Information processing device, operation method for information processing device, and operation program for information processing device

Abstract

An information processing device according to the present invention comprises a processor, wherein the processor inputs, into a trained large-scale language model, related data obtained in a manufacturing process for manufacturing a product, and outputs, from the trained large-scale language model, a cause that brought about a change in quality of the product and / or a countermeasure against the change in the quality of the product.

Description

Information processing device, method of operating the information processing device, and operating program for the information processing device.

[0001] The technology disclosed herein relates to an information processing device, a method for operating an information processing device, and an operating program for an information processing device.

[0002] In various production processes for manufacturing products, trained models are used to predict the quality of the products. For example, Japanese Patent Publication No. 2023-000828 describes a technique in which the difference in relevant data at two different points in time, before and after a process performed in the production process, is input into a trained model, and the trained model outputs the quality of the product. In Japanese Patent Publication No. 2023-000828, protein is given as the product, protein concentration as the quality, chromatography as the process, and spectral data as the relevant data.

[0003] In the production process, when the quality of a product improves or deteriorates, operators (including employees of the production company or researchers at research institutions) have traditionally investigated the root cause of such quality changes and devised countermeasures. However, for operators to investigate the cause of a quality change or devise countermeasures requires a vast amount of analysis time and highly specialized knowledge. Therefore, there has been a demand for the development of a technology that can easily predict the cause of a quality change and / or countermeasures for a quality change using a pre-trained model, similar to the technology described in Japanese Patent Publication No. 2023-000828, which uses a pre-trained model to predict the quality of a product.

[0004] One embodiment of the technology of this disclosure provides an information processing device, a method for operating the information processing device, and an operating program for the information processing device that can easily predict the causes of changes in product quality and / or countermeasures for changes in quality using a trained model.

[0005] The information processing device of the present disclosure includes a processor which inputs relevant data obtained in a production process for producing a product into a trained large-scale language model and causes the trained large-scale language model to output at least one of the causes that led to the change in the quality of the product and countermeasures to address the change in the quality of the product.

[0006] The relevant data preferably includes comparative data with reference data obtained in past production processes.

[0007] The comparison data has multiple variables, and it is preferable for the processor to selectively input the variables that satisfy a pre-set first criterion into the trained large-scale language model.

[0008] The reference data is preferably data that meets the pre-defined second criterion.

[0009] The relevant data preferably includes at least one of the following: physical property data of the substance including the product, condition data of the processing carried out in the production process, measurement data measured during the processing carried out in the production process, and prediction data of the product quality.

[0010] It is preferable that the physical property data include spectral data.

[0011] Spectroscopic data is preferably Raman spectral data.

[0012] The product is preferably a protein.

[0013] The pre-trained large-scale language model is preferably a model that has been trained on regions where characteristic changes appear in spectral data.

[0014] A pre-trained large-scale language model is preferably a model trained using a combination of relevant training data and ground truth data.

[0015] The related data is physical property data of the substance including the product, and preferably includes at least one of the following: physical property data including spectral data, setting condition data for the treatment carried out in the production process, measurement data measured during the treatment carried out in the production process, and prediction data for the quality of the product, and preferably includes comparison data with reference data obtained in past production processes.

[0016] The method of operating the information processing device of this disclosure includes inputting relevant data obtained in a production process for producing a product into a trained large-scale language model, and having the trained large-scale language model output at least one of the causes that led to the change in the quality of the product, and countermeasures to address the change in the quality of the product.

[0017] The operating program for the information processing device of this disclosure causes a computer to perform a process that includes inputting relevant data obtained in a production process for producing a product into a trained large-scale language model, and outputting from the trained large-scale language model at least one of the causes that led to the change in the quality of the product, and countermeasures to address the change in the quality of the product.

[0018] The technology of this disclosure provides an information processing device, a method for operating the information processing device, and an operating program for the information processing device that can easily predict the causes of changes in product quality and / or countermeasures for changes in quality using a trained model.

[0019] This is a diagram showing the culture unit, Raman spectrometer, and information processing unit. This is a diagram showing the excitation light and Raman scattered light. This is a diagram showing the target Raman spectrum data. This is a block diagram of the computer constituting the information processing unit. This is a block diagram of the CPU processing unit of the information processing unit. This is a diagram showing the processing of the first prediction unit. This is a diagram showing the reference data. This is a diagram showing the processing of the comparison unit. This is a diagram showing the processing of the comparison unit. This is a diagram showing the prompt input screen. This is a diagram showing the processing of the second prediction unit. This is a diagram showing the report display screen. This is a diagram showing the learning flow of a large-scale language model. This is a diagram showing the processing in the learning phase of a large-scale language model. This is a diagram showing the processing flow of the information processing unit. This is a table showing the culture conditions in the example. This is a table showing the Raman spectrum measurement conditions in the example. This is a diagram showing the prompts in the example. This is a diagram showing the prompts in the example. This is a diagram showing the report in the example. This is a diagram showing the report in the example.

[0020] As an example, as shown in Figure 1, the culture unit 10 is responsible for the cell culture process in the biopharmaceutical manufacturing system. The culture unit 10 has a culture tank 11. Cell culture medium 12 is stored in the culture tank 11. Antibody-producing cells 13 are seeded in the cell culture medium 12, and the antibody-producing cells 13 are cultured in the cell culture medium 12. The culture method can be either perfusion culture or fed-batch culture. Antibody-producing cells 13 are cells established by incorporating antibody genes into host cells, such as Chinese hamster ovary cells (CHO cells). Antibody-producing cells 13 produce immunoglobulins, i.e., antibodies 14, during the culture process. Therefore, not only antibody-producing cells 13 but also antibodies 14 are present in the cell culture medium 12. The antibody 14 is, for example, a monoclonal antibody and becomes the active ingredient of the biopharmaceutical. The cell culture medium 12 contains, in addition to antibody-producing cells 13 and antibodies 14, cell-derived proteins, cell-derived DNA (Deoxyribonucleic Acid), antibody fragments 14A, antibody aggregates 14B, or viruses, etc. The cell culture process is an example of a "production process" relating to the technology of this disclosure. Culture is an example of a "process performed in the production process" relating to the technology of this disclosure. The cell culture medium 12 is an example of a "substance containing a product" relating to the technology of this disclosure. Antibody 14 is an example of a "product" and "protein" relating to the technology of this disclosure.

[0021] The culture tank 11 is equipped with a gas supply pipe 15 and a stirring blade 16. The gas supply pipe 15 supplies gas 17, such as oxygen, from the outside into the culture tank 11. The stirring blade 16 rotates at a preset stirring speed (rotation speed) to agitate the cell culture medium 12 inside the culture tank 11.

[0022] A cell removal filter (not shown) is attached to the culture vessel 11. The cell removal filter uses a filter membrane to capture antibody-producing cells 13 in the cell culture medium 12, for example, using a tangential flow filtration (TFF) method or an alternating tangential flow filtration (ATF) method, thereby removing the antibody-producing cells 13 from the cell culture medium 12. The cell removal filter also allows antibodies 14 to pass through. Therefore, the cell culture medium 12 output from the culture unit 10 mainly contains antibodies 14. This cell culture medium 12 is called the culture supernatant. The culture supernatant is sent to a purification unit located downstream of the cell removal filter. There, it undergoes various chromatographic treatments and virus inactivation treatments to become a purified solution. The purified solution is further concentrated and filtered by ultrafiltration (UF) and diafiltration (DF). The active pharmaceutical ingredient for biopharmaceuticals is obtained from the purified solution processed in this way.

