Method for predicting generation stability of clone generating useful substance, information processing device, program, and prediction model generation method

Abstract

Provided are a method, an information processing device, a program, and a prediction model generation method with which it is possible to predict, with high precision and low cost, the generation stability of clones that generate useful substances. The one or more processors execute the following processes: acquiring culture data of one or more clones for clones in which useful substances are generated; analyzing the culture data to define a clone as a prediction object; and predicting the generation stability of the useful substance based on the clone to be predicted using the data measured for the clone to be predicted. The generation stability can be defined according to the presence or absence of a change in the amount of useful substance generated at the beginning of the culture and after the culture for a prescribed period.

Description

Technical Field

[0001] The present invention relates to an information processing technology and a machine learning technology for predicting the production stability of a clone producing a useful substance. Background Art

[0002] In recent years, the industrial utilization of manufacturing methods that enable cells to produce complex useful substances that were difficult to produce in previous chemical syntheses is being promoted. One example is biopharmaceuticals. Among the top 10 pharmaceutical sales in the world, more than half of the varieties account for about two-thirds of the sales. Compared with previous low-molecular-weight pharmaceuticals, biopharmaceuticals use complex proteins, etc., which are very difficult to synthesize artificially. Therefore, antibody pharmaceuticals, which are an example of biopharmaceuticals, for example, insert genes corresponding to desired human proteins into CHO cells (Chinese Hamster Ovary cells), etc., produce the desired proteins through cell functions, and extract and purify them to produce antibody pharmaceuticals. Production methods are widely popular.

[0003] As described above, the insertion of genes into cells cannot be precisely controlled, so genes are usually inserted into a large number of cells at the same time. At this time, considering that the gene insertion position of each cell generated is random, in order to stabilize antibodies as pharmaceuticals and ensure quality, many regulatory agencies require that the cells responsible for antibody production after gene insertion are derived from a single cell and that their properties do not change through subculture, which is the so-called monoclonality.

[0004] Therefore, a single cell is extracted from each cell in which the gene is inserted at a random position, and the single cell is proliferated to produce a cell clone (hereinafter referred to as a clone), and the clone is made to produce antibodies to ensure monoclonality. The clone of the present invention refers to a group of genetically identical cells or cells constituting the group.

[0005] On the other hand, in industrialization, clones with high-quality antibody production are required. Here, high-quality antibody production refers to having high antibody production ability at the current time point and stable antibody production ability during long-term culture. As mentioned above, there is a deviation in the antibody production ability of clones made by random cells with gene insertion positions, and it is necessary to distinguish whether each clone is high-quality antibody production ability. Whether it is a high-yield clone with high antibody production ability at the current time point can be distinguished by a 2-week standard test, but the determination of whether the antibody production ability is stable during long-term culture actually requires experimental verification (stability test) based on long-term culture of about several months.

[0006] In this context, Patent Document 1 proposes a method for predicting the production stability of a cloned recombinant protein several months later based on the gene expression data of the clone obtained at the current time point. In addition, Non-Patent Document 1 proposes a method for identifying a marker gene that can predict the stable expression of a recombinant protein in the early stage of clone development and predicting the production stability of a recombinant protein in the early stage of clone development.

[0007] Prior art literature

[0008] Patent Literature

[0009] Patent Document 1: International Publication No. 2016 / 075216

[0010] Non-patent literature

[0011] Non-patent literature 1: Uros Jamnikar, Petra Nikolic, Ales Belic, Marjanca Blas, Dominik Gaser, Andrej Francky, Holger Laux, Andrej Blejec, Spela Baebler and Kristina Gruden, "Transcriptome study and identification of potential markergenes related to the stable expression of recombinant proteins in CHO clones" BMC Biotechnology volume 15, Article number 98(2015). Summary of the invention

[0012] Technical issues to be solved by the invention

[0013] However, the method described in Patent Document 1 cannot be said to be sufficient in terms of prediction accuracy. In addition, since gene analysis of multiple clones usually requires high costs, there is also a problem that the cost reduction effect obtained by predicting the production stability of recombinant proteins is reduced due to the cost increase caused by the gene analysis used for prediction. In order to suppress costs, it is considered to reduce the number of clones as the prediction object of production stability, but in this way, the number of clones with high production stability among the prediction objects will also decrease, resulting in a decrease in the number of clones with high production stability, and it becomes difficult to simply reduce the number of clones as the prediction object.

[0014] The first problem to be solved by the present invention is to provide a method for predicting the production stability of a useful substance in a clone with high accuracy. The second problem is to provide a method for reducing the cost of predicting the production stability of a useful substance in a clone.

[0015] The present invention has been made in view of such circumstances, and an object of the present invention is to provide a method, an information processing device, a program, and a prediction model generation method capable of predicting the production stability of a clone producing a useful substance with high accuracy and at low cost.

[0016] Means for solving technical problems

[0017] The method involved in the first embodiment of the present invention is a method for predicting the production stability of a clone that produces a useful substance, wherein one or more processors perform the following processing: obtaining culture data of one or more clones; analyzing the culture data to limit the clones that are the prediction objects; and using the data measured for the clones that are the prediction objects to predict the production stability of the useful substance of the clones that are the prediction objects.

[0018] According to the first embodiment, since the prediction object is limited based on the information obtained from the culture data and the generation stability is predicted, the generation stability can be predicted with high accuracy compared to the case where the object is not limited. In addition, since only the clones as the prediction object need to be limited and the data required for the prediction are acquired, the cost can be suppressed.

[0019] The predicted production stability can represent the state of the clone in the future several months later in the same manner as the production stability actually verified experimentally by long-term cultivation of several months. For example, the production stability can be evaluated from the perspective of whether the initial production amount is still maintained after long-term cultivation. According to the first embodiment, the result of the stability test requiring long-term cultivation can be predicted with high accuracy and low cost.

[0020] The method according to the second aspect of the present invention may be configured such that, in the method according to the first aspect, the production stability is defined based on the presence or absence of a change in the production amount of the useful substance between the start of culture and after culture for a predetermined period of time.

[0021] The method involved in the third embodiment of the present invention may be a structure in which one or more processors perform the following processing in the method involved in the first embodiment or the second embodiment: including setting indicators obtained from cultivation data and thresholds related to the indicators, and limiting the prediction object based on the values ​​of the indicators and the thresholds.

[0022] The method according to a fourth aspect of the present invention may be configured such that, in the method according to the third aspect, the threshold is adjusted so that the prediction accuracy resulting in stability is higher than when the prediction target is not limited.

[0023] The method according to the fifth aspect of the present invention may be a structure in which the threshold is defined using the rank of the value of the indicator in the method according to the third aspect or the fourth aspect. In addition, the "rank" may be the rank when the values ​​of the indicators for a plurality of clones are arranged in descending order or in ascending order. For example, the threshold may be defined as the top 40% of the relative rank in a population including a plurality of clones.

[0024] A method according to a sixth aspect of the present invention is the method according to any one of the third to fifth aspects, wherein the prediction target may be a superordinate group of the index value.

[0025] In the method according to a seventh aspect of the present invention, in the method according to any one of the third to sixth aspects, the indicator may be the amount of generated useful substances.

[0026] In the method according to an eighth aspect of the present invention, in the method according to any one of the third to sixth aspects, the index may be an integrated viable cell density.

[0027] In the method according to a ninth aspect of the present invention, in the method according to any one of the third to sixth aspects, the indicator may be a lactic acid concentration.

[0028] The method according to a tenth aspect of the present invention may be configured such that in the method according to any one of the first to ninth aspects, data for generating a prediction of stability may include one or more gene expression levels.

[0029] The method involved in the 11th embodiment of the present invention may be the following structure in the method involved in any one of the 1st to 10th embodiments: one or more processors receive input of data of a prediction object and use a 2-category classification model of stability or instability to predict stability.

