Model generation method, model generation apparatus, model generation program, estimation method, estimation apparatus, and estimation program

Abstract

In particular, to provide novel technologies for improving accuracy in machine learning techniques related to predicting the properties of pharmaceutical formulations over time. [Solution] A model generation method comprising: an acquisition step of acquiring the blending ratio of each raw material in a formulation and time-dependent property information indicating the properties of the formulation related to the passage of time for each of several formulations; an item determination step of determining items to be used for model training from the blending information based on the blending ratio, based on the sparsity of the data based on the blending ratio of each raw material; a training data creation step of creating training data including feature quantities including the values ​​of the items and time-dependent property information; and a training step of training a model that outputs estimation results of the time-dependent property information of a formulation based on the blending information of the formulation using the training data.

Description

Technical Field

[0001] The present invention relates to a model generation method, a model generation device, a model generation program, an estimation method, an estimation device, and an estimation program.

Background Art

[0002] In recent years, as a model generation technology for predicting physical property values of formulations, an approach using machine learning has been utilized. In this technology, data related to the components and manufacturing processes of formulations is collected, and by generating a prediction model based on this data, it is possible to estimate the physical properties and quality of formulations. In particular, since the properties of formulations change over time, these predictions play an important role in long-term quality control and improvement of products.

[0003] In recent years, the accuracy of machine learning models has been improving, and it has become possible to train models using various datasets. For example, Patent Document 1 describes a learning method for a prediction model having physical property values of formulations as objective variables. With this method, a highly accurate prediction model can be generated by combining a plurality of selection methods.

[0004] Also, Patent Document 2 discloses a technique for efficiently determining the raw material formulation of cosmetics. Furthermore, a method for obtaining evaluation index values of sensory evaluation including stability from formulation feature amounts using a determination model learned based on teacher data is described.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0006] The raw materials used in pharmaceutical formulations are diverse, and for some raw materials, data such as their proportions may be scarce. In machine learning techniques, a certain amount of training data is necessary to achieve sufficient prediction accuracy, and there was room for improvement in accuracy when using machine learning based on information about the formulation of pharmaceuticals.

[0007] Therefore, the present invention aims to provide a novel technology for improving accuracy, particularly in machine learning techniques related to predicting the properties of pharmaceutical formulations over time. [Means for solving the problem]

[0008] [1] A model generation method comprising: an acquisition step of acquiring the blending ratio of each raw material in a formulation and time-dependent property information indicating the properties of the formulation in relation to the passage of time for each of several formulations; an item determination step of determining items to be used for model training from the blending information based on the blending ratios based on the sparsity of the data based on the blending ratios of each raw material; a training data creation step of creating training data including feature quantities including the values ​​of the items and the time-dependent property information; and a training step of training a model that outputs estimation results of the time-dependent property information of a formulation based on the blending information of the formulation using the training data.

[0009] This configuration allows for model generation using items with sufficient data, based on the data's sparsity. This is expected to improve accuracy.

[0010] [2] The model generation method according to [1], wherein the temporal property information includes physical property values ​​after a predetermined period has elapsed from a reference point, and in the learning step, the model is trained with the formulation information as explanatory variables and the physical property values ​​as the objective variable.

[0011] This configuration allows for the generation of a model that estimates physical properties after a predetermined period of time has elapsed, based on the formulation information.

[0012] [3] The model generation method according to [2], wherein in the learning data creation step, the learning data is created using converted physical property values ​​obtained by log transformation of the physical property values, and in the learning step, the model is trained using the learning data in which the converted physical property values ​​are the target variables.

[0013] For example, if there is variability in the region where the data resides, some data points may appear as outliers. To address this problem, the above configuration can be used to transform the data set so that it resides in a continuous region, thereby improving prediction accuracy.

[0014] [4] The model generation method according to any one of [1] to [3], wherein the temporal property information includes a deterioration classification indicating whether or not there is deterioration of the functional properties after a predetermined period has elapsed from a reference point, and the learning step involves learning a model that classifies the deterioration classification of the formulation based on the formulation information.

