Estimation device, learning device, optimization device, estimation method, learning method, and optimization method

Abstract

The estimation device (200) generates quality prediction data (540) indicating the quality of a drug substance of a biopharmaceutical manufactured by culturing the cells, by inputting assay data (510) containing an assay result obtained by assaying the substance in the culture vessel at at least one timing after a prescribed period has elapsed from inoculation of the cells into the culture medium, to a prediction model (420). The prediction model (420) is generated by performing a learning process using learning data (530) containing assay data containing assay results obtained by assaying the substance in the culture vessel at a plurality of timings from inoculation of the cells into the culture medium, and quality data obtained by analyzing the drug substance of the biopharmaceutical manufactured with the cells.

Description

Technical Field

[0001] This disclosure relates to an estimation device, a learning device, an optimization device, an estimation method, a learning method, and an optimization method. Background Technology

[0002] In recent years, research has been conducted on biopharmaceuticals manufactured through cell culture. Japanese Patent No. 4496086 (Patent Document 1) discloses a cell culture method for culturing host cells to generate target proteins.

[0003] Biopharmaceuticals are produced using cellular metabolism, which is difficult to control artificially, and sometimes contain components derived from living cells. Therefore, even when manufactured according to established protocols, it is difficult to ensure consistent quality in the produced biopharmaceuticals. Thus, generally, characterization of the active pharmaceutical ingredient (API) of biopharmaceuticals obtained through cell culture is performed to confirm the quality of the API.

[0004] Existing technical documents

[0005] Patent documents

[0006] Patent Document 1: Japanese Patent No. 4496086 Summary of the Invention

[0007] The problem the invention aims to solve

[0008] The manufacture of biopharmaceuticals, if involving cultivation, purification, and processing, can take a very long time, ranging from weeks to months. Furthermore, the active pharmaceutical ingredient (API) for characterization is obtained through purification after the cultivation process. In other words, characterization results cannot be obtained during the cultivation stage; therefore, it can only be determined after purification that the target API has been successfully obtained through cultivation.

[0009] One object of this disclosure is to predict the quality of the active pharmaceutical ingredient (API) obtained from the culture at a stage midway through the culture process. Another object of this disclosure is to optimize the conditions for improving the quality of the API obtained from the culture at a stage midway through the culture process.

[0010] Solution for solving the problem

[0011] The estimation apparatus disclosed herein comprises: a receiving unit that receives input measurement data; a prediction unit that generates quality prediction data representing the predicted quality of a drug substance by inputting the measurement data received by the receiving unit into a prediction model; and an output unit that outputs the quality prediction data generated by the prediction unit. The measurement data includes measurement results obtained from at least one timed measurement of at least one substance contained in a culture vessel containing the cells and the culture medium after a predetermined period has elapsed since cells were inoculated into the culture medium. The prediction model is a model used to predict the quality of a drug substance for a biopharmaceutical manufactured by culturing cells.

[0012] Furthermore, the learning apparatus of this disclosure includes: a receiving unit that receives learning data; and a model generation unit that generates a predictive model by performing learning processing using the learning data received by the receiving unit. The learning data includes measurement data and quality data. The measurement data includes measurement results obtained from multiple timed measurements of at least one substance in a culture vessel containing the cells and the culture medium, starting from when cells are seeded into the culture medium. The quality data is data obtained by analyzing the active pharmaceutical ingredient (API) of a biopharmaceutical manufactured using the cells. The predictive model is a model for generating quality prediction data representing the quality of the API of a biopharmaceutical manufactured using the cells contained in the culture vessel during the intermediate stage of cell culture, based on the measurement results obtained from measuring at least one substance in the culture vessel during the intermediate stage of cell culture.

[0013] Furthermore, the estimation method disclosed herein includes the following steps: culturing cells by placing cells and culture medium in a culture vessel; a measurement step, measuring at least one substance in the culture vessel at at least one time interval after a predetermined period has elapsed since the cells were seeded into the culture medium; generating quality prediction data representing the quality of the active pharmaceutical ingredient by inputting the measurement data, including the measurement results obtained in the measurement step, into a prediction model; and outputting the quality prediction data. The prediction model is a model used to predict the quality of the active pharmaceutical ingredient of a biopharmaceutical manufactured by culturing cells.

[0014] Furthermore, the learning method disclosed herein includes the following steps: culturing cells by placing cells and culture medium in a culture vessel; a measurement step, measuring at least one substance within the culture vessel at multiple time intervals from the time the cells are seeded into the culture medium; an analysis step, analyzing the quality of the active pharmaceutical ingredient (API) of a biopharmaceutical manufactured from the cells; and generating a predictive model by performing learning processing using learning data. The learning data includes measurement data and quality data, the measurement data including the measurement results obtained through the measurement step, and the quality data being the data obtained through the analysis step. The predictive model is a model used to generate quality prediction data representing the quality of the API of a biopharmaceutical manufactured from cells contained in the culture vessel during the intermediate stages of cell culture, based on the measurement results obtained from measuring at least one substance within the culture vessel during the intermediate stages of cell culture.

[0015] Furthermore, the optimization apparatus disclosed herein includes: a receiving unit that receives values ​​of multiple parameters for defining culture conditions when cells are seeded into a culture medium and quality data obtained by analyzing the active pharmaceutical ingredient of a biopharmaceutical manufactured by culturing cells; an estimation unit that takes the values ​​of the multiple parameters received by the receiving unit and the quality data as inputs to estimate an optimal combination of values ​​of the multiple parameters; and an output unit that outputs the combination of values ​​of the multiple parameters estimated by the estimation unit.

[0016] Furthermore, the optimization method disclosed herein includes the following steps: a cell culture step, in which cells and culture medium are placed in a culture container and the cells are cultured under culture conditions defined by the values ​​of multiple parameters; an analysis step, in which the quality of the active pharmaceutical ingredient (API) of the biopharmaceutical manufactured by culturing the cells is analyzed; and an estimation step, in which the optimal combination of the values ​​of the multiple parameters is estimated, using the values ​​of the multiple parameters and the quality data obtained through the analysis step as input. In the cell culture step, the optimal combination of the values ​​of the multiple parameters estimated through the estimation step is used as the new culture conditions for culturing the cells.

[0017] The effects of the invention

[0018] According to this disclosure, the quality of the active pharmaceutical ingredient (API) obtained from the culture can be predicted at a stage during the culture process. Alternatively, according to this disclosure, the user can optimize the conditions for improving the quality of the API obtained from the culture at a stage during the culture process. Therefore, based on the predicted quality of the API, the user can take various measures such as modifying the experimental protocol, extending the culture time, or terminating the culture at the stage of the culture process before the purification process to obtain the API, thereby reducing the manufacturing cost of biopharmaceuticals. Attached Figure Description

[0019] Figure 1 This is a diagram schematically illustrating the overall structure of the prediction system involved in Embodiment 1.

[0020] Figure 2 This is a block diagram showing the hardware structure of the estimation device.

[0021] Figure 3 This is a block diagram showing the hardware structure of the learning device.

[0022] Figure 4 This is a diagram used to illustrate the timing of the measurement.

[0023] Figure 5 This is a graph showing an example of measured data.

[0024] Figure 6 This is a block diagram illustrating an example of the structure of the prediction unit.

[0025] Figure 7 This is a diagram showing an example of the output result from the output unit.

[0026] Figure 8 This is a flowchart illustrating the process of the estimation method involved in Embodiment 1.

[0027] Figure 9 This is a flowchart illustrating the process of the learning method involved in Embodiment 1.

[0028] Figure 10 This is a block diagram illustrating the structure of the learning device involved in the modified example.

[0029] Figure 11 This is a diagram showing a variation of the model generation section.

[0030] Figure 12 This is a block diagram showing the structure of the estimation device involved in the modified example.

[0031] Figure 13 This is a diagram showing a variation of the prediction model.

[0032] Figure 14 This is a diagram schematically illustrating the overall structure of the prediction system involved in Embodiment 2.

[0033] Figure 15 This is a block diagram illustrating the hardware structure of the optimization device.

[0034] Figure 16 This is a block diagram illustrating an example of the structure of the optimization device.

[0035] Figure 17 This is a flowchart illustrating the processing flow of the optimization device.

[0036] Figure 18 This is a diagram illustrating a variation of the prediction system according to Embodiment 2. Detailed Implementation

[0037] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings. Furthermore, the same or equivalent parts in the drawings will be labeled with the same reference numerals, and their description will not be repeated.

[0038] Implementation method 1.

