the judgment unit 12. In this case, the threshold is simply defined as a quantity related to the distribution when each mass analysis data of a plurality of culture media is projected onto the state estimation space in which the principal component vectors are defined.

[0090] For example, when the average of the points obtained by projecting the mass spectrometry data of each culture medium onto a state estimation space defining multiple principal component vectors extracted based on the mass spectrometry data is μ, the abnormality threshold can be given as a quantity related to the Mahalanobis distance from μ.

[0091] Specifically, for example, if a point group onto which mass spectrometry data is projected is classified into a normal group and an abnormal group, and the mass spectrometry data of the normal group is projected onto a one-dimensional space and the distribution is approximated to a normal distribution, the mean value of the distribution is μ and the absolute value of the standard deviation of the normal distribution is σ, then the threshold can be set to μ±3σ. In other words, when the mass spectrometry data of a certain medium is projected onto the one-dimensional space and the coordinates are deviated from μ by 3σ or more, the threshold can be set so that the medium is judged to be abnormal.

[0092] The judgment unit 12 judges the presence or absence of an abnormality using the set threshold value and the coordinate x when the mass spectrometry data of the target culture medium acquired from the state estimation unit 10a is projected into the state estimation space. As an example, if the distance between the coordinate x and the average value μ of the distribution of the normal group is 3 or more when converted into the Mahalanobis distance of the distribution of the normal group, it is judged that there is an abnormality. If not, it is judged that there is no abnormality.

[0093] The degree of abnormality of the target culture medium may be determined by dividing the threshold into multiple stages. As an example, if the distance between the coordinate x and the average value μ of the distribution of the normal group is 2 or more and less than 3 when converted into the Mahalanobis distance of the distribution of the normal group, the degree of abnormality may be determined to be small, and if it is 3 or more, the degree of abnormality may be determined to be large. The determination unit 12 outputs a determination result of the presence or absence of abnormality in the target culture medium or the degree of abnormality.

[0094] (1-7) Other embodiments regarding the timing of measuring the characteristics of a biological sample (1-7-1) When measuring the characteristics of a culture medium during the culture process after the start of the culture process In the above-described embodiment, the case where the measurement unit 2 performs mass analysis of the culture medium to obtain mass analysis data before the start of the culture process has been described, but the present invention is not limited to this. In the above-described embodiment, for example, as shown in 10A of Fig. 10, after the start of the culture process, the measurement unit 2 may perform mass analysis of the culture medium in the middle of being used in the culture process to obtain mass analysis data.

[0095] In this case, as shown in 10A of Fig. 10, the state estimation system 1a performs mass analysis of the multiple culture media by the measurement unit 2 at a certain point Sa2 during the culture process (S in Fig. 10 indicates the start time of the culture process, as in Fig. 2), and acquires mass analysis data. In this way, the state estimation system 1a accumulates mass analysis data at a certain point Sa2, a predetermined period of time after the start of culture, for the multiple culture media currently culturing the cultured product in the culture process.

[0096] Then, in a given culture process, the culture is cultured until the end of the originally planned culture period, and when the culture of the culture in each medium is completed, the culture after the culture is completed contained in the medium is collected, and the yield of the culture that has been cultured by the culture process is measured (Fa1). In this way, the state of the culture after culture is measured, and the actual culture result is obtained for each medium.

[0097] In the state estimation system 1a, when the yield of the culture obtained for each medium (actual culture result) is input by an operator or the like, the result of the culture state is stored in the database 4. At this time, the database 4 stores mass spectrometry data of each medium measured at a certain point Sa2 during the culture process in association with the final culture result obtained using that medium.

[0098] In the state estimation unit 10a according to this embodiment, mass spectrometry data of each culture medium measured at a certain point Sa2 during the culture process and the final culture state obtained with that culture medium are read as sample data from the database 4 to generate a feature extraction model.

[0099] This feature extraction model takes as input mass spectrometry data of the target culture medium at a certain point Sa2 (e.g., the 10th day after the start of culture) after the start of cultivation of the culture using a specified cultivation process, and outputs an estimate of the future culture state that will ultimately be obtained if the culture is continued to be cultivated in the target culture medium.

[0100] As in the above-described embodiment, the feature extraction model may be, for example, various feature extraction models such as PCA, PLS, polynomial regression, Gaussian process regression, Random Forest Regression, etc., feature extraction models generated by machine learning of characteristic values ​​(mass spectrometry data) of culture media during culture and the culture results, etc. When generating a feature extraction model by supervised learning, it is desirable to learn the correspondence between mass spectrometry data of multiple culture media and the culture results when these culture media are used, and characterize the distribution of mass spectrometry data from the culture results.

[0101] For example, in the case of supervised learning, the state estimation unit 10a uses mass spectrometry data of the culture medium during the culture process as explanatory variables, and the actual final culture state obtained in this culture process (the yield of the culture contained in the medium at the end of the culture) as the objective variable, and generates in advance a feature extraction model that shows characteristic relationships such as correlation and inverse correlation between the mass spectrometry data and the culture state.

[0102] In this case, the feature extraction model may be based on unsupervised learning in addition to supervised learning as described above. When generating a feature extraction model by unsupervised learning, it is preferable to use only mass spectrometry data of a plurality of culture media and perform learning to characterize the distribution of the plurality of mass spectrometry data.

[0103] As a result, in the state estimation system 1a, while a culture process is being carried out using a target culture medium whose culture results are unknown, it is possible to use the feature extraction model to estimate the future culture results that will ultimately be obtained from the target culture medium during the culture process, before the culture process is completed.

[0104] For example, in the series of culture processes in which cell culture is induced to differentiate and the target cells are obtained, it takes a long time to obtain the target cells. In particular, the series of culture processes in which undifferentiated human stem cells are induced to differentiate and the target cells are obtained usually takes more than about three months. In such a culture process, for example, when trying to obtain optimal culture results (i.e., the target cells) by sequentially optimizing various parameters such as the components of the culture medium and the culture temperature, a method can be considered in which the actual culture results obtained are analyzed after the end of the culture process, and the next parameters (the components of the culture medium used, the culture temperature, etc.) are determined each time.

[0105] However, with this method, for example, if one cycle of the culture process from the start to the end of the culture takes about three months, only about four cycles can be performed per year, and trial and error of each parameter of the culture process is required in four cycles, which is inefficient in terms of time.

[0106] In contrast, in the state estimation system 1a according to this embodiment, since an estimation result of the culture state that will ultimately be obtained from the target culture medium during the culture process is obtained based on the feature extraction model, the estimation result can be presented to an operator during the culture, and various parameters of the culture process (such as the components of the culture medium used and the culture temperature) can be changed as necessary based on the obtained estimation result without waiting for the end of the culture process, and a new culture process can be performed. In this way, a new culture process can be started in the middle of the currently ongoing culture process, in which the parameters of the culture process are adjusted through trial and error so as to obtain the desired culture result, without waiting for the end of the culture process, and parameter optimization can be performed in a time-efficient manner.

[0107] The above-mentioned "(1-2) state estimation unit" uses mass spectrometry data of the culture medium before the start of the culture process, but this mass spectrometry data can be replaced with mass spectrometry data of the culture medium obtained during the culture process, and this embodiment can be applied. Similarly, the "(1-3) determination unit" and the "(1-4) measurement target optimization estimation unit" can also be applied by replacing the mass spectrometry data of the culture medium before the start of the culture process with mass spectrometry data of the culture medium obtained during the culture process.

[0108] For example, the determination unit 12 determines the state of the target culture medium based on the state estimation result obtained in the state estimation unit 10a from the mass spectrometry data of the culture medium during the culture process. Also, the measurement target optimization estimation unit 14 estimates additives to be added to the target culture medium to obtain a desired culture state based on the state estimation result obtained in the state estimation unit 10a from the mass spectrometry data of the culture medium during the culture process (measurement target optimization estimation process).

[0109] (1-7-2) When measuring medium characteristics multiple times during the culture process In the above-mentioned embodiment, the mass analysis of the culture medium is performed by the measuring unit 2 either before the start of the culture process or during the culture process to obtain mass analysis data, but the present invention is not limited to this. In the above-mentioned embodiment, for example, as shown in FIG. 10B, the mass analysis of the culture medium may be performed by the measuring unit 2 at a certain time point Sa1 before the start of the culture process and at certain times Sa2 and Sa3 during the culture process, and mass analysis data may be obtained multiple times. Also, the mass analysis of the culture medium may not be performed by the measuring unit 2 before the start of the culture process, but may be performed by the measuring unit 2 only at certain times Sa2 and Sa3 during the culture process, and mass analysis data may be obtained multiple times. That is, the mass analysis data of the culture medium may be measured multiple times at intervals in time series from the start of the culture, such as the mass analysis data of the culture medium on the 10th day after the start of the culture process and the mass analysis data of the culture medium on the 20th day after the start of the culture process.

[0110] 10B, the state estimation system 1a performs mass spectrometry on a plurality of culture media using the measurement unit 2 at a certain time point Sa1 before the start of the culture process and at certain times Sa2 and Sa3 during the culture process, and obtains mass spectrometry data. After the culture process is completed, the state estimation system 1a also measures the culture states (yield and quality of the cultures) of the cultures after they have been actually cultured in each medium (Fa1).

[0111] The state estimation unit 10a generates a feature extraction model based on the mass spectrometry data of the medium at a certain time point Sa1 before the start of the culture process and the mass spectrometry data of the medium at certain times Sa2 and Sa3 after the start of the culture process, which are obtained in this manner. That is, in the case of supervised learning, for example, the state estimation unit 10a generates a feature extraction model showing the relationship between the mass spectrometry data and the culture state, using the mass spectrometry data of the medium before the start of the culture process as an explanatory variable and the culture state of the culture contained in the medium after the end of the culture process (after use for culture) as a target variable. After that, when the state estimation unit 10a obtains mass spectrometry data measured before the start of the culture process for a target medium whose components are unknown, it estimates the culture state (yield and quality of the culture) of the culture to be obtained in the future when the target medium is used in the culture process from the mass spectrometry data based on the feature extraction model.

[0112] Similarly, the state estimation unit 10a generates a feature extraction model showing the relationship between each mass spectrometry data of the medium at a certain time point Sa2 after the start of the culture process and the culture state of the culture contained in the medium after the end of the culture process (after the culture use), and similarly generates a feature extraction model showing the relationship between each mass spectrometry data of the medium at a certain time point Sa3 after the start of the culture process and the culture state of the culture contained in the medium after the end of the culture process (after the culture use). The state estimation unit 10a estimates the culture state (yield and quality of the culture) of the culture to be obtained in the future when the target medium is used in the culture process from the mass spectrometry data measured at each time point (Sa2, Sa3) for the target medium currently being used for culture, whose components are unknown, based on the feature extraction model at the corresponding time points Sa2, Sa3.

[0113] In this way, mass analysis is performed on the target culture medium by the measurement unit 2 at a certain point Sa1 before the culture process is started and at certain points Sa2 and Sa3 during the culture process, and the culture process can be performed while sequentially estimating the culture state of the culture after the culture process is completed based on the feature extraction model corresponding to the certain points Sa1, Sa2, and Sa3 from the mass analysis data of the target culture medium that changes over time.

[0114] Note that, although the case where each feature extraction model is generated by supervised learning has been described here, the present invention is not limited to this, and each feature extraction model may be generated by unsupervised learning in accordance with the above-mentioned “(1-6) Generation of feature extraction models by unsupervised learning”.

[0115] (1-8) Another embodiment for estimating the culture state of a culture during the culture process before the end of the culture process In the above-described embodiment, the case where the culture state of the culture contained in the medium (e.g., the yield of the culture per unit volume of the medium contained in the target medium) is estimated after the culture process is completed (after the completion time F) has been described, but the present invention is not limited to this. In the above-described embodiment, for example, as shown in 11A of Fig. 11, the culture state of the culture at a certain time point Fa2 during the culture before the culture process is completed may be estimated.

[0116] In this case, as shown in 11A of Fig. 11, the state estimation system 1a performs mass analysis of multiple culture media by the measurement unit 2 at a certain time point Sa1 before the start of the culture process, and obtains mass analysis data. Also, the state estimation system 1a measures the culture state (yield and quality of the culture) of the culture that is actually being cultured in each culture medium at a certain time point Fa2 before the end of the culture process.

[0117] The state estimation unit 10a generates a feature extraction model by supervised learning based on the mass spectrometry data of the medium before the start of the culture process and the culture state of the culture during the culture, which are obtained in this manner. That is, the state estimation unit 10a generates a feature extraction model showing the relationship between the mass spectrometry data and the culture state, using the mass spectrometry data of the medium before the start of the culture process as an explanatory variable and the culture state of the culture contained in the medium at a certain time point Fa2 before the end of the culture process as a target variable. When the state estimation unit 10a subsequently obtains mass spectrometry data measured before the start of the culture process for a target medium whose components are unknown, the state estimation unit 10a can estimate the culture state (yield and quality of the culture) of the culture during the culture, which will be obtained in the future when the target medium is used in the culture process, from the mass spectrometry data based on the feature extraction model.

[0118] The above-mentioned "(1-2) state estimation unit" uses the measurement result of the culture state of the culture after the end of the culture process, but by replacing the measurement result of the culture state with the measurement result of the culture state of the culture obtained during the culture process, it can also be applied to this "(1-8) other embodiment of estimating the culture state of the culture during the culture before the end of the culture process". Similarly, the "(1-3) determination unit" and the "(1-4) measurement target optimization estimation unit" can also be applied by replacing the measurement result of the culture state of the culture after the end of the culture process with the measurement result of the culture state of the culture obtained during the culture process.