[0030] In the method involved in the 12th embodiment of the present invention, in the method involved in the 11th embodiment, the model is a model trained by machine learning using multiple training data, and the multiple training data are multiple training data that establish associations between data of training clones that have been subjected to the same restrictions as the clones as the prediction object and correct stability labels.

[0031] The method involved in the 13th embodiment of the present invention, in the method involved in the 12th embodiment, can be of the following structure: a plurality of training data includes training data for a plurality of types of clones that produce different useful substances, and one or more processors predict the production stability of clones that produce useful substances different from the useful substances used in the training of the model.

[0032] In the method according to a fourteenth aspect of the present invention, in the method according to any one of the first to thirteenth aspects, the useful substance may be any of proteins, peptides, and viruses that are raw materials for pharmaceuticals.

[0033] The method according to a fifteenth aspect of the present invention is the method according to any one of the first to fourteenth aspects, wherein the useful substance may be an antibody or an antibody-like protein.

[0034] The method according to a sixteenth aspect of the present invention is the method according to any one of the first to fifteenth aspects, wherein the clones may be cells derived from vertebrates.

[0035] The method according to a seventeenth aspect of the present invention is the method according to any one of the first to fifteenth aspects, wherein the clones may be mammal-derived cells.

[0036] According to an eighteenth aspect of the present invention, in the method according to any one of the first to fifteenth aspects, the clones may be CHO cells or HEK cells (Human Embryonic Kidney cells).

[0037] The information processing device involved in the 19th embodiment of the present invention comprises: one or more processors; and one or more storage devices storing instructions for the one or more processors to execute, and the one or more processors perform the following processing: obtaining culture data of one or more clones with respect to clones that produce useful substances; analyzing the culture data to limit the clones that are the objects of prediction; and using the data measured for the clones that are the objects of prediction, predicting the stability of the production of useful substances based on the clones that are the objects of prediction.

[0038] The information processing device according to the nineteenth aspect can be configured to include the same method as any one of the second aspect to the eighteenth aspect.

[0039] The program involved in the 20th mode of the present invention enables a computer to implement the following functions: obtaining culture data of one or more clones for clones that produce useful substances; analyzing the culture data to limit the clones that are prediction objects; and using the data measured for the clones that are prediction objects to predict the production stability of useful substances based on the clones that are prediction objects.

[0040] The program according to the twentieth aspect can be configured to include the same aspects as the method according to any one of the second aspect to the eighteenth aspect.

[0041] The prediction model generation method involved in the 21st embodiment of the present invention is a prediction model generation method for generating a prediction model that enables a computer to realize the function of predicting the production stability of clones that produce useful substances, wherein a system including one or more processors includes the following steps: obtaining culture data of one or more clones; analyzing the culture data to limit the clones that are prediction objects; and using multiple training data that establish associations between data measured for the clones belonging to the prediction objects and correct stability labels to perform machine learning, and train the prediction model in a manner that makes the output of the prediction model relative to the input data close to the correct stability label.

[0042] The prediction model generation method according to the 21st aspect can be configured to include the same aspects as the method according to any one of the 2nd to 18th aspects.

[0043] Effects of the Invention

[0044] According to the present invention, the prediction object can be appropriately limited based on the information obtained by analyzing the culture data, and the production stability of the clone producing the useful substance can be predicted with high accuracy. In addition, according to the present invention, by limiting the prediction object, the prediction cost of the production stability can be suppressed, and the prediction can be performed at a low cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is an explanatory diagram showing an overview of the production process of antibody pharmaceuticals.

[0046] Figure 2 This is a graph showing an example of changes in antibody production levels based on clones.

[0047] Figure 3 This is an explanatory diagram outlining the effects of stability prediction AI (Artificial Intelligence) realized by this embodiment.

[0048] Figure 4 A conceptual diagram of a machine learning model that predicts gene stability based on gene expression data.

[0049] Figure 5 It is an explanatory diagram showing the outline of the method for predicting the production stability of clones according to the present embodiment.

[0050] Figure 6 This is a table showing examples of data sets used for model training and evaluation.

[0051] Figure 7 This is a graph showing an example of narrowing down the objects based on a certain index of the cultivation data.

[0052] Figure 8This is a table showing the clone numbers of five antibody-producing CHO cells prepared as evaluation samples and examples of the assignment of stability labels.

[0053] Fig. 9 This is a table showing examples of the number of clones whose antibody production values ​​belong to the top 40% of the relative ranking in each antibody type and the assignment of stability labels.

[0054] Fig.10 This is a table showing the number of clones whose integrated viable cell density values ​​belong to the top 60% of the relative ranking in each antibody type and examples of assignment of stability labels.

[0055] Fig.11 This is a table showing the number of clones whose lactate concentration values ​​belong to the top 40% of the relative ranking in each antibody type and examples of the assignment of stability labels.

[0056] Fig.12 It is a block diagram showing the functional structure of the information processing device involved in the embodiment.

[0057] Fig.13 This is a block diagram showing an example of the hardware configuration of an information processing device.

[0058] Fig.14 This is a block diagram showing an example of the hardware structure of a machine learning device that performs machine learning processing for generating a stability prediction model.

[0059] Fig.15 This is a flowchart showing an example of a machine learning method performed by a machine learning device.

[0060] Fig.16 This is a flowchart showing an example of an information processing method executed by the information processing device according to the embodiment. DETAILED DESCRIPTION

[0061] Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings.

[0062] 《Overview of the production process of antibody pharmaceuticals》

[0063] Among biopharmaceuticals, antibody pharmaceuticals, whose market is expanding, are produced by cloning animal cells that can stably produce antibodies, proteins with complex structures, from the perspective of high compatibility in terms of efficacy and safety. Hereinafter, antibodies will be used as an example of useful substances. Figure 1 This is an explanatory diagram showing an overview of the production process of antibody pharmaceuticals. The process to produce antibody pharmaceuticals includes [1] the cloning stage, [2] the process development stage, and [3] the GMP (Good Manufacturing Practice) manufacturing stage.

[0064] The cloning production stage includes the following steps: adding vectors to animal cells suitable for the production of antibody pharmaceuticals to carry out genetic recombination and produce multiple cloning candidates; and selecting clones from these multiple candidates that are excellent in terms of antibody production, cell proliferation, and quality stability in which cell characteristics do not change even after repeated proliferation.

[0065] The process development stage is the stage in which the identified clones are used to develop the production process (culture conditions, purification conditions, etc.) required for GMP manufacturing.

[0066] In the GMP manufacturing stage, clones are cultured and multiplied based on established production processes to produce antibodies, and then the antibodies are purified and formulated to produce antibody pharmaceuticals.

[0067] When clones are made to produce antibodies, their productivity is required to remain constant (stable) over a long period of time. Therefore, multiple types of clones are prepared as much as possible, and clones with stable productivity are selected from them. However, in the past, experimental verification requiring continuous cultivation for several months was necessary, and thus the load became high.

[0068] Figure 2 This is a graph showing an example of changes in antibody production based on clones. The vertical axis represents antibody production, and the horizontal axis represents elapsed time (time point). "Antibody production" is represented by the antibody production per unit time of antibodies produced by a clone.

[0069] exist Figure 2 The graphs plotting how the amount of antibodies produced by clones changes over a long period of time (2 to 3 months) are shown. Graph G1 is a graph showing changes in antibody production for clones with stable production. Graph G2 is a graph showing changes in antibody production for clones with unstable production. As shown in Graph G1, the production of clones with stable production remains roughly unchanged even after 2 to 3 months from the current time point, and the production is maintained roughly the same as at the current time point. In contrast, as shown in Graph G2, the production of clones with unstable production gradually decreases during the 2 to 3 months period.

[0070] In the present invention, "the current time point" refers to the time point of the 2-week standard test or the time point at which the standard test ends, that is, the time point at which the culture for determining the production stability begins. In addition, "the production of antibodies at the current time point" refers to the amount of antibodies produced per unit time by the clone in the 2-week standard test.

