Quality monitoring device, method for operating quality monitoring device, and program for operating quality monitoring device

Abstract

A processor of the quality monitoring device is capable of executing a monitoring process for monitoring quality using a plurality of state prediction models for predicting the state of the liquid related to the quality using a spectral spectrum acquired online from the liquid generated in the manufacturing process as input data, and a model switching process for switching the state of the liquid related to the quality using a plurality of state prediction models for predicting the state of the liquid related to the quality as input data. According to process management information for managing at least one process included in the manufacturing process, the state prediction models are switched midway in the monitoring process, and the plurality of state prediction models are learned before the monitoring process starts. The process management information includes at least one of process identification information, use equipment information, and process quality information.

Description

Technical Field

[0001] The present invention relates to a quality monitoring device, a method for operating the quality monitoring device, and a procedure for operating the quality monitoring device. Background Technology

[0002] For example, there are known manufacturing processes for biopharmaceuticals with biomolecules such as monoclonal antibodies and proteins as active ingredients. In such processes, a suspension is often produced, in which various components, represented by the active ingredient, are dispersed in a liquid. To successfully guide the ongoing manufacturing process, it is important to monitor the state of the target components (e.g., proteins or protein-derived dopants) in this suspension within the production line.

[0003] Japanese Patent Application Publication No. 2022-512775 discloses a technique for predicting the concentration of a target component's state within a manufacturing process. Specifically, in Japanese Patent Application Publication No. 2022-512775, the Raman spectrum of a suspension is measured within a production line, and a state prediction model is used to predict the concentration of the target component based on the Raman spectrum. Summary of the Invention

[0004] The technical problem to be solved by the invention

[0005] To manage the quality of manufacturing processes, the prediction accuracy of state prediction models is crucial. In manufacturing processes, the state of liquids such as suspensions can change over time; if the state changes significantly, the prediction accuracy of the state prediction model may decrease. The technology described in Japanese Patent Application Publication No. 2022-512775 collects learning data from the state prediction model during the operation of the manufacturing process and updates the state prediction model with the collected learning data, thereby suppressing the decline in prediction accuracy.

[0006] However, the technology described in Japanese Patent Application Publication No. 2022-512775, which involves collecting learning data and updating the state prediction model during the operation of a manufacturing process, cannot cope with a decrease in prediction accuracy and results in inadequate quality management, for example, in situations where the manufacturing process has multiple steps and the state changes significantly between steps. If the time interval between steps is insufficient, the method in Japanese Patent Application Publication No. 2022-512775 may not be able to update the state prediction model in a timely manner. In this case, inadequate quality management occurs.

[0007] One embodiment of the present invention provides a quality monitoring device, a method for operating the quality monitoring device, and a procedure for operating the quality monitoring device that can appropriately manage the quality of the manufacturing process of biopharmaceuticals compared with the prior art.

[0008] means for solving technical problems

[0009] The quality monitoring device of the present invention includes a processor capable of performing monitoring processing and model switching processing. In the monitoring processing, the spectroscopic spectrum of a liquid obtained online from a manufacturing process of a biopharmaceutical is used as input data, and multiple state prediction models that predict the state of the liquid related to the quality of the manufacturing process are used to monitor the quality. In the model switching processing, the state prediction model is switched midway through the monitoring processing based on process management information for managing at least one step included in the manufacturing process. Moreover, the multiple state prediction models have been learned before the monitoring processing begins. The process management information includes at least one of process identification information for identifying at least one step included in the manufacturing process, equipment information for the equipment used in the step, and process quality information related to the quality of each step that can be obtained when at least one step has been performed.

[0010] The preferred model switching process includes switching the state prediction model based on process identification information or using equipment before a step in the manufacturing process begins.

[0011] The preferred model switching process includes switching the state prediction model based on quality information that is acquired midway through the process, which is part of the process quality information.

[0012] Preferably, when the manufacturing process includes a purification process of the target component contained in the liquid, the state of the liquid is the concentration of the target component contained in the liquid, and the quality information obtained during the process includes information obtained online during the process and related to the concentration of the target component in the liquid.

[0013] The preferred target component is an antibody.

[0014] Preferred spectroscopic spectra include any one of infrared spectroscopy, fluorescence spectroscopy, and Raman spectroscopy.

[0015] The preferred processor outputs a predicted value of the liquid's state at a pre-set first time interval based on the spectroscopic spectrum.

[0016] The first time interval is preferably less than 5 seconds.

[0017] Preferably, the first time interval is less than 1 second.

[0018] The preferred processor acquires the spectroscopic spectrum at a second time interval below the first time interval, and outputs the predicted value based on the moving average of the acquired spectroscopic spectra.

[0019] The preferred processor can output the final quality of the process, representing the final state of the overall liquid produced in the process, based on the predicted values.

[0020] The preferred processor performs a condition determination process to determine the process conditions for the next process based on the final quality of the previous process.

[0021] Preferably, when the manufacturing process includes a purification process for the target component contained in the liquid, the state of the liquid is the concentration of the target component contained in the liquid.

[0022] The preferred target component is an antibody.

[0023] The preferred state of the liquid is the state of the proteins contained within it.

[0024] The working method of the quality monitoring device of the present invention is capable of performing monitoring processing and model switching processing. In the monitoring processing, the spectroscopic spectrum of the liquid obtained online from the manufacturing process of biopharmaceuticals is used as input data, and multiple state prediction models that predict the state of the liquid related to the quality of the manufacturing process are used to monitor the quality. In the model switching processing, the state prediction model is switched midway through the monitoring processing according to process management information for managing at least one step included in the manufacturing process. Moreover, the multiple state prediction models have been learned before the monitoring processing begins. The process management information includes at least one of the following: process identification information for identifying at least one step included in the manufacturing process, equipment information for the equipment used in the step, and process quality information related to the state of the liquid in each step that can be obtained when at least one step has been performed.

[0025] The operating procedure of the quality monitoring device of the present invention enables a computer to function as a quality monitoring device. This operating procedure includes monitoring processing and model switching processing. In the monitoring processing, the spectroscopic spectrum of liquids obtained online from the manufacturing process of biopharmaceuticals is used as input data. Multiple state prediction models that predict the state of the liquids related to the quality of the manufacturing process are used to monitor quality. In the model switching processing, the state prediction models are switched midway through the monitoring process based on process management information for managing at least one step included in the manufacturing process. Furthermore, the multiple state prediction models are created before the monitoring process begins. The process management information includes at least one of the following: process identification information for identifying at least one step included in the manufacturing process, equipment information for the equipment used in the step, and process quality information related to the state of the liquid in each step that can be obtained when at least one step has been performed.

[0026] Invention Effects

[0027] According to the technology of the present invention, the quality of the manufacturing process of biopharmaceuticals can be appropriately monitored compared with the prior art. Attached Figure Description

[0028] Figure 1 It is a diagram that shows an overview of the manufacturing process of biopharmaceuticals.

[0029] Figure 2 This is a block diagram of the computer that constitutes the quality monitoring device.

[0030] Figure 3 This is a block diagram of the processor that constitutes the quality monitoring device.

[0031] Figure 4 It is a diagram representing process management information.

[0032] Figure 5 It is a graph representing spectral measurement data.

[0033] Figure 6 This is a diagram representing the neural network that constitutes the concentration prediction model.

[0034] Figure 7 This is a graph representing the concentration prediction model and the learning data.

[0035] Figure 8 It is a graph representing the situation where predictions are made at a specified period.

[0036] Figure 9 This is a graph showing a monitoring screen displaying the predicted antibody concentration in a time series.

[0037] Figure 10 This is a flowchart representing the monitoring and processing steps.

[0038] Figure 11 This is a diagram illustrating the method for obtaining the final antibody concentration.

[0039] Figure 12 This is a graph representing multiple items predicted by the concentration prediction model.

[0040] Figure 13 This is another example of a surveillance video.

[0041] Figure 14 This is a diagram showing process management information for a variation of the first embodiment.

[0042] Figure 15 This is a graph showing the concentration prediction model and learning data of a variation of the first embodiment.

[0043] Figure 16 This is a diagram illustrating the processing steps of a modified example of the first embodiment.

[0044] Figure 17 This is a diagram showing the process management information of the second embodiment.

[0045] Figure 18This is a diagram illustrating the processing steps of the second embodiment.

[0046] Figure 19 This is a graph showing the change in absorbance over time and the time of model switching.

[0047] Figure 20 This is a diagram showing process management information for a variation of the second embodiment.

[0048] Figure 21 This is a diagram illustrating the processing steps of a modified example of the second embodiment.

[0049] Figure 22 This is a graph showing the changes in antibody concentration over time and the timing of model switching.

[0050] Figure 23 This is a graph illustrating the method for determining the final antibody concentration based on predicted values.

[0051] Figure 24 It is a graph representing the moving average of spectral measurement data. Detailed Implementation

[0052] [First Implementation]

[0053] As an example, such as Figure 1 As shown, the manufacturing process 2 of the biopharmaceutical product using the quality monitoring device 41 applicable to the technology of this invention is roughly divided into a first process 10, a second process 11, and a third process 12. The first process 10 involves encoding the antibody gene 14 into host cells 13, such as Chinese hamster ovary cells (CHO), to establish antibody-producing cells 15. The second process involves culturing the antibody-producing cells 15 in a culture tank 16.

[0054] Process 3, 12, is a purification process for purifying the active pharmaceutical ingredient 18 from the culture supernatant 17. The culture supernatant 17 is a solution obtained by removing cells from the culture medium in the culture tank 16 after the completion of Process 2, 11. The culture supernatant 17 contains dispersed immunoglobulins produced by antibody-producing cells 15, i.e., antibodies 19. Antibody 19, for example, is a monoclonal antibody, which becomes the active ingredient of the biopharmaceutical. Furthermore, in addition to antibodies 19, the culture supernatant 17 also contains dispersed impurities such as cell-derived proteins / cell-derived DNA (Deoxyribonucleic acid) 20, aggregates of antibody 19 21, or viruses 22.

[0055] In process 3 12, immunoaffinity chromatography apparatus 25, cation chromatography apparatus 26 and anion chromatography apparatus 27 are used.