[0119] In addition, as shown in 11A of Fig. 11, the case has been described in which the measurement unit 2 performs mass analysis on a plurality of culture media at a certain time Sa1 before the start of the culture process, and mass analysis data is obtained, but the present invention is not limited to this. Even in this embodiment, for example, as shown in 11B of Fig. 11, the measurement unit 2 may perform mass analysis on a plurality of culture media at a certain time Sa2 during the culture process, and mass analysis data may be obtained.

[0120] In this case, the state estimation unit 10a uses mass spectrometry data of the medium at a certain time point Sa2 during the culture process as explanatory variables and the culture state of the culture contained in the medium at a certain time point Fa2 before the end of the culture process as a target variable to generate a feature extraction model that characterizes the relationship between the mass spectrometry data and the culture state. The state estimation unit 10a then obtains mass spectrometry data measured at a certain time point Sa2 during the culture process for a target medium with unknown components, and can estimate from the mass spectrometry data, based on the feature extraction model, the culture state of the culture at a certain time point Fa2 during the culture that will be obtained in the future when the target medium is used in the culture process.

[0121] 11C, mass spectrometry may be performed by the measurement unit 2 on the culture medium at a certain time point Sa1 before the start of the culture process and at certain times Sa2 and Sa3 during the culture process to obtain mass spectrometry data multiple times. Alternatively, mass spectrometry may not be performed by the measurement unit 2 on the culture medium before the start of the culture process, but may be performed by the measurement unit 2 on the culture medium only at certain times Sa2 and Sa3 during the culture process to obtain mass spectrometry data multiple times.

[0122] The state estimation unit 10a uses mass spectrometry data of the medium before the start of the culture process as explanatory variables, and the culture state of the culture contained in the medium at a certain time point Fa2 before the end of the culture process as a response variable, and generates a feature extraction model that characterizes the relationship between the mass spectrometry data and the culture state. When the state estimation unit 10a subsequently obtains mass spectrometry data measured before the start of the culture process for a target medium with unknown components, it can estimate from the mass spectrometry data, based on the feature extraction model, the culture state of the culture at a certain time point Fa2 during the culture that will be obtained in the future when the target medium is used in the culture process.

[0123] The state estimation unit 10a also generates a feature extraction model that characterizes the relationship between each mass spectrometry data of the medium at a certain time point Sa2 after the start of the culture process and the culture state of the culture contained in the medium at a certain time point Fa2 before the end of the culture process, and similarly generates a feature extraction model that characterizes the relationship between each mass spectrometry data of the medium at a certain time point Sa3 after the start of the culture process and the culture state of the culture contained in the medium at a certain time point Fa2 before the end of the culture process. The state estimation unit 10a then obtains mass spectrometry data measured at certain times (Sa2, Sa3) for a target medium in use for culture whose components are unknown, and can estimate the culture state (yield and quality of the culture) of the culture at a certain time point Fa2 during the culture that will be obtained in the future when the target medium is used in the culture process from the mass spectrometry data based on the feature extraction models at the corresponding certain times Sa2 and Sa3.

[0124] (1-9) Optimization and estimation of culture parameters Next, a state estimation system according to another embodiment, which performs optimization estimation processing of culture parameters, will be described. Fig. 12 is a block diagram showing the configuration of a state estimation system 1c according to another embodiment. The state estimation system 1c has a different configuration of a calculation processing unit 3c from the state estimation system 1a in Fig. 1 described above, and the following description will focus on this calculation processing unit 3c, and descriptions of the measurement unit 2, database 4, and notification unit 5 will be omitted to avoid duplication.

[0125] Here, for example, when culturing cells to be cultured in a medium, a situation may arise where culture conditions are considered to make the cells reach a certain culture state. For example, in the continuous culture of CHO cells, in which the cells are introduced into an initial medium and then grown to a certain number, it is necessary to consider optimal culture conditions for growing the CHO cells to a certain number after the CHO cells are introduced into the initial medium. In this case, if parameters (culture parameters) related to the culture conditions of the culture process are trial and error so as to obtain a desired culture state, and a new culture process is started from scratch each time, it takes time to identify the optimal culture conditions. Therefore, it is desirable to proceed with the optimization of the culture parameters in a time-efficient manner.

[0126] 12 according to this embodiment can estimate culture parameters that will result in a desired culture state with respect to parameters (culture parameters) related to culture conditions such as the culture temperature, culture time, culture medium pH, and culture medium stirring speed in the culture process, using a calculation processing unit 3c, and can present the estimation results to an operator or the like via a notification unit 5. This allows the culture process to be carried out using optimal culture parameters that will result in a desired culture state.

[0127] The arithmetic processing unit 3c according to this embodiment includes a state estimation unit 10c, a determination unit 12, and a culture parameter optimization estimation unit 15. When performing the culture parameter optimization estimation process in the arithmetic processing unit 3c, it is necessary to generate a feature extraction model for estimating the culture parameters in the state estimation unit 10c, which will be described below.

[0128] (1-9-1) Creation of feature extraction model for estimating culture parameters For example, a set of culture parameters that serve as a reference for culturing the culture is determined, such as the culture temperature, culture time, culture medium pH, and culture medium stirring speed. In addition, a plurality of correction candidate values ​​Δa for the culture parameters are determined based on the culture parameters that serve as a reference. Specifically, the correction candidate value Δa for the culture temperature is, for example, +2°C or -1°C, the correction candidate value Δa for the culture time is, for example, +30 minutes or -30 minutes, the correction candidate value Δa for the culture medium pH is, for example, +0.1 or -0.2, and the correction candidate value Δa for the culture medium stirring speed is, for example, +1s -1 Ya-1s -1 etc.

[0129] Here, as an example, a case will be described in which CHO cells are added as a culture to 20 culture media, and the culture (CHO cells) is started with a certain culture parameter a. Three hours after the start of the culture, the supernatant is taken, and mass analysis is performed on each medium using the measurement unit 2, which is a mass spectrometer, to obtain a mass spectrum. Then, three hours after the start of the culture, the culture is subcultured and dispensed.

[0130] Here, the culture medium of 20 culture media is divided into four portions and each portion is transferred to a different culture vessel. Four different culture parameters (a+△a0 (here, △a0=0)), (a+△a1), (a+△a2), (a+△a3) are set for each culture vessel, and the culture is continued with these culture parameters (a+△a0), (a+△a1), (a+△a2), (a+△a3). Here, △a1, △a2, △a3 indicate different correction candidate values ​​of the culture parameter, and for example, when the culture parameter is the culture temperature, they indicate the increase or decrease in temperature from the reference culture temperature.

[0131] Then, the actual yield of the culture per unit volume of medium (culture state) is obtained when the culture process is performed with each of the culture parameters (a+△a0), (a+△a1), (a+△a2), and (a+△a3), and this is stored as the actual result in database 4. In this way, 20 x 4 = 80 types of actual learning results are obtained, where all 20 media are cultured with the four culture parameters (a+△a0), (a+△a1), (a+△a2), and (a+△a3), respectively.

[0132] Then, the mass spectrometry data of the culture medium, the reference culture parameter a, the candidate correction values ​​for the culture parameter △a0, △a1, △a2, △a3, and the actual results showing the culture state of the culture when the culture was actually cultured with the culture parameters (a+△a0), (a+△a1), (a+△a2), (a+△a3) are matched, and these are each stored as a data set in database 4.

[0133] Next, the state estimation unit 10c reads out these data sets stored in the database 4, and generates a feature extraction model for estimating the culture parameters that characterizes the distribution of the input data, using the mass spectrometry data, the culture parameter a before correction that serves as a reference, and the correction candidate values ​​Δa0, Δa1, Δa2, and Δa3 of the culture parameter a as input data. Specifically, the state estimation unit 10c generates a feature extraction model for estimating the culture parameters that estimates what kind of correction should be made to the reference culture parameters to obtain a desired culture state for the culture 3 hours after the start of culture.

[0134] In this embodiment, the feature extraction model for estimating the culture parameters is generated using the mass spectrometry data, the culture parameter a before correction as a reference, and the correction candidate values ​​(Δa0, Δa1, Δa2, Δa3) of the culture parameter a as explanatory variables, but the present invention is not limited to this. The feature extraction model for estimating the culture parameters may be generated using the mass spectrometry data and the culture parameters (a+Δa0), (a+Δa1), (a+Δa2), and (a+Δa3) as explanatory variables. As described in "(1-1) Configuration of the state estimation system" above, the feature extraction model for estimating the culture parameters may be, for example, any of various feature extraction models such as PCA, PLS, polynomial regression, Gaussian process regression, convolutional neural network, and Bayesian neural network.

[0135] For example, in this case, the state estimation unit 10c uses the mass spectrometry data of the culture medium obtained by the measurement unit 2, the reference uncorrected culture parameter a, and the correction candidate values ​​△a0, △a1, △a1, △a1 as inputs (explanatory variables), and uses the culture state (here, the yield) of the culture contained in the culture medium after use for culture as an output (objective variable) to generate a feature extraction model for estimating the culture parameters by using a neural network or the like.

[0136] In this way, the state estimation system 1c generates a feature extraction model for estimating culture parameters by characterizing the distribution of input data based on the culture state of the cultured product cultivated in the culture process, using input data consisting of mass spectrometry data of the culture medium and culture parameters (a+△a0), (a+△a1), (a+△a2), and (a+△a3) set in the culture process.

[0137] (1-9-2) Application of feature extraction model for estimating culture parameters When the state estimation system 1c obtains mass spectrometry data of a target culture medium with unknown components via the measurement unit 2, the culture parameter optimization estimation unit 15 randomly selects a correction candidate value Δa of the culture parameter within a preset range, and outputs a plurality of correction candidate values ​​Δa to the state estimation unit 10c. Note that, although the case where the correction candidate value Δa of the culture parameter is randomly generated by the culture parameter optimization estimation unit 15 has been described here, the present invention is not limited to this, and for example, a plurality of correction candidate values ​​Δa stored in advance in the database 4 may be applied, or the correction candidate value Δa may be selected by an operator or the like.

[0138] The state estimation unit 10c inputs the mass spectrometry data of the target culture medium obtained via the measurement unit 2, the culture parameter a, and multiple correction candidate values ​​Δa of the culture parameter as input data into the feature extraction model for estimating the culture parameters described above and analyzes them, thereby estimating the culture state (yield) of the culture contained in the culture medium used for the culture.

[0139] The state estimation unit 10c outputs a plurality of estimation results obtained by changing the correction candidate value Δa to the culture parameter optimization estimation unit 15. The culture parameter optimization estimation unit 15 selects an estimation result that is closest to a preset optimum culture state result from the plurality of estimation results obtained by the state estimation unit 10c, and specifies the correction candidate value Δa of the culture parameter a that is set in the selected estimation result.

[0140] The culture parameter optimization estimation unit 15 outputs the identified correction candidate value Δa to the notification unit 5, and may present the correction candidate value Δa to an operator or the like via the notification unit 5. In this way, by presenting the correction candidate value Δa obtained by the culture parameter optimization estimation unit 15 to an operator or the like via the notification unit 5, the operator or the like can correct the culture parameters based on the correction candidate value Δa, and can cultivate a culture under optimal culture conditions (culture parameters).

[0141] (1-9-3) Action and Effects In the above configuration, the state estimation system 1c uses the characteristic values ​​of multiple culture media and multiple culture parameters measured by the measurement unit 2 as input data, and estimates the culture state of the culture from the characteristic values ​​and culture parameters of the target culture medium in the state estimation unit 10c based on a feature extraction model for estimating the culture parameters that characterizes the distribution of this input data. In this way, the state estimation system 1c can estimate in advance whether or not a desired culture state can be obtained in biological production and research and development using the target culture medium, thereby shortening the time required to properly determine the target culture medium. This can improve the efficiency of biological production and research and development using the culture medium.

[0142] In addition, the state estimation system 1c identifies a correction candidate value Δa of the culture parameter a that provides an optimal culture state based on the culture state estimation result obtained by the state estimation unit 10c. This allows the state estimation system 1c to have an operator or the like correct the culture parameter based on the obtained correction candidate value Δa of the culture parameter, thereby providing an optimal culture state.

[0143] In this embodiment, for example, the estimation result obtained by the state estimation unit 10c may be output to the judgment unit 12, and the judgment unit 12 may judge the culture parameters based on the estimation result. In this case, the culture parameters are judged based on the estimation result by the state judgment process or the abnormality judgment process described in "(1-3) Judgment unit" above. For example, when the judgment unit 12 performs the state judgment process, it sets a discrimination value capable of discriminating between a good estimation result and a bad estimation result in a stepwise manner in a state estimation space in which a principal component vector is defined as a feature in a feature extraction model for estimating the culture parameters, prior to the state judgment process. The judgment unit 12 selects a discrimination value close to the value of a state estimation vector obtained by projecting the mass spectrometry data and the culture parameters of the target medium into the state estimation space, and discriminates the estimation result from the selected discrimination value to judge the appropriateness of the culture parameters.

[0144] Furthermore, as an abnormality determination process, the determination unit 12 may set a threshold value for abnormality determination in a state estimation space of a feature extraction model for estimating culture parameters based on a certain rule for abnormality determination. The determination unit 12 identifies the estimation result from the value of a state estimation vector obtained by projecting the mass spectrometry data and culture parameters of the target medium onto a state estimation space in which a principal component vector is defined, and the threshold value, and determines the presence or absence of an abnormality in the culture parameters. Furthermore, the above threshold value may be set to be divided into multiple stages to determine the degree of abnormality in the culture parameters.

[0145] (1-10) State estimation system for estimating culture time In the above-described embodiments, the state estimation systems 1a and 1c that estimate the culture state of a culture based on a feature extraction model have been described, but the present invention is not limited to this. A state estimation system may also be provided that estimates a culture time until a culture being cultured in a culture process reaches a preset culture state based on a feature extraction model.