[0015] By using this configuration, it is possible to generate a model that estimates whether or not sensory properties deteriorate over time.

[0016] [5] The model generation method according to [4], wherein the properties include functional properties.

[0017] [6] The model generation method according to [4] or [5], wherein in the learning step, the learning data of the degradation category indicating that degradation exists is treated as an outlier, and the model is trained using an anomaly detection algorithm.

[0018] This configuration allows for more accurate predictions when there is limited data on formulations that degrade over time.

[0019] [7] In the item determination step, the sum of the blending ratios of raw materials belonging to the same classification is calculated as the blending information for each formulation, and the sparsity of the blending ratio for each raw material and the sum of the blending ratios for each classification are used as candidates for items to be used to train the model, and the items to be used to train the model are determined by calculating the sparsity of each. The model generation method described in any of [1] to [6].

[0020] By adding the sum of the blending ratios for each classification, in addition to the blending ratio for each raw material, as a candidate item for training the model, the number of candidate items that can be used as blending information increases. This increases the amount of information that can be used as blending information, and is expected to improve the stability of predictions.

[0021] [8] In the item determination step, raw materials with the same blending purpose are classified as the same category, and the blending information that shows the sum of the blending ratios of raw materials belonging to the same category is calculated, as described in [7].

[0022] [9] A model generation device comprising an acquisition unit, an item determination unit, a learning data creation unit, and a learning unit, wherein the acquisition unit acquires the blending ratio of each raw material in a formulation and time-dependent property information indicating the properties of the formulation in relation to the passage of time for each of a plurality of formulations; the item determination unit determines items to be used for learning the model from the blending information based on the blending ratios based on the sparsity of the data based on the blending ratios of each raw material; the learning data creation unit creates learning data including feature quantities including the values ​​of the items and the time-dependent property information; and the learning unit uses the learning data to learn a model that outputs an estimation result of the time-dependent property information of a formulation based on the blending information of the formulation.

[0023]

[10] A model generation program that causes a computer to function as an acquisition unit, an item determination unit, a learning data creation unit, and a learning unit, wherein the acquisition unit acquires the blending ratio of each raw material in a formulation and time-dependent property information indicating the properties of the formulation in relation to the passage of time for each of a plurality of formulations; the item determination unit determines items to be used for learning the model from the blending information based on the blending ratios based on the sparsity of the data based on the blending ratios of each raw material; the learning data creation unit creates learning data including feature quantities including the values ​​of the items and the time-dependent property information; and the learning unit uses the learning data to learn a model that outputs the estimation result of the time-dependent property information of a formulation based on the blending information of the formulation.

[0024]

[11] An estimation method for performing an estimation based on the relationship between the formulation properties related to the mixing ratio of the formulation raw materials and the passage of time, wherein formulation information related to the mixing ratio of the formulation raw materials is input into the model generated by the model generation method described in any one of [1] to [8], and an estimation result related to the time-dependent property information indicating the property of the formulation related to the passage of time is obtained as the output of the model, thereby performing an estimation of the time-dependent property information of the formulation from the formulation information.

[0025]

[12] An estimation device comprising a reception unit and an estimation unit, wherein the reception unit receives an input of formulation information related to the mixing ratio of the formulation raw materials, and the estimation unit inputs the formulation information related to the mixing ratio of the formulation raw materials into the model generated by the model generation method described in any one of [1] to [8], and an estimation result related to the time-dependent property information indicating the property of the formulation related to the passage of time is obtained as the output of the model, thereby performing an estimation of the time-dependent property information of the formulation from the formulation information.