[0023] A Raman spectrometer 20 is installed in the culture section 10. As an example, as shown in Figure 2, the Raman spectrometer 20 is an instrument that evaluates material M using the characteristics of Raman scattered light RSL. When excitation light EL is irradiated onto material M, the excitation light EL interacts with material M, generating Raman scattered light RSL with a different wavelength from the excitation light EL. The wavelength difference between the excitation light EL and the Raman scattered light RSL corresponds to the energy of the molecular vibrations of material M. Therefore, Raman scattered light RSL with different wavenumbers can be obtained between materials M with different molecular structures. Of the Stokes lines and anti-Stokes lines, it is preferable to use Stokes lines for the Raman scattered light RSL.

[0024] Returning to Figure 1, the Raman spectrometer 20 consists of a sensor unit 21 and an analyzer 22. The sensor unit 21 is immersed in the cell culture medium 12. The sensor unit 21 emits excitation light EL from its tip. The excitation light EL is irradiated onto the cell culture medium 12. The interaction between this excitation light EL and components such as antibodies 14 in the cell culture medium 12 generates Raman scattered light RSL. The sensor unit 21 receives the Raman scattered light RSL and outputs the received Raman scattered light RSL to the analyzer 22. Alternatively, a flow cell may be provided in the piping through which the cell culture medium 12 flows, and the sensor unit 21 may be set in the flow cell to measure the Raman scattered light RSL.

[0025] The analyzer 22 decomposes the Raman scattered light RSL into wavenumbers and derives the intensity value of the Raman scattered light RSL for each wavenumber, thereby generating target Raman spectrum data 23 representing the Raman spectrum, i.e., the Raman spectrum, as shown in Figure 3 as an example. The target Raman spectrum data 23 is data in which the intensity values ​​of the Raman scattered light RSL for each wavenumber are registered. In this example, the target Raman spectrum data 23 is for wavenumber 300 cm⁻¹. －１ ~3100cm －１ The intensity values ​​of Raman scattered light RSL in the range up to 1 cm －１ This data was derived in increments. The target Raman spectrum data 23 includes not only the intensity value of Raman scattered light RSL but also the number of culture days. In Figure 3, the graph shown below the target Raman spectrum data 23 plots the intensity values ​​of the target Raman spectrum data 23 for each wavenumber and connects them with lines. The target Raman spectrum data 23 is an example of "physical property data" and "spectral spectrum data" related to the technology of this disclosure. The intensity values ​​of the target Raman spectrum data 23 are an example of "variables" related to the technology of this disclosure.

[0026] In this way, the Raman spectrometer 20 irradiates the cell culture medium 12 stored in the culture vessel 11 with excitation light EL through the sensor unit 21, thereby measuring the Raman scattered light RSL of components such as antibodies 14 in the cell culture medium 12 and obtaining the target Raman spectral data 23.

[0027] Returning to FIG. 1 again, in the culture unit 10, target condition data 24 including various conditions of the culture performed in the cell culture process is generated. In the target condition data 24, for example, target values such as the hydrogen ion concentration (pH: potential of Hydrogen) of the cell culture solution 12, temperature, dissolved oxygen, stirring speed by the stirring blade 16, and nutrient addition amount are registered. In addition to the target values of the above various conditions, the number of culture days is also registered in the target condition data 24. The target condition data 24 is an example of the "condition data" according to the technology of the present disclosure. Further, the target condition data 24 is an example of the "related data" according to the technology of the present disclosure.

[0028] Also, in the culture unit 10, measurement data 25 of various conditions of the culture performed in the cell culture process is generated based on the measurement results of various measurement sensors (not shown). In the target measurement data 25, measured values such as the hydrogen ion concentration, temperature, dissolved oxygen, stirring speed by the stirring blade 16, and nutrient addition amount of the cell culture solution 12 are registered. These measured values include time-series data measured at set intervals from the start to the end of the culture for one day. In addition to the measured values of the above various conditions, the number of culture days is also registered in the target measurement data 25. The target measurement data 25 is an example of the "measurement data" according to the technology of the present disclosure. Further, each variable such as the pH of the target measurement data 25 is an example of the "variable" according to the technology of the present disclosure. Note that the target measurement data 25 can include image data obtained from various imaging sensors such as cameras and / or structural data representing the shapes of antibody-producing cells 13, antibodies 14, etc. Further, the target condition data 24 and the target measurement data 25 can include condition data and measurement data in processes related to the cell culture process, for example, condition data and measurement data of various chromatography processes in the purification process by the purification unit.

[0029] The Raman spectrometer 20 is connected to an information processing device 30 through a computer network such as a LAN (Local Area Network). Target Raman spectrum data 23 is transmitted from the Raman spectrometer 20 to the information processing device 30. Further, target condition data 24 and target measurement data 25 are transmitted from the culture unit 10 to the information processing device 30.

[0030] The information processing device 30 is, for example, a desktop personal computer, and includes a display 31 for displaying various screens and an input device 32 such as a keyboard, a mouse, a touch panel, and / or a microphone for voice input. The information processing device 30 is installed in, for example, a pharmaceutical company developing biopharmaceuticals or an organization that commissions the development of biopharmaceuticals from a pharmaceutical company, that is, a contract research organization (CRO: Contract Research Organization). The information processing device 30 is operated by an operator OP involved in the development of biopharmaceuticals in a pharmaceutical company or a contract research organization (hereinafter collectively referred to as a pharmaceutical facility).

[0031] As shown in FIG. 4 as an example, the computer constituting the information processing device 30 includes, in addition to the aforementioned display 31 and input device 32, a storage 35, a memory 36, a CPU (Central Processing Unit) 37, and a communication unit 38. These are interconnected via a bus line 39.

[0032] The storage 35 is a hard disk drive built into the computer constituting the information processing device 30 or connected through a cable or a network. Alternatively, the storage 35 is a disk array with multiple hard disk drives installed. The storage 35 stores control programs such as an operating system, various application programs, and various data associated with these programs. Note that a solid state drive may be used instead of the hard disk drive.

[0033] Memory 36 is a work memory for the CPU 37 to execute processing. The CPU 37 loads the program stored in storage 35 into memory 36 and executes processing according to the program. In this way, the CPU 37 comprehensively controls each part of the computer. CPU 37 is an example of a "processor" related to the technology of this disclosure. Note that memory 36 may be built into the CPU 37. The communication unit 38 controls the transmission of various information with external devices such as the Raman spectrometer 20.

[0034] As an example, as shown in Figure 5, the storage 35 stores an operating program 45. The operating program 45 is an application program that causes the computer to function as an information processing device 30. In other words, the operating program 45 is an example of an "operating program for an information processing device" related to the technology of this disclosure. In addition to the operating program 45, the storage 35 also stores a prediction model 46, reference data 47, selection criteria 48, and a trained large-scale language model 49, etc. The selection criteria 48 is an example of a "pre-set criterion" related to the technology of this disclosure.

[0035] When the operating program 45 is started, the CPU 37 of the computer constituting the information processing device 30 works in cooperation with the memory 36 and the like to function as an acquisition unit 55, a read / write (hereinafter abbreviated as RW (Read / Write)) control unit 56, a first prediction unit 57, a comparison unit 58, a display control unit 59, an instruction reception unit 60, and a second prediction unit 61.