[0071] In the case of producing antibody-producing cells by gene introduction, Figure 2As shown in FIG. 1 , both stable clones and unstable clones are produced. Therefore, in the clone production stage, multiple types of clones are produced, and a productively stable clone that exhibits a behavior as shown in FIG. G1 is selected from among them.

[0072] like Figure 2 The production behavior shown varies depending on the type of clone. Conventionally, whenever the type of antibody produced by a clone is changed, it is necessary to perform a Figure 2 The same experiment was performed to evaluate the production stability of each clone.

[0073] In contrast, in an embodiment of the present invention, a mechanism is proposed for accurately predicting the stability of antibody production several months later based on information obtained from clones at the current time point. Here, "information obtained from clones at the current time point" refers to information obtained from clones in a 2-week standard test. The stability of antibody production, which is the target variable for prediction, can be defined based on whether there is a change in the amount of antibody produced between the current time point and after a period of several months of culture. The "several months" here refers to, for example, a period of more than 2 months, and illustratively can be 2 to 3 months. Furthermore, it can also be set to a period until a specified number of passages are performed. Regarding the setting of the period, it can be determined based on the proliferation ability of the clone, and can also be determined based on the culture period of the clone when the antibody is actually manufactured. The "current time point" is Figure 2 The time point at the beginning of the culture shown on the left end of the graph, that is, the time point at which the 2-week standard test ends, is the time point at which the culture starts to determine the stability of antibody production. Productivity "stable" means that the amount of antibody produced does not change between the current time point and a few months later. "No change" includes situations where the amount of change is within the allowable range and can be considered to be substantially unchanged. Productivity "unstable" means that the amount of antibody produced changes between the current time point and a few months later, and in most cases refers to a decrease in production. The threshold value for a change in productivity can be set arbitrarily, for example, it can be ±30% or ±20% relative to the production at the current time point.

[0074] On the generalization performance to unknown clones

[0075] Figure 3 This is an explanatory diagram for explaining the role of the stability prediction AI (Artificial Intelligence) achieved by this embodiment. Figure 3As shown, in the clone production stage, a gene design drawing of a gene of a useful substance to be produced is introduced into a host cell. For example, when a design drawing for producing a useful substance A is introduced into a host cell, a cell producing the useful substance A can be obtained. Since such a producing cell can be formed probabilistically, a cell that does not produce the useful substance A or a cell that produces an insufficient amount of the useful substance A may be produced. Therefore, a simple test is first performed at this stage to select a high-yield clone that can produce the useful substance A sufficiently.

[0076] Afterwards, if we follow the previous practice, Figure 2 As described in , a stability test is performed for 2 to 3 months to confirm whether the useful substance A can be continuously produced over several months, and clones with stable production are selected.

[0077] In this embodiment, as a method to replace the previous stability test, a stability prediction AI is constructed, and based on the characteristic data (profile) obtained by measuring the status of the clones at the current time point, that is, the status of the clones in a 2-week standard test, the stability prediction AI predicts the status (productive changes) 2 to 3 months later.

[0078] The types of useful substances (e.g., antibodies) produced by cells vary depending on the purpose, and therefore it is desirable to construct a model that can predict the production stability regardless of the type of useful substances produced by cells. In other words, a model that robustly predicts the production stability of antibodies for unknown antibody types is preferred.

[0079] When learning a model suitable for stability prediction AI, the useful substances used as the object cannot be known in advance, and the type of useful substances used when learning the model may be different from the useful substances generated by the clone of the object predicted by the model after learning. That is, it is preferred to have a model that can predict the stability of unknown useful substance types robustly and accurately, and it is preferred to construct a prediction model with domain generalization using the useful substance types as the domain.

[0080] An overview of machine learning models for predictive stability

[0081] In the present embodiment, at the clone production stage, a stability prediction AI is constructed that can estimate (predict) whether there will be changes in production 2 to 3 months later based on the information of the clone at the current time point, that is, can predict the production stability of useful substances. More specifically, a model is constructed that receives the input of the gene expression data of the clone at the current time point (during the standard test) and outputs a stability label representing the production stability of the useful substance. In more detail, a part of the clone is used as a clone for standard testing, and another part is used as a clone for genetic analysis to obtain gene expression data, thereby obtaining the gene expression data of the clone used in the standard test. The stability label can be represented by two values, "1" representing a value of "stable" or "0" representing a value of "unstable". The prediction model for predicting production stability can be a 2-category classification model that classifies into "stable" or "unstable".

[0082] Gene expression data includes more than one gene level. The gene expression data used in the present embodiment includes data obtained by digitizing the gene expression level of each of multiple genes. Gene expression data can be obtained, for example, by RNA (ribonucleic acid: ribonucleic acid) sequence analysis. The value representing the gene expression amount is, for example, a count value taking a positive integer, which can be logarithmically transformed and used as a feature amount.

[0083] Figure 4 This is a conceptual diagram of the MLM machine learning model that predicts stability based on gene expression data. Figure 4 An example of a data set of training data is shown inside the rectangular frame RF1. Figure 4In the present invention, the gene expression pattern GEP is represented as a heat map that visualizes the gene expression data of each of the multiple clones A to N at the current time point (during the standard test). The horizontal axis of the gene expression pattern GEP represents the type of gene, and the gene expression level of each of the multiple genes is represented by a two-color grayscale (heat map). The number of types of genes a, b, c, d, etc. contained in the gene expression data is preferably as follows: all gene expression data are obtained for stable clones and unstable clones, and 300 to 400 types are selected using two groups of statistically significant probabilities between stable clones and unstable clones. Moreover, in the case of reducing the number of gene types, it is preferred to use the selected genes while increasing or decreasing the number of gene types while actually training the machine learning model MLM, searching for the number of types with higher prediction performance, and narrowing it down to, for example, 50 to 100 genes. In addition, all gene expression data are obtained here, but it is not necessarily necessary to obtain all gene expression data. A part of the genes can also be randomly selected and their gene expression data can be obtained. Due to the limitation of the diagram, the color of the heat map cannot be displayed, so red is displayed as "R", blue is displayed as "B", and white is displayed as "W". Red (R) indicates that the gene expression level is relatively high, and blue (B) indicates that the gene expression level is relatively low. White (W) indicates that the gene expression level is an intermediate value.

[0084] Regarding each of the multiple clones A to N, it can be confirmed as "stable" or "unstable" based on experimental verification based on cultivation over a period of several months after the standard test, and each clone A to N is assigned a stability label (correct answer label) indicating "stable" or "unstable". In this way, a data set is prepared that includes multiple training data that establish an association (relationship) between the gene expression data of the current time point of each of the multiple clones A to N and the stability label that is the correct answer. Thereafter, the machine learning model MLM is trained using the multiple training data, and the machine learning model MLM is made to learn stable or unstable gene patterns. If the gene expression data of the current time point (during the standard test) of the unknown clone X is input to the machine learning model MLM that has been trained and learned (trained), the machine learning model MLM predicts stability based on the input gene expression data, and outputs a "stable" or "unstable" label as the prediction result. In addition, in Figure 4 FIG. 4 shows an example in which the machine learning model MLM predicts that an unknown clone X is “stable”.

[0085] 《Outline of implementation method: Constructing a model for predicting the production stability of useful substances by limiting the prediction objects》

[0086] Since clones producing useful substances have various characteristics, it is difficult to predict the production stability of all clones with high accuracy regardless of type. In this embodiment, the prediction object is limited based on the index obtained from the culture data of each clone at the current time point (during the standard test), and a high-precision prediction is achieved. Here, the culture data refers to general data about the clone that can be measured using a culture device or a dedicated device for sampling a portion of the culture solution containing cells.

[0087] Figure 5 : is an explanatory diagram showing the outline of the method for predicting the production stability of useful substances cloned according to the present embodiment. Figure 5 The left figure F5A of FIG. 5A shows a comparative example of a case where the prediction object is not limited. Figure 5 An overview of the method based on this embodiment is shown in the right figure F5B.