[0026]

[13] An estimation program that causes a computer to function as a reception unit and an estimation unit, wherein the reception unit receives an input of formulation information related to the mixing ratio of the formulation raw materials, and the estimation unit inputs the formulation information related to the mixing ratio of the formulation raw materials into the model generated by the model generation method described in any one of [1] to [8], and an estimation result related to the time-dependent property information indicating the property of the formulation related to the passage of time is obtained as the output of the model, thereby performing an estimation of the time-dependent property information of the formulation from the formulation information. [Effect of the Invention]

[0027] According to the present invention, in the machine learning technology related to the prediction of formulation properties related to the passage of time, a novel technology for realizing accuracy improvement can be provided. [Brief Description of the Drawings]

[0028] [Figure 1] Configuration diagram of this embodiment. [Figure 2]Hardware configuration diagram. [Figure 3] A flowchart illustrating the procedure for model generation according to this embodiment. [Figure 4] An example of the classification of raw materials in this embodiment. [Figure 5] A flowchart detailing the item determination process of this embodiment. [Figure 6] An example of training data in Examples 1 and 2. [Figure 7] An example of training data from Examples 3-6. [Figure 8] A figure showing the experimental results in Example 3. [Figure 9] A figure showing the experimental results in Example 4. [Figure 10] A figure showing the experimental results in Example 5. [Figure 11] A figure showing the experimental results in Example 6. [Modes for carrying out the invention]

[0029] This invention relates to a technique for estimating the properties of a pharmaceutical formulation over time. In particular, this invention relates to a technique for determining the items to be used to train a model in a model for estimating the properties of a pharmaceutical formulation over time based on the blending ratio of each raw material of the formulation, according to the sparsity of the blending ratio data.

[0030] Two main types of properties related to the passage of time in a pharmaceutical product are the physical properties after a predetermined period has elapsed from a reference point, and the presence or absence of deterioration in functional properties after a predetermined period has elapsed from a reference point. The model generation and estimation for each of these will be explained in detail below. While preferred embodiments are shown in the drawings, many different forms are possible and the embodiments are not limited to those described herein.

[0031] For example, in this embodiment, the configuration and operation of the model generation device and estimation device are described, but similar effects can be achieved by a system or device having similar functions, a method performed by such a system or device, or a computer program that causes a computer device to perform the method. The program may be provided as a non-transient recording medium that can be read by a computer, or it may be provided so that it can be downloaded from an external server.

[0032] In the following embodiments, "part" may include, for example, hardware resources implemented by a circuit in a broad sense, and information processing of software that can be specifically realized by these hardware resources. In this embodiment, "information" can be represented, for example, by the physical value of a signal value representing voltage or current, the high or low value of a signal value as a set of binary bits composed of 0s or 1s, or by a quantum superposition (so-called qubit), and communication and calculations can be performed on a circuit in a broad sense.

[0033] In a broad sense, a circuit is a set of circuits (Circuitry) that are realized by appropriately combining circuits, processors, and memory. For example, it includes circuits that contain any of the following: CPU (Central Processing Unit), GPU (Graphics Processing Unit), LSI (Large Scale Integration), ASIC (Application Specific Integrated Circuit), FPGA (Field-Programmable Gate Array), etc.

[0034] <1.Definition> In this invention, a formulation refers to a product made by processing various ingredients into a form that is easy to use. For example, it includes pharmaceuticals, pesticides, cosmetics, etc. In the following embodiments, cosmetics such as lotions, creams, and gels will be the main formulations described.

[0035] In this invention, the blending ratio refers to the proportion of each raw material contained in the formulation. Hereinafter, the blending ratio will mainly be assumed to be the proportion by weight. Furthermore, formulation information refers to information about the formulation based on the proportion of each raw material in a pharmaceutical product. For example, it refers to information identified based on the proportion of each raw material, or the proportion of each raw material added together according to a certain standard.

[0036] In this invention, time-dependent property information refers to information that indicates the properties of a formulation in relation to the passage of time. For example, information regarding changes in the properties of a formulation over time is included in time-dependent property information. Properties include sensory properties such as color, aroma, and feel, mechanical properties such as physical properties, and other chemical properties.

[0037] In the following, among the temporal property information, the physical property values ​​after a predetermined period has elapsed from the reference point will be referred to as temporal property values. Furthermore, among the temporal property information, the classification indicating whether or not the properties of the formulation have deteriorated after a predetermined period has elapsed from the reference point will be referred to as the deterioration classification. Note that the deterioration classification may not only indicate whether or not deterioration has occurred, but may also further classify the degree of deterioration.