[0036] The acquisition unit 55 acquires target Raman spectrum data 23 from the Raman spectrometer 20, and target condition data 24 and target measurement data 25 from the culture tank 11. The acquisition unit 55 applies baseline correction to the target Raman spectrum data 23. The acquisition unit 55 outputs the baseline-corrected target Raman spectrum data 23, target condition data 24, and target measurement data 25 to the RW control unit 56.

[0037] The RW control unit 56 controls the storage of various data to the storage 35 and the reading of various data stored in the storage 35. The RW control unit 56 stores the target Raman spectrum data 23, target condition data 24, and target measurement data 25 from the acquisition unit 55 in the storage 35. The RW control unit 56 also reads the target Raman spectrum data 23, target condition data 24, and target measurement data 25 from the storage 35 and outputs the target Raman spectrum data 23 to the first prediction unit 57 and comparison unit 58, the target condition data 24 to the display control unit 59 (arrow omitted), and the target measurement data 25 to the comparison unit 58.

[0038] The RW control unit 56 reads the prediction model 46 from the storage 35 and outputs the prediction model 46 to the first prediction unit 57. The RW control unit 56 also reads the reference data 47 and selection criteria 48 from the storage 35 and outputs the reference data 47 and selection criteria 48 to the comparison unit 58. Furthermore, the RW control unit 56 reads the trained large-scale language model 49 from the storage 35 and outputs the trained large-scale language model 49 to the second prediction unit 61.

[0039] The first prediction unit 57 inputs the target Raman spectrum data 23 into the prediction model 46 and causes the prediction model 46 to output the target prediction data 65. The first prediction unit 57 outputs the target prediction data 65 to the RW control unit 56. The RW control unit 56 stores the target prediction data 65 in the storage 35. The RW control unit 56 also reads the target prediction data 65 from the storage 35 and outputs the target prediction data 65 to the comparison unit 58. The culture period is registered in the target prediction data 65, just like in the target Raman spectrum data 23.

[0040] The comparison unit 58 compares the target Raman spectrum data 23, target measurement data 25, and target prediction data 65 with reference data 47 whose culture period matches or is the closest to them, and generates comparison data 66. The reference data 47 with the closest culture period is, for example, if the culture period of the target Raman spectrum data 23, etc., is 5.6 days, then the reference data 47 with a culture period of 6 days. The comparison unit 58 outputs the comparison data 66 to the RW control unit 56. The RW control unit 56 stores the comparison data 66 in the storage 35. The RW control unit 56 also reads the comparison data 66 from the storage 35 and outputs the comparison data 66 to the display control unit 59. The comparison data 66 is an example of "related data" related to the technology of this disclosure.

[0041] The display control unit 59 controls the display of various screens on the display 31. These various screens include a prompt input screen 100 (see Figure 11) for inputting prompts 102 to be given to the trained large-scale language model 49, and a report display screen 105 (see Figure 13) for displaying reports 67 output from the trained large-scale language model 49.

[0042] The instruction receiving unit 60 receives various operation instructions input by the operator OP via the input device 32. These operation instructions include input instructions for the prompt 102, etc.

[0043] The second prediction unit 61 inputs the prompt 102 to the trained large-scale language model 49 and causes the trained large-scale language model 49 to output a report 67. The second prediction unit 61 outputs the report 67 to the display control unit 59.

[0044] As an example, as shown in Figure 6, the prediction model 46 consists of an antibody concentration prediction model 70, a glucose concentration prediction model 71, and a lactate concentration prediction model 72. These three models—antibody concentration prediction model 70, glucose concentration prediction model 71, and lactate concentration prediction model 72—are all based on machine learning models using algorithms such as decision trees, gradient boosting decision trees, naive Bayes, random forests, support vector machines, and neural networks. The antibody concentration prediction model 70, glucose concentration prediction model 71, and lactate concentration prediction model 72 predict the concentrations of antibody 14, glucose, and lactate in the cell culture medium 12. Glucose is the energy source for antibody-producing cells 13, and lactate is a waste product of antibody-producing cells 13. The concentrations of these antibodies 14, glucose, and lactate are important indicators for determining whether the culture of antibody-producing cells 13 is successful.

[0045] The first prediction unit 57 inputs the intensity values ​​of each wavenumber of the target Raman spectrum data 23 to the antibody concentration prediction model 70, the glucose concentration prediction model 71, and the lactate concentration prediction model 72, respectively. The antibody concentration prediction model 70 outputs antibody concentration prediction data 73, the glucose concentration prediction model 71 outputs glucose concentration prediction data 74, and the lactate concentration prediction model 72 outputs lactate concentration prediction data 75. These antibody concentration prediction data 73, glucose concentration prediction data 74, and lactate concentration prediction data 75 constitute the target prediction data 65.

[0046] Furthermore, the intensity values ​​of the wavenumber region in which characteristic changes of antibody 14 appear (hereinafter referred to as the characteristic wavenumber region) may be selectively input into the antibody concentration prediction model 70. Similarly, the intensity values ​​of the characteristic wavenumber regions for glucose and lactate may be selectively input into the glucose concentration prediction model 71 and the lactate concentration prediction model 72. In addition, target condition data 24 and / or target measurement data 25 may be input into the antibody concentration prediction model 70, glucose concentration prediction model 71, and lactate concentration prediction model 72, in addition to or instead of the target Raman spectrum data 23.

[0047] For the characteristic frequency region, in the case of antibody 14, it is a region attributed to the amide bond and / or disulfide bond of the protein, a region attributed to aromatic amino acids, and a region attributed to carbon-carbon bonds, carbon-nitrogen bonds, and carbon-oxygen bonds. More specifically, for the characteristic frequency region, in the case of antibody 14, for example, wavenumbers 757 cm －１ , 1003 cm －１ , 1029 cm －１ , 1208 cm －１ , 1236 cm<s －１ , 1337 cm －１ , 1447 cm －１ , 1553 cm －１ , and a region including 1673 cm －１ . For the characteristic frequency region in the case of lactic acid, for example, it is a region including wavenumbers 855 cm －１ , 930 cm －１ , 1044 cm －１ , 1420 cm －１ , and a region including 1457 cm －１ .

[0048] The characteristic frequency region may be a region attributed to at least any one of the amide bond, disulfide bond, aromatic amino acids, carbon-carbon bond, carbon-nitrogen bond, and carbon-oxygen bond of the protein. Therefore, for the characteristic frequency region, in the case of antibody 14, for example, it may be at least any one of the regions including wavenumbers 757 cm －１ , 1003 cm －１ , 1029 cm －１ , 1208 cm －１ , 1236 cm －１ , 1337 cm －１ , 1447 cm －１ , 1553 cm －１ , and a region including 1673 cm －１ . For the characteristic frequency region in the case of lactic acid, for example, it may be at least any one of the regions including wavenumbers 855 cm －１ , 930 cm －１ , 1044 cm －１ , 1420 cm －１ , and a region including 1457 cm －１ .

[0049] In the case of antibody 14, the characteristic wavenumber region is more preferably the region associated with the amide bond. That is, the characteristic wavenumber region is wavenumber 1236 cm⁻¹. －１ , 1337cm －１ , and 1673cm －１ It is more preferable that the region includes [the specified element].

[0050] As an example, as shown in Figure 7, reference data 47 represents data from a past cell culture process that achieved the best antibody production results. The best antibody production results refer to, for example, the case where the yield of antibody 14 was the highest. Alternatively, the best antibody production results may also refer to the case where the concentration of antibody 14 was the highest. The data representing the best antibody production results is an example of "data that meets the pre-set second criterion" relating to the technology of this disclosure. That is, the second criterion in this example is whether or not the best antibody production results were achieved. The second criterion may also be data in which the antibody production results are above a set level and the antibody produced is the same as or similar to the target antibody 14.