[0056] Culture supernatant 17 is introduced into the immunoaffinity chromatography apparatus 25. For example, a pump P is provided upstream of the immunoaffinity chromatography apparatus 25. Driven by the pump P, the culture supernatant 17 is introduced into the immunoaffinity chromatography apparatus 25. The immunoaffinity chromatography apparatus 25 extracts antibody 19 from the culture supernatant 17 using a chromatographic column in which a ligand such as protein A, which has affinity for antibody 19, is immobilized on a carrier, to generate a first purified solution 28. Here, the purification process based on the immunoaffinity chromatography apparatus 25 is referred to as the immunoaffinity chromatography process. A recovery container T is provided downstream of the immunoaffinity chromatography apparatus 25, and the first purified solution 28 is recovered into the recovery container T.

[0057] A valve V is installed on the pipeline between the immunoaffinity chromatography device 25 and the recovery container T. The valve V switches the flow path of the liquid flowing out of the immunoaffinity chromatography device 25 between the recovery pipeline CL and the waste liquid pipeline WL, which are directed toward the recovery container T.

[0058] The immunoaffinity chromatography process includes three sub-processes: adsorption, washing, and elution. The adsorption process involves introducing culture supernatant 17 into the column to specifically adsorb antibody 19 onto the column. The washing process involves introducing washing solution into the column to remove impurities other than antibody 19 that are not specifically adsorbed onto the column. The elution process involves introducing acidic eluent into the column to dissociate antibody 19 specifically adsorbed onto the column and elute it into the eluent. During the adsorption and washing processes, the flow path is switched via valve V to direct the liquid from the immunoaffinity chromatography apparatus 25 to the waste liquid line WL. The liquid exiting the immunoaffinity chromatography apparatus 25 during the adsorption and washing processes contains almost entirely impurities. This is discharged as waste liquid through the waste liquid line WL.

[0059] In the elution step, the flow path is switched via valve V to direct the liquid from the immunoaffinity chromatography unit 25 to the recovery line CL. The first purified solution 28 containing antibody 19, extracted by the chromatographic column, flows out of the immunoaffinity chromatography unit 25 in the elution step and is recovered into the recovery container T via the recovery line CL.

[0060] In addition, although the illustration is omitted, the first purification solution 28 was subjected to a treatment to inactivate the virus 22 (hereinafter referred to as virus inactivation treatment).

[0061] Furthermore, an ultraviolet detector 31 (hereinafter referred to as a UV (Ultraviolet) detector) and a Raman spectrometer 32 are installed in the pipeline between the immunoaffinity chromatography apparatus 25 and valve V. The UV detector 31 and the Raman spectrometer 32 are sensors that measure values ​​related to the state of the liquid flowing in the pipeline, and are sensors that sense characteristic values ​​related to the mass of the liquid in the pipeline online during the operation of the third process 12.

[0062] The Raman spectrometer 32 consists of a flow cell, a probe, and an analyzer. The flow cell of the Raman spectrometer 32 is connected to a pipe downstream of the UV detector 31. Liquid passing through the UV detector 31 flows through the flow cell. The probe is connected to the flow cell. The Raman spectrometer 32 is a device for evaluating substances using the characteristics of Raman scattering light. When excitation light is irradiated onto a substance, the excitation light interacts with the substance, thereby producing Raman scattering light with a wavelength different from the excitation light. The wavelength difference between the excitation light and the Raman scattering light corresponds to the energy portion of the molecular vibrations of the substance. Therefore, Raman scattering light with different wavenumbers can be obtained between substances with different molecular structures. Among Stokes lines and anti-Stokes lines, Stokes lines are preferred for Raman scattering light. The spectrum of Raman scattering light, i.e., the Raman spectrum, is an example of the "spectral spectroscopy" involved in the technology of this invention.

[0063] In the Raman spectrometer 32, an excitation light is emitted from the probe's outlet into a flow cell containing a liquid. Then, a light-receiving section located at the front end receives the Raman scattered light generated by the interaction between the excitation light and the substances within the liquid. The probe outputs the received Raman scattered light to the analyzer. In this example, a laser beam is used as the excitation light, with the laser beam output set to 500 mW, the excitation wavelength set to 785 nm, and the irradiation time set to 1 second. The Raman spectrometer 32 acquires Raman spectra at 1-second intervals. This 1-second interval is an example of the "second time interval" involved in the technology of this invention. The second time interval is a time interval less than or equal to the first time interval described later. There is no particular limitation, but for monitoring short-term concentration changes in the liquid, 5 seconds or less is preferred, and more preferably 1 second or less.

[0064] In the Raman spectrometer 32, the analyzer decomposes the Raman spectrum by wavenumber and derives the intensity value of the Raman spectrum for each wavenumber, thereby generating spectral measurement data 71 (reference). Figure 5 ).

[0065] As an example, such as Figure 5 As shown, spectral measurement data 71 contains the intensity values ​​of Raman scattered light for each wavenumber. Figure 5 In the middle, the spectral measurement data 71 were measured at 1 cm. -1 Scale-derived wavenumber 700cm-1 ~1800cm -1 Data on the intensity values ​​of scattered light over a range. Additionally, Figure 5 The graph shown at the bottom is a plot of the intensity values ​​of the spectral measurement data 71 by each wavenumber and connected by lines, which is a visualization of the Raman spectrum.

[0066] The UV detector 31 illuminates the liquid from the chromatographic column with detection light and measures the absorbance (light absorption) 64 of the substances in the liquid. The detection light is ultraviolet and / or visible light (light with wavelengths from 190 nm to 800 nm) of a wavelength matched to antibody 19 and aggregate 21; in this example, it is ultraviolet light with a wavelength of 280 nm. As an example, the acquisition interval of the UV detector 31 is also set to 1 second, the same as that for Raman spectroscopy. The absorbance 64 serves as a benchmark for antibody concentration and is effective as a quality monitoring item. However, in addition to aggregate 21 other than antibody 19, the detection light is also sensitive to cellular proteins / cellular DNA 20, thus making it impossible to determine the accurate antibody concentration. Therefore, as described later, antibody concentration prediction based on Raman spectroscopy is performed.

[0067] A first purified solution 28, after virus inactivation treatment, is introduced into a cation chromatography apparatus 26. The cation chromatography apparatus 26 extracts antibody 19 from the first purified solution 28 using a chromatographic column with a cation exchanger as the stationary phase to generate a second purified solution 29. Here, the purification process based on the cation chromatography apparatus 26 is referred to as the cation chromatography process. Similar to the immunoaffinity chromatography process, the cation chromatography process includes adsorption, washing, and elution processes as sub-processes. Furthermore, the structure and function of valve V, recovery line CL, waste liquid line WL, UV detector 31, and Raman spectrometer 32 are the same as in the immunoaffinity chromatography process in the cation chromatography process.

[0068] A second purified solution 29 is introduced into the anion chromatography apparatus 27. The anion chromatography apparatus 27 extracts antibody 19 from the second purified solution 29 using a chromatographic column with an anion exchanger as the stationary phase to generate a third purified solution 30. Here, the purification process based on the anion chromatography apparatus 27 is referred to as the anion chromatography process. Although not illustrated, the third purified solution 30 undergoes a virus removal process. Subsequently, the third purified solution 30 undergoes concentration / filtration based on ultrafiltration (UF) and diafiltration (DF), thereby purifying the original drug 18. In the anion chromatography process, similar to the immunoaffinity chromatography process, there are adsorption, washing, and elution processes as sub-processes. Furthermore, in the anion chromatography process, the structure and function of valve V, recovery line CL, waste liquid line WL, UV detector 31, and Raman spectrometer 32 are also the same as in the immunoaffinity chromatography process.

[0069] By sequentially performing this component separation process based on multiple chromatographic devices 25-27, impurities and viruses 22 are removed in stages, and the purity of antibody 19 is improved in stages. Antibody 19 is an example of the "target component" and "protein" involved in the technology of this invention. The target component refers to the component that is the object of purification in each step of the multiple chromatographic devices 25-27.

[0070] The quality monitoring device 41 is an example of a "quality monitoring device" according to the technology of this invention. The quality monitoring device 41 has the function of monitoring the quality of the third process 12 in manufacturing process 2. The quality of the third process 12 ultimately refers to the quality of the active pharmaceutical ingredient 18, but the quality of the third process 12 includes the quality of the liquids flowing out from each chromatography unit 25-27, such as the first purified solution 28, the second purified solution 29, and the third purified solution 30, which are equivalent to the active pharmaceutical ingredient 18. The quality monitoring device 41 performs a monitoring process to monitor the quality of the third process 12 by using a state prediction model that predicts the state of the liquids flowing out from each chromatography unit 25-27. "State of the liquid" is an indicator representing the physicochemical characteristics of the liquid; in this example, it is the antibody concentration Dp (reference) as the concentration of antibody 19. Figure 3 (etc.). The antibody concentration Dp is an example of the concentration of the target component contained in the liquid involved in the technology of this invention.

[0071] Quality monitoring device 41 uses concentration prediction model 96 (reference) Figure 3 and Figure 5 (etc.) as a state prediction model. The concentration prediction model 96 takes as input spectral measurement data 71 representing the online acquired spectroscopic spectrum, and predicts, for example, the state of the liquid related to the quality of the third process 12, namely the antibody concentration Dp. The concentration prediction model 96 is an example of a "state prediction model" involved in the technology of this invention.

[0072] As an example, such as Figure 2 As shown, the quality monitoring device 41 is composed of computers such as personal computers and server computers. The computer constituting the quality monitoring device 41 includes a memory 45, a RAM 46, a CPU (Central Processing Unit) 47, a communication unit 48, a display 49, and input devices 50. They are interconnected via a bus 51.

[0073] The memory 45 is a hard disk drive built into the computer or connected via cable or network. Alternatively, the memory 45 is a disk array consisting of multiple hard disk drives. The memory 45 stores control programs such as the operating system, various applications, and various data associated with these programs. Alternatively, a solid-state drive can be used instead of a hard disk drive.

[0074] Memory 46 is working memory used by CPU 47 for processing. CPU 47 loads programs stored in memory 45 into memory 46 and executes them accordingly. Thus, CPU 47 centrally controls the various parts of the computer. CPU 47 is an example of a "processor" according to the technology of this invention. Alternatively, memory 46 may also be built into CPU 47.

[0075] The communication unit 48 is a network interface for transmitting and controlling various information via the network 42, etc. The display 49 shows various screens. These screens have GUI (Graphical User Interface) based operation functions. The computer constituting the quality monitoring device 41 receives operation instructions from the input device 50 through these screens. The input device 50 includes a keyboard, mouse, touch panel, and microphone for voice input, etc.