[0039] [Overall structure of the prediction system]

[0040] Figure 1This diagram schematically illustrates the overall structure of the prediction system according to Embodiment 1. The prediction system SYS1 is a system for predicting the quality of active pharmaceutical ingredients (APIs) of biopharmaceuticals manufactured through cell culture. In Embodiment 1, "biopharmaceutical" refers to a drug manufactured using cells, such as antibody drugs and vaccines. Furthermore, in Embodiment 1, the biopharmaceutical may also include the cells themselves used in regenerative medicine. In Embodiment 1, "API" refers to the target substance obtained through cell culture, for example, obtained by a purification process after the cell culture process to extract the target component. The prediction system SYS1 includes a measuring device 100, an estimating device 200, a learning device 300, and a server 400.

[0041] The measuring device 100 measures the substance contained within the culture container 10, which includes cells 1 and culture medium 2, as the object of measurement. The substance being measured is preferably a substance that changes during the culture of cells 1, such as cells 1 themselves, cell metabolites, or nutrients of cells 1. The user can select any measuring device 100 depending on the object of measurement. For example, when cells 1 are the object of measurement, the user can select a microscope and software for analyzing images obtained from the microscope as the measuring device 100 to obtain measurement results such as the number of cells, viability, and cell shape (roundness, length of the long axis, length of the short axis, etc.) in the culture medium composed of cells 1 and culture medium 2 within the culture container 10. Furthermore, when cell metabolites or nutrients are the object of measurement, the user can select liquid chromatography-mass spectrometry (LC-MS), inductively coupled plasma (ICP) mass spectrometry, or similar instruments as the measuring device 100 to obtain measurement results such as the concentration of metabolites and nutrients in the culture medium.

[0042] The estimation device 200 uses measurement data 510, which includes measurement results for the test subject after a predetermined period has elapsed since the cells 1 were inoculated into the culture medium 2, to predict the quality of the active pharmaceutical ingredient. For example, the measurement data 510 is obtained by extracting the culture medium in the culture vessel 10 at least once after a predetermined period has elapsed since the cells 1 were inoculated into the culture medium 2 and measuring the substances in the extracted culture medium using the measurement device 100.

[0043] As passed by processor 21 (reference) Figure 2 ) Perform estimation procedure 560 (refer to) Figure 2 As an example of the software structure implemented, the estimation device 200 includes a receiving unit 210, a prediction unit 220, and an output unit 230. The receiving unit 210 receives the input of the measurement data 510.

[0044] The prediction unit 220 generates quality prediction data 540 representing the quality of the active pharmaceutical ingredient (API) by inputting the measurement data 510 received by the receiving unit 210 into the prediction model 420. The prediction model 420 is a model used to predict the quality of the API of a biopharmaceutical manufactured by culturing cells 1. It accepts the measurement data 510 as input and outputs a prediction result representing the quality of the API. The prediction model 420 is generated by supervised learning processing performed by the model generation unit 320 of the learning device 300.

[0045] The quality prediction data 540 can be the prediction result output from the prediction model 420 itself, or it can be data generated based on the prediction result output from the prediction model 420. The prediction result output from the prediction model 420 corresponds to the quality data 520 used as the positive solution data when the model generation unit 320 generates the model. For example, if the quality data 520 is data representing the analysis result obtained by arbitrarily analyzing the active pharmaceutical ingredient 3, the prediction model 420 outputs the analysis result predicted as the prediction result. If the quality data 520 is data representing the similarity between the target active pharmaceutical ingredient and the analysis result, the prediction model 420 outputs the similarity predicted as the prediction result.

[0046] The output unit 230 outputs the quality prediction data 540 generated by the prediction unit 220 to any output destination such as a monitor, printer, or server.

[0047] The learning device 300 uses the learning data 530 to generate a prediction model 420. This is achieved through the processor 31 (see reference). Figure 3 ) Execute learning program 580 (refer to) Figure 3 As an example of the software structure implemented, the learning device 300 has a receiving unit 310 and a model generation unit 320.

[0048] The receiving department 310 receives the input of learning data 530. Learning data 530 includes time series data 512 and quality data 520.

[0049] Time series data 512 is a type of measurement data 510, obtained by measuring at least one substance within the culture vessel 10 at multiple time points (time point t1, ..., time point tx) from the time the cells 1 are seeded into the culture medium 2. Time series data 512 is data representing the time-dependent changes of a substance within the culture vessel 10 containing cells 1 and culture medium 2 as the measurement object. Time series data 512 is obtained by extracting culture medium from the culture vessel 10 at multiple time points (time point t1, ..., time point tx) from the time the cells 1 are seeded into the culture medium 2, and measuring each of the extracted culture medium samples using the measuring device 100. For example, time series data 512 represents time-dependent changes in cell number, cell viability, cell shape, nutrient concentration, and metabolite concentration, etc.

[0050] Quality data 520 is obtained by analyzing the active pharmaceutical ingredient 3 (API 3) of the biopharmaceutical manufactured using cell 1. Quality data 520 is data obtained after culturing cell 1 through purification and analysis processes, and is data representing the quality of API 3 obtained by arbitrarily analyzing API 3 obtained after the purification process in the analysis process. The analysis performed in the analysis process is arbitrarily selected depending on the type of API 3. The quality of API 3 is evaluated from one or more perspectives, such as physical properties, chemical properties, biological activity, and immunochemical properties. Quality data 520 can be data representing the analysis results obtained by arbitrarily analyzing API 3, or it can be data representing the similarity to the target API based on the analysis results.

[0051] For example, when active pharmaceutical ingredient 3 is used in an antibody drug, the analytical process may involve analyzing the amino acid composition, amino acid sequence, peptide mapping, disulfide bond crosslinking positions, and glycosylation modification structures of active pharmaceutical ingredient 3. In this case, as an example, the results of the amino acid composition analysis and data indicating the degree of similarity between the amino acid composition and the target active pharmaceutical ingredient may also be included in the quality data 520.

[0052] The model generation unit 320 generates a prediction model 420 by performing learning processing on the quality data 520 contained in the learning data 530 as the positive solution data. For example, the model generation unit 320 inputs the time series data 512 contained in the learning data 530 into the prediction model 420, calculates the error between the output quality prediction result and the quality data 520 used as the positive solution data, and optimizes the prediction model 420 in a way that reduces the error.

[0053] Server 400 is one example of a storage device used to store the prediction model 420. Alternatively, the prediction model 420 may be stored in storage device 23 of the estimation device 200 (see reference). Figure 2 Alternatively, it can be stored in the storage device 33 of the learning device 300 (see reference). Figure 3 Alternatively, the estimation device 200 and the learning device 300 can be implemented using a single device. In the case where the estimation device 200 and the learning device 300 are implemented using a single device, the receiving unit 210 and the receiving unit 310 can be implemented as a single receiving unit.

[0054] Furthermore, in this embodiment, the culture conditions are assumed to be the same when generating the prediction model 420 and when using the generated prediction model 420 to generate the quality prediction data 540. In this embodiment, "same culture conditions" means that the culture is carried out according to the same experimental protocol.

[0055] Regarding active pharmaceutical ingredients (APIs) obtained through cell culture, even when cell culture is performed according to the same experimental protocol, they may sometimes exhibit characteristics different from the target API due to differences in the cells themselves, or the presence of components derived from living cells within the drug being used. Therefore, for biopharmaceuticals, characterization of APIs obtained through cell culture is generally performed to ensure quality.

[0056] The active pharmaceutical ingredient 3 is obtained through the activity of cell 1. Therefore, it is anticipated that there is a correlation between the properties of the active pharmaceutical ingredient 3 and the activity state of cell 1. Furthermore, it is anticipated that the substances within the culture vessel 10 containing cell 1 will exhibit different changes over time when cell 1 is actively or inactive, such as during proliferation and metabolism. Therefore, it is anticipated that there is a correlation between the state of the substances within the culture vessel 10 and the activity state of cell 1. In other words, it is anticipated that the state of the substances within the culture vessel 10 is indirectly related to the properties of the active pharmaceutical ingredient 3.

[0057] The learning device 300 generates a predictive model 420 by performing learning processing using learning data 530, which includes time-series data 512 representing the time-dependent changes of substances within the culture container 10 and quality data 520 representing the quality of the active pharmaceutical ingredient 3. In other words, the learning device 300 generates the predictive model 420 by learning the correlation between the time-dependent changes of substances within the culture container 10 and the quality of the active pharmaceutical ingredient 3.

[0058] The user repeatedly performs culture experiments, including culture and purification steps, while changing the culture conditions. Quality data 520 is obtained for each culture experiment. The measuring device 100 outputs time-series data 512 for each culture experiment. By repeating the culture experiments, a large amount of learning data 530 is collected corresponding to the number of culture experiments. The learning device 300 repeatedly optimizes the prediction model 420 based on the large amount of learning data.