[0051] Reference data 47 includes reference Raman spectrum data 80, reference measurement data 81, and reference prediction data 82. These reference Raman spectrum data 80, reference measurement data 81, and reference prediction data 82 are registered for each culture day, such as day 1, day 2, etc. Reference Raman spectrum data 80 is data measured by the Raman spectrometer 20 in past cell culture processes. Reference measurement data 81 is data measured in the culture vessel 11 in past cell culture processes. Reference prediction data 82 is data obtained by inputting the reference Raman spectrum data 80 into the prediction model 46. Specifically, reference Raman spectrum data 80 corresponds to the target Raman spectrum data 23, and reference measurement data 81 corresponds to the target measurement data 25. Also, reference prediction data 82 corresponds to the target prediction data 65. More specifically, reference prediction data 82, like the target prediction data 65, includes antibody concentration prediction data, glucose concentration prediction data, and lactate concentration prediction data. Reference data 47 may also include data from days 0.5 and 1.5 of culture.

[0052] As an example, as shown in Figure 8, the comparison unit 58 calculates the mean value AVE and standard deviation Σ of the intensity values ​​of the reference Raman spectrum data 80. Then, the target Raman spectrum data 23 is standardized using the mean value AVE and standard deviation Σ to become the standardized target Raman spectrum data 23S. Specifically, if the intensity value of the target Raman spectrum data 23 is X and the intensity value of the standardized target Raman spectrum data 23S is Y, standardization is performed by calculating the following equation (1): Y = (X - AVE) / Σ (1)

[0053] The comparison unit 58 selects intensity values ​​Y for each wavenumber of the standardized Raman spectral data 23S that satisfy the selection criteria 48. In this case, the selection criteria 48 is that the absolute value of Y is greater than 3 (|Y| > 3). The comparison unit 58 extracts the intensity values ​​of the standardized Raman spectral data 23S corresponding to the wavenumber of the selected intensity value Y and generates comparative Raman spectral data 86. Alternatively, instead of Y in the above formula (1), the ratio of the intensity values ​​of the target Raman spectral data 23 to the average value AVE may be calculated, and a ratio that satisfies the selection criteria may be selected as comparative Raman spectral data 86. Alternatively, the intensity value Y may be selected based on the coefficient of variation.

[0054] Similarly, as shown in Figure 9 as an example, the comparison unit 58 calculates the mean value ave and standard deviation σ of the time-series data for each variable, such as pH, in the reference measurement data 81. Then, the target measurement data 25 is standardized using the mean value ave and standard deviation σ to obtain the standardized target measurement data 25S. Specifically, if the variable of the target measurement data 25 is x and the variable of the standardized target measurement data 25S is y, standardization is performed by calculating the following equation (2): y = (x - ave) / σ (2)

[0055] The comparison unit 58 selects a value y from each variable y in the standardized measurement data 25S that satisfies the selection criterion 48. In this case, the selection criterion 48 is that the absolute value of y is greater than 3 (|y| > 3). The comparison unit 58 extracts the variables from the standardized measurement data 25S corresponding to the selected variable y and generates comparative measurement data 91. If there is no value y that satisfies the selection criterion 48, the comparison unit 58 generates comparative measurement data 91 with the content "Not applicable," as exemplified. Alternatively, instead of y in the above formula (2), the ratio of each variable to the mean value ave may be calculated, and the ratio that satisfies the selection criterion may be selected as comparative measurement data 91. Alternatively, the intensity value y may be selected based on the coefficient of variation.

[0056] As an example, as shown in Figure 10, the comparison unit 58 generates standardized prediction data (hereinafter referred to as comparative prediction data) 95 for each variable of the target prediction data 65 (antibody concentration prediction data 73, glucose concentration prediction data 74, and lactate concentration prediction data 75) and each variable of the reference prediction data 82. The comparison data 66 is composed of these comparative Raman spectrum data 86, comparative measurement data 91, and comparative prediction data 95 (see also Figure 12).

[0057] When comparison data 66 is input from the RW control unit 56, the display control unit 59 displays a prompt input screen 100 on the display 31, as shown in Figure 11 as an example. The prompt input screen 100 displays the target condition data 24. The prompt input screen 100 also displays the comparison data 66, which consists of comparison Raman spectrum data 86, comparison measurement data 91, and comparison prediction data 95. Insert buttons 101 are provided at the bottom of each of the display areas for the target condition data 24, comparison Raman spectrum data 86, comparison measurement data 91, and comparison prediction data 95. The operator OP selects the insert button 101 when inserting the target condition data 24, comparison Raman spectrum data 86, comparison measurement data 91, and comparison prediction data 95 into the prompt 102.

[0058] In the comparative measurement data 91, a variable labeled "no abnormality" means that there were no variables y that met the selection criteria 48. Similarly, in the comparative prediction data 95, a variable labeled "no abnormality" means that the value of the comparative prediction data 95 was below a predetermined threshold (for example, three times the standard deviation of the reference prediction data 82).

[0059] The operator OP inserts the target condition data 24 and comparison data 66 by selecting the insert button 101, etc., while inputting the prompt 102. The prompt 102 is a sentence to be input to the trained large-scale language model 49, and includes a sentence asking about the cause of the change in antibody quality 14 and the countermeasures to the change in antibody quality 14, which can be predicted based on the target condition data 24 and comparison data 66. Here, "countermeasures" includes not only measures to achieve the target quality, but also measures to further improve quality beyond the target quality.

[0060] An input button 103 is provided at the bottom of the input area for prompt 102. After entering prompt 102, the operator OP selects the input button 103. When the input button 103 is selected, the instruction receiving unit 60 receives an input instruction for prompt 102. The input instruction includes prompt 102. The instruction receiving unit 60 outputs prompt 102 to the second prediction unit 61.

[0061] When prompt 102 is input from the instruction receiving unit 60, the second prediction unit 61 inputs prompt 102 to the trained large-scale language model 49, as shown in Figure 12 as an example. The trained large-scale language model 49 then outputs a report 67. The second prediction unit 61 outputs the report 67 to the display control unit 59.

[0062] When a report 67 is input from the second prediction unit 61, the display control unit 59 displays a report display screen 105 on the display 31, as shown in Figure 13 as an example. The report display screen 105 literally displays the report 67. The report 67 includes text describing the cause of the change in the quality of the antibody 14 and countermeasures to address the change in the quality of the antibody 14. In Figure 13, an example of a report 67 is shown in which poor nutrient supply to antibody-producing cells 13 is cited as the cause, and countermeasures to address poor nutrient supply are listed.

[0063] A save button 106 and an OK button 107 are provided at the bottom of the report display screen 105. When the operator OP selects the save button 106, the report 67 is stored in the storage 35 under the control of the RW control unit 56. When the operator OP selects the OK button 107, the display on the report display screen 105 is turned off.

[0064] Figure 11 shows an example in which operator OP inputs prompt 102, but it is not limited to this. Alternatively, a template for prompt 102 may be prepared, and prompt 102 may be automatically generated by fitting the target condition data 24 and comparison data 66 into the template, and the automatically generated prompt 102 may be input into the trained large-scale language model 49. In this case, the instruction receiving unit 60 or the second prediction unit 61 is input with the target condition data 24 and comparison data 66 to be fitted into the prompt 102 template.