[0076] As an example, such as Figure 3 As shown, the memory 45 of the quality monitoring device 41 stores a working program 75. The working program 75 is an application program used to enable the computer to function as the quality monitoring device 41. That is, the working program 75 is an example of the "working program of the quality monitoring device" involved in the technology of this invention. In addition to storing the working program 75, the memory 45 also stores multiple concentration prediction models 96A, 96B, and 96C as process management information 85 and concentration prediction models 96.

[0077] Multiple concentration prediction models 96A, 96B, and 96C are used separately for each process step. Concentration prediction model 96A is used to predict the antibody concentration Dp in the immunoaffinity chromatography process. Concentration prediction model 96B is used to predict the antibody concentration Dp in the cation chromatography process. Concentration prediction model 96C is used to predict the antibody concentration Dp in the cation chromatography process. Hereinafter, without distinguishing between the three types of data, they will be referred to simply as concentration prediction model 96.

[0078] If the working program 75 is started, the CPU 47 and memory 46 of the computer constituting the quality monitoring device 41 cooperate to function as the processor 52, which has an acquisition unit 120, an RW control unit 121, a prediction unit 122, a display control unit 123, and a process management unit 124.

[0079] The acquisition unit 120 acquires spectral measurement data 71 from the Raman spectrometer 32. The acquisition unit 120 outputs the spectral measurement data 71 to the RW control unit 121. Figure 3In this document, the spectral measurement data 71 is recorded separately as spectral measurement data 71A, 71B, and 71C, distinguishing the data obtained in each chromatographic step. Spectral measurement data 71A is the data obtained in the immunoaffinity chromatography step, spectral measurement data 71B is the data obtained in the cation chromatography step, and spectral measurement data 71C is the data obtained in the anion chromatography apparatus 27. Hereinafter, unless there is a need to distinguish between the three types of data, they will be referred to simply as spectral measurement data 71.

[0080] Furthermore, the acquisition unit 120 acquires the absorbance 64 from the UV detector 31. The acquisition unit 120 outputs the absorbance 64 to the RW control unit 121. The absorbance 64 is recorded as absorbance 64A, 64B, and 64C, distinguishing the data acquired in each chromatographic step. Absorbance 64A is the data acquired in the immunoaffinity chromatography step, absorbance 64B is the data acquired in the cation chromatography step, and absorbance 64C is the data acquired in the anion immunoaffinity chromatography step. Hereinafter, unless it is necessary to distinguish between the three types of data, they will be simply referred to as absorbance 64.

[0081] Furthermore, the acquisition unit 120 acquires various operation instructions from the input device 50. These various operation instructions include start instructions for each step of the third process 12.

[0082] The RW control unit 121 controls the storage of various data to the memory 45 and the retrieval of various data stored in the memory 45. The RW control unit 121 stores the spectral measurement data 71 and absorbance 64 from the acquisition unit 120 into the memory 45. The spectral measurement data 71 and absorbance 64 are appended with a timestamp as a timestamp and stored.

[0083] Furthermore, the RW control unit 121 reads process management information 85 from memory 45 and outputs the read process management information 85 to the process management unit 124. The RW control unit 121 reads process management information 85, concentration prediction model 96, and spectral measurement data 71 from memory 45 and outputs the read concentration prediction model 96 and spectral measurement data 71 to the prediction unit 122. Concentration prediction models 96A-96C and spectral measurement data 71A-71C are selectively output to the prediction unit 122 according to the process. Figure 3 An example is shown where a concentration prediction model 96A and spectral measurement data 71A are output for the immunoaffinity chromatography process. Furthermore, the RW control unit 121 outputs absorbance 64 to the display control unit 123.

[0084] The prediction unit 122 applies the spectral measurement data 71 to the concentration prediction model 96 and outputs the antibody concentration Dp from the concentration prediction model 96. The prediction unit 122 outputs the predicted antibody concentration Dp to the display control unit 123.

[0085] The display control unit 123 controls the display of various screens on the monitor 49. For example, the display control unit 123 controls the display of the monitoring screen 91 (reference) Figure 9 The control (etc.) is displayed on the monitor 49. The antibody concentration Dp and absorbance 64, which are obtained online in real time during each step of the execution of the third process 12, are displayed on the monitoring screen 91.

[0086] When executing process 3 12, the process management unit 124, in addition to controlling the operation of equipment such as pump P and valve V, also centrally controls various parts of the processor 52, thereby performing monitoring processing and model switching processing. The monitoring processing involves acquiring spectral measurement data 71 and absorbance 64, predicting the antibody concentration Dp using the concentration prediction model 96, and displaying the predicted antibody concentration Dp and absorbance 64 on the display 49. This allows operators and other users to confirm the quality of process 3 12.

[0087] The process management department 124 performs model switching processing based on process management information 85, switching the concentration prediction model 96 midway through the monitoring process.

[0088] As an example, such as Figure 4 As shown, process management information 85 contains process management information for each chromatographic step included in process 3 12, namely, process management information for the immunoaffinity chromatography step (in... Figure 4 The process management information for the cation chromatography process (marked with "A") is shown in [the relevant document / section]. Figure 4 The process management information for the anion chromatography process (marked with "B") and other processes (in...) Figure 4 (marked with "C").

[0089] The management information for each process includes process identification information, equipment used information, process quality information, model used information, and other information. Process identification information is used to identify each chromatographic process included in Process 3 12, specifically the process name and identification information (ID). Equipment used information refers to the equipment used in each process of Process 3 12 (chromatographic columns of chromatographic apparatus 25-27, UV detector 31, Raman spectrometer 32, pump P, and valve V, etc.). Process quality information is quality-related information obtainable after each chromatographic process has been performed at least once. Specifically, it includes quality information acquired during the process and final quality information after the process is completed. Quality information acquired during the process includes online antibody concentration Dp and online absorbance 64, etc.

[0090] Final quality information, for example, is the final antibody concentration (Dpf) of the total liquid produced during the process (reference). Figure 11(etc.). More specifically, for an immunoaffinity chromatography step, it refers to the average antibody concentration of the total amount of the first purified solution 28 recovered into the recovery container T after the step is completed. The antibody concentration Dp acquired online can be considered an instantaneous value, varying depending on the acquisition time. In contrast, the final antibody concentration Dpf, as the final mass, is a constant final value, which is the average of the varying antibody concentrations Dp. Furthermore, if the first purified solution 28 is not recovered into the recovery container T and is sent to the next step, the final antibody concentration Dpf is the average antibody concentration Dp of the total amount of the first purified solution 28 sent to the next step.

[0091] The model information used is the information for using concentration prediction model 96 for each process. In this example, the model used in the immunoaffinity chromatography process is concentration prediction model 96A (named concentration prediction model "A" in the process management information). The model used in the cation chromatography process is concentration prediction model 96B (named concentration prediction model "B" in the process management information). The model used in the anion chromatography process is concentration prediction model 96C (named concentration prediction model "C" in the process management information).

[0092] Other information includes date and time information, as well as variety information. Variety information, for example, includes information about the variety of antibody 19 that is the target component.

[0093] Thus, the process management information 85 in this example is a mixture of information entered by the operator, such as process identification information, equipment used information, and model used information, and information obtained when a process is performed, such as process quality information.

[0094] Figure 4 The process management information, process identification information, equipment used information, and process quality information shown are examples of the "process management information," "process identification information," "equipment used information," and "process quality information" involved in the technology of this invention. Furthermore, this example illustrates the summarization of multiple process management information into a single process management information 85, but each process management information can be segmented according to each process.

[0095] In this example, the process management unit 124 performs model switching processing based on process identification information before the start of a step included in process 123. Specifically, before the immunoaffinity chromatography step begins, the process management unit 124 sets up concentration prediction model 96A with reference to the process identification information. Then, after the immunoaffinity chromatography step is completed and before the cation chromatography step begins, the concentration prediction model 96B is set up with reference to the process identification information. Moreover, after the cation chromatography step is completed and before the anion chromatography step begins, the concentration prediction model 96C is set up with reference to the process identification information.

[0096] As an example, such as Figure 5 and Figure 6 As shown, the concentration prediction model 96 is constructed from a neural network 105. Thus, the concentration prediction model 96 is a machine learning model. As is well known, the neural network 105 has an input layer 106, intermediate layers (also called hidden layers) 107, and an output layer 108. These input layers 106, intermediate layers 107, and output layers 108 each have multiple nodes ND. Coefficients representing the binding strength of each node ND are set between the nodes ND of the input layer 106 and the nodes ND of the intermediate layer 107, between the nodes ND within the intermediate layer 107, and between the nodes ND of the intermediate layer 107 and the nodes ND of the output layer 108. Appropriate activation functions, such as linear functions and ReLU (Rectified Linear Unit) functions, are set in the nodes ND of the output layer 108.

[0097] Input in each node ND of input layer 106 Figure 5 The intensity values ​​of each wavenumber of the spectral measurement data 71 shown are used as input data. Furthermore, the predicted concentration value of antibody 19, i.e., antibody concentration Dp, is output from node ND of output layer 108.

[0098] In concentration prediction model 96, if the predicted antibody concentration Dp varies greatly, the prediction accuracy may sometimes decrease. Process 12 is a process that increases the antibody concentration Dp in stages according to each step of the chromatographic process; therefore, the variation in antibody concentration Dp between steps is large. Therefore, in the quality monitoring device 10, by switching the concentration prediction model 96 for each step, the decrease in prediction accuracy is suppressed.

[0099] Moreover, as an example, such as Figure 7As shown, the multiple concentration prediction models 96A to 96C used in each chromatographic step have been trained before the start of the monitoring process of the third process 12. Each concentration prediction model 96A to 96C is trained with different training data 71LA to 71LC according to the predicted range of antibody concentration Dp. The training data 71LA of the concentration prediction model 96A is training data for the immunoaffinity chromatography step, and is training data that matches the range of antibody concentration Dp in the immunoaffinity chromatography step. The training data 71LA consists of the spectral measurement data 71 acquired for training and the corresponding positive solution data of antibody concentration Dp. As a method for creating the training data 71LA, for example, there is the method described in International Publication No. WO2022 / 209422, etc. That is, multiple first purification solutions 28 with different antibody concentrations Dp are created, and the spectral measurement data 71 of the multiple first purification solutions 28 created are acquired using a Raman spectrometer 32. Then, antibody concentration Dp is obtained, for example, using HPLC (high-performance liquid chromatography), as the positive solution data for multiple first purified solutions 28. Multiple learning data 71LA are created by combining these spectroscopic data 71 and multiple combinations of antibody concentration Dp that become the positive solution data. In order to improve the generalization performance of the concentration prediction model 96A to unknown liquids, the multiple first purified solutions 28 are preferably prepared with various changes in the concentration of dopants, etc., in addition to antibody concentration Dp, and corresponding learning data 71LA are created.