[0146] In this case, in each of the items "(1-1) Configuration of the state estimation system", "(1-2) State estimation unit", "(1-3) Determination unit", "(1-4) Measurement object optimization estimation unit", "(1-5) Action and effect", "(1-6) Generation of feature extraction model without a teacher", "(1-7) Other embodiments related to measurement timing of culture medium characteristics", and "(1-9) Optimization estimation process of culture parameters" described in "(1) First embodiment" above, the processing of each of these items can be executed by replacing the "culture state (yield) of the culture", which is the objective variable, with "culture time until the culture reaches a preset culture state".

[0147] (1-10-1) About the state estimation system shown in Figure 1 For example, if the state estimation system 1a shown in Fig. 1 is a state estimation system 1a that estimates a culture time, in the case of supervised learning, the state estimation unit 10a generates a feature extraction model using mass spectrometry data of multiple culture media obtained by the measurement unit 2 at a certain time as explanatory variables and the culture time at which a culture reaches a desired culture state when the culture is actually cultured using the culture medium as a response variable. Such a feature extraction model can analyze the components of the culture medium that affect the culture time at which a culture reaches a desired culture state when the culture is cultured using the culture medium (hereinafter, also simply referred to as the culture time).

[0148] In the case of supervised learning, the feature extraction model desirably applies a regression model or the like that inputs mass spectrometry data of the target medium and outputs an estimated culture time (hereinafter also referred to as time estimation result) required for the culture to reach a desired culture state when the culture is continued in the target medium. In addition, when generating a feature extraction model by supervised learning, it is desirable to learn the correspondence between the mass spectrometry data of multiple media and the culture time at which the culture reaches a desired culture state when these media are used, and characterize the distribution of the mass spectrometry data from the culture time.

[0149] As another example of supervised learning of feature extraction models, mass spectrometry data of multiple culture media, the respective incubation times obtained when using these culture media, and correct answer labels indicating whether or not the incubation times are the desired results (e.g., accuracy rates indicating whether or not the incubation times are the desired results) may be used, and when learning the mass spectrometry data of the culture media and their incubation times, the correct answer labels may be assigned and the learning may be performed to characterize the distribution of the mass spectrometry data.

[0150] As the feature extraction model, various feature extraction models can be applied, such as PCA, PLS, polynomial regression, Gaussian process regression, Random Forest Regression, etc. In addition, as the feature extraction model, for example, a machine learning model generated by machine learning the characteristic values ​​(mass spectrometry data) of the medium and the results of the culture time may be applied.

[0151] As the machine learning model, for example, a neural network (for example, a CNN (convolutional neural network) or a BNN (Bayesian neural network)) can be applied. In addition to the above-mentioned supervised learning, the feature extraction model may be unsupervised learning. Details of generating a feature extraction model by unsupervised learning will be described later, but when generating a feature extraction model by unsupervised learning, it is desirable to use mass spectrometry data of multiple culture media and perform learning to characterize the distribution of these multiple mass spectrometry data.

[0152] The judgment unit 12 judges the state of the target culture medium based on the estimation result (time estimation result) of the culture time required for the culture to reach the desired culture state, obtained by the state estimation unit 10a. The judgment unit 12 sets discrimination values ​​that can discriminate between good and bad culture mediums in a staged manner in a state estimation space in which the principal component vectors (features) are defined in the feature extraction model, projects the mass spectrometry data of the target culture medium into the state estimation space, and estimates whether the target culture medium is in a state in which the desired culture time can be achieved using the discrimination values ​​as a guide (state judgment process).

[0153] In this specification, a medium is referred to as a "good medium" if the culture time required for the mass of cultured matter per unit volume of medium obtained when culturing a culture using the medium to reach or exceed a predetermined optimal setting value (good judgment value) is within a preset arbitrary time, and a medium is referred to as a "poor medium" if the culture time exceeds the arbitrary time.

[0154] The determination unit 12 may set a threshold value for abnormality determination in the state estimation space of the feature extraction model based on a rule for the abnormality determination. In this case, the threshold value may be set in the state estimation space by an operator or the like, taking into consideration a desired culture time. For example, the threshold value is set so that if the threshold value is not exceeded, it can be determined that the medium has a short culture time and provides a good culture time, and on the other hand, if the threshold value is exceeded, it can be determined that the medium has a long culture time and provides a poor culture time.

[0155] The determination unit 12 determines the presence or absence of an abnormality in the target medium from the viewpoint of culture time based on the value of a state estimation vector obtained by projecting the mass spectrometry data of the target medium onto a state estimation space in which the principal component vectors are defined, and a threshold value. In addition, the degree of abnormality in the target medium may be determined by setting the threshold value to be divided into multiple stages.

[0156] Furthermore, when the state estimation unit 10a obtains an estimation result (time estimation result) of the culture time of the culture based on the feature extraction model from the mass spectrometry data of the target medium before the start of the culture process, the measurement object optimization estimation unit 14 receives the time estimation result from the state estimation unit 10a. The measurement object optimization estimation unit 14 analyzes the time estimation result obtained by the state estimation unit 10a and estimates an additive that can, for example, shorten the culture time of the culture by adding it to the target medium before use for culture. The measurement object optimization estimation unit 14 has an operator or the like add the estimated additive (also called an estimated additive) to the target medium to correct the components of the target medium used in the culture process.

[0157] In this case, the measurement object optimization estimation unit 14 calculates the relevance (similarity) between the additive measurement vector, which is mass spectrometry data on multiple types of additives, and the principal component vector defined in the state estimation space in the feature extraction model, according to the above-mentioned "(1-4) Measurement object optimization estimation unit". The measurement object optimization estimation unit 14 estimates the additive of the additive measurement vector that is most highly related to the basis of the feature extraction model from among the three types of additives as the optimal additive to be added to the target medium. In this way, the culture time can be improved by adding the additive of the additive measurement vector that is most highly related to the basis of the feature extraction model to the target medium before use for culture.

[0158] Here, an additive whose additive measurement vector is highly related to the basis of the feature extraction model is an additive whose additive measurement vector has a high similarity to the principal component vector defined in the state estimation space in the feature extraction model. In this embodiment, the similarity between the principal component vector of the feature extraction model and the additive measurement vector may be found by calculating the inner product between the principal component vector of the feature extraction model and the additive measurement vector, and when the feature extraction model has multiple principal component vectors, for example, the principal component vector with the largest eigenvalue may be selected and the inner product between the selected principal component vector and the additive measurement vector may be calculated.

[0159] In the above embodiment, the similarity between the principal component vector defined in the state estimation space in the feature extraction model and a plurality of types of additive measurement vectors is calculated, and additives of the additive measurement vector having a high similarity to the principal component vector are specified by the measurement target optimization estimation unit 14. However, the present invention is not limited to this. For example, information on one or more additives may be associated in advance with the estimated result of the culture time that will be obtained by the state estimation unit 10a, and when the estimated result of the culture time is obtained from the mass spectrometry data of the target culture medium based on the feature extraction model, one or more additives associated in advance with the estimated result may be estimated as the optimal additive to be added to the target culture medium.

[0160] (1-10-2) About the state estimation system shown in Figure 12 12 performs an optimization estimation process for the culture parameters, and the state estimation system 1c may also estimate optimal culture parameters based on the estimation result of the culture time according to the above-mentioned "(1-9) Optimization estimation process for the culture parameters". In this case, in the above-mentioned "(1-9-1) Generation of feature extraction model for estimating culture parameters", the mass spectrometry data of the culture medium, the reference culture parameter a, the correction candidate values ​​Δa0, Δa1, Δa2, Δa3 of the culture parameters, and the actual results indicating the culture time required for the culture to reach a desired culture state when the culture is actually cultured with the culture parameters (a+Δa0), (a+Δa1), (a+Δa2), (a+Δa3) are associated with each other, and these are stored in the database 4 as data sets.

[0161] The state estimation unit 10c reads out these data sets stored in the database 4, and generates a feature extraction model for estimating the culture parameters that characterizes the distribution of the input data, using the mass spectrometry data, the reference uncorrected culture parameter a, and the correction candidate values ​​Δa0, Δa1, Δa2, and Δa3 of the culture parameter a as input data. Specifically, the state estimation unit 10c generates a feature extraction model for estimating the culture parameters that estimates what kind of correction should be made to the reference culture parameters to achieve the desired culture time for the culture 3 hours after the start of culture.

[0162] In "(1-9-2) Application of feature extraction model for estimating culture parameters", in the state estimation unit 10c, the mass spectrometry data of the target culture medium measured via the measurement unit 2, the culture parameter a, and a plurality of correction candidate values ​​Δa of the culture parameter are input as input data to the feature extraction model for estimating the culture parameters described above and analyzed, thereby estimating the culture time required for the culture to reach the desired culture state (yield).

[0163] The culture parameter optimization estimation unit 15 selects, for example, an estimation result with the shortest culture time from among the multiple estimation results obtained by the state estimation unit 10c, and identifies a correction candidate value Δa of the culture parameter a set in the selected estimation result. In this way, by presenting the correction candidate value Δa obtained by the culture parameter optimization estimation unit 15 to an operator or the like via the notification unit 5, the operator or the like can correct the culture parameters based on the correction candidate value Δa, and the culture time can be shortened.

[0164] Also in this embodiment, for example, the time estimation result obtained by the state estimation unit 10c may be output to the judgment unit 12, and the judgment unit 12 may judge the culture parameter based on the time estimation result. In this case, the culture parameter is judged based on the time estimation result by the state judgment process or abnormality judgment process described above in "(1-3) Judgment unit". Note that since the state judgment process or abnormality judgment process has been described above in "(1-3) Judgment unit", a description thereof will be omitted to avoid duplication.

[0165] (1-11)Other In the above-described embodiment, a culture medium is used as the biological sample and the target biological sample. However, the present invention is not limited to this. For example, any of a culture medium, a culture cultivated in the culture medium, a culture medium extract extracted from the culture medium, and a culture extract extracted from the culture may be used as the biological sample and the target biological sample.

[0166] For example, when cultures of cells or the like are used as the biological sample and the target biological sample, the measurement unit 2 performs mass spectrometry on a plurality of cultures at a certain time and at least one target culture to measure mass spectrometry data. The state estimation unit 10a estimates the future culture state of the target culture from the characteristic values ​​of the target culture based on a feature extraction model that characterizes the distribution of mass spectrometry data, which are characteristic values ​​of the plurality of cultures. Similarly, the state estimation unit 10c may also use characteristic values ​​of the culture, medium extract, or culture extract instead of the characteristic values ​​of the medium.

[0167] The medium extract extracted from the medium includes, for example, substances contained in the original medium raw material, as well as metabolites of microorganisms or cells including target substances and by-products produced by the culture's metabolism of the substances (e.g., antibodies, proteins, carbohydrates, amino acids, lipids), and substances obtained by denaturing substances contained in the original medium raw material over time (e.g., Maillard reaction). When the medium extract is used as the biological sample and the target biological sample, the measurement unit 2 performs mass spectrometry on multiple medium extracts and at least one target medium extract at a certain time point, and measures the mass spectrometry data. The state estimation unit 10a estimates the future culture state of the culture from the characteristic values ​​of the target medium extract based on a feature extraction model that characterizes the distribution of mass spectrometry data, which is the characteristic values ​​of the multiple medium extracts.

[0168] The culture extract extracted from the culture is, for example, a substance contained in the culture or constituting the culture itself, and more specifically, includes metabolites (e.g., antibodies, proteins, carbohydrates, amino acids, lipids, etc.) that are metabolized in the culture by metabolizing the raw medium material, and substances originally contained in the raw medium that are taken in and accumulated in the body of the culture (e.g., heavy metals). When the culture extract is a biological sample and a target biological sample, the measurement unit 2 performs mass analysis on multiple culture extracts at a certain time and at least one target culture extract to measure mass analysis data. The state estimation unit 10a estimates the future culture state of the culture from the characteristic values ​​of the target culture extract based on a feature extraction model that characterizes the distribution of mass analysis data, which is the characteristic values ​​of the multiple culture extracts.

[0169] (2) Second embodiment In the above-described first embodiment, a case has been described in which, for example, a mass spectrometer that measures mass analysis data of a plurality of biological samples and a target biological sample is applied as a measurement unit that measures the characteristics of a plurality of biological samples and a target biological sample. In the second embodiment, a case in which a microscope device that measures the appearance of a plurality of biological samples and a target biological sample is applied will be described below.

[0170] (2-1) Overview of the state estimation unit and the determination unit according to the second embodiment In this case, in the state estimation system 1a shown in Fig. 1, a microscope device (e.g., a fluorescence microscope device) equipped with an imaging device such as a CCD (Charge Coupled Device) camera is applied as the measurement unit 2. The measurement unit 2, which is a microscope device, images the culture in order to measure the appearance state of the culture in the medium as a characteristic of the culture, and obtains image data of the culture. Note that in the second embodiment, the biological sample is a culture, and the culture will be described using, for example, cells used in a cell culture experiment as an example.

[0171] For example, in a cell culture experiment, undifferentiated cells are cultured in a medium through a specified culture process to obtain cultured cells. The cultured cells are then stained, for example, by immunostaining, and the expression state of light emitted from the cells is observed using a microscope device, thereby identifying the differentiation state of the cells. In this case, it is desirable to use a fluorescence microscope device capable of capturing an image of the fluorescence emitted from the culture to be measured as the microscope device. In addition, the differentiation state of the cultured cells can be identified, for example, by analyzing metabolites from the cells using a mass spectrometer, and the differentiation state of the cells can be estimated by measuring and analyzing the gene expression state in the cultured cells using RNA sequencing (RNA-Seq).

[0172] In the state estimation system 1a according to the second embodiment, image data is obtained by imaging undifferentiated target cells with a measurement unit 2, and this image data is used as an input for a feature extraction model, and the future culture state (differentiation state) of the undifferentiated target cells is estimated from the image data of the target cells based on the feature extraction model before the undifferentiated target cells become differentiated. The state estimation system 1a according to the second embodiment has a measurement unit 2 which is a microscope device, a calculation processing unit 3a, a database 4, and a notification unit 5, and is mainly different in that "image data" of the culture obtained by the measurement unit 2 is used instead of the "mass spectrometry data" of the culture medium described in the first embodiment described above.