[0059] In this embodiment 1, the measurement data 510 is obtained by extracting the culture medium in the culture vessel 10 at at least one time interval after a predetermined period has elapsed since the cells 1 were inoculated into the culture medium 2, and measuring the substances in the extracted culture medium using the measuring device 100. That is, the measurement data 510 represents the state of the substances in the culture medium at at least one time interval after a predetermined period has elapsed since the cells 1 were inoculated into the culture medium 2.

[0060] Since the measurement data 510 represents the state of the substance after a specified period has elapsed since the cells 1 were inoculated into the culture medium 2, a prediction result representing the quality of the active pharmaceutical ingredient 3 can be obtained by inputting the measurement data 510 into a prediction model 420 that has learned the correlation between the time-dependent changes of the substance and the quality of the active pharmaceutical ingredient 3.

[0061] As described above, the estimation device 200 of this embodiment 1 generates quality prediction data 540 based on measurement data 510, which includes measurement results obtained from the measurement object within the measurement culture container 10, indicating the quality of the active pharmaceutical ingredient (API) obtained by culturing cells 1 within the culture container 10. Therefore, the quality of the API obtained from the culture can be predicted at a stage midway through the culture process. Furthermore, the learning device 300 can generate a prediction model for predicting the quality of the API obtained from the culture at a stage midway through the culture process. Based on the quality prediction data 540, users can take various actions such as modifying experimental protocols, extending culture time, or terminating culture at stages before the purification process to obtain the API, thereby reducing the manufacturing cost of biopharmaceuticals.

[0062] Furthermore, the learning device 300 according to Embodiment 1 can generate a predictive model 420 that learns the correlation between the time-varying changes of substances in the culture container 10 and the quality of the active pharmaceutical ingredient 3 by performing learning processing using learning data 530, which includes time series data 512 representing the time-varying changes of substances in the culture container 10 and quality data 520 representing the quality of the active pharmaceutical ingredient 3.

[0063] [Hardware structure of the estimation device]

[0064] Figure 2This is a block diagram showing the hardware structure of the estimation device 200. The estimation device 200 includes a processor 21, a main memory 22, a storage device 23, a communication interface (I / F) 24, a USB (Universal Serial Bus) interface (I / F) 25, an input unit 26, and a display unit 27. These components are connected via a processor bus 28.

[0065] The processor 21 consists of a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), etc. It reads the program (for example, estimation program 560) stored in the storage device 23 and executes it in the main memory 22. The processor 21 performs a series of processes for predicting the quality of the active pharmaceutical ingredient by executing the program.

[0066] The main memory 22 is composed of volatile storage devices such as RAM (Random Access Memory) and DRAM (Dynamic RAM). The storage device 23 is composed of non-volatile storage devices such as HDD (Hard Disk Drive) and SSD (Solid State Drive).

[0067] The storage device 23 stores measurement data 510 transmitted via USB I / F 25, quality prediction data 540 generated based on the measurement data 510, and estimation procedures 560 for performing a series of processes to predict the quality of the active pharmaceutical ingredient.

[0068] The communication I / F 24 uses wired or wireless communication to exchange signals with the server 400 that stores the prediction model 420.

[0069] A USB memory (not shown) is detachably installed in the USB I / F 25, and the USB I / F 25 reads the measurement data 510 stored in the USB memory. Alternatively, the measurement data 510 can be transmitted from the measurement device 100 to the estimation device 200 by connecting the measurement device 100 and the estimation device 200 in a communicative manner.

[0070] The input unit 26 receives user input and typically consists of a touch panel, keyboard, mouse, etc. The display unit 27 is an example of the output destination of the quality prediction data 540 and consists of an LCD panel or similar device capable of displaying images.

[0071] [Hardware Structure of the Learning Device]

[0072] Figure 3This is a block diagram showing the hardware structure of the learning device 300. The learning device 300 includes a processor 31, a main memory 32, a storage device 33, a communication I / F 34, a USB I / F 35, an input unit 36, and a display unit 37. These components are connected via a processor bus 38.

[0073] The processor 31, consisting of a CPU, GPU, etc., reads the program (for example, learning program 580) stored in the storage device 33 and executes it in the main memory 32. The processor 31 performs a series of processes to generate the prediction model 420 by executing the program.

[0074] The main memory 32 is composed of volatile storage devices such as RAM and DRAM. The storage device 33 is composed of non-volatile storage devices such as HDD and SSD.

[0075] The storage device 33 stores learning data 530 transmitted via USB I / F 35 and a learning program 580 for performing a series of processes to generate a prediction model 420.

[0076] The communication I / F 34 uses wired or wireless communication to exchange signals with the server 400 that stores the prediction model 420.

[0077] A USB memory (not shown) is detachably installed in the USB I / F 35, and the USB I / F 35 reads the learning data 530 stored in the USB memory. Alternatively, an information processing device that accepts input of measurement data 510 and quality data 520 and generates learning data 530 by associating the measurement data 510 and quality data 520 can be communicatively connected to the learning device 300, and the learning data 530 can be sent from the information processing device to the estimation device 200.

[0078] The input unit 36 ​​receives user input and typically consists of a touch panel, keyboard, mouse, etc. The display unit 37 consists of a liquid crystal panel or similar device capable of displaying images.

[0079] [An example of measurement data 510 used in quality prediction]

[0080] The measurement data 510 received by the estimation device 200 for predicting the quality of the active pharmaceutical ingredient 3 is defined as the measurement results obtained by measuring the test subject at at least one time interval after a predetermined period has elapsed since the cells 1 were inoculated into the culture medium 2. Furthermore, the measurement data 510 used in predicting the quality of the active pharmaceutical ingredient 3 may also be time series data 512 showing the time-varying changes of the test subject, obtained by measuring the test subject at multiple time intervals since the cells 1 were inoculated into the culture medium 2.

[0081] Figure 4 This is a diagram used to illustrate the timing of the measurement. In Figure 4 The data includes typical cell proliferation curves. For example... Figure 4 As shown, generally, after cells 1 are inoculated into culture medium 2, they do not immediately proliferate. Instead, after an induction phase (or lag phase), they enter the logarithmic growth phase. Subsequently, nutrients decrease or excretions (metabolites) accumulate, ending the logarithmic growth phase and entering the stationary phase (or quiescent phase). During the stationary phase, a balance is maintained between the number of proliferating cells 1 and the number of dying cells 1. Shortly after the stationary phase begins, as the number of dying cells 1 increases, the cell death phase (or decline phase) begins.

[0082] The measurement data 510 may also include the measurement results obtained by measuring the test subject at time t0 when cell 1 begins to be cultured. Furthermore, the time when cell 1 begins to be cultured can include both the time before cell 1 is seeded into culture medium 2 and the time immediately after cell 1 is seeded into culture medium 2. The measurement results at time t0 represent the initial state within the culture vessel 10. When the measurement data 510 includes measurement results representing the initial state within the culture vessel 10, the measurement data 510 becomes a representation of changes from the initial state.

[0083] Furthermore, the measurement data 510 can also include the measurement results obtained by measuring the test subject during the time interval t1 to t2, which is the logarithmic proliferation period. That is, at least one time interval after a predetermined period has elapsed since the cells 1 were seeded into the culture medium 2 is the time interval t1 to t2, which is the logarithmic proliferation period. During the logarithmic proliferation period, the number of cells 1 changes drastically, and therefore the state of the substance within the culture vessel 10 changes drastically. By including the results obtained by measuring the test subject at the time interval during which such a large state change occurs in the time series data 512, the estimation device 200 can use data that more accurately represents the state change of the cells 1 to predict the quality of the active pharmaceutical ingredient 3.

[0084] Furthermore, the measurement data 510 can also include the measurement results obtained by measuring the test subject during the time period t2 to t3, which is the stabilization period. That is, at least one time period after a predetermined period has elapsed since the cells 1 were inoculated into the culture medium 2 is the time period t2 to t3, which is the stabilization period. During the stabilization period, a balance is maintained between the number of proliferating cells 1 and the number of dead cells 1, and the number of cells 1 is at its maximum. By including the results obtained by measuring the test subject at the time when the number of cells 1 is at its maximum in the time series data 512, the estimation device 200 can use the data representing the state of the substance in the culture container 10 when the number of cells 1 is at its maximum to predict the quality of the active pharmaceutical ingredient 3.