[0100] Furthermore, when multiple learning data sets 71LA are created, it is preferable to use a portion of them as training data and another portion as test data for verifying prediction accuracy. In the verification based on the test data, evaluation values ​​such as the predicted value of antibody concentration Dp and the RMSE (Root-Mean-Square Error) of the correct answer data are used, for example. Moreover, the evaluation values ​​of the verification results based on the test data can be associated with the learned concentration prediction models 96A to 96C respectively and recorded in the memory 45. For example, multiple concentration prediction models 96A can be prepared for use in the same process. In this case, for example, the evaluation value of prediction accuracy can be used as a selection criterion for the concentration prediction model 96A, such as using the one with the highest possible evaluation value.

[0101] The training data 71LB for concentration prediction model 96B is for cation chromatography and is matched to the range of antibody concentration Dp in cation chromatography. The training data LC for concentration prediction model 96C is for anion chromatography and is matched to the range of antibody concentration Dp in anion chromatography. The methods for creating training data 71LB and 71LC are the same as those for training data 71LA.

[0102] As an example, such as Figure 8 As shown, in the immunoaffinity chromatography process, the prediction unit 122 outputs the antibody concentration Dp at predetermined intervals using a concentration prediction model 96A. T1, T2, T3, T4, T5, T6, ... are the prediction times for the antibody concentration Dp, and the time interval between each prediction time is the first time interval. The first time interval is an example of the "first time interval" involved in the technology of this invention. Specifically, the first time interval is 1 second. The first time interval is not particularly limited, but as mentioned above, in order to monitor short-term concentration changes, it is preferably 5 seconds or less, and more preferably 1 second or less. The prediction interval of the antibody concentration Dp, i.e., the first time interval, is limited by the acquisition interval of the spectral measurement data 71A of the Raman spectrometer 32, i.e., the second time interval. Therefore, as mentioned above, the second time interval is set to a time of 5 seconds or less than the first time interval. Figure 8 In the example, the first time interval is the same as the second time interval, which is 1 second.

[0103] Thus, the antibody concentration Dp output by the prediction unit 122 at the first time interval is assigned a timestamp as the time of prediction and stored in the memory 45. The spectral measurement data 71A and absorbance 64A are also assigned timestamps and stored in the memory 45. Therefore, the spectral measurement data 71A, absorbance 64A, and antibody concentration Dp are stored in the memory 45 as time series data.

[0104] Furthermore, considering the ease of comparison between the two data points, the acquisition interval of the spectral measurement data 71, i.e., the second time interval, and the prediction interval of the antibody concentration Dp, i.e., the first time interval, are preferably the same, but they can also be different. In this case, it is preferable to perform extraction processing and other adjustments based on the time series data of the one with the shorter time interval. Moreover, if a timestamp exists, it can also be used as a reference to arrange the two data points in a time series manner so that they can be compared.

[0105] In the cation chromatography step, similarly to the immunoaffinity chromatography step, the prediction unit 122 uses the concentration prediction model 96B to output the antibody concentration Dp at a preset first time interval. Then, the spectral measurement data 71B, absorbance 64B, and antibody concentration Dp are stored as time-series data in the memory 45. The same prediction is performed in the anion chromatography step, and the spectral measurement data 71C, absorbance 64C, and antibody concentration Dp are stored as time-series data in the memory 45.

[0106] As an example, such as Figure 9 As shown, the display control unit 123 outputs a monitoring screen 91 displaying graphs showing the changes over time for the absorbance 64 and antibody concentration Dp based on time-series data. The graphs showing the changes over time are updated each time the latest antibody concentration Dp is acquired. For example, the monitoring screen 91 displays the changes over time for the antibody concentration Dp during the elution step in the immunoaffinity chromatography process. Figure 9 In this context, antibody concentration Dp is represented by mAb (monoclonal antibody). Furthermore, in... Figure 9 In the diagram, UV280 indicates the absorbance 64 of the UV detector 31, which uses ultraviolet light with a wavelength of 280 nm as the detection light. During the elution process, antibody 19 adsorbed on the column of the immunoaffinity chromatography apparatus 25 detaches and flows into the tubing, thus the antibody concentration DP begins to rise. Ultimately, the amount of antibody 19 adsorbed on the column decreases, and therefore antibody 19 also begins to decrease.

[0107] As shown in the solid line graph, absorbance 64 indicates a concentration higher than antibody concentration Dp. This is because the detection light of UV detector 31 is also sensitive to dopants other than antibody 19. Figure 9 In the example shown, the absorbance was saturated at 64 shortly after antibody 19 began to rise.

[0108] Next, as an example, refer to Figure 10 The flowchart shown illustrates the function of the above structure. Figure 10 The flowchart shown illustrates the steps of the monitoring process performed by the quality monitoring device 41.

[0109] The concentration prediction models 96A to 96C, which have been learned before the start of the third process 12, are stored in the memory 45 of the quality monitoring device 41. In this state, the third process 12 begins, and monitoring processing starts simultaneously. If monitoring processing begins, in step ST1000, the processor 52 determines whether a process should be started. The process determination is, for example, based on an operation instruction specifying the process name to be started by the operator. For example, in the case of starting an immunoaffinity chromatography process, the process name of the immunoaffinity chromatography process is specified.

[0110] In step ST1100, the processor 52 refers to the process management information and sets the concentration prediction model 96A corresponding to the specified process. In step ST1200, the processor 52 starts the process by starting the drive of pump P, etc., according to the operation instruction indicating the start of the process.

[0111] In step ST1300, the processor 52 acquires absorbance 64 and spectral measurement data 71, and predicts the antibody concentration Dp based on the spectral measurement data 71. Furthermore, the processor 52 displays the changes in absorbance 64 and antibody concentration Dp over time on the display 49 via a monitoring screen 91.

[0112] In step ST1400, processor 52 determines whether the process is complete. The determination of process completion is made, for example, based on whether the antibody concentration Dp falls below a pre-set threshold after the elution process begins. When the antibody concentration Dp falls below the threshold, it indicates that the elution of antibody 19 from the chromatographic column is nearing completion. If the elution process continues while the antibody concentration Dp is decreasing, the final antibody concentration Dpf of the first purification solution 28 will also decrease. Therefore, processor 52 determines the completion of the process based on the threshold. If the process is not complete ("No" in step ST1400), the process continues to step ST1300. If the process is complete ("Yes" in step ST1400), processor 52 proceeds to step ST1500 to determine whether a next process is possible.

[0113] If there is a next step in step ST1500 (in step ST1500 it is "Yes"), processor 52 returns to step ST1000 and repeats the processing after step ST1000. In processor 52, in step ST1200, model switching processing is performed to switch the concentration prediction model 96 according to the next step.

[0114] If there is no next process in step ST1500 (the value in step ST1500 is "No"), the monitoring process ends.

[0115] In addition, while the start of a process is determined based on the process name input by the operator in step ST1000, the processor 52 may also determine the start of a process independently of the operator's input. For example, the processor 52 may detect the residual amount of culture supernatant 17 introduced into the immunoaffinity chromatography apparatus 25 using a liquid level sensor or similar means, thereby enabling the processor 52 to determine the next process to begin without relying on the operator's input.

[0116] As explained above, the quality monitoring device 41 according to the present invention includes a processor 52. The processor 52 performs a monitoring process in which spectral measurement data 71, representing Raman spectra (an example of spectroscopic spectroscopy) obtained online from a liquid produced in a biopharmaceutical manufacturing process, is used as input data. Multiple concentration prediction models 96A to 96C (an example of multiple state prediction models) are used to monitor quality, predicting antibody concentration Dp (an example of liquid state) related to the quality of an example of the manufacturing process, namely, the third process 12. Then, the processor 52 can perform a model switching process, in which the concentration prediction models 96A to 96C are switched midway through the monitoring process based on process identification information (an example of process management information) used to manage the processes included in the third process 12. Furthermore, the multiple concentration prediction models 96A to 96C have been learned before the monitoring process begins.

[0117] Therefore, compared with existing technologies, the quality of the manufacturing process of biopharmaceuticals can be appropriately managed. Specifically, in the past, because learning data was collected and state prediction models were updated during the operation of the manufacturing process, the state prediction models might not be updated in a timely manner when there were large changes in state between processes, resulting in insufficient quality management. According to the technology of the present invention, multiple concentration prediction models 96A-96C that have been learned before the start of the monitoring process are used, and model switching is performed midway through the monitoring process. Therefore, even when there are large changes in state between processes, model switching can be performed quickly. Thus, compared with existing technologies, the quality of the manufacturing process of biopharmaceuticals can be appropriately managed.

[0118] The operator can monitor the changes in absorbance 64 and antibody concentration Dp over time in process 3 12 via screen 91. Absorbance 64 and antibody concentration Dp are displayed in real-time, time-series, allowing for quick assessment of whether process 3 12 is proceeding normally.

[0119] Furthermore, in the above embodiment, the model switching process involves switching the concentration prediction model 96 based on the process name (an example of process identification information) before the start of a step included in the third process 12 (an example of a manufacturing process). Therefore, an appropriate state prediction model can be used in each step.

[0120] Alternatively, the state prediction model can be switched based on the equipment used, instead of the process name. The equipment used could be, for example, chromatographic columns of 25 to 27 mm in diameter.

[0121] Furthermore, in the above embodiment, Raman spectroscopy was used as the spectroscopic spectrum. Predicting antibody concentration Dp based on Raman spectroscopy is a well-known method, therefore, reliable and stable prediction accuracy can be expected compared to using a new prediction method.

[0122] Furthermore, as a spectroscopic method, it can be either Raman spectroscopy, infrared spectroscopy, or fluorescence spectroscopy. This is a well-known prediction method, thus reliable and stable prediction accuracy can be expected.

[0123] Alternatively, multiple spectroscopic methods of different types, such as Raman spectroscopy and infrared spectroscopy, can be used. For example, a method of using spectroscopic methods differently for each state prediction model can be considered.

[0124] Furthermore, in the above embodiment, the processor 52 outputs a predicted value of the antibody concentration Dp (an example of the liquid state) at a preset first time interval based on the spectral measurement data 71 representing Raman spectroscopy (an example of spectroscopic spectroscopy). Therefore, it is possible to monitor changes in the state of the liquid in real time.