[0088] The case of not limiting the prediction object in the left figure F5A is explained. The data set DSc containing training data of various types of clones producing useful substances A to D is schematically shown in the rectangular frame RF2 of the left figure F5A. The data set DSc contains training data of 20 clones in total, 5 clones for each useful substance A to D. Here, the training data is data that establishes an association (connection) between the gene expression data of each of the 20 clones during the standard test and the correct stability label. Figure 5 The values ​​"9", "7", "6" and the like displayed at the bottom of each clone represent the measured values ​​of a certain culture data of each clone during the standard test. In addition, instead of indicating the measured values, the relative levels in each clone that can be obtained from the measured values ​​may be indicated. Here, the left figure F5A is described, but the right figure F5B is the same.

[0089] In the left figure F5A, it is shown that the machine learning model MLMc is trained using all the training data of the data set DSc without limiting the training data, and the learned (trained) model is used to predict the production stability of multiple types of clones that produce an unknown useful substance X. In this case, there is no particular limitation on the multiple types of clones that produce the unknown useful substance X as the prediction object, and as shown in the rectangular frame RF3, the production stability of all five clones that produce the unknown useful substance X is predicted. The prediction of production stability is performed by obtaining the gene expression data of all five clones that produce the unknown useful substance X at the current time point (during the standard test) and inputting it into the learned (trained) model, but its prediction accuracy is low.

[0090] Next, the method based on the present embodiment of the right figure F5B is described. Compared with the method of not limiting the prediction object of the left figure F5A, in the method shown in the right figure F5B, the prediction object is limited by the value of a certain culture data during the standard test as an indicator. First, the value of a certain culture data is focused on to determine the threshold, and the clone population contained in the data set DSd is grouped. For example, the value of the culture data as an indicator is divided into two groups that are relatively large and small relative to the threshold. Here, an example is shown in which the threshold is set to "5", the population with a value of "5" or more of the culture data as an indicator is set as the training object, and the population with a value of culture data less than "5" is excluded from the object. Through this threshold processing, as shown in the rectangular frame RF4, the training data of 3 clones each for each useful substance A~D, a total of 12 clones, is retained as an object, and the data set DSe including the training data of these limited populations is used for the training of the machine learning model MLMe. On the other hand, the training data of the 8 clones shown in the dotted rectangular frame RF5, that is, the training data of the clones that do not meet the threshold condition, are excluded from the processing object.

[0091] In this way, the data set DSe with a limited object is used to train the machine learning model MLMe. In addition, when the learned model is used to predict the production stability of clones that produce unknown useful substances X, the clones that are the prediction objects are the same as the clones of the data set DSe used in the training of the model. The threshold is applied to the value of the culture data as an indicator, and the prediction is limited to the group that satisfies the limiting conditions based on the threshold (the group whose indicator value is higher than the threshold). The three clones shown in the rectangular frame RF6 represent clones corresponding to the prediction object. In addition, the two clones shown in the dotted rectangular frame RF7 represent clones that are excluded from the prediction object. In this way, by limiting the prediction object for prediction, high prediction accuracy can be achieved. Moreover, the clones excluded from the prediction object shown in the dotted rectangular frame RF7 do not need to obtain gene expression data, so the cost of gene analysis can be suppressed.

[0092] Examples of datasets used in training and evaluation

[0093] Figure 6 Examples of datasets used for model training and evaluation are shown in Figure 6 The upper part shows an example of a data set DSA for clones producing antibody A as a useful substance, and the lower part shows an example of a data set DSB for clones producing antibody B as a useful substance. Although not shown in the figure, the data sets for clones producing other types of antibodies as useful substances are the same.

[0094] The data set DSA includes culture data during standard tests measured for each of multiple clones ACLj, gene expression data during standard tests, and stability labels obtained as correct answers through stability tests. The suffix j represents the index number that identifies the clone. The culture data may include, for example, one or more items such as antibody production, integral viable cell density (IVCD), lactate concentration, pH, etc. The culture data may be general data that can be sampled from a portion of a culture device or a culture solution containing cells and measured using a dedicated device, for example, it may include one or more of the total number of cells, the amount of cell secretory substances, the amount of cell products, the amount of cell metabolites, and the amount of culture medium components. Figure 6 The character symbols (symbols with the suffix "j") in each cell of the table shown represent the value of the corresponding data item.

[0095] The same applies to the data set DSB. The number na of clones ACLj included in the data set DSA and the number nb of clones BCLj included in the data set DSB may be different.

[0096] From the data sets of a plurality of domains (types of useful substances) prepared in this way, the target is narrowed down (restricted) by focusing on an index of a certain cultivation data.

[0097] 《Example of reducing the prediction target》

[0098] Figure 7 This is a graph showing a narrowed example of a prediction target based on a certain indicator of culture data. On the horizontal axis, multiple types of clones that produce multiple useful substances A to E are arranged. The vertical axis is the value of a certain indicator obtained from the culture data during the standard test. In addition, Figure 7 The clones shown are the clones used for training (learning) of the model.

[0099] like Figure 7 As shown in FIG. 1 , sometimes the distribution range of a certain index obtained from the culture data is different for clones that produce different types of useful substances. In this case, if Figure 5 As described in , a threshold is determined relative to the value of the index, clones are divided into two groups according to the relative size relationship with the threshold, and a group of clones used in training and a group of clones excluded from training are determined. Then, depending on the type of useful substance produced, the number of clones that become training objects may deviate. For example, when the threshold of the index is set to 2.5 and a clone group with a value above the threshold is used for training, clones that produce useful substance B are not used for training.

[0100] So, for example Figure 7As shown, with respect to clones producing each useful substance A to D, the training objects can be limited to the top X% (Top-X%) of a certain indicator obtained from the culture data. Here, the top X% refers to the top X% (Top-X%) of the population of clones producing each useful substance A to D when arranged in descending order for a certain indicator obtained from the culture data. The benchmark "X%" which is equivalent to the threshold value of the limiting condition is preferably adjusted so that the number of samples from each useful substance A to E becomes approximately the same. The top X% is an example of a "threshold defined by the rank of the value of the usage indicator" in the present invention.

[0101] Even if the training objects are limited in this way, there may be clones with stable production and clones with unstable production. In addition, when using the learned (trained) model to predict the production stability of clones that produce unknown useful substance Y, the clones used as prediction objects are also predicted by limiting a certain indicator obtained from the culture data to the top X% of clones, just like the clones used in the model training.

[0102] Here, multiple types of clones that produce multiple useful substances A to E are used for model training, but it is not necessary to use multiple clones that produce different useful substances. For example, only clones that produce useful substance A may be used for training. In this case, the method for limiting the clone population used in training can focus on the value of a certain culture data during the standard test to set a threshold value, and perform it based on the relative size relationship with the threshold value, or the value of a certain indicator obtained from the culture data during the standard test can be set to the top X%. Moreover, it is set to the relative top X%, but it can also be set to the relative bottom X% based on a certain indicator obtained from the culture data.

[0103] The index and threshold of the culture data when limiting the prediction object can be determined by repeatedly performing hypothesis and verification operations on the prepared data set in a trial and error manner. Alternatively, the index and threshold of the culture data when limiting the prediction object can be determined by performing exploratory analysis on the prepared data set.

[0104] For example, in the presence of Figure 7In the case of the data set of the five useful substances A to E (domains) shown, an information processing device including a processor uses a feature selection method such as a filter method to evaluate the correlation between each feature quantity and the target variable (stability label) in each of the five domains, and sets the feature quantity with high correlation in, for example, four or more of the five domains as a feature quantity with high domain universality. The information processing device extracts data that meet specific conditions as a subset from all the data with a focus on a certain indicator, and evaluates the domain generalization of the extracted subset based on the number of feature quantities with high domain universality. In the case where the number of feature quantities with high domain universality is large, it is evaluated as a subset with high domain generalization. By using the data of the subset with high domain generalization as training data to learn (train) the prediction model, the learned model can also robustly predict the stability of other domains (types of useful substances) for the group (subset) of objects limited under the same conditions as when learning.