[0085] Measurement data 510 (time series data 512) includes results obtained during the cell death phase after time t3. Due to cell death, intracellular enzymes and other substances are exposed to the culture medium, potentially affecting the quality of the active pharmaceutical ingredient. In this case, the results obtained during the cell death phase can potentially be effectively utilized in machine learning as examples of data indicating poor quality.

[0086] Figure 5 This is a graph showing an example of measurement data 510. Furthermore, in Figure 5 The example shown illustrates time series data 512a-512f obtained from multiple timed measurements starting from the inoculation of cell 1 into culture medium 2. However, each time series data 512a-512f only needs to include the measurement results obtained from measuring the subject at at least one timed interval after a specified period has elapsed since the inoculation of cell 1 into culture medium 2.

[0087] The measurement data 510 includes measurement results obtained from the measured cells 1, such as time series data 512a of cell number, time series data 512b of viable cell rate, time series data 512c of cell shape, or time series data 512d of microscope image, etc.

[0088] Additionally, the measurement data 510 includes, for example, time-series data 512e of metabolites, which are measurement results obtained from measuring metabolites produced by the metabolism of cell 1. Furthermore, the metabolites include at least one of metabolites secreted externally by cell 1 and metabolites contained within cell 1 itself.

[0089] Additionally, the measurement data 510 includes time-series data 512f of the culture medium components, which are the measurement results obtained as a result of the nutrients taken up by the measurement cell 1. Furthermore, if the cell 1 produces the same components as those contained in the culture medium through metabolism, the time-series data 512f of the culture medium components can become time-series data of the metabolites.

[0090] That is, the substance being measured is at least one of the nutrients taken up by cell 1, metabolites produced by the metabolism of cell 1, and substances present in cell 1. Furthermore, the measurement data 510 may also include information such as the pH and temperature of the culture medium within the culture vessel 10.

[0091] Reference Figure 4 and Figure 5 The measurement data 510 received by the estimation device 200 has been described. In addition, the time series data 512 included in the learning data 530 received by the learning device 300 can also contain multiple types of time series data, just like the measurement data 510 received by the estimation device 200.

[0092] For example, the time series data 512 included in the learning data 530 may also be data showing the time-varying changes of the nutrients taken up by cell 1, the metabolites produced by the metabolism of cell 1, and at least one substance in cell 1 as the measurement object.

[0093] More specifically, the time-series data 512 included in the learning data 530 includes time-series data on cell count, viability, cell morphology, and microscopic images obtained by cell measurement. Additionally, the time-series data 512 included in the learning data 530 includes time-series data on metabolites obtained by metabolite measurement. Furthermore, the time-series data 512 included in the learning data 530 includes time-series data on the composition of the culture medium obtained by nutrient measurement. Moreover, the time-series data 512 included in the learning data 530 can include information such as pH and temperature of the culture medium within the culture vessel 10.

[0094] Furthermore, the time series data 512 used in generating the prediction model 420 and the measurement data 510 input to the prediction model 420 when generating quality prediction data 540 using the prediction model 420 both contain measurement results obtained by using the same substance as the measurement object. That is, when time series data representing the time-varying changes in cell number is used in generating the prediction model 420, the measurement data 510 input to the prediction model 420 contains at least the measurement results representing the cell number.

[0095] [Structure of the prediction unit 220 of the estimation device 200]

[0096] Figure 6This is a block diagram showing an example of the structure of the prediction unit 220. The prediction unit 220 includes an estimation processing unit 222 and a determination unit 224. The estimation processing unit 222 inputs the measurement data 510 received by the receiving unit 210 into the prediction model 420 to obtain a prediction result. The prediction result is, for example, a predicted value 542.

[0097] Generally, the quality of active pharmaceutical ingredient 3 is evaluated from one or more viewpoints. Therefore, the prediction model 420 can also be configured to output a predicted value 542 as a prediction result when evaluated from a specific viewpoint. As an example, the predicted value 542 is an estimate of the result obtained when analyzing the active pharmaceutical ingredient 3 obtained from cells during culture from a specific viewpoint, or a similarity indicating how similar the obtained estimate is to the result of the target active pharmaceutical ingredient.

[0098] As a perspective for evaluating the quality of active pharmaceutical ingredient 3, there are, for example, physical properties, chemical properties, biological activity, immunochemical properties, etc. The prediction model 420 can also be configured to output multiple predicted values ​​obtained when the active pharmaceutical ingredient 3 is evaluated from multiple perspectives. For example, the prediction model 420 can also be configured to output a first predicted value obtained from the evaluation from a first perspective (e.g., physical properties) and a second predicted value obtained from the evaluation from a second perspective (e.g., biological activity).

[0099] Furthermore, the prediction model 420 may also be configured to include a first prediction model for outputting a first prediction value when evaluating the quality of the active pharmaceutical ingredient from a first perspective and a second prediction model for outputting a second prediction value when evaluating the quality of the active pharmaceutical ingredient from a second perspective.

[0100] The judgment unit 224 determines the quality of the active pharmaceutical ingredient 3 based on the predicted value 542 and generates a judgment result 544. The judgment result 544 can be a two-level evaluation result (e.g., good / poor), or a multi-level evaluation result (e.g., A, B, C, etc.), or a score-based evaluation result (e.g., 50 points, 95 points, etc.).

[0101] When the prediction model 420 is configured to output multiple predicted values ​​obtained from evaluating the active pharmaceutical ingredient 3 from multiple perspectives, the determination unit 224 determines the quality of the active pharmaceutical ingredient 3 based on the multiple predicted values ​​and generates a determination result 544. Furthermore, in this case, the determination unit 224 may also weight each of the multiple predicted values ​​according to the importance of the perspective.

[0102] The quality prediction data 540 includes the prediction value 542 generated by the estimation processing unit 222 and the judgment result 544 generated by the judgment unit 224. For example, if the prediction model 420 is configured to output multiple prediction values ​​obtained when evaluating the active pharmaceutical ingredient 3 from multiple perspectives, the quality prediction data 540 includes multiple prediction values ​​(first prediction value, second prediction value, etc.).

[0103] By including multiple predicted values ​​obtained from evaluating API 3 from multiple perspectives in the quality prediction data 540, the quality of API can be evaluated from a variety of perspectives.

[0104] Furthermore, since the judgment result 544 is included in the quality prediction data 540, the prediction results for the quality of the active pharmaceutical ingredient can be displayed more easily and understandably.

[0105] Output quality prediction data for output unit 230: 540. Figure 7 This is a diagram showing an example of the output result of the output unit 230.

[0106] like Figure 7 As shown, the output unit 230 can also be configured to output the judgment result 544 in accordance with the prediction value 542 obtained for each viewpoint.

[0107] Furthermore, the prediction unit 220 may not have a determination unit 224 that comprehensively determines the quality of the active pharmaceutical ingredient.

[0108] [Estimation Method]

[0109] Figure 8 This is a flowchart illustrating the estimation method involved in Embodiment 1. The estimation method includes a cell culture step (S100), a measurement step of measuring substances in the culture container (S200), a prediction step of predicting the quality of the active pharmaceutical ingredient 3 based on the measurement data 510 containing the measurement results (S300), and an output step of outputting the quality prediction data 540 obtained through the prediction step (S400).

[0110] The culture step (S100) includes the step of seeding cells 1 onto culture medium 2 (S110). Furthermore, depending on the culture conditions (a predetermined experimental protocol), the culture step (S100) may include steps such as changing culture medium 2 (S112) and adding reagents (S114). In addition, culture medium 2 can be either liquid or solid.

[0111] One or more measurement steps (S200) can be performed in the culture step (S100). For example, the measurement step (S200) can be performed at any time, such as when the step of seeding cells 1 into culture medium 2 (S110) is timed, when the step of changing culture medium 2 (S112) is timed, and when the step of adding reagents (S114) is timed.

[0112] That is, the timing of the measurement can also be the timing of the step of seeding cells 1 into culture medium 2 (S110), the timing of changing culture medium 2 (S112), and the timing of adding reagents (S114), etc. In this way, by aligning the timing of the operations occurring in the culture step (S100) with the timing of the measurement, contamination can be prevented. Furthermore, as referred to... Figure 4 As explained, the timing of the measurement can also be during the logarithmic growth phase, the stationary phase, or the death phase.

[0113] The measurement results are obtained by performing one or more measurement steps (S200) in the culture process (S100), thereby measuring the substances in the culture container 10 at least once after a predetermined period has elapsed since the cells 1 were inoculated into the culture medium 2.

[0114] Furthermore, the cultivation process (S100) and the measurement process (S200) can be carried out manually or automatically or semi-automatically using machinery.