[0065] In addition to making report 67 available for viewing by operators (OP) via the report display screen 105 shown in Figure 13, or alternatively, it may be handled as follows: Report 67 may be uploaded to a cloud server or the like, and operators (OP) and other personnel involved in the development of biopharmaceuticals may access the server to view, download, etc.

[0066] Figures 11 and 13 show the case where the culture period is 10 days, but this is merely an example. The input of prompt 102 and the output of report 67 may be limited to specific days such as the 10th day or the final day of culture, or they may be performed regularly, such as every hour, every 12 hours, daily, or every 5 days. When the input of prompt 102 and the output of report 67 are performed regularly, it is preferable to retain the contextual information of previously inputted prompts 102 and reports 67, and to reflect this contextual information when outputting report 67 in response to a new prompt 102 input.

[0067] As an example, as shown in Figure 14, the pre-trained large-scale language model 49 is obtained by pre-training an untrained natural language model 110 and then fine-tuning the resulting pre-trained natural language model 111. The natural language model 110 is based on a model such as BERT (Bidirectional Encoder Representations from Transformers), which uses a transformer encoder.

[0068] Pre-training includes MLM (Masked Language Modeling) and NSP (Next Sentence Prediction). MLM is a learning method that masks parts of the input text and asks the model to predict what words will go in the masked areas; it is a so-called fill-in-the-blank problem. NSP is a learning method that asks the model to determine whether two different texts are related or not. For pre-training, a vast amount of previously published text is used, such as dictionaries, novels, newspaper articles, magazine articles, news articles, or various research papers, including those related to cell culture. In addition, drawing data of production facilities used in various production processes, including cell culture processes (data including where and what kind of equipment is located, how they are connected, and the size of the equipment, etc.), and / or process data of various production processes (data including the content of the process and its order, etc.) may also be used. Fine-tuning is learning that is performed by the trained large-scale language model 49 according to the desired processing task.

[0069] As an example, as shown in Figure 15, the trained large-scale language model 49 after fine-tuning is subjected to additional training using training data 115. The training data 115 is a pair of a first training prompt 1021L and a correct answer report 67CA. The first training prompt 1021L includes training condition data 24L and training comparison data 66L obtained in past cell culture processes. The training comparison data 66L consists of training comparison Raman spectrum data 86L, ​​training comparison measurement data 91L, and training comparison prediction data 95L. The correct answer report 67CA is the report 67 that should be output from the trained large-scale language model 49 when the first training prompt 1021L is input, and is, so to speak, data for checking the answer. The training condition data 24L and the training comparison data 66L are examples of "training-related data" related to the technology of this disclosure. The correct answer report 67CA is an example of "correct answer data" related to the technology of this disclosure.

[0070] In the additional training, the first training prompt 1021L is input to the trained large-scale language model 49, causing the trained large-scale language model 49 to output a training report 67L. Then, this training report 67L is compared with the ground truth report 67CA, and based on the comparison result, a loss calculation of the trained large-scale language model 49 is performed using a loss function. Subsequently, the update settings for various coefficients of the trained large-scale language model 49 are set according to the result of the loss calculation, and the trained large-scale language model 49 is updated according to the update settings.

[0071] In additional training, the above series of processes—inputting the first training prompt 1021L into the trained large-scale language model 49, outputting the training report 67L from the trained large-scale language model 49, loss calculation, update settings, and updating the trained large-scale language model 49—are repeated while the training data 115 is exchanged. The repetition of the above series of processes is terminated when the prediction accuracy of the training report 67L, that is, the prediction accuracy of the cause and countermeasures for the change in the quality of the antibody 14, reaches a predetermined set level. Alternatively, additional training may be terminated after the above series of processes has been repeated a set number of times, regardless of the prediction accuracy of the training report 67L.

[0072] As an example, as shown in Figure 16, in additional learning, a second learning prompt 1022L is input to the trained large-scale language model 49. The second learning prompt 1022L includes the feature wavenumber regions of each component in the cell culture medium 12, such as the antibody 14. These feature wavenumber regions have been identified through experiments and simulations and can be easily obtained. In this way, the trained large-scale language model 49, in which the prediction accuracy of the cause and countermeasures for the change in the quality of the antibody 14 has reached a set level and in which the feature wavenumber regions of each component in the cell culture medium 12 have been taught, is stored in storage 35 and used in the second prediction unit 61. Note that pre-training, fine-tuning, and additional learning may be performed in the information processing device 30 or in a device other than the information processing device 30. Furthermore, pre-training, fine-tuning, and additional learning may be continued even after the trained large-scale language model 49 is stored in storage 35.

[0073] Next, the operation of the above configuration will be explained by referring to the flowchart shown in Figure 17 as an example. As shown in Figure 5, the CPU 37 of the information processing device 30 functions as an acquisition unit 55, an RW control unit 56, a first prediction unit 57, a comparison unit 58, a display control unit 59, an instruction receiving unit 60, and a second prediction unit 61 when the operation program 45 is activated.

[0074] As shown in Figure 1, antibody-producing cells 13 are cultured in a cell culture medium 12 in the culture tank 11 of the culture section 10, and antibodies 14 are produced from the antibody-producing cells 13.

[0075] The Raman scattering light RSL of components such as antibodies 14 in the cell culture medium 12 is measured by the sensor unit 21, and target Raman spectrum data 23 is generated in the analyzer 22. In addition, target condition data 24 is generated in the culture unit 10. The target Raman spectrum data 23 and target condition data 24 are transmitted to the information processing device 30.

[0076] In the information processing device 30, the acquisition unit 55 acquires the target Raman spectrum data 23, the target condition data 24, and the target measurement data 25 (step ST100). The target Raman spectrum data 23, the target condition data 24, and the target measurement data 25 are output from the acquisition unit 55 to the RW control unit 56, and stored in the storage 35 by the RW control unit 56.

[0077] The target Raman spectrum data 23, target condition data 24, and target measurement data 25 are read from the storage 35 by the RW control unit 56. The target Raman spectrum data 23 is output from the RW control unit 56 to the first prediction unit 57 and the comparison unit 58. The target condition data 24 is output from the RW control unit 56 to the display control unit 59, and the target measurement data 25 is output from the RW control unit 56 to the comparison unit 58.

[0078] The prediction model 46 is read from storage 35 by the RW control unit 56 and output to the first prediction unit 57. Reference data 47 and selection criteria 48 are also read from storage 35 by the RW control unit 56 and output to the comparison unit 58. Furthermore, the trained large-scale language model 49 is read from storage 35 by the RW control unit 56 and output to the second prediction unit 61.

[0079] As shown in Figure 6, in the first prediction unit 57, the target Raman spectrum data 23 is input to the prediction model 46 (antibody concentration prediction model 70, glucose concentration prediction model 71, and lactate concentration prediction model 72), and the prediction model 46 outputs target prediction data 65 (antibody concentration prediction data 73, glucose concentration prediction data 74, and lactate concentration prediction data 75) (step ST110). The target prediction data 65 is output from the first prediction unit 57 to the RW control unit 56, and stored in the storage 35 by the RW control unit 56.

[0080] The target prediction data 65 is read from the storage 35 by the RW control unit 56 and output to the comparison unit 58. As shown in Figures 8 to 10, the comparison unit 58 compares the target Raman spectrum data 23 with the reference Raman spectrum data 80, the target measurement data 25 with the reference measurement data 81, and the target prediction data 65 with the reference prediction data 82, respectively, and outputs the comparison Raman spectrum data 86, comparison measurement data 91, and comparison prediction data 95 (step ST120). The comparison data 66, which consists of the comparison Raman spectrum data 86, comparison measurement data 91, and comparison prediction data 95, is output from the comparison unit 58 to the RW control unit 56 and stored in the storage 35 by the RW control unit 56.