[0105] In the case of antibody-producing clones, indicators of culture data that are effective for limiting the prediction object include, for example, antibody production amount, integrated viable cell density, lactate concentration, etc. It was confirmed that by taking the superordinate group of the value of any one of these indicators as the object, high-precision production stability prediction can be performed.

[0106] Examples of Useful Substances

[0107] The useful substance is not limited to an antibody, but may be an antibody-like protein. The useful substance may be any of proteins, peptides, and viruses that are raw materials for pharmaceuticals.

[0108] The Example of Cloning

[0109] The clones producing useful substances may be cells derived from vertebrates. The clones may be cells derived from mammals, for example. The clones may be CHO cells or HEK cells.

[0110] 《Implementation Examples》

[0111] Hereinafter, embodiments 1 to 3 to which the technology of the present invention is applied are described. The common structure in each embodiment 1 to 3 is as follows. That is, the useful substance is set as an antibody, and the producing cell is set as a CHO cell. The following example is shown: as evaluation samples, multiple clones are prepared for 5 clones of antibody-producing CHO cells, and 100 gene expression levels are selected as explanatory variables from all gene expression levels measured in a 2-week standard test by RNA sequencing (RNA-Seq) analysis, and a 5-fold cross-validation of a logistic regression model classified as two categories of stable or unstable is implemented as a learner to train (learn) the prediction model, and the performance evaluation uses PRAUC (Area Under the Precision-Recall Curve: Area Under the Precision-Recall Curve). Regarding the types of gene expression levels used for explanatory variables, in embodiments 1 to 3, 300 to 400 genes selected by statistically significant probability are used, and the number of types is increased or decreased while the prediction model is actually trained (learned) to search for the number of types with higher prediction performance, thereby setting it to 100 types. In addition, for the standard test, the clone (CHO cells) was seeded at a number of 5×10^5 cells / mL and suspension culture was performed in a 40 mL flask.

[0112] The 5-fold cross validation was divided into five antibody types, and the performance based on the unlearned antibody types was evaluated. That is, the data sets of four antibody types were used as training (learning) data, and the data set of the remaining one antibody type was used as test data for performance evaluation.

[0113] Figure 8 This is a table showing the number of clones of 5 antibody-producing CHO cells prepared as evaluation samples and examples of assigning stability labels. As evaluation samples, 182 clones of 5 antibody-producing cells were prepared, and cell culture was performed for 2 months under the same conditions as the standard test, thereby assigning stability labels ("stable" or "unstable") to each clone (reference Figure 8 ). For example, there are 24 clones producing antibody A, of which 7 clones are assigned a "stable" label and 17 clones are assigned an "unstable" label. In addition, for each of the 182 clones, culture data and gene expression data are obtained during the standard test, and the gene expression data is associated with the stability label, etc. for each clone to form training data.

[0114] [Example 1]

[0115] In Example 1, an example of stability prediction is described in which the prediction targets are limited to “relatively high-producing clones.” Here, “relatively high-producing clones” refer to clones that produce a relatively high amount of a useful substance.

[0116] The method of limiting clones that are training objects for training a prediction model when performing stability prediction in which the prediction objects are limited to "relatively high-yield clones" is described. In addition, limiting the clones that are training objects is equivalent to limiting the clones that are the objects for prediction by the prediction model during training, that is, equivalent to limiting the clones that are the prediction objects based on the prediction model.

[0117] The method for limiting the clones used as training objects is set as follows: focus on the "antibody production" from the culture data during the standard test of all 182 clones, search for a threshold in a way that the prediction performance in the prediction model becomes higher, and limit it to clones with a relative ranking of the top 40% in each antibody type. Here, the "antibody production" can be set, for example, as the cumulative amount of antibody production in 2 weeks (14 days) in the standard test. Alternatively, it can also be set to a certain period in the standard test, such as a period, such as the cumulative amount of antibody production for 10 days, or it can be set to the antibody production per unit time obtained by dividing by the measurement period. "Top 40%" is an example of a threshold. Fig. 9 This is a table showing examples of the number of clones whose antibody production values ​​belong to the top 40% of the relative ranking in each antibody type and the assignment of stability labels.

[0118] exist Fig. 9 A total of 73 clones belonging to the top 40% of the relative ranking in each antibody class are shown in FIG. Fig. 9 A 5-fold cross validation was performed on the data set of each antibody type of training data in which the gene expression data of the standard test of the 73 clones shown was associated with the stability label. As a result of limiting the training object in this way, the PRAUC value of the prediction performance of the prediction model after learning became 0.743. In addition, regarding the clones that are the prediction objects when predicting the production stability of unknown useful substances using the prediction model after learning, they are also set as the culture data during the analysis of the standard test (current time point) in the same way as the limitation of the training object, and are limited to the top 40% of the "antibody production amount" for prediction.

[0119] [Comparative Example]

[0120] In contrast, the training subjects are not limited and the use includes Figure 8 The same learning was performed on the data set of all data of the 182 clones shown, and the PRAUC value of the prediction performance of the prediction model involved in the comparative example obtained when 5-fold cross validation was performed was 0.503. In addition, the prediction object was set to be unrestricted as the training object. It was confirmed that the performance of the prediction model in which the prediction object was limited by Example 1 was higher than the performance of the prediction model involved in the comparative example.

[0121] This result indicates that the prediction model generated by the method of Example 1 can be used to predict unknown useful substances with high accuracy. At the same time, limiting the clones to relatively high-yielding clones will not be an obstacle at all in the selection process of clones that produce useful substances. By limiting the objects to those that can be predicted with high accuracy, it can be implemented at low cost. Therefore, it is believed that the stability prediction based on the present invention can be used in practice.

[0122] [Example 2]

[0123] In Example 2, an example of stability prediction in which the prediction object is limited to "clones with high relative cell density" is described. First, a method for limiting clones as training objects for training a prediction model when stability prediction is limited to "clones with high relative cell density" is described. As in Example 1, the following method is used: Figure 8 The culture data of all 182 clones in the standard test are shown, focusing on the "integrated viable cell density (IVCD)", searching for a threshold in a way that the prediction performance in the prediction model becomes higher, and limiting it to the top 60% of clones in relative ranking in each antibody type. Here, "clones with relatively high cell density" can be obtained, for example, based on the "integrated viable cell density (IVCD)" of 2 weeks (14 days) in the standard test. Alternatively, it can also be obtained based on a certain period in the standard test, such as 10 days of "integrated viable cell density (IVCD)". "Top 60%" is an example of a threshold. Fig.10 This is a table showing the number of clones whose integrated viable cell density values ​​belong to the top 60% of the relative ranking in each antibody type and examples of assignment of stability labels.

[0124] exist Fig.10 A total of 109 clones belonging to the top 60% of the relative ranking in each antibody class are shown in FIG. Fig.10 A 5-fold cross validation was performed on the data set of each antibody type of training data in which the gene expression data during the standard test of the 109 clones shown was associated with the stability label. As a result of limiting the training object in this way, the PRAUC value of the prediction performance of the prediction model after learning became 0.647. That is, it was confirmed that the performance of the prediction model in which the prediction object was limited by Example 2 was highly accurate compared with the PRAUC (0.503) of the prediction model involved in the comparative example in which the object was not limited. In addition, regarding the clones that are the prediction objects when using the prediction model after learning to predict the production stability of unknown useful substances, the culture data during the analysis of the standard test (current time point) are set in the same way as the limitation of the training object, and are limited to the top 60% of the "integrated viable cell density (IVCD)" for prediction.

[0125] This result indicates that the prediction model generated by the method of Example 2 can be used to predict unknown useful substances with high accuracy. At the same time, limiting the number of clones to clones with relatively high viable cell density will not become an obstacle in the selection process of clones that produce useful substances. By limiting the number of objects to be predicted with high accuracy, it can be implemented at low cost. Therefore, it is believed that the stability prediction based on the present invention can be actually used.