[0115] The prediction process (S300) includes a process of receiving the input of measurement data 510 (S310), a process of inputting the measurement data 510 into the prediction model 420 (S312), a process of obtaining the prediction result from the prediction model 420 (S314), and a process of generating quality prediction data 540 (S316). That is, in the prediction process (S300), quality prediction data 540 is generated by inputting the measurement data 510, which includes the measurement result obtained in the measurement process (S200), into the prediction model 420. Furthermore, S310 to S316 and S400 are specifically performed by the estimation device 200.

[0116] [Learning Methods]

[0117] Figure 9 This is a flowchart illustrating the process of the learning method according to Embodiment 1. The learning method includes a cell culture step (S10), a measurement step of measuring substances in a culture container (S20), a purification step of obtaining a raw material drug (S30), an analysis step of analyzing the quality of the raw material drug (S40), and a learning step of performing learning processing using learning data 530 to generate a prediction model 420 (S50).

[0118] The culture step (S10) is the same as the culture step (S100) included in the estimation method. Specifically, the culture step (S10) includes the step of seeding cells 1 onto culture medium 2 (S11). Furthermore, depending on the culture conditions (a predetermined experimental protocol), the culture step (S10) may include steps such as replacing culture medium 2 (S12) and adding reagents (S13). In addition, culture medium 2 can be either liquid or solid.

[0119] Multiple measurement steps (S20) are performed in the culture process (S10). As a result, time series data 512 representing the time-dependent changes of the measured object is obtained. For example, the measurement steps (S20) are performed at arbitrary times, such as when the step of seeding cells 1 into culture medium 2 (S11), when the step of changing culture medium 2 (S12), and when the step of adding reagents (S13).

[0120] That is, the timing of the measurement can also be the timing of the step of seeding cells 1 into culture medium 2 (S11), the timing of changing culture medium 2 (S12), and the timing of adding reagents (S13), etc. In this way, by aligning the timing of the operations occurring in the culture step (S10) with the timing of the measurement, contamination can be prevented. Furthermore, as referred to... Figure 4 As explained, the timing of the measurement can also be during the logarithmic growth phase, the stationary phase, or the death phase.

[0121] By performing multiple measurement steps (S20) in the culture process (S10), measurement results are obtained by measuring substances in multiple timed culture containers 10 from the time the cells 1 are seeded into the culture medium 2.

[0122] Furthermore, the cultivation process (S10) and the measurement process (S20) can be carried out manually or automatically or semi-automatically using machinery.

[0123] The purification process (S30) is an operation that extracts only the target component from the culture medium within the culture vessel 10. Through the purification process (S30), a high-purity active pharmaceutical ingredient containing the target component can be obtained.

[0124] In the analytical step (S40), one or more analyses are performed to evaluate the active pharmaceutical ingredient from one or more perspectives, thereby obtaining quality data 520. When multiple analyses are performed, the quality data 520 can contain multiple analytical results (analytical results).

[0125] The learning process (S50) includes a process of accepting the input of learning data 530 (S51), a process of performing learning processing using the learning data 530 (S52), and a process of outputting the prediction model 420 generated by the learning processing (S53). Furthermore, in this embodiment 1, the output destination of the prediction model 420 is the server 400. Alternatively, the prediction model 420 may not be output but stored in the storage device 33 of the learning device 300. Specifically, S51 to S53 are performed by the learning device 300.

[0126] <Variation Example>

[0127] [Examples of learning devices]

[0128] In Embodiment 1 described above, the learning device 300 is configured to acquire learning data 530 from an external source. Alternatively, the learning device may also have the function of generating learning data 530. Furthermore, in Embodiment 1 described above, referring to... Figure 4 The measurement timing for obtaining the measurement data 510 received by the estimation device 200 is described. The proliferation curve of cell 1 varies depending on the type of cell 1. Therefore, the learning device may also have the function of determining the measurement timing for predicting the generation quality prediction data 540.

[0129] Figure 10 This is a block diagram illustrating the structure of the learning device involved in the modified example. Figure 10 The diagram shows a structure that the learning device 300 according to Embodiment 1 does not possess. The learning device 300a also includes a learning data generation unit 312 and a timing determination unit 330, which is different from the learning device 300.

[0130] The learning data generation unit 312 generates learning data 530 based on time series data 512 and quality data 520. Although not shown, the learning data generation unit 312 sends the generated learning data 530 to the receiving unit 310. The receiving unit 310 accepts the input of the sent learning data 530.

[0131] The timing determination unit 330 determines the timing of the measurement when predicting the generation quality prediction data 540 based on the time series data 512a of cell number. More specifically, the timing determination unit 330 determines the induction phase, logarithmic proliferation phase, stationary phase, and death phase (see reference) based on the time series data 512a. Figure 4 This generates information that can determine time intervals t1 to t3. This information, for example, includes the time from when cells 1 are seeded onto culture medium 2. Based on this information, the user can determine the measurement timing for prediction.

[0132] [Regarding training conditions]

[0133] In this embodiment described above, the culture conditions are assumed to be the same when generating the prediction model 420 and when using the generated prediction model 420 to generate the quality prediction data 540. However, the culture conditions may differ when generating the prediction model 420 and when using the generated prediction model 420 to generate the quality prediction data 540. (Refer to...) Figures 11-13 This will illustrate the learning and estimation methods in this case.

[0134] Figure 11 This is a diagram illustrating a modified example of the model generation unit. In this modified example, the model generation unit 320a generates a predictive model 420a based on multiple types of learning data 530a and 530b.

[0135] Learning data 530a and 530b each contain conditional data 550. Conditional data 550 represents information about culture conditions. For example, changing culture conditions to repeat culture, assay, purification, and analysis processes (see reference). Figure 9 This allows us to obtain learning data such as 530a and 530b generated under different training conditions.

[0136] Different culture conditions can include differences in the composition of culture medium 2, the types of reagents used, the types of chemicals used, the temperature conditions, and the timing of reagent addition.

[0137] Condition data 550 can include information such as the chemicals used, temperature conditions, and timing of reagent addition. Furthermore, condition data 550 may not be information about a specific culture method, but rather simply indicate that each set of learning data 530a and 530b was obtained under different culture conditions (e.g., condition A, condition B, etc.).

[0138] In addition to learning the correlation between the changes in substances within the culture container 10 over time and the quality of the active pharmaceutical ingredient 3, the model generation unit 320a also learns the correlation between different culture conditions and the changes in substances within the culture container 10 over time, or the correlation between different culture conditions and the quality of the active pharmaceutical ingredient 3, to generate a predictive model 420a.

[0139] Figure 12 This is a block diagram illustrating the structure of the estimation device involved in the modified example. Figure 12 In this text, the description of the structure of the output unit according to Embodiment 1 above is omitted. The estimation device 200a is different from the estimation device 200 according to Embodiment 1 above by having a receiving unit 210a instead of a receiving unit 210 and a prediction unit 220a instead of a prediction unit 220.

[0140] In addition to accepting the input of measurement data 510, the receiving unit 210a also accepts the input of condition data 550 indicating culture conditions. The prediction unit 220a also inputs the condition data 550 in addition to the measurement data 510. Figure 13 The model generation unit 320a generates a prediction model 420a, thereby obtaining a prediction result representing the quality of the active pharmaceutical ingredient.

[0141] Prediction model 420a, because it considers the correlation between different culture conditions and the time-dependent changes of substances within the culture container 10, or the correlation between different culture conditions and the quality of the active pharmaceutical ingredient 3, can predict the quality of the active pharmaceutical ingredient 3 more accurately than prediction model 420. Even when the culture conditions for generating prediction model 420a differ from those for generating quality prediction data 540 using the generated prediction model 420a, because prediction model 420a considers the different culture conditions, prediction unit 220a can obtain a prediction result representing the quality of the active pharmaceutical ingredient by inputting condition data 550 in addition to the measurement data 510 into prediction model 420a. In other words, prediction unit 220a can obtain a prediction result that considers the correlation between different culture conditions and quality.

[0142] Figure 13 This is a diagram illustrating variations of the prediction model. For example, such as... Figure 13 As shown, the model generation unit can also generate prediction models for each culture condition, such as prediction model 420-1 for culture condition A and prediction model 420-2 for culture condition B.

[0143] In this case, the estimation device can also use a prediction model that matches the culture conditions based on the condition data 550.

[0144] Implementation method 2.

[0145] [Overall structure of the prediction system]

[0146] Figure 14 This diagram schematically illustrates the overall structure of the prediction system according to Embodiment 2. Like Prediction System SYS1, Prediction System SYS2 has the function of predicting the quality of the active pharmaceutical ingredient (API) of a biopharmaceutical manufactured through cell culture. In addition to the functions of Prediction System SYS1, Prediction System SYS2 also has the function of optimizing culture conditions. Prediction System SYS2 includes an optimization device 600 for optimizing culture conditions.