[0081] The comparison data 66 is read from the storage 35 by the RW control unit 56 and output to the display control unit 59. Then, under the control of the display control unit 59, the prompt input screen 100 shown in Figure 11 is displayed on the display 31 (step ST130).

[0082] On the prompt input screen 100, the operator OP inputs prompt 102 and selects the input button 103. As a result, the instruction to input prompt 102 is received by the instruction receiving unit 60 (step ST140). Prompt 102 is output from the instruction receiving unit 60 to the second prediction unit 61.

[0083] As shown in Figure 12, in the second prediction unit 61, the prompt 102 is input to the trained large-scale language model 49, and as a result, the trained large-scale language model 49 outputs a report 67 (step ST150). The report 67 is output from the second prediction unit 61 to the display control unit 59.

[0084] Under the control of the display control unit 59, the report display screen 105 shown in Figure 13 is displayed on the display 31 (step ST160). The operator OP reviews and examines the report 67 to determine the cause of the change in antibody quality 14 and to devise countermeasures.

[0085] As described above, the CPU 37 of the information processing device 30 is equipped with a second prediction unit 61. As shown in Figure 12, the second prediction unit 61 inputs a prompt 102 containing comparative data 66 obtained in the cell culture process that produces antibody 14 to a trained large-scale language model 49, and causes the trained large-scale language model 49 to output a report 67 including the cause and countermeasures that led to the change in the quality of antibody 14. Therefore, it is possible to easily predict the cause and countermeasures that led to the change in the quality of antibody 14 using the trained large-scale language model 49. Furthermore, this makes it possible to predict the optimal values ​​of conditions such as pH, temperature, and dissolved oxygen. Note that it is also possible to predict either the cause or the countermeasure, rather than both.

[0086] As shown in Figures 8 to 10 and Figure 12, the relevant data includes comparative data 66 with reference data 47 obtained in past cell culture processes. Therefore, based on the degree of deviation from the reference data 47, it is possible to predict the cause and countermeasures that led to the change in the quality of the antibody 14, thereby improving the accuracy of the prediction of the cause and countermeasures that led to the change in the quality of the antibody 14. In particular, if the reference data 47 is data that meets a pre-set second criterion as shown in Figure 7, the comparative data 66 shows how much the data deviates from the second criterion, thus further improving the accuracy of the prediction of the cause and countermeasures that led to the change in the quality of the antibody 14.

[0087] The comparative Raman spectrum data 86 and comparative measurement data 91 among the comparative data 66 have multiple variables. As shown in Figures 8 and 9, the comparison unit 58 selects variables that satisfy the selection criteria 48 from among the multiple variables. The second prediction unit 61 selectively inputs the variables selected by the comparison unit 58 into the pre-trained large-scale language model 49. In this way, the input to the pre-trained large-scale language model 49 is narrowed down to variables that are likely to contribute to the prediction of the cause of the change in antibody 14 quality, thereby further improving the prediction accuracy of the cause of the change in antibody 14 quality and the countermeasures.

[0088] As shown in Figure 12, the comparative data 66 includes comparative Raman spectral data 86 obtained by comparing the target Raman spectral data 23 of the cell culture medium 12 containing the antibody 14 with the reference Raman spectral data 80, target culture condition data 24 performed in the cell culture process, comparative measurement data 91 obtained by comparing the target measurement data 25 measured in the culture performed in the cell culture process with the reference measurement data 81, and comparative prediction data 95 obtained by comparing the target prediction data 65 of the quality of the antibody 14 with the reference prediction data 82. Therefore, sufficient information can be provided to the trained large-scale language model 49 to predict the causes and countermeasures that led to the change in the quality of the antibody 14. The accuracy of the prediction of the causes and countermeasures that led to the change in the quality of the antibody 14 can be improved. According to the comparative measurement data 91, the actual state of the culture performed in the cell culture process can be reflected in the prediction of the causes and countermeasures that led to the change in the quality of the antibody 14.

[0089] Note that the comparison data 66 does not need to include all of the comparison Raman spectrum data 86, comparison measurement data 91, and comparison prediction data 95, but only needs to include at least one of these comparison Raman spectrum data 86, comparison measurement data 91, and comparison prediction data 95. Also, while the target condition data 24 and comparison data 66 are given as examples of related data, the data is not limited to these. At least one of the target Raman spectrum data 23 and target prediction data 65 may be input into the trained large-scale language model 49 as related data.

[0090] The comparative Raman spectral data 86, and consequently the target Raman spectral data 23, allows us to identify the physical properties of the cell culture medium 12 containing the antibody 14, that is, the types and amounts of chemical substances in the cell culture medium 12. Therefore, it can be said that this data is appropriate as relevant data.

[0091] Raman scattering light (RSL) readily reflects information derived from the functional groups of amino acids in proteins. Therefore, by converting spectral data to Raman spectral data, as in this example, it is possible to obtain target Raman spectral data 23 that accurately reflects the physical properties of the antibody 14, which is a protein.

[0092] Biopharmaceuticals containing antibody 14 are called antibody drugs and are widely used to treat chronic diseases such as cancer, diabetes, and rheumatoid arthritis, as well as rare diseases such as hemophilia and Crohn's disease. Therefore, in this example, where the product is a protein and the protein is antibody 14, it is possible to promote the development of antibody drugs that are widely used to treat a variety of diseases.

[0093] As shown in Figure 16, the pre-trained large-scale language model 49 is a model that has been trained on the feature wavenumber region, which is the region in the Raman spectral data where characteristic changes in intensity values ​​appear. Therefore, compared to the case where training on the feature wavenumber region is not performed, the accuracy of predicting the cause of the change in the quality of the antibody 14 and the countermeasures can be improved.

[0094] As shown in Figure 15, the trained large-scale language model 49 is a model trained using a set of a first training prompt 1021L containing training comparison data 66L and a ground truth report 67CA. Therefore, compared to the case where such training is not performed, the accuracy of predicting the causes and countermeasures for changes in the quality of the antibody 14 can be improved.

[0095] [Example] In this example, assuming an abnormality occurred in the production of antibody 14, we verified whether it was possible to propose an appropriate cause to the operator OP that caused the change in the quality of antibody 14. In order to determine whether the proposed cause was appropriate, the antibody-producing cells 13 were intentionally induced to develop abnormalities due to genetic factors.

[0096] Cell culture was performed using a fed-batch culture method, and as an example, the cell culture process was carried out under the culture conditions shown in Table 120, which is shown in Figure 18. More specifically, in both the reference batch from which reference data 47 was obtained and the target batch from which target Raman spectral data 23 etc. were obtained, CHO cells were used as antibody-producing cells 13, and IgG (Immunoglobulin G) 1-λ class antibodies 14 were produced. In both the reference batch and the target batch, the amount of culture medium 12 was 340 mL, the stirring speed was 300 rpm (rotation per minute), the culture period was 12 days, the temperature was 37 ± 0.5 °C, the dissolved oxygen was 33% or higher, and the pH was 6.8 to 7.3. Temperature, dissolved oxygen, and pH were managed and controlled in the culture unit 10 and measured, for example, at 1-minute intervals using a measurement sensor. In addition, the necessary nutrients for culture, such as glucose, were added to the culture vessel 11 in appropriate amounts based on the offline quality measurement results for the cell culture medium 12 on each culture day.