[0126] [Example 3]

[0127] In Example 3, an example of stability prediction limited to "clones with relatively high lactate concentrations" is described. First, a method of limiting clones as training objects for training a prediction model when performing stability prediction limited to "clones with relatively high lactate concentrations" is described. Figure 8 The culture data of the 2-week standard test of a total of 182 clones shown above focus on the "lactate concentration" of the culture solution of the cultured clones, and the median value of the "lactate concentration" of the culture solution measured at each time point within 2 weeks (14 days), for example, every day, is used as a representative value to obtain the "lactate concentration" of each clone. After that, a method is set to search for a threshold value in a manner that increases the prediction performance in the prediction model, and limit it to the top 40% of clones in relative ranking in each antibody type. "Top 40%" is an example of a threshold value. Fig.11 This is a table showing the lactate concentration values ​​and the relative rankings of the top 40% of clones in each antibody type and examples of the assignment of stability labels.

[0128] exist Fig.11 The examples of 72 clones in total corresponding to the top 40% of the relative ranking in each antibody type are shown in FIG. Fig. 9 The reason why the number of clones is one less than that of the control group is that the data of one clone is missing in the determination of lactate concentration.

[0129] Use includes Fig.11 A 5-fold cross validation was performed on the data set of each antibody type of training data in which the gene expression data during the standard test of the 72 clones shown was associated with the stability label. As a result of limiting the object in this way, the PRAUC value of the prediction performance of the prediction model after learning became 0.613. That is, it was confirmed that the performance of the prediction model in which the prediction object was limited by Example 3 was highly accurate compared with the PRAUC (0.503) of the prediction model involved in the comparative example in which the object was not limited. In addition, with regard to the clones that are the prediction objects when using the prediction model after learning to predict the production stability of unknown useful substances, the culture data during the analysis of the standard test (current time point) is set in the same way as the limitation of the training objects, and the prediction is limited to the top 40% of the "lactic acid concentration" for prediction.

[0130] This result indicates that unknown useful substances can be predicted with high accuracy. At the same time, limiting the number of clones to clones with relatively high lactic acid concentrations will not become an obstacle in the selection process of clones producing useful substances. By limiting the number of objects to be predicted with high accuracy, it can be implemented at low cost. Therefore, it is believed that the stability prediction based on the present invention can be used in practice.

[0131] 《Configuration example of information processing device》

[0132] Fig.12 1 is a block diagram showing the functional structure of the information processing device 10 involved in the embodiment. The information processing device 10 includes a data acquisition unit 12, a prediction object limiting unit 14, a stability prediction model generation unit 16, and a processing result output unit 18. The various functions of the information processing device 10 can be realized by a combination of computer hardware and software. The physical form of the information processing device 10 is not particularly limited, and it can be a server computer, a workstation, a personal computer, a tablet terminal, etc.

[0133] The data acquisition unit 12 acquires various data including culture data and gene expression data on one or more clones of the clones producing the useful substance.

[0134] The prediction object limitation unit 14 includes a culture data analysis unit 20 and a limiting condition determination unit 22, which analyzes the culture data of one or more clones input to limit the clones that are the prediction objects. The culture data analysis unit 20 analyzes the culture data. The limiting condition determination unit 22 limits the object by a threshold value based on the analysis results of the culture data. In addition, for ease of explanation, the culture data analysis unit 20 and the limiting condition determination unit 22 are recorded separately, but the limiting condition determination unit 22 can also be included in the culture data analysis unit 20. In addition, it can also be understood that the culture data analysis unit 20 functions as the prediction object limitation unit 14.

[0135] The culture data analysis unit 20 is capable of executing a process of determining an index and a threshold value for limiting a prediction object from an input data set. In addition, the index and the threshold value that become limiting conditions for the prediction object can be set based on the analysis results based on the culture data analysis unit 20, or can be set in the prediction object limiting unit 14 as known information that is previously known by using the results of a search process using other information processing devices not shown in the figure.

[0136] A machine learning model is applied to the production stability prediction model 16. The production stability prediction model 16 may be a two-class classification model that receives the input of the gene expression data of the clone at the current time point as the prediction target, predicts the production stability of the clone based on the input gene expression data, and outputs a stability label. The production stability prediction model 16 uses the Figure 5The training data of the object is limited by the method described in the right figure F5B of the present invention. The gene expression data input to the stability prediction model 16 includes more than one gene expression level. The gene expression data input to the stability prediction model 16 may include data on the expression levels of multiple genes. The feature quantity used as an explanatory variable can be selected by a known feature quantity selection method.

[0137] The processing result output unit 18 outputs processing results including prediction results generated by the stability prediction model 16. The processing result output unit 18 may be configured to perform at least one of displaying the processing results, recording the processing results in a database, etc., and printing the processing results.

[0138] Fig.13 1 is a block diagram showing an example of the hardware configuration of the information processing device 10. Here, an example is described in which the processing functions of the information processing device 10 are realized by using one computer, but the processing functions of the information processing device 10 may also be realized by a computer system composed of a plurality of computers.

[0139] The information processing device 10 includes a processor 102 , a non-transitory tangible computer readable medium 104 , a communication interface 106 , an input / output interface 108 , and a bus 110 . The processor 102 is connected to the computer readable medium 104 , the communication interface 106 , and the input / output interface 108 via the bus 110 .

[0140] The processor 102 includes a CPU (Central Processing Unit). The processor 102 may also include a GPU (Graphics Processing Unit). The computer-readable medium 104 includes a memory 112 as a main storage device and a storage device 114 as an auxiliary storage device. The computer-readable medium 104 may be, for example, a semiconductor memory, a hard disk (HDD) device, or a solid state drive (SSD) device, or a combination of multiple thereof. The computer-readable medium 104 is an example of a "storage device" in the present invention.

[0141] The computer-readable medium 104 includes a data storage area 120 for storing various data such as culture data and gene expression data of more than one clone. In addition, a plurality of programs including a prediction object limiting program 140, a stability prediction model generation 16, a processing result output program 180, and a display control program 190, as well as data and the like are stored in the computer-readable medium 104. The term "program" includes the concept of a program module and includes commands that follow the program. The processor 102 functions as various processing units by executing the commands of the programs stored in the computer-readable medium 104.

[0142] The prediction object limiting program 140 includes a command to perform the following processing: analyzing the culture data to limit the prediction object. The prediction object limiting program 140 can be configured to include a culture data analysis program 142 and a limiting condition determination program 144. The culture data analysis program 142 includes a command to perform the following processing: analyzing the culture data of one or more clones. The culture data analysis program 142 may include a command to perform the following processing: searching for indicators and thresholds for narrowing down the prediction object from the data set.

[0143] The restriction condition determination program 144 includes instructions for executing a process of restricting prediction targets based on the index and threshold value set as restriction conditions by using the analysis results of the cultivation data analysis program 142 .

[0144] The production stability prediction model 16 includes commands for executing the following process: receiving input of gene expression data of clones related to prediction targets that satisfy defined conditions, and predicting production stability.

[0145] The processing result output program 180 includes commands for executing the following processing: outputting the processing result including the production stability predicted by the production stability prediction model 16. The display control program 190 includes commands for executing the following processing: generating a display signal required for display output of the display device 154 and performing display control of the display device 154.

[0146] The communication interface 106 performs communication processing with an external device via wired or wireless communication, and exchanges information with the external device. The information processing device 10 is connected to a communication line (not shown) via the communication interface 106. The communication line can be a local area network, a wide area network, or a combination thereof. The communication interface 106 can serve as a data acquisition unit that receives data input.