[0125] In the above embodiment, the first time interval is 1 second. Thus, by setting the first time interval to 5 seconds or less, and more preferably to 1 second or less, it is possible to grasp the changes in the state of the liquid that occur within a short period of time.

[0126] Furthermore, in the above embodiment, manufacturing process 2 includes, for example, a purification process of antibody 19 (an example of the target component) contained in the liquid, namely, process 12. Moreover, the state of the liquid refers to the concentration of antibody 19 contained within it. In the case of process 12 being a purification process for antibody 19, the concentration of antibody 19 is an important quality-related indicator. Therefore, compared to not monitoring the concentration of antibody 19, quality can be appropriately monitored.

[0127] Furthermore, in the above embodiment, the target component is antibody 19. Antibody 19 is an active ingredient in biopharmaceuticals, thus enabling appropriate monitoring of the quality of biopharmaceuticals.

[0128] Furthermore, in the above embodiments, the state of the liquid being predicted is the state of the proteins contained within it. Proteins are a major component of liquids produced during the manufacturing process of biopharmaceuticals; therefore, by monitoring the state of proteins, the quality of biopharmaceuticals can be appropriately monitored.

[0129] (Method for obtaining the final antibody concentration Dpf)

[0130] As an example, such as Figure 11As shown, an example of the final quality, namely the final antibody concentration Dpf, is obtained offline after the process is completed. The immunoaffinity chromatography process is used as an example. If the immunoaffinity chromatography process is completed, the total amount of the first purified solution 28 purified in this process is contained in the recovery container T. The first purified solution 28 is sampled from the recovery container T, and the sampled first purified solution 28 is analyzed offline to obtain the final antibody concentration Dpf. In the offline analysis, methods such as HPLC (high performance liquid chromatography) are used to obtain the final antibody concentration Dpf.

[0131] (Forecasts for multiple projects)

[0132] As an example, such as Figure 12 As shown, multiple items can be predicted using the concentration prediction model 96 based on the spectral measurement data 71. These multiple items include, in addition to antibody concentration Dp, aggregate concentration Dhw and HCP (Host Cell Protein) concentration Dhcp, etc. Raman spectroscopy contains information beyond antibody 19, thus enabling the prediction of these multiple items.

[0133] (Items displayed on the surveillance screen)

[0134] Then, as an example, such as Figure 13 As shown, multiple predicted items can be displayed on the monitoring screen 91. If so, similarly to the antibody concentration Dp, the changes in aggregate concentration Dhw and HCP concentration Dhcp over time can be displayed in real time.

[0135] in addition, Figure 13 and Figure 9 Similarly, the changes in antibody concentration Dp over time during the elution process are shown. Aggregates and HCP are dopants. During the elution process, it is assumed that HCP undergoes non-specific adsorption on the column, therefore elution is faster compared to antibody 19. In contrast, aggregate 21 (in...) Figure 13 The group represented by HMWS is a collection of antibodies 19, and therefore is eluted using a process that is almost identical to that of antibody 19. Figure 13 This situation is illustrated.

[0136] (Switch models within sub-processes)

[0137] As an example, such as Figures 14-16 The variation of the first embodiment shown can also switch the concentration prediction model 96 between multiple sub-processes within a single process included in the third process 12. Figures 14-16 The example shown is an example of using concentration prediction model 96A1 in the adsorption and cleaning processes and concentration prediction model 96A2 in the elution process within the three sub-processes of immunoaffinity chromatography: adsorption, cleaning and elution.

[0138] like Figure 14 As shown, in the process management information of the immunoaffinity chromatography process in process management information 85, the model names of concentration prediction model 96A1 (named "A1" in the process management information) and concentration prediction model 96A2 (named "A2" in the process management information) are recorded as model information. Furthermore, the adsorption and washing processes are recorded as applicable conditions for concentration prediction model 96A1, and the elution process is recorded as applicable conditions for concentration prediction model 96A2. In addition, the order of the adsorption, washing, and elution processes is recorded as sub-process information in the process identification information. The processor 52 performs model switching processing between the two concentration prediction models 96A1 and 96A2 based on the sub-process information and applicable conditions included in the process identification information.

[0139] As described above, the adsorption step involves introducing the culture supernatant 17 into the immunoaffinity chromatography apparatus 25 to adsorb antibody 19 onto the chromatographic column, and the washing step involves introducing a washing solution to wash away non-specifically adsorbed impurities other than antibody 19 that are specifically adsorbed onto the chromatographic column. Therefore, in both the adsorption and washing steps, the antibody concentration Dp of the liquid effluent from the immunoaffinity chromatography apparatus 25 is very low. In contrast, the elution step involves introducing an eluent into the chromatographic column to detach antibody 19 that is specifically adsorbed onto the column, resulting in a higher antibody concentration Dp in the liquid effluent from the immunoaffinity chromatography apparatus 25. Thus, the antibody concentration Dp varies significantly between the adsorption, washing, and elution steps. Furthermore, the liquid introduced into the immunoaffinity chromatography apparatus 25 changes during the adsorption, washing, and elution steps, leading to substantial changes in the composition of the effluent.

[0140] As an example, such as Figure 15 As shown, concentration prediction model 96A1 is a completed learning model trained using learning data 71LA1 from the adsorption and washing processes. Concentration prediction model 96A2 is a completed learning model trained using learning data 71LA2 from the elution process. Learning data 71LA1 is for relatively low concentrations and is suitable for the components of the liquid being measured in the adsorption and washing processes. Learning data 71LA2 is for relatively high concentrations and is suitable for the components of the liquid being measured in the elution process. Therefore, concentration prediction model 96A1 has high prediction accuracy in the low concentration region, while concentration prediction model 96A2 has high prediction accuracy in the high concentration region.

[0141] Figure 16This is an example of a model switching step within the immunoaffinity chromatography process. For example... Figure 16 As shown, when the processor 52 begins monitoring the immunoaffinity chromatography process, it first refers to the process identification information in the process management information to determine that the initial sub-process of the immunoaffinity chromatography process is the adsorption process. Then, the processor 52 refers to the usage model information in the process management information and sets a low-concentration concentration prediction model 96A1. In the adsorption and washing processes, the antibody concentration Dp is predicted using the concentration prediction model 96A1, and the changes in the predicted antibody concentration Dp over time are displayed in real time on the monitoring screen 91.

[0142] Then, after the adsorption and washing processes are completed and before the elution process begins, the processor 52 switches from the low-concentration concentration prediction model 96A1 to the high-concentration concentration prediction model 96A2. During the elution process, the antibody concentration Dp is predicted using the concentration prediction model 96A2, and the changes in the predicted antibody concentration Dp over time are displayed in real time on the monitoring screen 91.

[0143] Thus, the model switching process also includes switching the state prediction model based on process identification information before the sub-process begins. Figures 14-16 In the example shown, the concentration prediction model 96A1 is used for low concentrations and the concentration prediction model 96A2 is used for high concentrations during the monitoring process, thus enabling appropriate quality monitoring compared to existing technologies.

[0144] exist Figure 16 In the processing steps shown, such as Figure 10 As shown in step ST1000, the determination of the starting sub-process is based on the processor 52's input of the process name by the operator. Alternatively, the processor 52 may determine the starting sub-process independently of the operator's input of the process name. For example, the processor 52 determines that the cleaning process is complete by detecting the residual amount of cleaning fluid. Then, regarding the next process to begin, it can be determined to be an elution process based on the process identification information.

[0145] [Second Implementation]

[0146] As an example, Figures 17-19 The model switching process in the second embodiment shown is to switch the concentration prediction model 96 based on quality information acquired midway through the process, which is one type of process quality information. In the first embodiment, the concentration prediction model 96 is switched before the process starts based on process identification information. In contrast, the second embodiment switches the concentration prediction model 96 based on quality information acquired midway through the process after it has started.

[0147] like Figure 17The process management information shown in the model information describes how the second embodiment distinguishes between the low-concentration concentration prediction model 96A1 and the high-concentration concentration prediction model 96A2. Figure 14 The variations of the first embodiment shown are the same. However, the difference lies in the use of model information: the applicable conditions for each concentration prediction model 96A1 and 96A2 are not process names, but are defined by the relationship between absorbance 64A and a pre-set threshold. Absorbance 64A is the absorbance of antibody 19 and aggregate 21 in the liquid, obtained online by UV detector 31. Absorbance 64A is stored as time-series data in memory 45, but can also be used as process quality information in process management information 85.

[0148] In addition, in the second embodiment, the same low-concentration concentration prediction model 96A1 and high-concentration concentration prediction model 96A2 as in the first embodiment are used, but a concentration prediction model 96A that is learned using learning data more suitable for the second embodiment can also be used.

[0149] like Figure 18 and Figure 19 As shown, processor 52 starts the elution process with the low-concentration prediction model 96A1 used in the adsorption and washing processes enabled. Processor 52 predicts the antibody concentration Dp using the low-concentration prediction model 96A1 and displays the change in antibody concentration Dp over time on monitoring screen 91. After starting the elution process, processor 52 compares the absorbance 64 over time with a threshold value midway through the process. Then, during the period when absorbance 64 is below the threshold value, the low-concentration prediction model 96A1 is used.

[0150] like Figure 19 As shown, the threshold (in) Figure 19 (represented by Th) For example, it is set to the value before absorbance 64 saturates. As mentioned above, the detection light of the UV detector 31 is also sensitive to aggregates 21 other than antibody 19 and dopants containing HCP, so absorbance 64 will saturate before the antibody concentration Dp shows a peak.

[0151] At the beginning of the elution process, the antibody concentration Dp is low. Eventually, antibody 19 detaches from the column, and the antibody concentration Dp begins to rise sharply. When the absorbance 64A exceeds the threshold, the processor 52 switches to the high-concentration concentration prediction model 96A2. The processor 52 uses the high-concentration concentration prediction model 96A2 to predict the antibody concentration Dp and displays the change in antibody concentration Dp over time on the monitoring screen 91. The processor 52 further continues to compare the absorbance 64A with the threshold, and when the absorbance 64A falls below the threshold, it switches back to the low-concentration concentration prediction model 96A1. Then, the concentration prediction model 96A1 is used until the elution process is completed.

[0152] Thus, in Figure 19 In the example shown, the elution process begins with a low concentration using concentration prediction model 96A1. Then, during the elution process, based on the change in absorbance 64A, it switches to a high concentration using concentration prediction model 96A2 before absorbance 64A becomes saturated. If absorbance 64A decreases due to a decrease in antibody concentration Dp, it switches back to a low concentration using concentration prediction model 96A1.