[0038] In this invention, sparsity refers to the small number of elements that determine the essential characteristics of a data set. That is, the higher the sparsity, the fewer elements that determine the essential characteristics are included in the data set. In the following embodiments, the small number of non-zero data points (the large number of zero data points) in a data set of the total blending ratios for each raw material or each classification is used as an indicator of sparsity. More specifically, the proportion of data points with a value of 0 in the entire data set of blending information is used as the sparsity.

[0039] <2. Configuration of the Model Generation System> Figure 1 is a block diagram showing the configuration of the model generation system of this embodiment. The model generation system shown in Figure 1 comprises a model generation device 1, a database DB, and an estimation device 2. Here, we will first describe the model generation device 1.

[0040] The model generation device 1 comprises an acquisition unit 11, an item determination unit 12, a learning data creation unit 13, and a learning unit 14. This represents a concrete implementation of software-based information processing by hardware. These components do not necessarily need to be implemented by a single computer; they may be implemented through the collaboration of multiple computers. The model generation device 1 is configured to communicate with a database DB via wired or wireless connection, and the generated models are stored in the database DB.

[0041] The acquisition unit 11 performs an acquisition process to acquire the blending ratio of each raw material in the formulation and the time-dependent property information of the formulation for multiple formulations. Specifically, it measures the actual values ​​of the blending ratio and time-dependent property information for multiple formulations and records the results. Note that the blending ratio may be information from the formulation design stage rather than analyzing the completed formulation again.

[0042] When generating a model to estimate physical properties over time as information on properties over time, the physical properties of the formulation after a predetermined period has elapsed from a reference point should be measured, and these measured values ​​should be used as the actual physical properties over time. The reference point can be set arbitrarily, but for example, it can be the time of manufacture. The elapsed period can also be set arbitrarily, but it is preferable to set a period that corresponds to the storage and use conditions of the formulation, for example, three months after the reference point.

[0043] When generating a model to estimate deterioration categories as information on properties over time, it is sufficient to evaluate the properties of the formulation after a predetermined period has elapsed from the reference point to determine whether or not deterioration has occurred. For example, when determining whether or not functional properties have deteriorated, it is conceivable to evaluate the functional properties at the reference point and after a predetermined period has elapsed, and then determine whether or not deterioration has occurred by comparing them. Alternatively, for example, the reference point could be set to immediately after manufacturing, and formulations with the same composition immediately after manufacturing and after a predetermined period has elapsed from manufacturing could be compared and evaluated to determine whether or not deterioration in functional properties has occurred.

[0044] Sensory properties can include multiple types, such as color, scent, and texture, and a degradation category may be specified for each of these. Furthermore, in evaluating sensory properties, it is preferable to have multiple subjects evaluate each formulation individually and use a combined result. On the other hand, when determining whether or not physical properties have deteriorated, the physical properties should be measured at a reference point and after a predetermined period, and it should be determined whether the degree of change falls within the reference range.

[0045] The item determination unit 12 performs an item determination process to determine items to be used for model training from the blending information based on the blending ratios, based on the sparsity of the data based on the blending ratios of each raw material. In this embodiment, the blending ratio for each raw material and the sum of the blending ratios for each classification are candidates for items to be used for model training, and the items to be used for model training are determined based on the sparsity of the data for these items.

[0046] The learning data creation unit 13 executes a learning data generation process that generates learning data including feature quantities containing values ​​for items determined by the item determination unit 12, and temporal property information. In addition to feature quantities containing values ​​for items determined by the item determination unit 12, other features such as the manufacturing method (stirring method, stirring time, stirring rotation speed, and emulsification temperature, etc.) may also be used as feature quantities.

[0047] When generating a model to estimate time-dependent physical property values ​​as time-dependent property information, the training data consists of a set of feature quantities that include values ​​for items determined by the item determination unit 12, and measured values ​​of time-dependent physical properties in the formulation that exhibit those feature quantities.