[0173] In the second embodiment, for example, a plurality of undifferentiated cells are imaged by the measurement unit 2, and a feature extraction model is generated based on the plurality of image data obtained. The state estimation system 1a then images an undifferentiated target cell whose state is unknown by the measurement unit 2, and can estimate the differentiation state (culture state) of the target cell when it is cultured in the future based on the image data obtained as a result and the above-mentioned feature extraction model. In this way, a biological sample whose culture state is estimated using a previously generated feature extraction model is referred to as a "target biological sample". Note that in the second embodiment, the case where a cell (culture) is applied as an example of a biological sample will be described below, and a cell whose culture state is estimated using a previously generated feature extraction model is referred to as a "target culture" or a "target cell".

[0174] In this case, the measurement unit 2 captures images of each of the undifferentiated cells contained in each of the multiple culture media, and outputs the multiple image data obtained to the calculation processing unit 3a and the database 4. When the database 4 receives image data from the measurement unit 2, it stores each image data in association with the date and time when the image data was obtained. The database 4 also stores information on the identity of the captured cells (identity includes, for example, the type of cell (spec information), model number, lot number, date of manufacture, date of arrival, and past usage history), as well as information on parameters related to the culture conditions (i.e., culture parameters) such as the culture temperature, culture time, culture medium pH, and culture medium stirring speed in the culture process, in association with the image data.

[0175] The cells imaged by the measurement unit 2 are actually cultured using a medium through a predetermined culture process. Then, for example, the cultured cells are stained with a staining substance, and the differentiation state of the actually cultured cells is confirmed based on image data obtained by imaging the cells using a microscope device.

[0176] The state estimation unit 10a according to the second embodiment performs, for example, multivariate analysis using image data of a plurality of undifferentiated cells obtained by the measurement unit 2 at a certain point in time and the differentiation state of the cells when the cells are actually cultured using a culture medium, i.e., the result of the culture state (culture result), to generate a feature extraction model. Such a feature extraction model can analyze the external characteristics of the cells, which may affect the culture result when the cells are cultured using a culture medium.

[0177] In the case of supervised learning, it is desirable to apply a regression model or the like to the feature extraction model, which inputs image data of the target cells and outputs an estimate of the future culture state (state estimation result) that will ultimately be obtained when the target cells are assumed to be continued to be cultured in the medium. In addition, when generating a feature extraction model by supervised learning, it is desirable to learn the correspondence between image data of multiple cells and the culture state (differentiation state) of these cells, and characterize the distribution of the external features of the cells in the image data from the culture state.

[0178] In addition, as a supervised learning of other feature extraction models, for example, image data of multiple cells, the culture state of each cell, and a correct answer label indicating whether or not the culture state is the desired result (for example, a correct answer rate indicating whether or not the culture state is the desired result) may be used, and when learning the image data of the cells and their culture states, the correct answer label may be assigned and the learning may be performed to characterize the distribution of external features of the cells in the image data.

[0179] As in the first embodiment, the feature extraction model according to the second embodiment may be any of various feature extraction models, such as principal component analysis (PCA), partial least squares (PLS), polynomial regression, Gaussian process regression, random forest regression, etc. In addition, as the feature extraction model, for example, a machine learning model generated by machine learning characteristic values ​​(image data) of cells and the culture results of the cells may be applied.

[0180] As the machine learning model, for example, a neural network (for example, a CNN (Convolutional Neural Network) or a BNN (Bayesian Neural Network)) can be applied.

[0181] When generating a feature extraction model by supervised learning, it is desirable to apply data showing the differentiation state of each cell actually cultured as an index as a target variable. Also, the characteristic state of each cell in the image data obtained by staining each cell actually cultured with a staining substance and capturing an image of the cell by the measuring unit 2, which is a microscope device, may be applied as a target variable (i.e., the image data itself capturing an image of the cell, without showing the differentiation state of the cell actually cultured as an "index").

[0182] In addition to the above-mentioned supervised learning, the feature extraction model may be unsupervised learning. When generating a feature extraction model for unsupervised learning, it is preferable to use image data of a plurality of cells and perform learning to characterize the distribution of the appearance of each cell in the plurality of image data.

[0183] In this second embodiment, by treating the "mass spectrometry data" (or measurement vector) acquired via the measurement unit 2 as "image data" in each of the items "(1-1) Configuration of the state estimation system", "(1-2) State estimation unit", "(1-3) Determination unit", "(1-5) Action and effect", "(1-6) Generation of feature extraction model without a teacher", "(1-7) Other embodiments relating to the timing of measurement of characteristics of a biological sample", "(1-8) Other embodiments for estimating the culture state of a culture during the culture process before the end of the culture process", "(1-9) Optimization estimation process of culture parameters", "(1-10) State estimation system for estimating culture time", and "(1-11) Others" described in the above "(1) First embodiment", the processing of each of these items can be applied.

[0184] As described above, in the second embodiment, each process can be explained by replacing “mass spectrometry data” (measurement vector) with “image data” in “(1) first embodiment”; therefore, detailed explanations of each item will be omitted to avoid duplication, and only an outline will be explained below.

[0185] For example, in the "(1-3) Determination unit" described above in the second embodiment, the determination unit 12 determines the state of the target cells based on the state estimation result obtained by the state estimation unit 10a. The determination process in the determination unit 12 includes a state determination process that simply determines in a stepwise manner whether the cultured target cells are in an appropriate culture state, and an abnormality determination process that determines whether the cultured target cells are in an abnormal culture state, such as not in a desired differentiation state.

[0186] When performing state determination processing by applying a regression model as a feature extraction model, the determination unit 12 sets, prior to the state determination processing, for example, a discrimination value capable of discriminating between good cells and bad cells in a stage-by-stage manner in a state estimation space in which a principal component vector (supervised basis) is defined as a feature in the feature extraction model. In addition, this discrimination value may be associated with a discrimination display that indicates the degree of goodness of the differentiation state of the cells in a stage-by-stage manner, for example, based on the culture result when the cells are cultured using a culture medium.

[0187] As in the first embodiment described above, the discrimination value for the state estimation space may be arbitrarily set by an operator or the like, or the discrimination value may be automatically set by automatically dividing the principal component vectors defined in the state estimation space at predetermined intervals. The determination unit 12 determines the state of the target cell from the value of the state estimation vector obtained by projecting the image data of the target cell onto the state estimation space and the discrimination value.

[0188] The determination unit 12 may set a threshold value for abnormality determination in the state estimation space of the feature extraction model based on a rule for the abnormality determination. In this case, the threshold value may be set in the state estimation space by an operator or the like, taking into consideration a desired culture result. For example, the threshold value is set so that if the threshold value is not exceeded, a good culture result equal to or greater than the reference value is obtained, and if the threshold value is exceeded, a poor culture result below the reference value is obtained.

[0189] The determination unit 12 inputs the image data of the target cell into a feature extraction model, and determines the presence or absence of abnormality in the target cell from the value of a state estimation vector obtained by projecting the feature extracted from the image data onto a state estimation space in which the principal component vectors are defined, and a threshold value. The above threshold value may be set to be divided into multiple stages to determine the degree of abnormality in the target cell.

[0190] As in the first embodiment described above, the notification unit 5 receives the determination result of the target cell by the state determination process, and notifies the operator, etc. of the determination result of the state of the target cell before use for culture based on the identification value. In addition, the notification unit 5 receives the determination result of the presence or absence of an abnormality or the degree of the abnormality by the abnormality determination process, and notifies the operator, etc. of the presence or absence of an abnormality or the degree of the abnormality.

[0191] In the above configuration, the state estimation system 1a according to the second embodiment includes a measurement unit 2 that measures image data of a plurality of cultures and at least one target culture as characteristic values ​​(input data), and a state estimation unit 10a that estimates the culture state of a target culture from the characteristic values ​​of the target culture based on a feature extraction model that characterizes the distribution of the characteristic values ​​(input data) of the plurality of cultures. In this way, the state estimation system 1a can estimate in advance whether or not a desired culture state can be obtained in biological production and research and development using the target culture medium, thereby shortening the time required to properly determine the target culture medium. This can improve the efficiency of biological production and research and development using the culture medium.

[0192] (2-2) Generation of unsupervised feature extraction model according to the second embodiment For example, we will describe a state estimation system 1a that performs PCA (principal component analysis) processing using image data of multiple cells, extracts multidimensional principal component vectors (unsupervised basis) that characterize the distribution of the apparent states of the cells in the image data, and obtains a feature extraction model.

[0193] In this case, image data of a plurality of cells is obtained by the measurement unit 2, and the obtained plurality of image data is used to construct a feature extraction model that can clearly detect differences between cells, for example, by performing multivariate analysis. Specifically, when the state estimation unit 10a receives image data from the measurement unit 2 or the database 4, it performs PCA processing on the image data to extract principal component vectors (unsupervised basis) that characterize the distribution of the image data as features, and generates a feature extraction model.

[0194] In the state estimation system 1a, when image data of an unknown target cell is subsequently obtained, it is possible to estimate which cell the unknown target cell is most similar to in comparison with past cells based on the image data of the target cell and the feature extraction model.

[0195] In this way, even in the case of a feature extraction model generated by PCA processing, a threshold for abnormality judgment can be set in the state estimation space of the feature extraction model based on a certain rule for abnormality judgment in the judgment unit 12. In this case, the threshold is simply defined as a quantity related to the distribution when image data of multiple cells are projected onto the state estimation space in which the principal component vectors are defined.

[0196] For example, the judgment unit 12 can determine the presence or absence of an abnormality using a threshold value set in the state estimation space and the coordinate x when the image data of the target cell is projected into the state estimation space. The above threshold value may be set to be divided into multiple stages to judge the degree of abnormality of the target cell. The judgment unit 12 outputs the judgment result of the presence or absence of an abnormality in the analysis or the degree of the abnormality to the notification unit 5. In this way, the notification unit 5 can present the judgment result of the judgment unit 12 to an operator or the like.

[0197] (2-3) Other embodiments regarding the timing of measuring the characteristics of the biological sample (2-3-1) Measuring cell characteristics during the culture process after the culture process has started

[0198] In the above-mentioned second embodiment, the case where image data of the cells is obtained by the measurement unit 2 before the start of the culture process has been described, but the present invention is not limited to this. Even in the second embodiment, for example, as shown in 10A of Fig. 10, after the start of the culture process, the measurement unit 2 may capture an image of the cells in the middle of being cultured in the culture process to obtain image data.

[0199] In this case, in the state estimation system 1a, as shown in 10A of Fig. 10, at a certain time point Sa2 during the culture process, the measurement unit 2 captures an image of the cells and acquires image data of the cells. In this way, the state estimation system 1a accumulates image data at the certain time point Sa2, a predetermined period of time after the start of culture, for the plurality of cells currently being cultured in the culture process.

[0200] Then, in a specified culture process, the cells of the culture are cultured until the originally planned end of culture, and when the cell culture is completed, the cells after culture contained in the medium are stained, the light emitted from the cells is measured, and the differentiation state of the cells is confirmed from the expression state of the light emitted from the cells (Fa1). In this way, the state of each cell after culture is measured, and the actual culture result is obtained for each cell.

[0201] In the state estimation system 1a, when an index (actual culture result) indicating the differentiation state of a cell obtained for each cell is input by an operator or the like, the result of the culture state is stored in the database 4. At this time, the database 4 stores image data of each cell measured at a certain time point Sa2 during the culture process in association with the final culture result obtained when the cell was cultured.

[0202] In the state estimation unit 10a according to the second embodiment, image data of each cell measured at a certain time point Sa2 during the culture process and the final culture state of the cell are read as sample data from the database 4 to generate a feature extraction model.

[0203] This feature extraction model takes as input image data of the target cells at a certain point Sa2 (e.g., the 10th day after the start of culture) after the start of culture of the target cells through a specified culture process, and outputs an estimate of the future culture state of the target cells that will ultimately be obtained if the culture of the target cells is continued as is.

[0204] As the feature extraction model here, as in the above-mentioned embodiment, for example, various feature extraction models such as PCA, PLS, polynomial regression, Gaussian process regression, Random Forest Regression, etc., feature extraction models generated by machine learning of characteristic values ​​(image data) of cells during culture and the culture results, etc. can be applied. When generating a feature extraction model for supervised learning, it is desirable to learn the correspondence between image data of multiple cells and the culture results of these cells, and to characterize the distribution of the image data from the culture results. Also, in this case, as described above, in addition to supervised learning, unsupervised learning may be applied as the feature extraction model. When generating a feature extraction model for unsupervised learning, it is desirable to use only image data of multiple cells and to learn to characterize the distribution of the cells of these multiple image data.

[0205] Here, in the case of supervised learning, for example, the state estimation unit 10a uses image data of cells during the culture process as explanatory variables and the actual final culture state obtained in this culture process (the differentiated state of the cells at the end of the culture) as the objective variable, and generates in advance a feature extraction model that shows a characteristic relationship, such as correlation or inverse correlation, between the image data and the culture state.

[0206] As a result, in the state estimation system 1a, by using the feature extraction model while the undifferentiated target cells are being cultured in a culture process, it becomes possible to estimate the future culture results of the target cells that will ultimately be obtained by culturing, during the culture process, before the culture process is completed.

[0207] For example, in the series of culture processes in which cells are induced to differentiate through culture to obtain the desired cells, a long time is required to obtain the desired cells. In particular, the series of culture processes in which undifferentiated human stem cells are induced to differentiate to obtain the desired cells usually takes more than about three months. In such a culture process, for example, when trying to obtain optimal culture results (i.e., the desired cells) by sequentially optimizing various parameters such as the components of the culture medium and the culture temperature, a conceivable method is to wait for the end of the culture process, analyze the actual culture results obtained, and determine the next parameters (the components of the culture medium used, the culture temperature, etc.) each time.