[0147] The information processing device 10 may include an input device 152 and a display device 154. The input device 152 may be, for example, a keyboard, a mouse, a multi-touch panel or other pointing device or a voice input device, or a suitable combination thereof. The display device 154 may be, for example, a liquid crystal display, an organic EL (organic electro-luminescence: OEL) display or a projector, or a suitable combination thereof. The input device 152 and the display device 154 are connected to the processor 102 via the input / output interface 108. In addition, the input device 152 and the display device 154 may be integrated as a touch panel, or the information processing device 10, the input device 152 and the display device 154 may be integrated as a touch panel, or the information processing device 10, the input device 152 and the display device 154 may be integrated as a touch panel type tablet terminal.

[0148] 《Configuration example of a machine learning device》

[0149] Fig.14 This is a block diagram showing an example of the hardware structure of a machine learning device 300 that performs machine learning processing for generating a stability prediction model 16. Here, an example of using one computer to implement the processing function of the machine learning device 300 is described, but the processing function of the machine learning device 300 can also be implemented by using a computer system composed of multiple computers.

[0150] The machine learning device 300 includes a processor 302, a non-transitory tangible computer-readable medium 304, a communication interface 306, an input / output interface 308, and a bus 310. The computer-readable medium 304 includes a memory 312 and a storage device 314. The processor 302 is connected to the computer-readable medium 304, the communication interface 306, and the input / output interface 308 via the bus 310. An input device 352 and a display device 354 are connected to the bus 310 via the input / output interface 308.

[0151] The hardware structure of the machine learning device 300 can be the same as that in Figure 6 The corresponding elements of the information processing device 10 described in are the same. The machine learning device 300 may be a server computer, a personal computer, or a workstation. The machine learning device 300 is an example of a "system including one or more processors" in the present invention.

[0152] The machine learning device 300 is connected to a communication line (not shown) via the communication interface 306, and is connected to an external device such as a data storage unit 550 in a manner capable of communicating. The data storage unit 550 includes a storage device that stores a data set including a plurality of training data. The data storage unit 550 may store data including the following: Figure 6The data set of all data of the multiple domains illustrated in the example may also store a data set including only data of samples of objects limited as prediction objects. In addition, the data storage unit 550 may also be constructed in the storage device 314 in the machine learning device 300.

[0153] The computer readable medium 304 stores a plurality of programs including a prediction object definition program 320, a learning processing program 330, and a display control program 340, as well as data and the like. The prediction object definition program 320 can be used in conjunction with the computer readable medium 304. Fig.12 The display control program 340 may be the same as the prediction object limiting program 140 described in Fig.12 The display control program 190 described in is the same.

[0154] The computer-readable medium 304 includes a prediction object data storage area 322. The training data corresponding to the defined prediction object is stored in the prediction object data storage area 322. The corresponding training data can be appropriately sampled from the data set stored in the data storage unit 550 by the prediction object definition program 320, or a data set containing only the prediction object can be pre-extracted as a subset.

[0155] The learning processing program 330 includes a data acquisition program 400, a prediction model 410 as a machine learning model, a loss calculation program 430, and an optimizer 440. The data acquisition program 400 includes a command for executing the following process: acquiring training data from the prediction target data storage area 322. The training data acquired by the data acquisition program 400 is input to the prediction model 410.

[0156] The loss calculation program 430 includes a command for executing the following processing: calculating the loss representing the error between the predicted value of the stability label output from the prediction model 410 and the correct stability label. The optimizer 440 includes a command for executing the following processing: calculating the update amount of the parameters of the prediction model 410 based on the calculated loss, and updating the parameters of the prediction model 410. The optimizer 440 can optimize the parameters by a method such as stochastic gradient descent (SGD).

[0157] Flowchart of machine learning methods

[0158] Fig.15 is a flowchart showing an example of a machine learning method performed by the machine learning device 300. Figure 6 The case of a data set for machine learning is described in detail with reference to FIG. 1. In step S102, the processor 302 obtains the cultivation data from the prepared data set.

[0159] In step S104, the processor 302 analyzes the culture data and defines the training object. The processor 302 can screen the data of the object sample that meets the definition condition or the data of the non-object sample that does not meet the definition condition according to the index and threshold of the culture data specified in advance, or can search the index and threshold as the definition condition according to the culture data and screen the data of the object sample and the data of the non-object sample.

[0160] In step S106, the processor 302 performs machine learning using only the data of clones that meet the defined conditions, and trains the prediction model 410. That is, the processor 302 inputs the gene expression data of the samples that meet the defined conditions into the prediction model 410, and calculates the loss representing the error between the predicted value of the stability label output from the prediction model 410 and the correct stability label. The processor 302 calculates the update amount of the parameters of the prediction model 410 based on the calculated loss, and updates the parameters. In this way, the processor 302 trains the prediction model 410 in a manner that makes the output (prediction value) from the prediction model 410 close to the correct stability label relative to the data input into the prediction model 410. In addition, the update of the parameters of the prediction model 410 can be implemented in small batches.

[0161] In step S108, processor 302 determines whether to end learning. The end condition of learning can be determined based on the loss value or the number of parameter updates. As a method based on the loss value, for example, the loss converging within a specified range can be used as a learning end condition. And, as a method based on the number of updates, for example, the number of updates reaching a specified number can be used as a learning end condition. Alternatively, in addition to the training data, a data set for performance evaluation of the model can be prepared, and whether learning can be terminated can be determined based on the evaluation value using the evaluation data.

[0162] If the result of the determination in step S108 is "No", the processor 302 returns to step S106 to continue the learning process. On the other hand, if the result of the determination in step S108 is "Yes", the processor 302 ends the learning process. Fig.12 Flowchart of the process.

[0163] The learned prediction model 410 is incorporated into the information processing device 10 as the generated stability prediction model 16. The machine learning method executed by the machine learning device 300 can be understood as a method for generating the generated stability prediction model 16, which is an example of the prediction model generation method in the present invention.

[0164] "Flowchart of information processing method for making predictions of generating stability"

[0165] Fig.162 is a flowchart showing an example of an information processing method executed by the information processing device 10. In step S202, the processor 102 acquires culture data measured for a clone producing a useful substance. The processor 102 may automatically acquire data from a data storage server (not shown) or the like, or may receive a designated input of data via a user interface and acquire data about the designated clone.

[0166] In step S204, the processor 102 analyzes the culture data and limits the prediction object. The processor 102 limits the prediction object by applying the same limiting conditions as the conditions for limiting the training object when training the stability prediction model 16. In addition, after limiting the prediction object in step S204, the gene expression data of the clones belonging to the prediction object is measured, thereby reducing the workload and cost compared with the case of performing gene analysis on all clones.

[0167] In step S206 , the processor 102 inputs the gene expression data of the clones belonging to the prediction target into the generation stability prediction model 16 , and predicts the stability by using the generation stability prediction model 16 .

[0168] In step S208, the processor 102 outputs the prediction result output from the production stability prediction model 16. Based on the prediction result of the production stability, selection of a production clone can be performed.

[0169] After step S208, the processor 102 ends Fig.16 Flowchart of the process.

[0170] About the Programs That Run Computers

[0171] A program that enables a computer to implement part or all of the processing functions in each device of the information processing device 10 and the machine learning device 300 involved in the implementation method can be recorded on a CD, a disk, a semiconductor memory, or other computer-readable medium that is a tangible non-temporary information storage medium, and the program can be provided through the information storage medium.

[0172] Furthermore, instead of providing the program by storing it in such a tangible non-transitory computer-readable medium, it is also possible to provide the program signal as a download service using a telecommunication line such as the Internet.

[0173] Furthermore, part or all of the processing functions in the above-mentioned devices may be realized by cloud computing, and may also be provided as SaaS (Software as a Service).

[0174] 《About the hardware structure of each processing unit》

[0175] The hardware structure of the processing unit (processing unit) that performs various processing, such as the data acquisition unit 12 in the information processing device 10, the prediction object limitation unit 14, the stability prediction unit that includes generating the stability prediction model 16, the processing result output unit 18, the training data analysis unit 20, the limitation condition determination unit 22, the learning unit that includes the prediction model 410 in the machine learning device 300, the loss calculation unit, the parameter update amount calculation unit, the parameter updating unit, etc., is, for example, the various processors (processor) shown below.