[0147] The optimization device 600 estimates optimal culture conditions to obtain high-quality active pharmaceutical ingredient 3 based on the time-series data 512 and quality data 520 contained in the learning data 530. The optimization device 600 outputs the estimated optimal culture conditions. The user performs a new culture experiment based on the culture conditions output from the optimization device 600. The culture experiment includes a culture step (step S10), a purification step (step S30), and an analysis step (step S40). These steps are the same as those described in Embodiment 1. Details of each step are provided in [the original text]. Figure 9 As shown in the diagram, its description will not be repeated here.

[0148] When a new culture experiment is conducted, new learning data 530 is input into the optimization device 600. The optimization device 600 uses the previously input learning data 530 and the newly input learning data 530 to re-estimate the optimal culture conditions. The optimization device 600 outputs the newly estimated culture conditions. The user conducts the next culture experiment based on the culture conditions output from the optimization device 600.

[0149] Therefore, each increase in the number of culture experiments leads to further optimization of the culture conditions. By optimizing the culture conditions, the properties of the learning data 530 are observed to change in the direction of producing high-quality active pharmaceutical ingredient 3. In other words, by optimizing the culture conditions, the quality of the learning data 530 is improved. This improvement in the quality of the learning data 530, in turn, improves the accuracy of the prediction model 420.

[0150] The optimization device 600 outputs optimized culture conditions based on user instructions. When using the estimation device 200 to predict the quality of the active pharmaceutical ingredient (API), the user can culture cells based on the culture conditions output from the optimization device 600. Therefore, the user can predict the quality of the API in an environment capable of producing high-quality APIs.

[0151] Therefore, the optimization device 600 functions very effectively in both the process of generating the prediction model 420 and the process of using the prediction model 420 to predict the quality of the active pharmaceutical ingredient. According to the prediction system SYS2 of Embodiment 2, the possibility of obtaining higher quality active pharmaceutical ingredients is significantly increased.

[0152] [Optimized device hardware structure]

[0153] Figure 15 This is a block diagram showing the hardware structure of the optimization device 600. The optimization device 600 includes a processor 61, a main memory 62, a storage device 63, a communication I / F 64, a USB I / F 65, an input unit 66, and a display unit 67. These components are connected via a processor bus 68.

[0154] The processor 61, consisting of a CPU, GPU, etc., reads the program (e.g., optimization program 630) stored in the storage device 63 and executes it in the main memory 62. By executing the program, the processor 61 performs a series of processes to search for the optimal values ​​of the parameters for the cultivation conditions.

[0155] The main memory 62 is composed of volatile storage devices such as RAM and DRAM. The storage device 63 is composed of non-volatile storage devices such as HDD and SSD.

[0156] The storage device 63 stores learning (estimation) data 530 transmitted via USB I / F 65, an optimization program 630 for searching for optimal values ​​620 of the parameters of the culture conditions, and optimal values ​​620 of the parameters of the culture conditions newly determined by the processor 61.

[0157] The Communication I / F 64 uses wired or wireless communication to exchange signals with other communication devices.

[0158] A USB memory device (not shown) is removably installed in the USB I / F 65, and the USB I / F 65 reads the learning data 530 stored in the USB memory.

[0159] The input unit 66 handles user operations and typically consists of a touch panel, keyboard, mouse, etc. The display unit 67 consists of a liquid crystal panel or similar device capable of displaying images.

[0160] [Structure of Optimized Device 600]

[0161] Figure 16 This is a block diagram illustrating an example of the structure of the optimization device 600. As shown by the processor 61 (refer to...) Figure 15 ) Execute optimization procedure 630 (refer to) Figure 15 As an example of the software structure implemented, the optimization device 600 includes a receiving unit 601, an estimation unit 602, an output unit 603, and a storage unit 604.

[0162] The optimization device 600 utilizes Bayesian optimization to search for culture conditions (hereinafter also referred to as the optimal solution) that improve the quality of the active pharmaceutical ingredient 3. Here, culture conditions are a combination of values ​​for multiple parameters. Bayesian optimization is a method that estimates the optimal event from observed events using statistical methods based on Bayesian probability. In Bayesian optimization, trial experiments are performed based on the set optimal solution, and other optimal solutions are searched based on the results of these trial experiments. The optimal solution found is set as the optimal solution to be used in the next trial experiment.

[0163] The reception department 601 receives estimated data. The estimated data includes... Figure 14 The learning data 530 is shown. The estimation data includes at least a portion of the time series data 512 and the quality data 520. The estimation unit 602 uses Bayesian optimization to search for the optimal solution of the culture conditions by executing the estimation program. The storage unit 604 includes multiple memories for storing the estimation program, the input estimation data, and the optimal solution found. The output unit 603 is, for example, a display unit 67 that displays the optimal solution of the culture conditions. Figure 15 The output unit 603 can also be an interface for outputting the optimal solution of the culture conditions to a display or printer.

[0164] As shown in window W10, an estimated data set consists of data from the culture conditions used in a single culture experiment and quality data. The culture condition data comprises multiple parameters p1, p2, ..., pn that define the culture conditions. These parameters can be, for example, any of the following: nutrient concentration, pH, temperature, humidity, culture medium composition, type of reagent added, and timing of reagent addition.

[0165] Various parameters, such as the timing of cell seeding into the culture medium and the conditions of the culture medium extraction, are used. This timing corresponds to... Figure 4 The measurement time t0 is shown. That is, the values ​​of various parameters are... Figure 14 The time series data 512 shown contains the values ​​of parameters measured at the measurement time t0. The receiving unit 601 can also input the values ​​of... Figure 4 The values ​​of various parameters measured at various time intervals are shown. For example, values ​​representing changes in nutrient concentration over time, changes in metabolite concentration over time, changes in pH of the culture medium in the culture vessel, changes in temperature, changes in humidity, etc., can also be input into the receiving unit 601 as estimation data. The estimation unit 602 can also search for optimal culture conditions based on these estimation data.

[0166] [Methods for optimizing culture conditions]

[0167] Figure 17 This is a flowchart illustrating the processing flow of the optimization device 600. Below, refer to... Figure 16 and Figure 17 This will illustrate the process of optimizing the 600 device to estimate the optimal culture conditions.

[0168] First, the estimation unit 602 determines whether estimation data has been input to the receiving unit 601 (step S600). If estimation data has been input to the receiving unit 601, the estimation unit 602 saves the input estimation data in the storage unit 604 (step S601). By repeatedly executing step S601, the estimation data input to the receiving unit 601 is accumulated in the storage unit 604. Next, the estimation unit 602 uses all the estimation data stored in the storage unit 604 to determine the values ​​of the parameters p1, p2, ..., pn of the optimal culture conditions (step S602). In step S602, the estimation unit 602 uses Bayesian optimization to search for culture conditions that provide optimal values ​​for the quality level of the active pharmaceutical ingredient 3.

[0169] Regarding the search for culture conditions, it is desirable to perform the search periodically when multiple estimated data corresponding to multiple culture experiments are input, rather than performing the search every time an estimated data point is input. For example, consider performing the processing in step S602 whenever 4 to 10 estimated data points corresponding to 4 to 10 culture experiments are input.

[0170] Next, the estimation unit 602 stores the optimal value of the newly determined parameter in the storage unit 604. Since the optimal value of the parameter determined through the previous search is stored in the storage unit 604, the estimation unit 602 updates these values ​​with the optimal value of the newly determined parameter (step S603). Next, the estimation unit 602 outputs the value of the newly determined parameter from the output unit 603 (step S604).

[0171] The parameter values ​​output from the output unit 603 represent the latest parameter values ​​indicating optimal culture conditions. The user performs one or more new culture experiments based on the output parameter values. When estimation data corresponding to a new culture experiment is input to the optimization device 600, the estimation unit 602 searches again for optimal culture condition parameter values. The search result is stored in the storage unit 604 as the optimal parameter values ​​for the current time point. Therefore, by repeatedly performing culture experiments and optimizing the culture conditions using the optimization device 600, the parameter values ​​for the culture conditions stored in the optimization device 600 are optimized.

[0172] If the estimation unit 602 determines "no" in step S600, it determines whether there is a request for an output parameter value (step S605). For example, if a user wants to predict the quality of a drug substance obtained through a certain culture experiment using the estimation device 200, it is necessary to determine the culture conditions for the culture experiment. When determining the culture conditions for the culture experiment, it is useful to use the parameter values ​​stored in the optimization device 600 as a reference. If the estimation unit 602 determines that there is a request for an output parameter value in step S600, it outputs the parameter values ​​stored in the storage unit 604 from the output unit 603 (step S604) and ends the processing based on this flowchart.