[0208] However, with this method, for example, if one cycle of the culture process from the start to the end of the culture takes about three months, only about four cycles can be performed per year, and trial and error of each parameter of the culture process is required in four cycles, which is inefficient in terms of time.

[0209] In contrast, the state estimation system 1a according to the second embodiment also obtains an estimation result of the culture state that will ultimately be obtained during the culture process based on the feature extraction model, and by presenting the estimation result to an operator or the like during the culture, it is possible to change various parameters of the culture process (such as the components of the culture medium used and the culture temperature) as necessary, change the cultured matter (target cells), or perform a new culture process based on the obtained estimation result, without waiting for the end of the culture process. In this way, a new culture process can be started in the middle of the currently ongoing culture process, where the parameters of the culture process and the cultured matter are trial and error-tested so as to obtain the desired culture result, and parameter optimization can be carried out in a time-efficient manner, without waiting for the end of the culture process.

[0210] In this case, the determination unit 12 determines the state of the target cell based on the state estimation result obtained in the state estimation unit 10a from the image data of the cell during the culture process.

[0211] (2-3-2) When measuring cell characteristics multiple times during the culture process In the second embodiment, as in the first embodiment described above, for example, as shown in 10B of FIG. 10, the cells may be imaged by the measurement unit 2 at a certain time point Sa1 before the start of the culture process and at certain times Sa2 and Sa3 during the culture process, and image data may be obtained multiple times. Alternatively, the cells may not be imaged by the measurement unit 2 before the start of the culture process, but may be imaged by the measurement unit 2 only at certain times Sa2 and Sa3 during the culture process, and image data of the cells may be obtained multiple times. That is, image data of the cells may be obtained multiple times at intervals in time series from the start of the culture, such as image data of the cells on the 10th day after the start of the culture process and image data of the cells on the 20th day after the start of the culture process.

[0212] In this case, as shown in 10B of Fig. 10, the state estimation system 1a captures images of a plurality of cells by the measurement unit 2 at a certain time point Sa1 before the start of the culture process and at certain times points Sa2 and Sa3 during the culture process, and acquires image data. In addition, the state estimation system 1a measures the culture state (differentiation state of the culture) of the target cells after they have been actually cultured in the medium after the culture process is completed (Fa1).

[0213] The state estimation unit 10a generates a feature extraction model based on the image data of the cell at a certain time point Sa1 before the start of the culture process and the image data of the cell at certain times Sa2 and Sa3 after the start of the culture process, which are obtained in this manner. That is, in the case of supervised learning, for example, the state estimation unit 10a generates a feature extraction model showing the relationship between the image data and the culture state, using the image data of the cell before the start of the culture process as an explanatory variable and the differentiation state of the cell after the end of the culture process as a target variable. When the state estimation unit 10a subsequently obtains image data of the target cell in an undifferentiated state before the start of the culture process, the state estimation unit 10a estimates from this image, based on the feature extraction model, the culture state of the target cell (the differentiation state of the target cell) that will be obtained in the future when the target cell is cultured by the culture process.

[0214] Similarly, the state estimation unit 10a generates a feature extraction model showing the relationship between the image data of the cell at a certain time point Sa2 after the start of the culture process and the culture state of the cell after the end of the culture process, and similarly generates a feature extraction model showing the relationship between the image data of the cell at a certain time point Sa3 after the start of the culture process and the culture state of the cell after the end of the culture process. The state estimation unit 10a estimates the culture state (differentiation state of the target cell) of the target cell that will be obtained in the future when the target cell is cultured by the culture process from the image data acquired at each of the certain times (Sa2, Sa3) for the target cell in the undifferentiated state and in use for culture, based on the feature extraction model for the corresponding certain times Sa2, Sa3.

[0215] In this way, the target cells are imaged by the measurement unit 2 at a certain point Sa1 before the culture process is started and at certain points Sa2 and Sa3 during the culture process, and the culture process can be performed while sequentially estimating the culture state of the target cells after the culture process is completed based on the feature extraction model corresponding to the certain points Sa1, Sa2, and Sa3 from the image data of the target cells that changes over time.

[0216] Note that, although the case where each feature extraction model is generated by supervised learning has been described here, the present invention is not limited to this, and each feature extraction model may be generated by unsupervised learning in accordance with the above-mentioned “(2-2) Generation of feature extraction models by unsupervised learning”.

[0217] (2-3-3) Another embodiment for estimating the culture state of cells during the culture process before the end of the culture process In the above-mentioned second embodiment, a case has been described in which the culture state (e.g., the differentiation state of the target cell) contained in the culture medium is estimated after the culture process is completed (after the completion time F), but the present invention is not limited thereto. In the above-mentioned second embodiment, for example, as shown in 11A of FIG. 11, the culture state of the target cell at a certain time point Fa2 during the culture before the culture process is completed may be estimated.

[0218] In this case, in the state estimation system 1a, as shown in 11A of Fig. 11, at a certain time point Sa1 before the start of the culture process, the measurement unit 2 captures an image of a plurality of cells and acquires image data. Also, in the state estimation system 1a, at a certain time point Fa2 before the end of the culture process, the culture state of the cells actually being cultured (the differentiation state of the cells at that time point) is measured.

[0219] The state estimation unit 10a generates a feature extraction model by supervised learning based on the image data of the undifferentiated cells before the start of the culture process and the culture state of the cells during the culture, which are obtained in this manner. That is, the state estimation unit 10a generates a feature extraction model that characterizes the relationship between the image data and the culture state, using the image data of the cells before the start of the culture process as an explanatory variable and the culture state of the cells at a certain time point Fa2 before the end of the culture process as a target variable. When the state estimation unit 10a subsequently obtains image data of the target cells in an undifferentiated state before the start of the culture process, it can estimate the culture state of the target cells during the culture (the differentiated state of the target cells) that will be obtained in the future by culturing them through the culture process from the image data based on the feature extraction model.

[0220] When generating a feature extraction model by supervised learning, it is desirable to apply data indicating, as an objective variable, the differentiation state of each cell actually cultured up to a certain time point Fa2 before the end of the culture process, as an indicator. Also, the characteristic state of each cell itself in the image data obtained by staining each cell actually cultured up to the time point Fa2 with a staining substance and capturing an image with the measuring unit 2, which is a microscope device, may be applied as the objective variable (i.e., the differentiation state of the actually cultured cell is not indicated by an "indicator", but the image data itself capturing the image of the cell).

[0221] In addition, here, as shown in 11A of Fig. 11, a case has been described in which a plurality of undifferentiated cells are imaged by the measurement unit 2 at a certain time Sa1 before the start of the culture process and image data is acquired, but the present invention is not limited to this. Even in this second embodiment, for example, as shown in 11B of Fig. 11, a plurality of cells may be imaged by the measurement unit 2 at a certain time Sa2 during the culture process and image data may be acquired.

[0222] In this case, the state estimation unit 10a uses image data of the cells at a certain time point Sa2 during the culture process as an explanatory variable and the culture state of the cells at a certain time point Fa2 before the end of the culture process as a target variable to generate a feature extraction model that characterizes the relationship between the image data and the culture state. When the state estimation unit 10a subsequently obtains image data of the target cells at a certain time point Sa2 during the culture process, it can estimate the culture state (differentiation state of the target cells) of the target cells at a certain time point Fa2 during the culture, which will be obtained in the future, from the image data based on the feature extraction model.

[0223] 11C, the cells may be imaged by the measurement unit 2 at a certain time point Sa1 before the start of the culture process and at certain times Sa2 and Sa3 during the culture process, and image data may be obtained multiple times. Also, the cells may not be imaged by the measurement unit 2 before the start of the culture process, but may be imaged by the measurement unit 2 only at certain times Sa2 and Sa3 during the culture process, and image data may be obtained multiple times.

[0224] The state estimation unit 10a generates a feature extraction model that characterizes the relationship between the image data and the culture state, for example, by using image data of the cells before the start of the culture process as an explanatory variable and the culture state of the cells contained in the medium at a certain time point Fa2 before the end of the culture process as a target variable. When the state estimation unit 10a subsequently obtains image data of the undifferentiated target cells before the start of the culture process, it can estimate from this image data, based on the feature extraction model, the culture state (differentiation state of the target cells) of the target cells at a certain time point Fa2 during the culture, which will be obtained in the future.

[0225] The state estimation unit 10a also generates a feature extraction model that characterizes the relationship between the image data of the cell at a certain time point Sa2 after the start of the culture process and the culture state of the cell contained in the medium at a certain time point Fa2 before the end of the culture process, and similarly generates a feature extraction model that characterizes the relationship between the image data of the cell at a certain time point Sa3 after the start of the culture process and the culture state of the cell contained in the medium at a certain time point Fa2 before the end of the culture process. When the state estimation unit 10a subsequently obtains image data of the target cell being cultured at certain times (Sa2, Sa3), it can estimate the culture state (differentiation state of the target cell) of the target cell at a certain time point Fa2 during the culture, which will be obtained in the future, from the image data based on the feature extraction models of the corresponding times Sa2, Sa3.

[0226] In addition, as described above, within each of the items "(1-9) Optimization estimation process for culture parameters" and "(1-10) State estimation system for estimating culture time", the processing of each of these items can be applied by treating the "mass spectrometry data" (or measurement vector) acquired via the measurement unit 2 as "image data".

[0227] For example, in "(1-9) Optimization and estimation process of culture parameters", the state estimation unit 10c uses image data of multiple cultures and culture parameters related to the culture conditions of a culture process using the cultures as input data to generate a feature extraction model for estimating the culture parameters that characterizes the distribution of the image data and the culture parameters.

[0228] More specifically, the state estimation unit 10c uses the cell image data obtained by the measurement unit 2, the reference uncorrected culture parameter a, and the correction candidate values ​​△a0, △a1, △a1, △a1 as inputs (explanatory variables), and generates a feature extraction model for estimating culture parameters by using a neural network or the like, with the culture state of the cells after culture (here, the differentiation state) as output (objective variable).

[0229] The state estimation unit 10c estimates the culture state (differentiation state) of the target cell by inputting the image data of the target cell obtained via the measurement unit 2, the culture parameter a, and multiple correction candidate values ​​Δa of the culture parameter as input data into the feature extraction model for estimating the culture parameters described above and analyzing them.

[0230] The state estimation unit 10c outputs a plurality of estimation results obtained by changing the correction candidate value Δa to the culture parameter optimization estimation unit 15. The culture parameter optimization estimation unit 15 selects an estimation result that is closest to a preset optimum culture state result from the plurality of estimation results obtained by the state estimation unit 10c, and specifies the correction candidate value Δa of the culture parameter a that is set in the selected estimation result.

[0231] Furthermore, in "(1-10) State Estimation System for Estimating Culture Time", for example, in the case of supervised learning, the state estimation unit 10a generates a feature extraction model using image data of a plurality of cells obtained by the measurement unit 2 at a certain time as an explanatory variable and the culture time at which the culture reaches a desired culture state when the cells are actually cultured by the culture process as a target variable. Such a feature extraction model can analyze the external characteristics of cells that affect the culture time at which the culture reaches a desired culture state when the culture is cultured using a medium (hereinafter also simply referred to as culture time).

[0232] In the second embodiment, the supervised learning of other feature extraction models may also be performed by using, for example, image data of a plurality of cells, the culture times of these cells, and correct answer labels indicating whether or not the culture times result in the desired results (for example, a correct answer rate indicating whether or not the culture times are the desired results), and when learning the image data of the cells and their culture times, the correct answer labels may be assigned to the cells and the cells may be learnt, thereby characterizing the distribution of the mass spectrometry data.

[0233] As the feature extraction model, various feature extraction models can be applied, such as PCA, PLS, polynomial regression, Gaussian process regression, Random Forest Regression, etc. In addition, as the feature extraction model, for example, a machine learning model generated by machine learning the characteristic values ​​(mass spectrometry data) of the medium and the results of the culture time may be applied.

[0234] As the machine learning model, for example, a neural network (for example, a CNN (convolutional neural network) or a BNN (Bayesian neural network)) can be applied. In addition to the above-mentioned supervised learning, the feature extraction model may be unsupervised learning. Details of generating a feature extraction model by unsupervised learning will be described later, but when generating a feature extraction model by unsupervised learning, it is desirable to use image data of a plurality of cells and perform learning to characterize the distribution of these multiple image data.

[0235] In addition, in the above-mentioned "(1-10-2) Regarding the state estimation system shown in FIG. 12", the optimal culture parameters are estimated based on the estimated culture time using mass spectrometry data, but it goes without saying that the "mass spectrometry data" can also be applied as the "image data" in this case.

[0236] That is, in this case, the state estimation unit 10c inputs the image data of the target cells measured via the measurement unit 2, the culture parameter a, and multiple correction candidate values ​​Δa of the culture parameter as input data into a feature extraction model for estimating the culture parameters, and analyzes the input data, thereby making it possible to estimate the culture time required for the culture to reach the desired culture state (differentiation state).

[0237] (3) Third embodiment In the above-described first embodiment, a case has been described in which, for example, a mass spectrometer that measures mass analysis data of a plurality of biological samples and a target biological sample is used as a measurement unit that measures the characteristics of a plurality of biological samples and a target biological sample. In the third embodiment, however, an outline of a case in which a measurement device that measures the impedance of a plurality of biological samples and a target biological sample is used will be described below.

[0238] In this case, for example, in the state estimation system 1a, any one of the culture medium, the culture target cultured in the culture medium, the culture medium extract extracted from the culture medium, and the culture target extract extracted from the culture target is taken as the biological sample and the target biological sample, and the impedances of multiple biological samples at a certain point in time and at least one target biological sample of the same type as the biological samples at that point in time are measured as characteristic values ​​by the measurement unit 2.