[0176] Various processors include general-purpose processors such as CPUs, GPUs, and FPGAs (Field Programmable Gate Arrays) that execute programs and function as various processing units; processors whose circuit structures can be changed after manufacturing, such as programmable logic devices (PLDs) and ASICs (Application Specific Integrated Circuits); and processors with circuit structures specifically designed to perform specific processing, such as dedicated circuits.

[0177] One processing unit may be composed of one of these various processors, or may be composed of two or more processors of the same type or different types. For example, one processing unit may be composed of a plurality of FPGAs or a combination of a CPU and an FPGA or a combination of a CPU and a GPU. Furthermore, multiple processing units may be composed of one processor. As an example of multiple processing units composed of one processor, first, there is a method in which one processor is composed of a combination of one or more CPUs and software, as represented by computers such as clients and servers, and the processor functions as multiple processing units. Second, there is a form of a processor device that uses one IC (Integrated Circuit) chip to implement the functions of the entire system including multiple processing units, as represented by a system on chip (SoC). In this way, various processing units are composed of more than one of the above-mentioned various processors as a hardware structure.

[0178] Furthermore, more specifically, the hardware structure of these various processors is a circuit (circuitry) formed by combining circuit elements such as semiconductor elements.

[0179] 《Advantages of implementation method》

[0180] According to the method of predicting the generation stability of a generated clone and the information processing device 10 for executing the method according to the above-mentioned embodiment, the following effects can be obtained.

[0181] [1] By appropriately limiting the clones to be predicted based on the indicators of the culture data at the current time point (during the standard test), it is possible to predict the stability of the clones to be predicted with high accuracy.

[0182] [2] Since genetic analysis (RNA-Seq analysis) is performed only on clones that are prediction targets, costs can be reduced compared to performing genetic analysis on all clones.

[0183] [3] By applying the method according to this embodiment instead of the conventional stability test, it is possible to shorten the period of the development process of the production cell and reduce the cost.

[0184] "other"

[0185] The present invention is not limited to the above-described embodiment, and various modifications are possible without departing from the gist of the technical concept of the present invention.

[0186] Explanation of symbols

[0187] 10 Information processing device

[0188] 12 Data Acquisition Department

[0189] 14 Prediction target limitation section

[0190] 16. Generate Stability Prediction Model

[0191] 18 Processing result output unit

[0192] 20. Cultivate the Data Analysis Department

[0193] 22. Limitation Condition Judgment Unit

[0194] 102 Processor

[0195] 104 Computer readable medium

[0196] 106 Communication Interface

[0197] 108 Input and output interfaces

[0198] 110 Bus

[0199] 112 Memory

[0200] 114 Storage Device

[0201] 120 Data storage area

[0202] 140 Prediction object limitation procedure

[0203] 142 Cultivation Data Analysis Program

[0204] 144 Condition determination procedure

[0205] 152 Input Devices

[0206] 154 Display device

[0207] 180 Processing result output program

[0208] 190 Display Control Program

[0209] 300 Machine Learning Devices

[0210] 302 Processor

[0211] 304 Computer readable medium

[0212] 306 Communication Interface

[0213] 308 Input and Output Interface

[0214] 310 Bus

[0215] 312 Memory

[0216] 314 Storage Devices

[0217] 320 Prediction Object Limitation Procedure

[0218] 322 Prediction object data storage area

[0219] 330 Learning Processor

[0220] 340 Display Control Program

[0221] 352 Input Device

[0222] 354 Display Device

[0223] 400 Data Acquisition Program

[0224] 410 Prediction Model

[0225] 430 Loss Calculation Procedure

[0226] 440 Optimizer

[0227] 550 Data Storage Department

[0228] DSA, DSB datasets

[0229] DSc, DSd, DSe datasets

[0230] F5A Left

[0231] F5B right

[0232] G1 Chart

[0233] G2 Chart

[0234] GEP gene expression pattern

[0235] MLM Machine Learning Model

[0236] MLMc, MLMe machine learning models

[0237] RF1~RF7 Rectangular frame

[0238] S102~S108 Steps of machine learning method

[0239] S202-S208 Steps of the information processing method for predicting generation stability

Claims

1. A method for predicting the production stability of a clone producing a useful substance, wherein: One or more processors perform the following processing: Obtaining culture data of one or more of the clones; Analyzing the culture data to define clones to be predicted; and The production stability of the useful substance based on the clone as the prediction target is predicted using the data measured for the clone as the prediction target.

2. The method according to claim 1, wherein: The production stability is defined by whether or not there is a change in the production amount of the useful substance between the start of culture and after culture for a predetermined period of time.

3. The method according to claim 1, wherein: The one or more processors perform the following processing: setting an indicator obtained from the culture data and a threshold value associated with the indicator; and The prediction object is limited according to the value of the indicator and the threshold.

4. The method according to claim 3, wherein: The threshold is adjusted so that the prediction accuracy of the generation stability is higher than when the prediction target is not limited.

5. The method according to claim 3, wherein: The threshold is defined using a ranking with respect to the value of the indicator.

6. The method according to claim 3, wherein: The prediction target is a superordinate group of the value of the indicator.

7. The method according to any one of claims 3 to 6, wherein: The indicator is the production amount of the useful substance.

8. The method according to any one of claims 3 to 6, wherein: The index is the integrated viable cell density.

9. The method according to any one of claims 3 to 6, wherein: The indicator is lactate concentration.

10. The method according to any one of claims 1 to 6, wherein: The data used to predict the production stability include expression levels of more than one gene.

11. The method according to any one of claims 1 to 6, wherein: The one or more processors perform the following processing: An input of the data of the prediction object is received, and the production stability is predicted using a two-class classification model for stability or instability.

12. The method according to claim 11, wherein: The model is a model trained by machine learning using a plurality of training data, wherein the plurality of training data is a plurality of training data in which the data of training clones having the same restrictions as the clone to be predicted are associated with the stability labels of the correct answers.

13. The method according to claim 12, wherein: The plurality of training data includes the training data for a plurality of clones having different useful substances produced, The one or more processors predict the production stability of a clone that produces a useful substance different from a useful substance used in training the model.

14. The method according to any one of claims 1 to 6, wherein: The useful substance is any one of proteins, peptides and viruses that are raw materials for pharmaceuticals.

15. The method according to any one of claims 1 to 6, wherein: The useful substance is an antibody or an antibody-like protein.

16. The method according to any one of claims 1 to 6, wherein: The clones are cells derived from vertebrates.

17. The method according to any one of claims 1 to 6, wherein: The clones are cells derived from mammals.

18. The method according to any one of claims 1 to 6, wherein: The clones are CHO cells or HEK cells.

19. An information processing device comprising: More than 1 processor; and one or more storage devices storing instructions for the one or more processors to execute, The one or more processors perform the following processing: For clones that produce useful substances, obtain culture data of one or more clones; analyzing the culture data to define clones to be predicted; and The production stability of the useful substance based on the clone as the prediction target is predicted using the data measured for the clone as the prediction target.

20. A program that enables a computer to implement the following functions: Acquiring culture data of one or more clones for clones that produce useful substances; Analyzing the culture data to define clones for prediction; and The production stability of the useful substance based on the clone as the prediction target is predicted using the data measured for the clone as the prediction target. 21 . A recording medium which is non-transitory and computer-readable and has the program according to claim 20 recorded thereon.

22. A prediction model generation method, which generates a prediction model that enables a computer to realize a function of predicting the production stability of a clone that produces a useful substance, wherein: A system including more than one processor includes the following steps: Obtaining culture data of one or more of the clones; Analyzing the culture data to define clones for prediction; and Machine learning is performed using multiple training data that establish associations between data measured for clones belonging to the prediction object and correct stability labels, and the prediction model is trained in such a way that the output of the prediction model relative to the input of the data is close to the correct stability label.

Images