[0097] The number of cell passages up to the seeding point was set at 19 for the target batch, more than double the reference batch's 8 passages. The reason for increasing the cell passage number in the target batch was based on the property that the antibody-producing capacity of antibody-producing cells 13 decreases as the cell passage number increases. Therefore, the target batch was designed to have a reduced antibody-producing capacity of antibody-producing cells 13.

[0098] A PI-200 Lab Specifications Raman spectrometer manufactured by HORIBA Process Instruments was used for the Raman spectrometer 20. The measurement conditions for the target Raman spectrum data 23 and the reference Raman spectrum data 80 were set as shown in Table 125 in Figure 19, as an example. Specifically, the laser output of the excitation light EL was set to 450 mW, the exposure time to 0.7 seconds, the number of integrations to 100, and the measurement interval to 5 minutes.

[0099] Baseline correction was applied to the target Raman spectrum data 23 and the reference Raman spectrum data 80. For the reference Raman spectrum data 80, the mean AVE and standard deviation Σ of the intensity values ​​were calculated as shown in Figure 8. Then, the target Raman spectrum data 23 was standardized using the mean AVE and standard deviation Σ to obtain the standardized target Raman spectrum data 23S. Finally, based on the selection criteria 48, comparative Raman spectrum data 86 was generated from the standardized target Raman spectrum data 23S. In addition, the processing shown in Figure 9 was applied to the target measurement data 25 and the reference measurement data 81, and the processing shown in Figure 10 was applied to the target prediction data 65 and the reference prediction data 82, respectively, to generate comparative measurement data 91 and comparative prediction data 95.

[0100] As an example, as shown in Figures 20 and 21, comparative data 66 consisting of target condition data 24, comparative Raman spectrum data 86, comparative measurement data 91, and comparative prediction data 95 was created, along with a prompt 102 containing a sentence asking about the cause of the change in antibody quality 14, which could be predicted based on the comparative data 66. This prompt 102 was then input into a pre-trained large-scale language model 49. The pre-trained large-scale language model 49 used was based on GPT-4o (Generative Pre-trained Transformer-4omni).

[0101] Figures 22 and 23 show the report 67 output from the trained large-scale language model 49 when the prompt 102 shown in Figures 20 and 21 was input. Predictions for the cause of the change in antibody 14 quality include poor nutrient supply, abnormal metabolic activity, and fluctuations in process factors, as well as cellular senescence or damage, which is thought to be the actual cause. Therefore, it has been confirmed that the technology of this disclosure has the effect of making it possible to easily predict the cause of the change in antibody 14 quality using the trained large-scale language model 49.

[0102] Furthermore, report 67 also describes the attribution of the Raman spectral data. This is the result of additional training using the second learning prompt 1022L, which includes the feature wavenumber region shown in Figure 16.

[0103] Physical property data is not limited to spectroscopic spectral data such as the example Raman spectral data. Chemical structural information of the product, such as the amino acid sequence information of antibody 14, is also acceptable. Chemical composition information of the culture medium and / or additives such as nutrients is also acceptable.

[0104] Antibody 14 may be a bispecific antibody, an antibody-drug conjugate, a small molecule antibody, a glycosylated antibody, or the like.

[0105] Proteins are not limited to antibody 14. They may also be peptides, nucleic acids (DNA, RNA (Ribbonucleic Acid)), lipids, viruses, viral subunits, and virus-like particles. Cytokines (interferon, interleukin, etc.), hormones (insulin, glucagon, follicle-stimulating hormone, erythropoietin, etc.), growth factors (IGF (Insulin-Like Growth Factor)-1, bFGF (Basic Fibroblast Growth Factor), etc.), blood coagulation factors (factor VII, factor VIII, factor IX, etc.), enzymes (lysosomal enzymes, DNA-degrading enzymes, etc.), Fc (Fragment Crystallizable) fusion proteins, receptors, albumin, and protein vaccines.

[0106] Electromagnetic waves are not limited to Raman scattered light (RSL), and therefore the spectral data is not limited to Raman spectral data such as the target Raman spectral data 23 and the reference Raman spectral data 80. Infrared absorption spectral data, near-infrared absorption spectral data, nuclear magnetic resonance spectral data, ultraviolet-visible (UV-Vis) spectral data, emission spectral data, or fluorescence spectral data may also be used. Furthermore, the spectral data may be multidimensional data such as hyperspectral data.

[0107] The prediction model 46 is not limited to machine learning models. It may also be a model generated by multivariate analysis or statistical analysis. Examples of multivariate analysis and statistical analysis include linear regression, multiple regression, principal component regression, partial least squares regression, logistic regression, Lasso regression, ridge regression, support vector regression, and Gaussian process regression. Among these, principal component regression is preferred. Furthermore, the prediction model 46 may also be rule-based.

[0108] The data predicted by the prediction model 46 is not limited to those exemplified above. For example, the purity of antibody 14 could be predicted instead of its concentration.

[0109] The substance used to measure the target Raman spectral data 23 is not limited to the example cell culture medium 12. The culture supernatant obtained after cell removal from the cell culture medium 12 may also be used. Alternatively, the purified solution obtained after various chromatographic treatments in the purification unit may also be used.

[0110] Although antibody 14 was given as the product and a cell culture process as the production process, the process is not limited to these. Other processes may include a flow synthesis process in which multiple raw materials are mixed to obtain the target compound, an atmospheric pressure plasma treatment process, and even a deposition process or sputtering process in which a thin film layer is formed on a substrate in a vacuum chamber.

[0111] The additional training shown in Figures 15 and 16 is not mandatory.

[0112] The information processing device 30 may be a personal computer installed in a production facility such as a pharmaceutical facility, as shown in Figure 1, or it may be a server computer installed in a data center independent of the production facility.

[0113] When the information processing device 30 is configured as a server computer, the target Raman spectrum data 23 and target condition data 24 are transmitted to the server computer from personal computers installed in each production facility, such as each pharmaceutical facility, via a network such as the Internet. The server computer distributes various screens, such as the prompt input screen 100 and the report display screen 105, to the personal computers in the form of web-distribution screen data created using a markup language such as XML (Extensible Markup Language). The personal computers reproduce the screen to be displayed on a web browser based on the screen data and display it on the display. Note that other data description languages ​​such as JSON (JavaScript® Object Notation) may be used instead of XML.

[0114] The hardware configuration of the computer constituting the information processing device 30 according to the technology disclosed herein can be modified in various ways. For example, the information processing device 30 can be configured with multiple computers separated as hardware, for the purpose of improving processing power and reliability. For example, the functions of the acquisition unit 55, RW control unit 56, and first prediction unit 57 and the functions of the comparison unit 58, display control unit 59, instruction reception unit 60, and second prediction unit 61 can be distributed among two computers. In this case, the information processing device 30 is configured with two computers.

[0115] Thus, the hardware configuration of the computer in the information processing device 30 can be appropriately changed according to the required performance, such as processing power, safety, and reliability. Furthermore, not only the hardware, but also application programs such as the operating program 45 can, of course, be duplicated or distributed and stored on multiple storage devices for the purpose of ensuring safety and reliability.

[0116] In the above embodiment, each processing unit, such as the acquisition unit 55, RW control unit 56, first prediction unit 57, comparison unit 58, display control unit 59, instruction receiving unit 60, and second prediction unit 61, is executed on any computer. Furthermore, any computer may execute these processes using a processor as hardware, a program as software, or a combination thereof. In this case, the processor is configured to cooperate with the program to execute the various processes in the above embodiment and can function as each unit or means in the above embodiment. Also, the execution order of the processor's processes is not limited to the order described and may be changed as appropriate. Any computer may be a general-purpose computer, a computer designed for a specific application, a workstation, or any other system capable of executing each process.