[0173] [Variations on the prediction system]

[0174] Figure 18 This is a diagram illustrating a modified example of the prediction system according to Embodiment 2. In the modified example, the learning device 300 includes a... Figure 16 The optimization unit 6000, which is part of the optimization device 600, is shown as an optimization unit 6000. Like the optimization device 600, the optimization unit 6000 includes a receiving unit 601, an estimation unit 602, an output unit 603, and a storage unit 604. The learning device 300 optimizes the prediction model 420 based on the learning data 530 and searches for the optimal values ​​of the parameters of the cultivation conditions. Therefore, according to the modified example, the added value of the learning device 300 can be increased. Furthermore, according to the modified example, since the optimization device 600 does not need to be provided separately from the learning device 300, cost reduction can be achieved.

[0175] In addition, the optimization device 600 or the optimization unit 6000 can also use a grid search method instead of Bayesian optimization to optimize the values ​​of the parameters of the culture conditions.

[0176] [Way]

[0177] Those skilled in the art will understand that the above-described embodiments are specific examples of the following methods.

[0178] (Item 1) An estimation apparatus according to one method comprises: a receiving unit that receives input measurement data; a prediction unit that generates quality prediction data representing the predicted quality of a drug substance by inputting the measurement data received by the receiving unit into a prediction model; and an output unit that outputs the quality prediction data generated by the prediction unit. The measurement data includes measurement results obtained from at least one timed measurement of at least one substance in a culture vessel containing the cells and the culture medium after a predetermined period has elapsed since the cells were inoculated into the culture medium. The prediction model is a model used to predict the quality of a drug substance of a biopharmaceutical manufactured by culturing cells.

[0179] The estimation device described in item 1 can predict the quality of the active pharmaceutical ingredient (API) obtained from the culture at a stage midway through the culture process. Therefore, based on the predicted API quality, users can take various actions such as modifying the experimental protocol, extending the culture time, or terminating the culture at the stage of the culture process before the purification process to obtain the API, thereby reducing the manufacturing cost of biopharmaceuticals.

[0180] (Item 2) In the estimation apparatus described in Item 1, the measurement data includes the measurement results obtained by measuring at least one substance at multiple time intervals from the time the cells are seeded into the culture medium.

[0181] (Item 3) In the estimation apparatus described in Item 1 or Item 2, the measurement data includes the measurement results obtained by measuring the at least one substance at the beginning of cell culture.

[0182] (Item 4) In the estimation apparatus described in any of Items 1 to 3, the measurement data includes the measurement results obtained by measuring at least one substance during the logarithmic proliferation phase of cells.

[0183] During the logarithmic growth phase, the number of cells changes dramatically, and therefore the state of the substances within the culture vessel is expected to change dramatically. The results obtained by measuring the sample at key points where large state changes occur are included in the measurement data; therefore, the estimation device described in section 4 can use data that more accurately represents changes in cell state to predict the quality of the active pharmaceutical ingredient.

[0184] (Item 5) In the estimation apparatus described in any of items 1 to 4, the measurement data includes the measurement results obtained by measuring at least one substance during the stationary phase of the cell.

[0185] During the stationary phase, a balance is maintained between the number of proliferating cells and the number of dead cells, with the cell count reaching its maximum. The results obtained by measuring the test subject at the time when the cell count is at its maximum are included in the measurement data. Therefore, the estimation device described in item 5 can use data representing the state of the substance within the culture vessel at the time when the cell count is at its maximum to predict the quality of the active pharmaceutical ingredient.

[0186] (Item 6) In the estimation apparatus described in any of items 1 to 4, the measurement data includes the measurement results obtained by measuring the at least one substance during the cell death period. Figure 5 ).

[0187] During cell death, intracellular enzymes and other substances are exposed to the culture medium, potentially affecting the quality of the active pharmaceutical ingredient. In this case, the measurement results obtained during cell death can potentially be effectively utilized in machine learning as examples of defective quality data.

[0188] (Item 7) In the estimating apparatus described in any one of items 1 to 6, at least one substance is a nutrient taken up by the cell, a metabolite produced by the cell’s metabolism, and at least one of the following in the cell.

[0189] (Item 8) In the estimation apparatus described in any one of items 1 to 7, by inputting measurement data received by the receiving unit, a first predicted value is output from the prediction model when the quality of the active pharmaceutical ingredient is evaluated from a first perspective, and a second predicted value when the quality of the active pharmaceutical ingredient is evaluated from a second perspective. The quality prediction data includes the first predicted value and the second predicted value.

[0190] The estimation device described in item 8 is capable of evaluating the quality of the active pharmaceutical ingredient from a variety of perspectives.

[0191] (Item 9) In the estimation apparatus described in any of items 1 to 7, a predicted value is output from the prediction model when the quality of the active pharmaceutical ingredient is evaluated from a prescribed perspective, by inputting measurement data received by the receiving unit. The prediction unit includes a determination unit that determines the quality of the active pharmaceutical ingredient based on the predicted value. The quality prediction data includes the determination result obtained from determining the quality of the active pharmaceutical ingredient.

[0192] The estimation device described in item 9 can more easily display the predicted results for the quality of the active pharmaceutical ingredient.

[0193] (Item 10) In the estimation device described in any of items 1 to 8, the receiving unit also receives input of condition data representing the cell culture conditions. The prediction unit generates quality prediction data by inputting the measurement data and condition data received by the receiving unit into the prediction model.

[0194] The estimation apparatus described in item 10 can obtain prediction results that take into account the correlation between different culture conditions and quality, thereby enabling more accurate prediction results.

[0195] (Item 11) Additionally, one method of the learning apparatus includes: a receiving unit that receives learning data; and a model generation unit that generates a predictive model by performing learning processing using the learning data received by the receiving unit. The learning data includes measurement data and quality data. The measurement data includes measurement results obtained from multiple timed measurements of at least one substance in a culture vessel containing the cells and the culture medium, starting from when cells are seeded into the culture medium. The quality data is data obtained by analyzing the active pharmaceutical ingredient (API) of a biopharmaceutical manufactured using the cells. The predictive model is a model for generating quality prediction data representing the quality of the API of a biopharmaceutical manufactured using the cells contained in the culture vessel during the intermediate stage of cell culture, based on the measurement results obtained from measuring at least one substance in the culture vessel during the intermediate stage of cell culture.

[0196] The learning device described in item 11 can generate a predictive model for predicting the quality of the active pharmaceutical ingredient (API) obtained from the culture at a stage midway through the culture process. Therefore, based on the predicted API quality, users can take various actions at the stage of the culture process before the purification process to obtain the API, such as modifying the experimental protocol, extending the culture time, or terminating the culture, thereby reducing the manufacturing cost of biopharmaceuticals.

[0197] (Item 12) In addition, in one of the learning devices, the learning data further includes condition data representing the cell culture conditions. The condition data includes the values ​​of multiple parameters for defining the culture conditions. The learning device also includes an optimization unit that optimizes the combination of the values ​​of the multiple parameters. The optimization unit takes the values ​​of the multiple parameters and quality data obtained when the cells are cultured based on the values ​​of the multiple parameters as input, and estimates the optimal combination of the values ​​of the multiple parameters.

[0198] The learning device described in item 12 is capable of estimating a combination of values ​​for multiple parameters defining the culture conditions of cells at a stage midway through the culture process. Therefore, the user can optimize the conditions to improve the quality of the active pharmaceutical ingredient obtained from the culture at a stage midway through the process.

[0199] (Item 13) In the learning apparatus described in Item 11, the learning data further includes condition data representing the conditions for culturing cells contained in a culture container. The learning data received by the receiving unit includes first learning data obtained when cells are cultured under first conditions and second learning data obtained when cells are cultured under second conditions different from the first conditions. The model generation unit performs learning processing using the first learning data and the second learning data to generate a predictive model.

[0200] The learning device described in item 13 can generate predictive models that take into account the correlation between different culture conditions and the changes in substances within the culture container over time, or the correlation between different culture conditions and the quality of the active pharmaceutical ingredient. Therefore, the learning device can generate predictive models for obtaining more accurate prediction results.