[0239] In the state estimation system 1a, the impedances of multiple biological samples acquired via the measurement unit 2 are used as input data, a feature extraction regression model that characterizes the distribution of this input data is generated, and this is stored in the database 4. Then, in the state estimation system 1a, the state estimation unit 10a estimates the culture state of the culture from the impedances of the target biological samples based on the feature extraction model.

[0240] In such a third embodiment, by setting the "mass spectrometry data" (or the measurement vector) acquired via the measurement unit 2 as "impedance" in each of the items described in the above-mentioned "(1) First embodiment", namely, "(1-1) Configuration of the state estimation system", "(1-2) State estimation unit", "(1-3) Determination unit", "(1-4) Measurement object optimization estimation unit", "(1-5) Action and effect", "(1-6) Generation of feature extraction model without a teacher", "(1-7) Other embodiments related to the measurement timing of the characteristics of a biological sample", "(1-8) Other embodiments for estimating the culture state of a culture during the culture before the end of the culture process", "(1-9) Optimization estimation process of culture parameters", and "(1-10) State estimation system for estimating the culture time", the processing of each of these items can be applied.

[0241] For example, in the above-mentioned "(1-4) measurement object optimization estimation unit", when the state estimation unit 10a obtains an estimation result of the culture state of the culture based on the feature extraction model from the impedance of the target medium before the start of the culture process, the state estimation unit 10a outputs the state estimation result to the measurement object optimization estimation unit 14.

[0242] The measurement target optimization estimation unit 14 analyzes the state estimation result obtained by the state estimation unit 10a, and estimates an additive that can be added to the target medium before use for culture to maximize the yield of the culture per unit volume of the medium contained in the target medium after use for culture. The measurement target optimization estimation unit 14 has an operator or the like add the estimated additive (also called an estimated additive) to the target medium, thereby correcting the components of the target medium used in the culture process.

[0243] In addition, in "(1-9) Optimization and estimation process of culture parameters", for example, impedances of a plurality of biological samples and culture parameters related to culture conditions of a culture process using the biological samples are used as input data, and a culture state is estimated from the input data of a target biological sample based on a feature extraction model that characterizes the distribution of the impedance and culture parameters as characteristic values.

[0244] Furthermore, in "(1-10) State estimation system for estimating culture time", in state estimation system 1a, the impedances of multiple biological samples acquired via measurement unit 2 are used as input data, a feature extraction regression model that characterizes the distribution of this input data is generated, and this is stored in database 4. Then, in state estimation system 1a, state estimation unit 10a estimates the culture time for the culture to reach a predetermined culture state from the impedance of the target biological sample based on the feature extraction model.

[0245] In addition, the measurement unit may be a Raman spectrometer, a chromatograph, a digital PCR measurement device, a nuclear magnetic resonance device, an antibody quantification kit using an antibody quantification method, and a nucleic acid sequencer. Examples of the nucleic acid sequencer include, but are not limited to, a nucleic acid sequencer using a known DNA sequence determination method such as Sanger sequencing or Next Generation Sequencing (NGS). Moreover, the flow cytometer described in the fourth embodiment may be applied as the measurement unit 2.

[0246] In this case, any of the Raman spectrum obtained from the Raman spectroscopy device, the chromatogram obtained from the chromatography device, the amount of a specific substance obtained from the digital PCR measurement device, the analysis data obtained from the nuclear magnetic resonance device, the amount of a specific antigen obtained from the antibody quantification kit, and the gene sequence analysis data obtained from the nucleic acid sequence determination device may be used in the same manner as described in "(1) First embodiment" above, in the "(1-1) Configuration of the state estimation system", "(1-2) State estimation unit", "(1-3) Determination unit", "(1-4) Measurement object optimization estimation unit", "(1-5) Actions and effects", and "(1-6) In each of the items "(1-7) Generation of a feature extraction model without a teacher," "(1-8) Other embodiments relating to the timing of measuring the characteristics of a biological sample," "(1-9) Optimization estimation process for culture parameters," "(1-10) State estimation system for estimating culture time," and "(1-11) Other," the processing of each of these items can be applied by changing the "mass spectrometry data" (or measurement vector) acquired via measurement unit 2 to, for example, "Raman spectrum," "gene sequence analysis data," etc.

[0247] In the above embodiments, a state estimation unit has been described that, for example, estimates the yield of a culture based on a feature extraction model from input data of the culture medium at a certain point before the end of the culture, or estimates the culture time at which the culture will have a predetermined yield. However, as described above, the present invention is not limited to the "yield of a culture", and the "quality of a culture", "yield (quality) of a culture extract extracted from a culture", and "yield (quality) of a component in a culture medium secreted by a cultured culture" can be applied as the "culture state of a culture".

[0248] For example, the state estimation unit may estimate the quality of the culture based on a feature extraction model from input data obtained from the medium or culture at a certain point before the end of the culture, or estimate the culture time until the culture reaches a predetermined quality.

[0249] Alternatively, the state estimation unit may be configured to estimate the yield (quality) of a culture extract extracted from a culture based on a feature extraction model from input data obtained from the medium or culture at a certain time before the end of the culture, or to estimate the culture time at which the culture extract will have a predetermined yield (quality). Furthermore, the state estimation unit may be configured to estimate the yield (quality) of a component in the medium secreted by the cultured product based on a feature extraction model from input data obtained from the medium or culture at a certain time before the end of the culture, or to estimate the culture time at which the component in the medium secreted by the culture will have a predetermined yield (quality).

[0250] (4) Fourth embodiment (4-1) State Estimation System According to the Fourth Embodiment Next, a fourth embodiment will be described. The state estimation system according to the fourth embodiment uses as input data the measurement results of the characteristic values ​​of a culture at a certain time point measured by a measurement unit, and estimates which of a plurality of culture states obtained when the culture is cultured according to a culture process will be in the future from the input data based on a feature extraction model. Then, the state estimation system determines the next processing step to be performed on the culture at the certain time point according to the type of the estimated culture state of the culture.

[0251] 13 is a schematic diagram for explaining the types of culture states of culture A estimated by the state estimation system according to the fourth embodiment, showing an example in which culture A is induced to differentiate into culture B, culture C, or non-target cells by culturing. Here, a case will be described as an example in which, when culture A (e.g., human iPS cells (hereinafter simply referred to as iPS cells)) shown in step S10 is cultured in a medium containing a predetermined cytokine, depending on the state and quality of culture A, the quality of the medium, and the like, culture A is induced to differentiate into culture B of paraxial mesoderm cells (step S11), culture A is induced to differentiate into culture C of lateral plate mesoderm cells (step S12), or culture A becomes non-target cells (step S13), such as other cells, dead cells, or cells in poor condition.

[0252] In the fourth embodiment, the optimal culture conditions for inducing differentiation of culture A into culture B and the optimal culture conditions for inducing differentiation of culture A into culture C are different culture conditions. In the state estimation system according to the fourth embodiment, based on the estimation result of culture A, one of a processing step of culturing culture A under optimal culture conditions for inducing differentiation into culture B, a processing step of culturing culture A under optimal culture conditions for inducing differentiation into culture C, and a processing step of discarding culture A is appropriately selected, and the optimal processing step according to the future culture state of culture A is executed. Here, the culture conditions include, for example, the type and quality of the culture medium, as well as various parameters (culture parameters) related to the culture conditions, such as the culture temperature, culture time, culture medium pH, and culture medium stirring speed of the culture medium in the culture process.

[0253] 14, in which the same components as those in FIG. 1 are denoted by the same reference numerals, is a block diagram showing the configuration of a state estimation system 1d according to a fourth embodiment. In the fourth embodiment, a culture is shown as an example of a biological sample. The fourth embodiment also shows an example in which a flow cytometer using a flow cytometry method is applied as the measurement unit 2.

[0254] The measurement unit 2 suspends culture A in a liquid at a certain point in time before the culture is completed, such as before the start of culture or immediately after the start of culture, and while flowing the liquid in a flow path so that the cultures A are aligned in a line, it irradiates laser light on each culture A and measures the reflected light such as scattered light and fluorescence. The measurement unit 2 obtains characteristic values ​​for each culture A by analyzing the measurement results of the reflected light, and generates analysis data indicating the characteristic values ​​measured for each culture A as input data for each culture A.

[0255] The state estimation system 1d according to the fourth embodiment, for example, measures a plurality of cultures A at a certain time point by the measurement unit 2, and generates a feature extraction model based on the obtained analysis data. The analysis data here refers to the reflected light intensity, scattered light intensity, fluorescence intensity, etc., measured when the culture A is analyzed by the flow cytometer, which is the measurement unit 2, and can be a characteristic value of the culture A. When the state estimation system 1d obtains the feature extraction model, it measures the culture A, which is not yet determined to be differentiated into a certain culture state, by the measurement unit 2, and can estimate, based on the obtained analysis data and the above-mentioned feature extraction model, what kind of culture state the culture A, which is not yet determined at present, will have if it is continued to be cultured as it is. The culture A, whose culture state is estimated using the generated feature extraction model, is referred to as a "target culture".

[0256] The measurement unit 2, which is a flow cytometer, measures each characteristic value of the plurality of cultures A, and outputs the obtained measurement results to the calculation processing unit 3d and the database 4. The calculation processing unit 3d has a state estimation unit 10a, a measurement object optimization estimation unit 14, and a next step determination unit 22. Note that the measurement object optimization estimation unit 14 is similar to that in the above-mentioned embodiment, so its description is omitted here, and the following description focuses on the state estimation unit 10a and the next step determination unit 22. In this case, when generating a feature quantity extraction model, the state estimation unit 10a receives, as analysis data, measurement results of the plurality of cultures A by the flow cytometer (measurement unit 2) from the measurement unit 2 or the database 4.

[0257] In the state estimation unit 10a according to the present embodiment, a feature extraction model is generated by performing, for example, multivariate analysis using the analysis data of a plurality of cultures A measured by the measurement unit 2 at a certain time point and the result of the culture state when each culture A is actually cultured using a predetermined medium (hereinafter, also referred to as the culture result). In the present embodiment, there are a plurality of culture states when culture A is cultured, including a case where the culture becomes culture B, a case where the culture becomes culture C, and a case where the culture becomes a non-target cell. Therefore, as an example, an example of generating a feature extraction model for estimating the culture state for each culture state will be described here. An example of estimating a plurality of types of culture states based on one feature extraction model will be described later.

[0258] That is, the state estimation unit 10a generates a feature extraction model for culture B that estimates whether culture A will be cultured to become culture B, a feature extraction model for culture C that estimates whether culture A will be cultured to become culture C, and a feature extraction model for non-target cells that estimates whether culture A will be cultured to become non-target cells. Such feature extraction models can analyze the state, quality, etc. of culture A that affect the culture result when culture A is cultured using a medium.

[0259] The feature extraction model generated for each culture result of culturing culture A may be generated by supervised learning or unsupervised learning. For example, as a feature extraction model generated by supervised learning, it is desirable to apply a regression model or the like, which inputs analysis data obtained as a result of measuring target culture A at a certain point in time by measurement unit 2, and outputs a state estimation result that estimates the future culture state that will ultimately be obtained if the culture of target culture A is continued as is.

[0260] In this case, for example, when the state estimation unit 10a generates a feature extraction model for culture B by supervised learning, the state estimation unit 10a learns the correspondence between multiple analysis data obtained by measuring multiple cultures A at a certain point in time using the measurement unit 2 and the culture results when each of these cultures A is actually cultured to become culture B, and characterizes the distribution of the analysis data from the culture results when culture B is produced.

[0261] Similarly, when the state estimation unit 10a generates a feature extraction model for culture C by supervised learning, it learns the correspondence between multiple analysis data obtained by measuring multiple cultures A at a certain point in time using the measurement unit 2 and the culture results obtained by actually culturing each of these cultures A to become culture C, and characterizes the distribution of the analysis data from the culture results when culture C is produced.

[0262] Similarly, when the state estimation unit 10a generates a feature extraction model for non-target cells by supervised learning, it learns the correspondence between multiple analysis data obtained by measuring multiple cultures A at a certain point in time using the measurement unit 2 and the culture results when each of these cultures A is actually cultured and results in non-target cells, and characterizes the distribution of the analysis data from the culture results when non-target cells result.

[0263] In addition, as another example of supervised learning of a feature extraction model for culture B, analytical data of multiple cultures A, each culture result obtained by actually culturing each of these cultures A, and a correct answer label indicating whether or not the culture result has become the desired culture B (e.g., the accuracy rate indicating whether or not the culture result has become the desired culture B) may be used, and when learning the analytical data of culture A and its culture results, the correct answer label indicating that culture A has become culture B may be added and the model may be trained to characterize the distribution of the analytical data.

[0264] In this case, supervised learning of a feature extraction model for culture C (or for non-target cells) is also performed in a similar manner, using analysis data of multiple cultures A, each culture result obtained by actually culturing each of these cultures A, and a correct answer label indicating whether the culture result has become the desired culture C (or non-target cells) (e.g., the accuracy rate indicating whether the culture result has become the desired culture B (or non-target cells)).When learning the analysis data of culture A and its culture results, the correct answer label indicating that culture A has become culture C (or non-target cells) is assigned and the model is trained, thereby characterizing the distribution of the analysis data.

[0265] As the feature extraction model, various feature extraction models can be applied, such as principal component analysis (PCA), partial least squares (PLS), polynomial regression, Gaussian process regression, random forest regression, etc. In addition, as the feature extraction model, for example, a machine learning model generated by machine learning the characteristic values ​​(analysis data by a flow cytometer) of culture A and the culture results may be applied.