[0117] A processor may consist of one or more hardware components, and the type of hardware is not limited. For example, a processor may consist of hardware such as the example CPU 37, a programmable logic device such as an MPU (Micro Processing Unit), an FPGA (Field Programmable Gate Array), a dedicated circuit for executing a specific process such as an ASIC (Application Specific Integrated Circuit), a GPU (Graphic Processing Unit), or an NPU (Neural Processing Unit). Furthermore, the type of hardware may be a combination of different types of hardware. When multiple hardware components are configured to execute one or more processes of a processor, these multiple hardware components may reside in physically separate devices or in the same device. Furthermore, in any embodiment, the order of each process performed by the processor is not limited to the order described above and may be changed as appropriate. The hardware is composed of an electrical circuit (circuitry) or the like, which is a combination of circuit elements such as semiconductor elements.

[0118] Furthermore, the program may be firmware or software such as microcode. Alternatively, the program may be, for example, a set of program modules, each function of which may be implemented by a processor configured to perform its respective function. The program may be program code or multiple code segments stored on one or more non-temporary computer-readable media (e.g., storage media or other storage devices). The program may be divided and stored on multiple non-temporary computer-readable media located in physically separate devices. Program code or code segments may represent any combination of procedures, functions, subprograms, routines, subroutines, modules, software packages, classes, or instructions, data structures, or program statements. Program code or code segments may be connected to other code segments or hardware circuits by sending and receiving information, data, arguments, parameters, or memory contents.

[0119] From the above description, the technology described in the following supplementary information can be understood.

[0120] [Appendix 1] An information processing device comprising a processor, wherein the processor inputs relevant data obtained in a production process for producing a product into a trained large-scale language model, and causes the trained large-scale language model to output at least one of the causes of the change in the quality of the product and countermeasures for the change in the quality of the product. [Appendix 2] The information processing device according to Appendix 1, wherein the relevant data includes comparison data with reference data obtained in past production processes. [Appendix 3] The information processing device according to Appendix 2, wherein the comparison data has a plurality of variables, and the processor selectively inputs variables that satisfy a predetermined first criterion from the plurality of variables into the trained large-scale language model. [Appendix 4] The information processing device according to Appendix 2 or Appendix 3, wherein the reference data is data that satisfies a predetermined second criterion. [Appendix 5] The information processing apparatus according to any one of Appendix 2 to Appendix 4, wherein the related data includes at least one of the following: physical property data of a substance including the product, condition data of the processing carried out in the production process, measurement data measured in the processing carried out in the production process, and prediction data of the quality of the product. [Appendix 6] The information processing apparatus according to Appendix 5, wherein the physical property data includes spectral data. [Appendix 7] The information processing apparatus according to Appendix 6, wherein the spectral data is Raman spectral data. [Appendix 8] The information processing apparatus according to Appendix 7, wherein the product is a protein. [Appendix 9] The information processing apparatus according to any one of Appendix 6 to Appendix 8, wherein the trained large-scale language model is a model trained on regions in which characteristic changes appear in the spectral data. [Appendix 10] The information processing apparatus according to any one of Appendix 2 to Appendix 9, wherein the trained large-scale language model is a model trained on a set of training related data and ground truth data.[Appendix 11] The information processing apparatus according to any one of Appendix 1 to Appendix 10, wherein the related data is physical property data of a substance including the product, and includes at least one of the following: physical property data including spectral data, setting condition data for processing carried out in the production process, measurement data measured in processing carried out in the production process, and prediction data for the quality of the product, and includes comparison data with reference data obtained in past production processes.

[0121] The technology of this disclosure can be appropriately combined with the various embodiments and / or variations described above. Furthermore, it is understood that various configurations can be adopted without departing from the spirit of the invention, and the invention is not limited to the embodiments described above. In addition, the technology of this disclosure extends not only to programs, but also to storage media for non-temporarily storing programs, and to computer program products containing programs.

[0122] The descriptions and illustrations presented above are detailed explanations of the technical aspects of this disclosure and are merely examples of the technical aspects. For example, the above descriptions of the structure, function, operation, and effect are examples of the structure, function, operation, and effect of the technical aspects of this disclosure. Therefore, it goes without saying that you may delete unnecessary parts, add new elements, or replace elements in the descriptions and illustrations presented above, as long as you do not deviate from the essence of the technical aspects of this disclosure. Furthermore, in order to avoid confusion and facilitate understanding of the technical aspects of this disclosure, explanations of common technical knowledge and the like that do not require special explanation to enable the implementation of the technical aspects of this disclosure have been omitted from the descriptions and illustrations presented above.

[0123] In this specification, "A and / or B" is synonymous with "at least one of A and B." That is, "A and / or B" means that it may be A alone, or B alone, or a combination of A and B. Furthermore, in this specification, the same concept as "A and / or B" applies when expressing three or more things linked by "and / or."

[0124] All documents, patent applications, and technical standards described herein are incorporated by reference to the same extent as if each individual document, patent application, and technical standard were specifically and individually noted to be incorporated by reference.

Claims

1. An information processing device comprising a processor, wherein the processor inputs relevant data obtained in a production process for producing a product into a pre-trained large-scale language model, and outputs from the pre-trained large-scale language model at least one of the causes of the change in the quality of the product and countermeasures to address the change in the quality of the product.

2. The information processing apparatus according to claim 1, wherein the related data includes comparison data with reference data obtained in past production processes.

3. The information processing apparatus according to claim 2, wherein the comparison data has a plurality of variables, and the processor selectively inputs variables that satisfy a predetermined first criterion from among the plurality of variables into the trained large-scale language model.

4. The information processing apparatus according to claim 2, wherein the reference data is data that satisfies a pre-set second criterion.

5. The information processing apparatus according to claim 2, wherein the related data includes at least one of the following: physical property data of a substance including the product, condition data of a process carried out in the production process, measurement data measured in a process carried out in the production process, and prediction data of the quality of the product.

6. The information processing apparatus according to claim 5, wherein the physical property data includes spectral data.

7. The information processing apparatus according to claim 6, wherein the spectral data is Raman spectral data.

8. The information processing apparatus according to claim 7, wherein the product is a protein.

9. The information processing apparatus according to claim 6, wherein the trained large-scale language model is a model that has been trained on regions in which characteristic changes appear in the spectral data.

10. The information processing apparatus according to claim 2, wherein the pre-trained large-scale language model is a model trained using a set of training-related data and ground truth data.

11. The information processing apparatus according to claim 1, wherein the related data is physical property data of a substance including the product, and includes at least one of the following: physical property data including spectral data, setting condition data for processing carried out in the production process, measurement data measured in processing carried out in the production process, and prediction data for the quality of the product, and includes comparison data with reference data obtained in past production processes.

12. A method for operating an information processing device, comprising: inputting relevant data obtained in a production process for producing a product into a pre-trained large-scale language model; and outputting from the pre-trained large-scale language model at least one of the causes that led to a change in the quality of the product, and countermeasures to address the change in the quality of the product.

13. An operating program for an information processing device that causes a computer to perform a process including inputting relevant data obtained in a production process for producing a product into a pre-trained large-scale language model, and outputting from the pre-trained large-scale language model at least one of the causes that led to a change in the quality of the product, and countermeasures to address the change in the quality of the product.

Images