[0201] (Item 14) Additionally, one method of estimation includes the following steps: culturing cells by placing cells and culture medium in a culture vessel; a measurement step, measuring at least one substance in the culture vessel at at least one time interval after a predetermined period has elapsed since the cells were seeded into the culture medium; generating quality prediction data representing the quality of the active pharmaceutical ingredient by inputting the measurement data, including the measurement results obtained in the measurement step, into a prediction model; and outputting the quality prediction data. The prediction model is a model used to predict the quality of the active pharmaceutical ingredient of a biopharmaceutical manufactured by culturing cells.

[0202] The estimation method described in item 14 enables the prediction of the quality of the active pharmaceutical ingredient (API) obtained from the culture at a stage midway through the culture process. Therefore, users can adjust experimental protocols, extend culture time, or terminate the culture at stages of the culture process prior to the purification of the API based on the predicted API quality, thereby reducing the manufacturing cost of biopharmaceuticals.

[0203] (Item 15) Additionally, one method of learning includes the following steps: culturing cells by placing cells and culture medium in a culture vessel; a measurement step, measuring at least one substance in the culture vessel at multiple time intervals from when the cells are seeded into the culture medium; an analysis step, analyzing the quality of the active pharmaceutical ingredient (API) of a biopharmaceutical manufactured from the cells; and generating a predictive model by performing learning processing using learning data. The learning data includes measurement data and quality data, the measurement data including the measurement results obtained through the measurement step, and the quality data being the data obtained through the analysis step. The predictive model is a model for generating quality prediction data representing the quality of the API of a biopharmaceutical manufactured from cells contained in the culture vessel during the intermediate stage of cell culture, based on the measurement results obtained from measuring at least one substance in the culture vessel during the intermediate stage of cell culture.

[0204] In the learning method described in item 15, a predictive model is generated to predict the quality of the active pharmaceutical ingredient (API) obtained from the culture at a stage midway through the culture process. Therefore, users can take various actions based on the predicted API quality, such as modifying the experimental protocol, extending the culture time, or terminating the culture at a stage in the culture process before the purification process to obtain the API, thereby reducing the manufacturing cost of biopharmaceuticals.

[0205] (Item 16) In addition, one method of the optimization apparatus includes: a receiving unit that receives values ​​of multiple parameters for defining culture conditions when cells are seeded into a culture medium and quality data obtained by analyzing the active pharmaceutical ingredient of a biopharmaceutical manufactured by culturing cells; an estimation unit that takes the values ​​of the multiple parameters received by the receiving unit and the quality data as inputs to estimate an optimal combination of the values ​​of the multiple parameters; and an output unit that outputs the combination of the values ​​of the multiple parameters estimated by the estimation unit.

[0206] In the optimization apparatus described in item 16, a combination of values ​​for multiple parameters defining cell culture conditions can be estimated at a stage midway through the culture process. Therefore, the user can optimize conditions to improve the quality of the active pharmaceutical ingredient obtained from the culture at a stage midway through the process. Furthermore, the estimation unit can perform Bayesian optimization using the values ​​of the multiple parameters and quality data obtained through an analysis step as input, thereby estimating the optimal combination of the values ​​of the multiple parameters.

[0207] (Item 17) Additionally, one method of optimization includes the following steps: a cell culture step, in which cells and culture medium are placed in a culture vessel and the cells are cultured under culture conditions defined by values ​​of multiple parameters; an analysis step, in which the quality of the active pharmaceutical ingredient of the biopharmaceutical manufactured by culturing the cells is analyzed; and an estimation step, in which the optimal combination of values ​​of the multiple parameters is estimated, using the values ​​of the multiple parameters and the quality data obtained through the analysis step as input. In the cell culture step, the optimal combination of values ​​of the multiple parameters estimated through the estimation step is used as the new culture conditions for culturing the cells.

[0208] In the estimation method described in item 17, the user is able to optimize the conditions for improving the quality of the active pharmaceutical ingredient obtained from the culture at a stage midway through the culture process.

[0209] (Item 18) In addition, the prediction system according to one method includes an estimation device as described in any one of items 1 to 10, a learning device as described in any one of items 11 to 13, and a measuring device for measuring at least one substance in a culture container.

[0210] Furthermore, in the prediction system described in item 18, the estimation device and the measuring device can also be implemented using an information processing device. In this case, the receiving unit of the estimation device and the receiving unit of the measuring device can also be implemented as a single receiving unit. That is, the receiving unit receives the measurement data obtained by measuring the substance through the measuring device.

[0211] (Item 19) In the prediction system described in Item 18, the measuring apparatus includes a first measuring apparatus for measuring a first substance and a second measuring apparatus for measuring a second substance.

[0212] The various embodiments disclosed herein are also intended to be appropriately combined and implemented to a extent that they are not technically contradictory. Furthermore, the embodiments disclosed herein should be considered illustrative rather than restrictive in all respects. The scope of the invention is not indicated by the description of the above embodiments, but by the claims, and is intended to include all modifications within the meaning and scope equivalent to the claims.

[0213] Explanation of reference numerals in the attached figures

[0214] 1: Cell; 2: Culture medium; 3: Active pharmaceutical ingredient; 10: Culture vessel; 21, 31, 61: Processor; 22, 32, 62: Main memory; 23, 33, 63: Storage device; 24, 34, 64: Communication I / F; 25, 35, 65: USB I / F; 26, 36, 66: Input unit; 27, 37, 67: Display unit; 28, 38, 68: Processor bus; 100: Measurement device; 200, 200a: Estimation device; 210, 210a, 310: Receiving unit; 220, 220a: Prediction unit; 222: Estimation processing unit; 224: Decision unit; 230: Output unit; 300, 300a: Learning device; 312: Learning data generation unit; 320, 320a: Model generation unit; 330: Timing determination unit; 400: Server; 420, 420a: Prediction model 510: Measurement data; 512: Time series data; 520: Quality data; 530, 530a, 530b: Learning data; 540: Quality prediction data; 542: Predicted value; 544: Judgment result; 550: Conditional data; 560: Estimation program; 580: Learning program; 600: Optimization device; 601: Receiving unit; 602: Estimation unit; 603: Output unit; 604: Storage unit; 620: Optimal value of parameter; 630: Optimization program; 6000: Optimization unit; SYS1, SYS2: Prediction system; W10: Window.

Claims

1. An estimation device comprising: The receiving department accepts input of measurement data, which includes the measurement results obtained from at least one timed measurement of at least one substance in a culture vessel containing the cells and the culture medium after a specified period has elapsed since the cells were seeded into the culture medium. The prediction unit generates quality prediction data representing the predicted quality of the active pharmaceutical ingredient (API) by inputting the measurement data received by the receiving unit into a prediction model for predicting the quality of the API manufactured by culturing the cells; and The output unit outputs the quality prediction data generated by the prediction unit. in, The measurement data includes the results obtained from measuring at least one substance during the logarithmic proliferation phase of the cells. The measurement data includes the results obtained by measuring the at least one substance in the culture vessel at multiple time intervals after the cells are seeded into the culture medium.

2. The estimation device according to claim 1, wherein, The measurement data includes the results obtained from measuring the at least one substance at the beginning of cell culture.

3. The estimation device according to claim 1, wherein, The measurement data includes the results obtained from measuring at least one substance during the stationary phase of the cells.

4. The estimation apparatus according to claim 1, wherein, The measurement data includes the measurement results obtained by measuring the at least one substance during the cell's death period.

5. The estimation apparatus according to claim 1, wherein, The at least one substance is a nutrient taken up by the cell, a metabolite produced by the cell's metabolism, or at least one of the substances in the cell.

6. The estimation apparatus according to claim 1, wherein, By inputting the measurement data received by the receiving department, the prediction model outputs a first predicted value when the quality of the active pharmaceutical ingredient is evaluated from a first perspective and a second predicted value when the quality of the active pharmaceutical ingredient is evaluated from a second perspective. The quality prediction data includes the first prediction value and the second prediction value.

7. The estimation apparatus according to claim 1, wherein, By inputting the measurement data received by the receiving department, the prediction model outputs a predicted value that evaluates the quality of the active pharmaceutical ingredient from a specified perspective. The prediction unit includes a determination unit that determines the quality of the active pharmaceutical ingredient based on the predicted value. The quality prediction data includes the judgment results obtained from determining the quality of the active pharmaceutical ingredient.

8. The estimation apparatus according to claim 1, wherein, The receiving department also accepts input of condition data indicating the culture conditions of the cells. The prediction unit generates the quality prediction data by inputting the measurement data and the condition data received by the acceptance unit into the prediction model.

9. The estimation apparatus according to claim 1, wherein, The plurality of timings includes a plurality of timings during the logarithmic proliferation phase, a plurality of timings during the stationary phase of the cell, and a plurality of timings during the cell death phase.

Images