[0153] Absorbance 64A is an example of "quality information acquired midway through the process". Based on this quality information, by switching a state prediction model, namely concentration prediction model 96A, an appropriate concentration prediction model 96A can be used even if the process quality changes midway through the process.

[0154] Furthermore, absorbance 64A, acquired online during the immunoaffinity chromatography process, is information related to the concentration of antibody 19 in the liquid, and is an example of "information related to the concentration of the target component in the liquid acquired online during the process." Manufacturing process 2 includes a purification process for the target component contained in the liquid, namely process 3 12. Moreover, the state of the liquid predicted by the concentration prediction model 96A is the concentration of the target component contained in the liquid, namely antibody 19. In the case of predicting the antibody concentration Dp, the information related to the concentration of antibody 19, namely absorbance 64A, is one of the important indicators related to process quality. Therefore, by switching the concentration prediction model 96A based on absorbance 64A, an appropriate model can be selected.

[0155] (A variation of the second embodiment)

[0156] Figures 20-22 The variations of the second embodiment shown are as follows: Figures 17-19 The second embodiment shown is largely the same, so only the differences will be described. The difference lies in that... Figures 17-19 In the second embodiment shown, as "information related to the concentration of the target component of the liquid acquired online during the process," the absorbance 64 Å measured by the UV detector 31 is used. In contrast, Figures 20-22In the variant shown, the antibody concentration Dp is used, which is the predicted value based on the concentration prediction model 96A.

[0157] like Figure 20 As shown, the relationship between antibody concentration Dp and the threshold is specified as an applicable condition in the usage model information of process management information. For example... Figure 21 and Figure 22 As shown, after the elution process begins, processor 52 compares the time-varying antibody concentration Dp with a threshold midway through the process. Then, during the period when the antibody concentration Dp is below the threshold, a low-concentration concentration prediction model 96A1 is used.

[0158] If the antibody concentration Dp predicted by the low-concentration concentration prediction model 96A1 exceeds a threshold, the processor 52 switches to the high-concentration concentration prediction model 96A2. The processor 52 further compares the antibody concentration Dp with the threshold, and if the antibody concentration Dp falls below the threshold, it switches back to the low-concentration concentration prediction model 96A1. Then, the concentration prediction model 96A1 is used until the elution process is completed.

[0159] As mentioned above, when predicting antibody concentration Dp, the predicted value of antibody concentration Dp is one of the important indicators related to process quality, just like absorbance 64A. Therefore, by switching the concentration prediction model 96A according to antibody concentration Dp, an appropriate model can be selected.

[0160] Furthermore, as in comparison Figure 19 and Figure 22 It is known that, regarding "information related to the concentration of the target component of the liquid acquired online in the process", regardless of whether absorbance 64 or antibody concentration Dp is used, both can be set to the same switching time based on the threshold setting.

[0161] (Other variation 1: How to determine the final antibody concentration Dpf)

[0162] As an example, such as Figure 23 As shown, the final antibody concentration Dpf can be calculated based on the predicted value of the concentration prediction model 96A, i.e., the antibody concentration Dp. For example, in the immunoaffinity chromatography step, the processor 52 uses the concentration prediction model 96A to predict the antibody concentration Dp at a predetermined time interval based on the spectral measurement data 71A acquired at a predetermined time interval. The processor 52 integrates the predicted antibody concentration Dp over time. The time integral value of the antibody concentration Dp is an estimate of the total amount of antibody 19 contained in the total amount of the first purified solution 28 that is ultimately recovered. The final antibody concentration Dpf can be derived by dividing the total amount of antibody 19 by the total amount of the first purified solution 28. The total amount of the first purified solution 28 can be obtained, for example, using a gravimeter.

[0163] Furthermore, the total amount of the first purified solution 28 can also be obtained by using a liquid level sensor installed in the recovery container T, and by measuring the height of the liquid level of the first purified solution 28 finally recovered to the recovery container T. Alternatively, the total amount of the first purified solution 28 can be determined using a flow meter installed in the recovery line CL. Specifically, the flow rate of the first purified solution 28 through the recovery line CL during the elution step of the immunoaffinity chromatography process is measured, and the time integral value of this flow rate is calculated. Moreover, the total amount of the first purified solution 28 can also be calculated based on the rotational speed of the pump P during the elution step.

[0164] pass Figure 23 Although the final antibody concentration Dpf obtained by the method shown is an estimate, it is still possible to obtain a more accurate and accurate result. Figure 11 As shown, the final antibody concentration (Dpf) can be determined immediately after the immunoaffinity chromatography step without offline analysis. Furthermore, the intermediate progress of the final antibody concentration (Dpf) can be calculated in real time during the elution process and displayed on the monitoring screen 91. Additionally, the recovery rate of antibody 19 can be calculated based on the intermediate progress of the final antibody concentration (Dpf), and the switching time of valve V for ending the elution process can be determined based on these values.

[0165] like Figure 23 As shown, the final antibody concentration Dpf, determined immediately after the completion of the process, can be used to determine the process conditions for the next process. Process conditions refer to various conditions that can be set during the execution of a process using equipment settings and parameters. Specifically, the process conditions for the next process are as follows. In this example, the next process after the immunoaffinity chromatography process is a cation chromatography process. In this example, in the cation chromatography apparatus 26 used in the cation chromatography process, the amount of antibody 19 that can be added is set according to the size of the column and the properties of the resin packed within the column. Therefore, if the final antibody concentration Dpf can be determined immediately after the completion of the immunoaffinity chromatography process, the amount of the first purification solution 28 to be added to the column of the cation chromatography apparatus 26 for the next process can be determined based on the final antibody concentration Dpf.

[0166] The final antibody concentration Dpf calculated based on the predicted value from concentration prediction model 96A is an example of "the final quality of a process, representing the final state of the liquid aggregate produced in the process," as per the technology of this invention. Thus, by calculating the final quality of a process based on the predicted value, even when the time interval between processes is short, the final quality of the previous process can be used, for example, to determine the process conditions of the next process.

[0167] Furthermore, according to the technology of the present invention, the state prediction model is appropriately switched based on process management information during the monitoring and processing of manufacturing process 2, thereby improving the prediction accuracy of the final quality estimate. Therefore, the process conditions for the next process can also be appropriately determined.

[0168] (Other variations 2)

[0169] Furthermore, as an example, such as Figure 24 As shown, the processor 52 can acquire spectral measurement data 71 (an example of spectroscopic spectrum) at a second time interval below the first time interval, and output antibody concentration Dp (an example of a predicted value of the liquid state) based on the moving average of the acquired multiple spectral measurement data 71.

[0170] like Figure 24 As shown, processor 52, for example, derives a moving average of 10 spectral measurements 71 acquired at a second time interval (1 second in this example). Averages 1, 2, 3, 4… are moving averages calculated by shifting the spectral measurements 71 of the averaged object by one interval. Based on this moving average of the spectral measurements 71, processor 52 outputs, for example, a predicted value of the liquid's state, namely, the antibody concentration Dp.

[0171] The noise in a single measurement of spectral data 71 is relatively high. Averaging it ten times can suppress the noise. Then, by using a moving average, the acquisition interval can be shortened. If the noise in the data used for prediction is suppressed, the prediction accuracy will also improve.

[0172] Thus, the processor 52 outputs a moving average of the spectroscopic spectrum for prediction, thereby enabling high-accuracy predictions with suppressed noise at shorter time intervals.

[0173] The reason why the quality monitoring device 41 uses a state prediction model that has been learned before the monitoring process begins is that the state prediction model can be switched midway through the monitoring process. That is, as shown in the above embodiment, the quality monitoring device 41 envisions repeatedly making predictions, such as real-time prediction of antibody concentration Dp within the process, at short time intervals. Therefore, using the moving average of the spectroscopic spectrum during prediction is a technique aimed at improving the prediction accuracy in short time interval predictions, and a synergistic effect can be expected in relation to the model switching process involved in the technology of this invention.

[0174] (Example 1)

[0175] Hereinafter, as Example 1, it is shown that... Figures 14-16 The modified example of the first embodiment shown, determining the process conditions for the next step based on the final antibody concentration Dpf. Figure 23The variations shown and the output moving average are as follows. Figure 24 Examples of variations in the combination shown are embodiments.

[0176] In this embodiment, as shown in the above embodiments, the culture supernatant 17 is the culture supernatant of CHO cells, and a protein A column (manufactured by Cytiva, product name MabSelectSuRe) is used as the chromatographic column of the immunoaffinity chromatography apparatus 25. A UV detector 31 and a Raman spectrometer 32, as shown in the above embodiments, are installed downstream of the chromatographic column. The Raman spectroscopy measurement conditions during the immunoaffinity chromatography process are also as described in the above embodiments: laser output 500mW, laser wavelength 785nm, and exposure time 1 second. Furthermore, as... Figure 24 As shown, in the concentration prediction model 96A, the moving average of 10 spectral measurements 71 acquired at 1-second intervals was used. Figure 14 As shown, the model switching process is performed based on the process name of the sub-process. Additionally, valve V is switched based on whether the absorbance 64A exceeds a preset threshold of 1000 mAU. Furthermore, to calculate the final antibody concentration Dpf immediately after the process is completed, a flow meter is installed on the recovery line CL.

[0177] The immunoaffinity chromatography process was started with the low-concentration concentration prediction model 96A1 set. Monitoring was also initiated, and the over-time change in antibody concentration Dp was displayed. If the elution process was initiated, the quality monitoring device 41, referring to the process management information, switched to the high-concentration concentration prediction model 96A2 corresponding to the process name.

[0178] Then, the immunoaffinity chromatography apparatus 25 switches the eluent introduced into the column to an acidic buffer. With the acidic buffer, the antibody 19 adsorbed on the column begins to detach and flow downstream. The absorbance 64A, measured by the UV detector 31, begins to rise after a period of time following the start of the elution process. Simultaneously, the predicted value, i.e., the antibody concentration Dp, also begins to rise. When the absorbance 64A exceeds the preset threshold of 1000 mAU, valve V activates, switching the flow path of the liquid from the immunoaffinity chromatography apparatus 25 from the waste line WL to the recovery line CL. The antibody concentration Dp is recorded from this point. If almost all of the antibody 19 adsorbed on the column has eluted, the absorbance 64A begins to gradually decrease. If it falls below the threshold of 1000 mAU, valve V activates, switching the flow path from the recovery line CL back to the waste line WL. At this point, the recording of the flow rate of the first purified solution 28 through the recovery line CL and the antibody concentration Dp ends.