[0048] When generating a model to estimate deterioration classifications as temporal property information, the training data consists of a set of features including values ​​for items determined by the item determination unit 12, and evaluation results of deterioration classifications for formulations that exhibit those features.

[0049] The learning unit 14 uses the learning data created by the learning data creation unit 13 to perform a learning process in which it takes the formulation information of the pharmaceutical product as input and trains a model that outputs the estimated results of time-dependent property information. In the case of a model that estimates time-dependent physical property values ​​as time-dependent property information, the output of the model is the estimated value of the time-dependent physical property value. In the case of a model that estimates degradation classification as time-dependent property information, the output of the model is one of the degradation classifications.

[0050] <3. Hardware Configuration> Next, the hardware configuration of the model generation system in this embodiment will be described. One or more general-purpose servers or personal computers can be used as the model generation device 1, the estimation device 2, and the database DB. In this embodiment, the model generation device 1 is an information processing device 10 on which a computer program (model generation program) that executes the model generation method is installed. Similarly, in this embodiment, the estimation device 2 is an information processing device 10 on which a computer program (estimation program) that executes the estimation method is installed.

[0051] Figure 2 is a hardware configuration diagram of the information processing device 10. As shown in Figure 2, the information processing device 10 has a control unit 101, a storage unit 102, and a communication unit 103, which are used to perform the functions of each unit and each process.

[0052] The control unit 101 has a processor such as a CPU that can execute instruction sets, and executes the OS and programs. The memory unit 102 includes volatile memory such as RAM capable of storing instruction sets, and non-volatile recording media such as HDDs or SSDs capable of storing the OS, model generation programs, DBMS, etc. The communication unit 103 has an interface for physically connecting to the network and performs communication control with the network NW to input and output information.

[0053] <4. Model Generation Method> Next, the model generation procedure in this embodiment will be described with reference to Figure 3. Figure 3 is a flowchart of the model generation method according to this embodiment. First, in step S11, the acquisition unit 11 acquires time-dependent property information and measured data of the blending ratio for each raw material for each of the multiple formulations.

[0054] Next, in step S12, the item determination unit 12 determines items to be used for model training from the blending information based on the blending ratio, based on the sparsity of the data based on the blending ratio for each raw material. In this embodiment, multiple pieces of blending information that satisfy the sparsity condition are identified as items to be used for model training. Details of the procedure for determining the items to be used for model training will be described later.

[0055] In step S13, the training data creation unit 13 creates training data according to the items determined in step S12. Specifically, it creates a set of feature quantities and temporal property information that include multiple values ​​for the relevant items as training data.

[0056] When estimating time-dependent physical properties as time-dependent property information, using measured values ​​directly may prevent proper learning due to bias in the region where the data values ​​exist. Therefore, it is also possible to use transformed physical properties as time-dependent physical properties. For example, values ​​obtained by logarithmically transforming physical properties can be used as transformed physical properties. In this case, the output of the model generated using the transformed physical properties will be an estimate of the transformed physical properties.

[0057] Then, in step S14, the learning unit 14 executes the model learning process based on the training data created in step S13. For example, when generating a model to estimate time-dependent physical property values ​​as time-dependent property information, the values ​​of several determined items are used as explanatory variables, and the time-dependent physical property values ​​are used as the dependent variable, and the model is trained accordingly. Furthermore, when generating a model to estimate degradation categories as temporal property information, the model is trained to classify the values ​​of multiple determined items into one of the degradation categories, using these values ​​as features.

[0058] In this embodiment, a model is generated using machine learning and statistical analysis with the above-mentioned training data. The model configuration and algorithm can be arbitrarily determined, and for example, regression models such as Ridge Regression, ElasticNet Regression, and Support Vector Regression (SVR), or neural networks can be used. Alternatively, methods derived from decision trees, such as Random Forest, LightGBM, and XGboost, may be used. These methods are particularly useful when generating models to estimate physical properties over time.

[0059] Furthermore, in generating a model for estimating degradation classifications, this embodiment employs an anomaly detection algorithm. An anomaly detection algorithm is a method for detecting anomaly data that differs in trend from a large amount of other data.