[0266] As in the above-described embodiment, in addition to the above-described polynomial regression, Gaussian process regression, and Random Forest Regression, for example, neural networks (CNN and BNN) can be applied as the machine learning model. When generating feature extraction models for culture B, culture C, and non-target cells by unsupervised learning, it is desirable to use multiple analysis data of culture A as in the above-described embodiment, and to learn to characterize the distribution of these multiple analysis data for each culture result that is culture B, culture C, or non-target cells.

[0267] Specifically, when generating a feature extraction model by unsupervised learning, the state estimation unit 10a approximates the distribution of the analysis data group of the multiple cultures A that have been cultured to become cultures B and the distribution of the analysis data group of the multiple cultures A that have been cultured to become cultures C, for example, by a multidimensional Gaussian distribution. A feature extraction model that distinguishes non-target cells from the rest (here, cultures B and C) can be constructed by performing a threshold calculation according to the Mahalanobis distance in the state estimation space from the distribution center of the multidimensional Gaussian distribution generated by the above process. Next, a cluster corresponding to culture B and a cluster corresponding to culture C can be obtained by clustering the distribution in the state estimation space of the analysis data group of the multiple cultures A that have been cultured to become cultures B and the distribution in the state estimation space of the analysis data group of the multiple cultures A that have been cultured to become cultures C, for example, by using the k-means method. This is a feature extraction model that distinguishes cultures B and C.

[0268] When estimating the future culture state of the target culture A, the state estimation unit 10a inputs the analysis data of the target culture A into a feature extraction model for culture B, and outputs an estimation result estimating whether or not the culture state of the target culture A will become culture B in the future when the target culture A is continued to be cultured as it is, based on the feature extraction model for culture B. Similarly, the state estimation unit 10a also inputs, for example, the analysis data of the target culture A into a feature extraction model for culture C, and outputs an estimation result estimating whether or not the culture state of the target culture A will become culture C in the future when the target culture A is continued to be cultured as it is, based on the feature extraction model for culture C. Furthermore, the state estimation unit 10a inputs, for example, the analysis data of the target culture A into a feature extraction model for non-target cells, and outputs an estimation result estimating whether or not the culture state of the target culture A will become non-target cells in the future when the target culture A is continued to be cultured as it is, based on the feature extraction model for non-target cells.

[0269] The state estimation unit 10a selects the estimation result with the highest accuracy rate from among the estimation results obtained from the feature extraction model for culture B, the estimation result obtained from the feature extraction model for culture C, and the estimation result obtained from the feature extraction model for non-target cells, as the estimation result that represents the future culture state when the target culture A is continued to be cultured.

[0270] In the above-mentioned fourth embodiment, a case has been described in which feature extraction models for culture B, culture C, and non-target cells are generated, and input data is input to each feature extraction model to obtain an estimation result for each feature extraction model, but the present invention is not limited to this. For example, it is desirable to generate one feature extraction model, input input data to the feature extraction model, and obtain an estimation result of the future culture state of culture A based on the one feature extraction model.

[0271] For example, when the state estimation unit 10a generates one feature extraction model by supervised learning, the state estimation unit 10a learns the correspondence between multiple analysis data obtained by measuring multiple cultures A at a certain point in time using the measurement unit 2 and the culture results obtained by actually culturing each of these cultures A and resulting in culture B, culture C, or non-target cells, and generates one feature extraction model that characterizes the distribution of the analysis data from these culture results.

[0272] Furthermore, when generating one feature extraction model by unsupervised learning, the state estimation unit 10a performs clustering using the k-means method on the analysis data of culture A that will become culture B, the analysis data of culture A that will become culture C, and the analysis data of culture A that will become non-target cells, and performs classification optimization without labels on these multiple analysis data in a state estimation space in which the features are defined. After that, an operator or the like labels the obtained clusters with culture B, culture C, and non-target cells (assigning identification values ​​such as lines and curves that can be identified in stages) to generate a feature extraction model. As a result, the state estimation unit 10a projects the analysis data of the target culture A into the state estimation space of the feature extraction model, and can estimate whether the culture state of the target culture A will become culture B, culture C, or non-target cells in the future when the target culture A is continued to be cultured, based on the identification values ​​(labels, etc.).

[0273] When the state estimation unit 10a obtains an estimation result of the future culture state of the target culture A, it outputs the estimation result to the next step determination unit 22. The next step determination unit 22 determines one of a processing step of culturing the target culture A under optimal culture conditions for inducing differentiation of the target culture A into culture B (hereinafter also referred to as a processing step for culture B), a processing step of culturing the target culture A under optimal culture conditions for inducing differentiation of the target culture A into culture C (hereinafter also referred to as a processing step for culture C), and a processing step of discarding the target culture A (hereinafter also referred to as a discarding processing step) based on the estimation result of the target culture A, and outputs the result to the fractionation unit 25 as next step information.

[0274] Specifically, when the next process determination unit 22 receives from the state estimation unit 10a an estimation result that the future culture state of the target culture A will be culture B, it determines a processing process for culture B, whereas when the next process determination unit 22 receives from the state estimation unit 10a an estimation result that the future culture state of the target culture A will be culture C, it determines a processing process for culture C. In addition, when the next process determination unit 22 receives from the state estimation unit 10a an estimation result that the future culture state of the target culture A will be non-target cells, it determines a disposal processing process.

[0275] The fractionating section 25 is configured to automatically fractionate the target culture A based on next process information received from the next process determination section 22, and send the target culture A to either the first culture section 26a, the second culture section 26b, or the disposal section 27. As shown in Fig. 15, the fractionating section 25 is provided with a substrate 30 on which a main channel 35, sub-channels 38a, 38b, and branch channels 37a, 37b, 37c are formed.

[0276] Main flow channel 35 formed in substrate 30 is configured such that the upstream side is connected to the flow channel of a flow cytometer which is measurement unit 2, and liquid 40 containing a plurality of cultures A arranged in a line flows into main flow channel 35 from the flow channel of the flow cytometer. The flow rate of liquid 40 flowing into main flow channel 35 from the flow channel of the flow cytometer is controlled by a syringe pump, rotary pump, centrifugal pump, air pressure pump, or the like (not shown).

[0277] Branch channels 37a, 37b, and 37c are connected to the downstream side of the main channel 35, and one end of each of the sub-channels 38a and 38b is connected between the flow channel of the flow cytometer and the branch channels 37a, 37b, and 37c. The sub-channels 38a and 38b are arranged to face each other and to be perpendicular to the flow direction of the main channel 35. The sub-channels 38a and 38b adjust the flow direction of the liquid 40 in the main channel 35 by controlling the inflow of the control liquid into the main channel 35, and cause each culture A contained in the liquid 40 to flow out to the optimal branch channels 37a, 37b, and 37c, respectively. The inflow of the control liquid from the sub-channels 38a and 38b to the main channel 35 is controlled by a fluid control unit 38, such as a syringe pump, a rotary pump, a centrifugal pump, or an air pressure pump, provided in the sub-channels 38a and 38b.

[0278] Here, the downstream of the branch flow path 37a is connected to the first culture section 26a which performs a processing step for culture B, the downstream of the branch flow path 37b is connected to the second culture section 26b which performs a processing step for culture C, and the downstream of the branch flow path 37c is connected to the waste section 27 which performs a waste processing step.

[0279] In the sorting section 25, the fluid control section 38 receives next process information from the next process determination section 22, and in accordance with the next process information, the fluid control section 38 adjusts the inflow of control liquid from the sub-flow paths 38a, 38b to the main flow path 35, and causes the target culture A for which the next process information has been obtained to flow into the branch flow paths 37a, 37b, 37c corresponding to the next process information.

[0280] For example, when target culture A, for which a processing step for culture B has been determined by next process determination unit 22, flows through main channel 35, sorting unit 25 controls the inflow of liquid from sub-channels 38a, 38b into main channel 35 by fluid control unit 38, guides target culture A to branch channel 37a, and guides it to first culture unit 26a via branch channel 37a. As a result, target culture A, for which a processing step for culture B has been determined, is cultured in first culture unit 26a under optimal culture conditions for inducing differentiation into culture B.

[0281] In addition, when the target culture A, for which a processing step for culture C has been determined by the next process determination unit 22, flows through the main channel 35, the sorting unit 25 controls the inflow of liquid from the sub-channels 38a, 38b to the main channel 35 by the fluid control unit 38, guides the target culture A to the branch channel 37b, and guides it to the second culture unit 26b via the branch channel 37b. As a result, the target culture A, for which a processing step for culture C has been determined, is cultured in the second culture unit 26b under optimal culture conditions for inducing differentiation into culture C.

[0282] When the target culture A for which a waste processing step has been determined by the next process determination unit 22 flows through the main channel 35, the sorting unit 25 controls the inflow of liquid from the sub-channels 38a, 38b to the main channel 35 by the fluid control unit 38, guides the target culture A to the branch channel 37c, and guides it via the branch channel 37c to the waste unit 27. As a result, the target culture A for which a waste processing step has been determined is discarded by the waste unit 27.

[0283] (4-2) Actions and Effects In the above configuration, in the state estimation system 1d according to the fourth embodiment, the characteristic values ​​of a plurality of cultures A measured by the measurement unit 2 are used as input data, and based on a feature extraction model that characterizes the distribution of this input data, the state estimation unit 10a estimates which of a plurality of culture states (culture B, culture C, or non-target cells) the culture A will have when it is cultured in the future, from the characteristic values ​​of the target culture A. In this way, the state estimation system 1d can also estimate in advance whether or not a desired culture state can be obtained from the target culture A, as in the above-mentioned embodiment, and therefore the time required for proper judgment of the target culture A can be shortened. This can improve the efficiency of bioproduction and research and development using the culture A.

[0284] In addition, in the state estimation system 1d according to the fourth embodiment, when culturing culture A, it is possible to separate at an early stage only culture A that is likely to reach a specific culture state in the future, and to proceed with culturing under optimal culture conditions for bringing only the separated culture A into the specific culture state.

[0285] For example, when performing differentiation-inducing culture of iPS cells to obtain target cells (e.g., retina, nerve, red blood cell, cardiac muscle, etc.), by using iPS cells as culture A and using the above-mentioned state estimation system 1d, it is possible to estimate in advance whether the iPS cells will become target cells when cultured, without actually culturing the iPS cells. Also, when culturing iPS cells to obtain iPS cell-derived cardiac muscle cells as target cells, successively via mesodermal cells and cardiac muscle precursor cells, the above-mentioned state estimation system 1d can be used to estimate early, before or during the culture of the iPS cells, whether the iPS cells will first become mesodermal cells through culture, thereby significantly shortening the time required to determine the suitability of the iPS cells.

[0286] Therefore, when iPS cells are cultured to produce the desired cells, that is, iPS cell-derived cardiomyocytes, the state estimation system 1d can distinguish between iPS cells that have a high probability of producing the desired cells in the future and iPS cells that may not be able to produce the desired cells even if cultured, and can select only the iPS cells that have a high probability of producing the desired cells in the future and continue culturing them, thereby eliminating unnecessary culturing work and allowing the culturing work to be performed efficiently. (4-3) Other embodiments In the above-mentioned fourth embodiment, a case has been described in which the cells (cultures B and C) obtained by culturing culture A are estimated as the future culture state of culture A based on a feature extraction model from the measurement results of culture A, but the present invention is not limited to this. For example, various blood cells such as red blood cells, white blood cells, and platelets obtained by culturing culture A, or substances and components other than cells, may be estimated as the future culture state of culture A based on a feature extraction model from the measurement results of culture A.

[0287] In the above-described embodiment, as an example of generating a feature extraction model, for example, a case has been described in which a feature extraction model is generated using the measurement result of culture A by the measurement unit 2 and the culture result that culture A has become culture B, culture C, or non-target cells when cultured. However, the present invention is not limited to this, and the feature extraction model may be generated using only the culture results when culture B and culture C are obtained, without using the culture result when culture A has become non-target cells when cultured. In this case, the measurement result of measuring the characteristic value of target culture A by the measurement unit 2 is used as input data, and when estimating the future culture state of target culture A from the input data based on the feature extraction model, the estimation result that culture A will not become culture B or culture C is output as the estimation result that culture A will become non-target cells.

[0288] In the above embodiment, the measurement result of the target culture A by the measurement unit 2 is used as input data, and the culture state when the target culture A is cultured is estimated based on the input data and the feature extraction model as to whether the culture state when the target culture A is cultured will be culture B, culture C, or non-target cells, but the present invention is not limited to this. For example, as the culture state when the target culture A is cultured, in addition to setting a discrimination value (straight line, curve, etc.) that defines culture B and culture C in the state estimation space of the feature extraction model, a discrimination value that further subdivides the non-target cells may be defined, and an estimation result that distinguishes the possibility that the target culture A will induce differentiation into dead cells (cells in poor condition) and other cells among the non-target cells may be obtained based on the feature extraction model.

[0289] As described above, the feature extraction model obtained from the measurement results (input data) of the target culture A by the measurement unit 2, which distinguishes between the possibility of inducing differentiation of non-target cells, such as dead cells (cells in poor condition), and other cells, can be generated by supervised learning or unsupervised learning in the same manner as for the above-mentioned cultures B and C.

[0290] In addition, in the above-mentioned fourth embodiment, when estimating the culture state when culture A is cultured based on the measurement results of the target culture A by the measurement unit 2 and based on the feature extraction model, it is possible to estimate the future culture state of the target culture A (cultures B, C, and non-target cells) while detecting abnormalities in the target culture A in the state estimation space of the feature extraction model based on, for example, the local outlier factor method or the like.

[0291] In the above-mentioned fourth embodiment, as shown in Fig. 13, the case where culture A is cultured and then induced to differentiate into one of three types of culture B, culture C, and non-target cells is described, but the present invention is not limited thereto. For example, as shown in Fig. 16A in 16, the present invention may be applied to the case where culture A is cultured (step S10) and then induced to differentiate into culture E (e.g., red blood cells) shown in step S15, or non-target cells (various blood cells such as red blood cells, white blood cells, and platelets) shown in step S13, or the case where culture A is induced to differentiate into four or more cells.