[0179] If the record ends, then... Figure 23Following the same procedure, processor 52 calculates the time integral value of antibody concentration Dp to determine the total amount of antibody 19. Furthermore, processor 52 calculates the total amount of the first purified solution 28 based on the time integral value of the flow rate measured by the flow meter, and determines the final antibody concentration Dpf by dividing the total amount of antibody 10 by the first purified solution 28. The final antibody concentration Dpf is 23.9 g / L. This is an estimated value; the actual measured value of the final antibody concentration Dpf, determined through offline analysis, is confirmed to be 24.1 g / L. The estimated final antibody concentration Dpf is close to the measured value.

[0180] As a comparative example, simulations of predicted values ​​were performed without model switching during the immunoaffinity chromatography process. Specifically, the concentration prediction model 96A was trained using learning data 71LA, which indiscriminately mixed spectral measurement data 71A obtained in the adsorption and washing processes and spectral measurement data 71A obtained in the elution process, representing different ranges of antibody concentration Dp. Then, the antibody concentration Dp was predicted using the concentration prediction model 96A, and... Figure 23 The method shown yielded the final antibody concentration Dpf. In the comparative example simulation, the final antibody concentration Dpf was 23.4 g / L.

[0181] As described above, in Example 1, where the model was switched between low and high concentrations based on the process name, the difference between the predicted and measured final antibody concentration Dpf was 0.2 g / L (24.1-23.9). In contrast, in the simulation of the comparative example, the difference between the predicted and measured final antibody concentration Dpf was 0.7 g / L (24.1-23.4), indicating a decrease in prediction accuracy compared to Example 1. This confirms the effectiveness of the model switching.

[0182] Furthermore, in order to add the recovered first purified solution 28 to the cation exchange chromatography in the next step, neutralization was performed with Tris-HCl (Tris Hydrochloride Acid Buffer). At this time, the final antibody concentration Dpf was diluted and decreased, but the final antibody concentration Dpf after neutralization was calculated to be 22.8 g / L based on the amount of neutralizing solution added. Then, the neutralized first purified solution 28 was introduced into the cation chromatography apparatus 26. The column adsorption capacity of the cation chromatography was 50 g / L LResin. By determining the final antibody concentration Dpf of the first purified solution 28 immediately after the completion of the previous step, it is possible to immediately calculate and determine the amount of neutralized first purified solution 28 that needs to be added in the next step, and the cation chromatography step of the next step can be started under these process conditions. As described above, Example 1 improved the prediction accuracy of the final antibody concentration Dpf by performing model switching, and therefore, compared with Comparative Example 1, which had low prediction accuracy, the process conditions of the next step can be appropriately determined.

[0183] (Example 2)

[0184] Example 2 is based on the concentration prediction model 96A used for automatic switching based on the absorbance 64A of the UV detector 31. Figures 17-19 Corresponding implementation examples.

[0185] In Example 2, the threshold for absorbance 64 used as the model switching point in the immunoaffinity chromatography step was set to 2500 mAU, based on the point where absorbance 64 of the UV detector 31 begins to saturate. As applicable conditions, the low-concentration concentration prediction model 96A1 was used when "absorbance 64A ≤ 2500 mAU", and the high-concentration concentration prediction model 96A2 was used when "absorbance 64A > 2500 mAU". The low-concentration concentration prediction model 96A1 was a model trained using learning data of antibody concentration Dp in the range of 0–7 g / L. The high-concentration concentration prediction model 96A2 was prepared using learning data of antibody concentration Dp in the range of 0–25 g / L. Thus, the learning data of the high-concentration concentration prediction model 96A2 includes high antibody concentration Dp learning data not included in the learning data of the low-concentration concentration prediction model 96A1.

[0186] The elution process was performed using a fraction collector, with the first purified solution 28 being aliquoted. During the elution process, the absorbance at 64A showed a tendency to temporarily rise to around 3000 mAU, eventually saturating, and then decreasing to 0 mAU. Therefore, model switching was performed in the following order: low concentration using concentration prediction model 96A1, high concentration using concentration prediction model 96A2, and low concentration using concentration prediction model 96A1 again. After the elution process, the first purified solution 28 aliquoted using the fraction collector was analyzed offline, and the prediction accuracy was evaluated using this analysis. The RMSE was 0.30. On the other hand, as always, when only high concentration using concentration prediction model 96A2 was used in the elution process, the RMSE was 0.39, and the prediction accuracy was improved by approximately 25%. Furthermore, if we only consider the predicted values ​​in the low-concentration region with absorbance ≤ 2500 mAU (64A), the RMSE of the low-concentration prediction model 96A1 is 0.16, while the RMSE of the previous high-concentration prediction model 96A2 is 0.33, indicating an improvement in prediction accuracy of over 50%. This confirms the effectiveness of the model switching process.

[0187] In the above example, the quality of manufacturing process 2, i.e., the liquid state, was described using antibody concentration as an example, but other examples are also possible. Besides antibody concentration, other factors include, for example, live cell density, cell state, glucose concentration, glutamate concentration, amino acid concentration, solvent component concentration, additive concentration, lactic acid concentration, ammonia concentration, concentration of cell-derived proteins and nucleic acids, concentration of other cell metabolites, antibody aggregate concentration, antibody fragment concentration, and charge isomer concentration. Among these, antibody concentration and antibody aggregate concentration are preferred, and antibody concentration is particularly preferred.

[0188] In the above example, Raman spectroscopy was used as an example of spectroscopic spectroscopy, but other examples are also possible, such as infrared spectroscopy, near-infrared spectroscopy, ultraviolet-visible (UV-Vis) spectroscopy, or fluorescence spectroscopy. Raman spectroscopy is particularly preferred because it readily reflects information about the functional groups of amino acids derived from proteins. This is to help predict the accuracy of the concentration of the protein, i.e., antibody 19.

[0189] Manufacturing process 2, which is the object of quality monitoring, will be described using purification process, i.e., process 12 (third process). Examples of steps included in process 12 include immunoaffinity chromatography, cation exchange chromatography, and anion exchange chromatography, but it is not limited to these. For example, manufacturing process 2, which is the object of quality monitoring, can include cell culture processes (cooked flow culture, perfusion culture) and cell separation processes (centrifugation, filtration, etc.) in biopharmaceutical manufacturing. Furthermore, as a step in the purification process, chromatographic steps can include size exclusion chromatography and hydrophobic interaction chromatography. In addition, virus inactivation processes, virus filtration processes, and filtration processes (ultrafiltration membranes, permeation membranes) can be cited. Among these, chromatographic steps are preferred for manufacturing process 2, especially immunoaffinity chromatography and cation exchange chromatography.

[0190] As a state prediction model, a machine learning model is illustrated, and as an algorithm for the machine learning model, a neural network 105 is illustrated, but it is not limited to this. Machine learning algorithms can be random forests, support vector machines, etc. Furthermore, it is not limited to machine learning models; it can also be multivariate analysis such as linear regression and logistic regression.

[0191] The process quality information, which serves as process management information, exemplifies the predicted antibody concentration and absorbance, but other examples are also possible, such as a range of antibody concentration and a range of aggregate concentration. Preferably, it includes the measured values ​​from a sensor such as a UV detector 31 that measures absorbance, and information calculated based on those measured values.

[0192] As an example of final quality, the final quality of antibody concentration is shown, but other examples of final quality can also be given. For example, final quality can include live cell density, cell state, glucose concentration, glutamate concentration, amino acid concentration, solvent component concentration, additive concentration, lactate concentration, ammonia concentration, concentration of cell-derived proteins and nucleic acids, concentration of other cellular metabolites, antibody aggregate concentration, antibody fragment concentration, and charge isomer concentration. Among these, the final quality of antibody concentration and antibody aggregate concentration is preferred, and the final quality of antibody concentration is particularly preferred.

[0193] In monitoring processing, the acquisition interval of data acquired as time series data can be output at intervals such as every hour, every minute, or every second, but preferably in the range of every minute to every second, and most preferably every second.

[0194] Biopharmaceuticals containing antibody 19 are called antibody pharmaceuticals, and they are widely used not only to treat chronic diseases such as cancer, diabetes, and rheumatoid arthritis, but also to treat rare diseases such as hemophilia and Crohn's disease. Therefore, targeting antibody 19 as the active ingredient could facilitate the development of antibody pharmaceuticals that can be widely used to treat various diseases.

[0195] Furthermore, the target component is not limited to antibodies 19. It can be aggregates 21, cytokines, hormones, etc. Additionally, cellular proteins / cellular DNA can also be used as target components.

[0196] In the above embodiments, for example, the hardware structure of a processing unit such as processor 52 that performs various processes can use various processors as shown below. As described above, the various processors include general-purpose processors such as CPU 47 that execute software (working program 75) to function as various processing units, as well as processors such as FPGA (Field Programmable Gate Array) whose circuit structure can be changed after manufacturing, i.e., programmable logic devices (PLDs), and processors such as ASIC (Application Specific Integrated Circuit) with circuit structures specifically designed for performing specific processes, i.e., dedicated circuits.

[0197] A processing unit can consist of one of these various processors, or it can consist of a combination of two or more processors of the same or different types (e.g., a combination of multiple FPGAs and / or a combination of a CPU and an FPGA). Furthermore, a single processor can also constitute multiple processing units.

[0198] As examples of a single processor comprising multiple processing units, firstly, as exemplified by client and server computers, there are processors that consist of a combination of more than one CPU and software, and that processor functions as multiple processing units. Secondly, as exemplified by System-on-Chip (SOC), there are processors that implement the overall system functionality including multiple processing units using a single integrated circuit (IC) chip. In these cases, various processing units are constructed as hardware structures using one or more of the aforementioned processors.

[0199] Furthermore, the hardware structure of these various processors, more specifically, can be a circuit composed of circuit elements such as semiconductor components.

[0200] Based on the above records, one can master the techniques described in the following notes.

[0201] [Note 1]

[0202] A quality monitoring device, comprising a processor,

[0203] The processor is capable of performing monitoring and model switching processes.

[0204] In this monitoring process, the spectroscopic spectra of liquids acquired online during the manufacturing process of biopharmaceuticals are used as input data. Multiple state prediction models that predict the state of the liquids in relation to the quality of the manufacturing process are then used to monitor the quality.

[0205] In this model switching process, based on process management information used to manage at least one process included in the manufacturing process, the state prediction model is switched midway through the monitoring process.

[0206] Furthermore, multiple state prediction models have completed learning before the monitoring process begins, and the process management information includes at least one of the following: process identification information for identifying at least one process included in the manufacturing process, equipment information for the equipment used in the process, and process quality information related to the quality of each process that can be obtained if at least one process has been performed.

[0207] [Note 2]

[0208] According to the quality monitoring device described in Appendix 1, wherein,

[0209] Model switching processing includes the process of switching the state prediction model based on process identification information or using equipment before a step in the manufacturing process begins.

[0210] [Note 3]

[0211] According to the quality monitoring device described in Appendix 1 or Appendix 2, wherein,

[0212] Model switching processing includes switching the state prediction model based on quality information that is acquired midway through the process, as part of the process quality information.

[0213] [Note 4]

[0214] According to the quality monitoring device described in Appendix 3, wherein...

[0215] In cases where the manufacturing process includes a purification process for the target components contained in the liquid,

[0216] The state of a liquid is the concentration of the constituent objects contained within it.

[0217] Quality information acquired midway through the process includes information acquired online during the process and related to the concentration of the target component in the liquid.

[0218] [Note 5]

[0219] According to the quality monitoring device described in Appendix 4, wherein,

[0220] The target component is an antibody.

[0221] [Note 6]

[0222] The quality monitoring device according to any one of appendices 1 to 5, wherein,

[0223] Spectroscopic spectroscopy includes any one of infrared spectroscopy, fluorescence spectroscopy, and Raman spectroscopy.

[0224] [Note 7]

[0225] The quality monitoring device according to any one of appendices 1 to 6, wherein,

[0226] The processor outputs a predicted value of the liquid's state at a pre-set first time interval based on the spectroscopic spectrum.

[0227] [Note 8]

[0228] According to the quality monitoring device described in Appendix 7, wherein,

[0229] The first time interval is less than 5 seconds.

[0230] [Note 9]

[0231] According to the quality monitoring device described in Appendix 8, wherein...

[0232] The first time interval is less than 1 second.

[0233] [Note 10]

[0234] The quality monitoring device according to any one of appendices 7 to 9, wherein,

[0235] The processor acquires the spectroscopic spectrum at a second time interval below the first time interval, and outputs the predicted value based on the moving average of the acquired spectroscopic spectra.

[0236] [Note 11]

[0237] The quality monitoring device according to any one of appendices 7 to 10, wherein,

[0238] The processor can output the final quality of the process, representing the final state of the overall liquid produced in the process, based on the predicted values.

[0239] [Note 12]

[0240] According to the quality monitoring device described in Appendix 11, wherein,

[0241] The processor performs a condition determination process to determine the process conditions for the next process based on the final quality of the previous process.

[0242] [Note 13]

[0243] The quality monitoring device according to any one of appendices 1 to 12, wherein,

[0244] In cases where the manufacturing process includes a purification process for the target components contained in the liquid,

[0245] The state of a liquid is the concentration of the constituent objects contained within it.

[0246] [Note 14]

[0247] According to the quality monitoring device described in Appendix 13, wherein,

[0248] The target component is an antibody.

[0249] [Note 15]

[0250] The quality monitoring device according to any one of appendices 1 to 14, wherein,

[0251] The state of a liquid refers to the state of the proteins contained within it.

[0252] [Note 16]

[0253] A method for operating a quality monitoring device, capable of performing monitoring processing and model switching processing.

[0254] In this monitoring process, the spectroscopic spectra of liquids acquired online during the manufacturing process of biopharmaceuticals are used as input data. Multiple state prediction models that predict the state of the liquids in relation to the quality of the manufacturing process are then used to monitor the quality.

[0255] In this model switching process, based on process management information used to manage at least one process included in the manufacturing process, the state prediction model is switched midway through the monitoring process.

[0256] Moreover, multiple state prediction models have completed learning before the monitoring process begins, and the process management information includes at least one of the following: process identification information for identifying at least one process included in the manufacturing process, equipment information for the equipment used in the process, and process quality information related to the state of the liquid in each process that can be obtained if at least one process has been performed.

[0257] [Note 17]

[0258] A working procedure for a quality monitoring device enables a computer to function as a quality monitoring device. This procedure includes the ability to perform monitoring processing and model switching processing.

[0259] In this monitoring process, the spectroscopic spectra of liquids acquired online during the manufacturing process of biopharmaceuticals are used as input data. Multiple state prediction models that predict the state of the liquids in relation to the quality of the manufacturing process are then used to monitor the quality.

[0260] In this model switching process, based on process management information used to manage at least one process included in the manufacturing process, the state prediction model is switched midway through the monitoring process.

[0261] Furthermore, multiple state prediction models have been created before the monitoring process begins, and the process management information includes at least one of the following: process identification information for identifying at least one process included in the manufacturing process, equipment information for the equipment used in the process, and process quality information related to the state of the liquid in each process that can be obtained if at least one process has been performed.

[0262] The technology of the present invention can also be appropriately combined with the various embodiments and / or variations described above. Furthermore, it is not limited to the embodiments described above; various structures can be adopted as long as they do not depart from the spirit of the invention. Moreover, the technology of the present invention relates not only to programs but also to storage media for non-temporary program storage.

[0263] The above descriptions and illustrations are detailed explanations of the parts related to the technology of this invention, and are merely examples of the technology of this invention. For example, the descriptions of the structure, function, effect, and effect described above are examples of the structure, function, effect, and effect of the parts related to the technology of this invention. Therefore, it is natural that unnecessary parts can be deleted and new elements can be added or replaced in the above descriptions and illustrations without departing from the spirit of this invention. Furthermore, to avoid complexity and facilitate understanding of the parts related to the technology of this invention, descriptions related to technical common sense that do not require special explanation have been omitted in the above descriptions and illustrations, based on the technology that enables the implementation of this invention.

[0264] In this specification, "A and / or B" has the same meaning as "at least one of A and B". That is, "A and / or B" means that it can be only A, only B, or a combination of A and B. Furthermore, in this specification, when more than three cases are combined using "and / or", the same approach as "A and / or B" applies.

[0265] The entire contents of Japanese Patent Application No. 2023-059234, filed on March 31, 2023, are incorporated herein by reference. Furthermore, all documents, patent applications, and technical standards described in this specification are incorporated herein by reference to the same extent as those specifically and individually described and incorporated by reference for individual documents, patent applications, and technical standards.

Claims

1. A quality monitoring device, comprising a processor, The processor is capable of performing monitoring and model switching processes. In the monitoring process, the spectroscopic spectrum of the liquid generated in the manufacturing process of biopharmaceuticals is acquired online as input data, and multiple state prediction models that predict the state of the liquid related to the quality of the manufacturing process are used to monitor the quality. In the model switching process, the state prediction model is switched midway through the monitoring process based on process management information used to manage at least one process included in the manufacturing process. Furthermore, multiple state prediction models have completed learning before the monitoring process begins, and the process management information includes at least one of the following: process identification information for identifying at least one process included in the manufacturing process, equipment information for the equipment used in the process, and process quality information related to the quality of each process that can be obtained if the process has been performed at least once.

2. The quality monitoring device according to claim 1, wherein, The model switching process includes switching the state prediction model based on the process identification information or the equipment used before the start of one of the steps in the manufacturing process.

3. The quality monitoring device according to claim 1, wherein, The model switching process includes switching the state prediction model based on quality information that is one type of process quality information and is acquired midway through the process.

4. The quality monitoring device according to claim 3, wherein, In the case where the manufacturing process includes a purification process for the target components contained in the liquid. The state of the liquid refers to the concentration of the target component contained in the liquid. The quality information acquired during the process includes information acquired online during the process and related to the concentration of the target component in the liquid.

5. The quality monitoring device according to claim 4, wherein, The object component is an antibody.

6. The quality monitoring device according to claim 1, wherein, The spectroscopic spectrum includes any one of infrared spectrum, fluorescence spectrum and Raman spectrum.

7. The quality monitoring device according to claim 1, wherein, The processor outputs a predicted value of the state of the liquid at a preset first time interval based on the spectroscopic spectrum.

8. The quality monitoring device according to claim 7, wherein, The first time interval is less than 5 seconds.

9. The quality monitoring device according to claim 8, wherein, The first time interval is less than 1 second.

10. The quality monitoring device according to claim 7, wherein, The processor acquires the spectroscopic spectrum at a second time interval below the first time interval, and outputs the predicted value based on the moving average of the acquired spectroscopic spectra.

11. The quality monitoring device according to claim 7, wherein, The processor is able to output the final quality of the process, representing the final state of the liquid as a whole produced in the process, based on the predicted value.

12. The quality monitoring device according to claim 11, wherein, The processor performs a condition determination process to determine the process conditions for the next process based on the final quality of the process.

13. The quality monitoring device according to claim 1, wherein, In the case where the manufacturing process includes a purification process for the target components contained in the liquid. The state of the liquid refers to the concentration of the object component contained in the liquid.

14. The quality monitoring device according to claim 13, wherein, The object component is an antibody.

15. The quality monitoring device according to claim 1, wherein, The state of the liquid refers to the state of the proteins contained within it.

16. A method for operating a quality monitoring device, capable of performing monitoring processing and model switching processing. In the monitoring process, the spectroscopic spectrum of the liquid generated in the manufacturing process of biopharmaceuticals is acquired online as input data, and multiple state prediction models that predict the state of the liquid related to the quality of the manufacturing process are used to monitor the quality. In the model switching process, the state prediction model is switched midway through the monitoring process based on process management information used to manage at least one process included in the manufacturing process. Furthermore, multiple state prediction models have completed learning before the monitoring process begins, and the process management information includes at least one of the following: process identification information for identifying at least one process included in the manufacturing process, equipment information for the equipment used in the process, and process quality information related to the state of the liquid in each process that can be obtained if the process has been performed at least once.

17. A working procedure for a quality monitoring device, enabling a computer to function as a quality monitoring device, wherein the working procedure of the quality monitoring device is capable of performing monitoring processing and model switching processing. In the monitoring process, the spectroscopic spectrum of the liquid generated in the manufacturing process of biopharmaceuticals is acquired online as input data, and multiple state prediction models that predict the state of the liquid related to the quality of the manufacturing process are used to monitor the quality. In the model switching process, the state prediction model is switched midway through the monitoring process based on process management information used to manage at least one process included in the manufacturing process. Furthermore, multiple state prediction models have been created prior to the start of the monitoring process, and the process management information includes at least one of the following: process identification information for identifying at least one process included in the manufacturing process, equipment information for the equipment used in the process, and process quality information related to the state of the liquid in each process that can be obtained if the process has been performed at least once.

Images