[0292] In addition, for example, as shown in 16B of FIG. 16, the state estimation system 1d according to the fourth embodiment may also be applied to a case where, as culture A is cultured, differentiation is induced in multiple stages to a culture E via a culture D. In this case, 16B of FIG. 16 shows an example where, when a culture A (e.g., human iPS cells (iPS cells)) shown in step S10 is cultured in a medium containing a predetermined cytokine, the culture A is induced to differentiate into a culture D of blood progenitor cells (step S15) due to the influence of the state of the culture A, or becomes non-target cells such as other cells, dead cells, cells in poor condition, etc. (step S13). In addition, in this example, when the obtained culture D is further cultured thereafter, the culture D is induced to differentiate into a culture E (red blood cells), or becomes non-target cells such as other cells, dead cells, cells in poor condition, etc. (step S13).

[0293] In such a case, in the state estimation system 1d, for example, the state estimation unit 10a generates a first feature extraction model that estimates whether culture A will be cultured to become culture D or whether culture A will be cultured to become non-target cells. Furthermore, in the state estimation system 1d, the state estimation unit 10a generates a second feature extraction model that estimates whether culture D will be cultured to become culture E or whether culture D will be cultured to become non-target cells.

[0294] Note that the generation of the feature extraction model here is generated in the same manner as described above, and therefore the description thereof will be omitted here. As for the feature extraction models, as described above, a feature extraction model for culture D that estimates whether culture A will be cultured to become culture D, a feature extraction model that estimates whether culture A will be cultured to become non-target cells, a feature extraction model for culture E that estimates whether culture D will be cultured to become culture E, and a feature extraction model that estimates whether culture D will be cultured to become non-target cells may be generated individually.

[0295] The measurement unit 2, which is a flow cytometer, measures each characteristic value of a plurality of target cultures A and obtains the measurement results as analysis data. The state estimation unit 10a inputs the analysis data of the target culture A to a first feature extraction model, and obtains an estimation result based on the first feature extraction model as to whether or not the culture state of the target culture A will become culture D or non-target cells in the future if the target culture A is continued to be cultured as is.

[0296] When the state estimation unit 10a obtains an estimation result of the future culture state for each target culture A, it outputs the estimation result to the next step determination unit 22. The next step determination unit 22 determines a corresponding processing step based on the estimation result of the target culture A from among a processing step for culturing the target culture A under optimal culture conditions for inducing differentiation of the target culture A into a culture D (hereinafter also referred to as a processing step for culture D) and a processing step for discarding the target culture A (hereinafter also referred to as a first discard processing step), and outputs the determination result to the fractionation unit 25 as next step information.

[0297] The sorting unit 25 automatically sorts the target culture A based on the next step information received from the next step determination unit 22, and selects and sends only the target culture A for which the next step information of the processing step for culture D has been obtained to the first culture unit 26a. Note that the sorting unit 25 sends the target culture A for which the next step information of the first disposal processing step has been obtained based on the estimation result that the target culture A will become non-target cells to the disposal unit 27, and causes the disposal unit 27 to discard the target culture A.

[0298] Next, the measurement unit 2, which is a flow cytometer, treats the culture D obtained by culturing the target culture A in the first culture unit 26a as the target culture D, measures the characteristic values ​​of each target culture D, and obtains the obtained measurement results as analysis data. The state estimation unit 10a inputs the analysis data of the target culture D into a second feature extraction model, and estimates, based on the second feature extraction model, whether the culture state of the target culture D will become culture E or non-target cells in the future if the target culture D is continued to be cultured as is.

[0299] When the state estimation unit 10a obtains an estimation result of the future culture state for each target culture D, it outputs the estimation result to the next step determination unit 22. The next step determination unit 22 determines a corresponding processing step based on the estimation result of the target culture D from among a processing step for culturing the target culture D under optimal culture conditions for inducing differentiation of the target culture D into a culture E (hereinafter also referred to as a processing step for culture E) and a processing step for discarding the target culture D (hereinafter also referred to as a second discard processing step), and outputs the determination result to the fractionation unit 25 as next step information.

[0300] The fractionating unit 25 automatically fractionates the target culture D based on the next step information received from the next step determination unit 22, and selects and sends only the target culture D for which the next step information of the processing step for culture E has been obtained to the second culture unit 26b. Note that the fractionating unit 25 sends the target culture D for which the next step information of the second disposal processing step has been obtained based on the estimation result that the target culture D will become non-target cells to the disposal unit 27, and causes the disposal unit 27 to discard the target culture D.

[0301] (4-3)Other In addition, in the state estimation systems 1a, 1c, and 1d according to the first to fourth embodiments described above, for the purpose of cultivating the culture, the culture state of the culture may be estimated from the input data obtained by the measurement unit based on a feature extraction model.

[0302] In this case, for example, in the example shown in 16B of FIG. 16, the target culture A estimated to become culture D by the state estimation step is fractionated (fractionation step), cultured in a first culture section to become culture D (culture step), and a mutagen that induces mutation is added to the obtained culture D, genetic manipulation is performed, or the like to obtain culture D' (not shown) (mutation step). Then, a feature extraction model that estimates the future culture state of culture D' is generated, and the culture state of culture D' is estimated using the feature extraction model (state estimation step). In this way, in the state estimation systems 1a, 1c, and 1d, the state estimation step, fractionation step, culture step, and mutation step are repeated to select and mutate the culture state of the culture, and the selected culture (for example, a group of cultures consisting of only mutated cultures, or a group of cultures in a state including both mutated and non-mutated cultures) is cultured, whereby a culture with a new function or the like modified can be obtained, and culture breeding can also be performed.

[0303] In addition, in the above breeding, the above mutation step does not necessarily have to be performed, and the above culture step, state estimation step, separation step, etc. may be repeated according to a known breeding process, for example, by separating several types of target culture A whose future culture state has been estimated based on the estimation result of the feature extraction model. In addition, as another breeding, the above culture step, state estimation step, and separation step may be repeated by culturing the target culture A whose culture state has been estimated based on the estimation result of the feature extraction model, separating a culture D'' in which a mutation has occurred, etc. Furthermore, the above culture step, state estimation step, and separation step may be repeated by separating a target culture A that may become a culture state of a mutation in the future based on the estimation result of the feature extraction model.

[0304] In the above-mentioned fourth embodiment, a flow cytometer is applied as the measurement unit, and the analysis data obtained from the flow cytometer is used as the input data, but the present invention is not limited to this, and for example, a mass spectrometer, a microscope, a Raman spectrometer, a chromatography, a digital PCR measurement device, a nuclear magnetic resonance device, an antibody quantification kit, and a nucleic acid sequence determination device may be applied as the measurement unit. In this case, the input data may be a mass spectrum obtained from a mass spectrometer, image data obtained from a microscope, a Raman spectrum obtained from a Raman spectrometer, a chromatogram obtained from chromatography, the amount of a specific substance obtained from a digital PCR measurement device, analysis data obtained from a nuclear magnetic resonance device, the amount of a specific antigen obtained from an antibody quantification kit, and gene sequence analysis data obtained from a nucleic acid sequence determination device.

[0305] In the above-mentioned fourth embodiment, the analysis data of a plurality of cultures acquired by a flow cytometer is used as input data including at least characteristic values ​​of a plurality of cultures acquired via the measurement unit 2, and a feature extraction model is generated using the input data, and an estimation result is obtained by the feature extraction model. However, the present invention is not limited to this. For example, in addition to the analysis data of a plurality of cultures acquired by a flow cytometer, a feature extraction model may be generated and an estimation result may be obtained by the feature extraction model using input data including parameters (culture parameters) related to culture conditions of the culture process of the culture, such as the culture temperature, culture time, culture medium pH, and culture medium stirring speed.

[0306] Similarly, when the mass spectrum obtained from the above-mentioned mass spectrometer, the image data obtained from the microscope, the Raman spectrum obtained from the Raman spectrometer, the chromatogram obtained from the chromatography, the amount of a specific substance obtained from a digital PCR measurement device, the analysis data obtained from a nuclear magnetic resonance device, the amount of a specific antigen obtained from an antibody quantification kit, and the gene sequence analysis data obtained from a nucleic acid sequence determination device are used as input data, the input data may also include parameters related to the culture conditions (culture parameters) in addition to these, and the input data may be used to generate a feature extraction model or obtain an estimation result using the feature extraction model. [Explanation of symbols]

[0307] 1a, 1c, 1d State Estimation System 2 Measuring part 10a, 10c State Estimation Unit 12 Judgment section 14 Measurement object optimization estimation section 15. Culture parameter optimization and estimation section

Claims

1. a measurement unit for measuring characteristics of a plurality of biological samples at a certain time point before the start of a culture process and / or during use in the culture process, and at least one of the target biological samples of the same type as the biological samples at the certain time point, the plurality of biological samples being selected from a culture medium, a culture cultured in the culture medium, a culture extract extracted from the culture, and a culture extract extracted from the culture, the plurality of biological samples being selected from a culture medium, a culture extract extracted from the culture, and a target biological sample being selected from the target biological sample. a database that stores a feature extraction model that characterizes a distribution of input data including at least characteristic values ​​of the plurality of biological samples acquired via the measurement unit; and a state estimation unit that estimates a culture state of the culture from input data including at least a characteristic value of the target biological sample acquired via the measurement unit based on the feature extraction model; a measurement target optimization estimation unit that selects an additive to be added to the culture medium based on the estimation result of the culture state obtained by the state estimation unit; Equipped with The characteristic value measured by the measurement unit is one or more of a mass spectrum obtained from a mass spectrometer, image data obtained from a microscope, a Raman spectrum obtained from a Raman spectrometer, a chromatogram obtained from a chromatography, an amount of a specific substance obtained from a digital PCR measurement device, analysis data obtained from a nuclear magnetic resonance device, an amount of a specific antigen obtained from an antibody quantification kit, and gene sequence analysis data obtained from a nucleic acid sequence determination device; The culture condition is the yield of the culture, the quality of the culture, the yield of the medium extract, or the quality of the culture extract; Culture medium state estimation system.

2. the measurement object optimization estimation unit calculates a similarity between a principal component vector defined in a state estimation space in the feature extraction model and each additive measurement vector obtained by measuring each of the plurality of types of additives by the measurement unit, and selects the additive of the additive measurement vector having a high similarity to the principal component vector; The state estimation system according to claim 1 .

3. The state estimation unit is Estimating the culture state after a culture period determined in the culture process has elapsed and the culture has been terminated. The state estimation system according to claim 1 or 2.

4. The state estimation unit is Estimating the culture state during the culture process. The state estimation system according to claim 1 or 2.

5. The measurement unit includes: measuring characteristics of the plurality of biological samples and the target biological sample in a time series during the culture process; The state estimation unit is the characteristic values ​​are the input data, and the culture state is estimated from the characteristic values ​​of the target biological specimen for each measurement time of the characteristic values ​​based on the feature extraction model that characterizes a distribution of the characteristic values ​​of the plurality of biological specimens for each measurement time of the characteristic values; The state estimation system according to any one of claims 1 to 4.

6. a determination unit for determining a state of the target biological sample based on the estimation result of the culture state obtained by the state estimation unit; The state estimation system according to any one of claims 1 to 5.

7. The determination unit is determining whether or not the target biological sample has an abnormality based on the culture state estimation result obtained by the state estimation unit; The state estimation system according to claim 6 .

8. The state estimation unit is using the characteristic values ​​of the plurality of biological specimens and culture parameters related to culture conditions of a culture process using the biological specimens as the input data, and estimating the culture state from the input data of the target biological specimen based on the feature extraction model that characterizes the distribution of the characteristic values ​​and the culture parameters. The state estimation system according to any one of claims 1 to 4.

9. a determination unit that determines appropriateness of the culture parameters based on the estimation result of the culture state obtained by the state estimation unit; The state estimation system according to claim 8 .

10. The state estimation unit is estimating a culture state of the culture from the input data based on the feature extraction model for breeding the culture; The state estimation system according to any one of claims 1 to 9.

11. a measuring step in which a medium, a culture cultivated in the medium, a medium extract extracted from the medium, or a culture extract extracted from the culture is used as a biological sample and a target biological sample, and characteristics of a plurality of the biological samples at a certain time point before the start of a culture process and / or during use in the culture process, and at least one target biological sample of the same type as the biological samples at the said time point are measured by a measuring unit; a storage step of storing in a database a feature extraction model that characterizes a distribution of input data including at least characteristic values ​​of the plurality of biological samples acquired via the measurement unit; a state estimating step of estimating a culture state of the culture by a state estimating unit from input data including at least a characteristic value of the target biological sample acquired via the measuring unit based on the feature extraction model; a measurement target optimization estimation step of selecting an additive to be added to the culture medium based on the estimation result of the culture state obtained by the state estimation unit; Including, The characteristic value measured by the measurement unit is one or more of a mass spectrum obtained from a mass spectrometer, image data obtained from a microscope, a Raman spectrum obtained from a Raman spectrometer, a chromatogram obtained from a chromatography, an amount of a specific substance obtained from a digital PCR measurement device, analysis data obtained from a nuclear magnetic resonance device, an amount of a specific antigen obtained from an antibody quantification kit, and gene sequence analysis data obtained from a nucleic acid sequence determination device; The culture condition is the yield of the culture, the quality of the culture, the yield of the medium extract, or the quality of the culture extract; A method for estimating the state of the medium.

12. In the measurement object optimization estimation step, a similarity between a principal component vector defined in a state estimation space in the feature extraction model and each additive measurement vector obtained by measuring each of the plurality of types of additives by the measurement unit is calculated, and the additive of the additive measurement vector having a high similarity to the principal component vector is selected. The state estimation method according to claim 11.

Images