[0060] In this embodiment, we assume that the degradation categories include categories indicating significant degradation and categories indicating no significant degradation. However, in this case, the number of formulations with degradation may be extremely small compared to those without significant degradation. When there is such a bias in the amount of training data for each category, general classification algorithms may not be able to perform appropriate classification. Therefore, in this embodiment, an anomaly detection algorithm is used to generate a model that estimates the degradation categories by learning the training data of the degradation category indicating degradation as anomalies.

[0061] Specifically, methods such as OneClass SVM (OCSVM), Local Outlier Factor (LOF), and Isolation Forest (IF) can be used.

[0062] It should be noted that, depending on the data set used and the degradation classification criteria, it is possible that a standard classification algorithm may also be able to perform an appropriate classification. Therefore, a classification method may be arbitrarily used instead of an anomaly detection algorithm.

[0063] As described above, the model generation device 1 executes the acquisition process, item determination process, training data creation process, and training process to generate a model that estimates temporal property information from the formulation information. The model thus generated is transmitted to the database DB by the model generation device 1 and stored in the database DB.

[0064] <5. Classification of raw materials> Regarding the determination of items to be used in model training, the prerequisite is that blending information, including the blending ratio for each raw material and the sum of the blending ratios for each classification of raw materials, will be considered as candidate items to be used in model training. Here, we will explain the classification of raw materials. In this embodiment, raw materials are classified according to their blending purpose, and the sum of the blending ratios of raw materials belonging to the same classification is used as blending information. In particular, in this embodiment, a classification with multiple levels is defined, and each raw material is classified accordingly. For example, as shown in Figure 4, a classification called "water-soluble components" is defined, and below that, subcategories such as "mono-dihydric alcohols," "tertiary-tetrahydric alcohols," "pentahydric or higher alcohols," "sugar alcohols," and "polymers" are defined. Each raw material is then classified into one of these subcategories. Furthermore, a higher-level classification (major classification) that combines multiple classifications may be defined, and blending information may be calculated for each of these higher-level classifications. If a raw material has multiple blending purposes, the raw material may belong to multiple classifications.

[0065] In this embodiment, the total blending ratio for each classification is calculated and used as blending information. Alternatively, or in addition to this, the total blending ratio for each sub-classification or the total blending ratio for each major classification may also be used as blending information.

[0066] <6. Determining the items to be used for learning> For example, in formulations such as cosmetics, a wide variety of raw materials are used, resulting in a vast array of available formulation information. In this case, for raw materials (or classifications) with low infrequent use, the proportion of data with a value of 0 may be high. Such formulation information may not be suitable for proper learning. Therefore, in this invention, the items of formulation information used for model training are determined based on the sparsity of the data.

[0067] The item determination process will be explained in detail below with reference to Figure 5. First, in step S21, the sparsity of each blending information in the data group acquired in the acquisition process (step S11) is identified. In this embodiment, the item determination unit 12 calculates the sparsity for each blending information (each raw material or classification) based on the percentage of data with a value of 0.

[0068] Next, in step S22, the sparsity is compared to a baseline value. If the sparsity is smaller than the baseline value (Y in step S22), that is, if a certain percentage or more of the target blending information contains data other than 0, then in step S23, the blending information is determined to be used as an item for learning. In other words, the blending information is made one of the input variables to the model.

[0069] On the other hand, if the sparsity is above a certain threshold (N in step S22), that is, if the proportion of non-zero data in the target formulation information is below a certain level, then the formulation information will not be used as an item for learning, and the process will proceed to step S24.

[0070] In step S24, it is checked whether the sparsity of all candidate breeding information has been confirmed, and steps S21 to S23 are repeated until all breeding information has been confirmed. As a result, from among the numerous combinations of information, those that satisfy the sparsity condition are selected as items to be used for model training (input features to the model).

[0071] Here, the blending information used as candidates for items to be used in training the model may differ depending on the sparsity threshold. For example, if the threshold is set to 99%, the sparsity threshold will be met for blending ratios of many raw materials. Therefore, instead of using the sum of the blending ratios for each classification, items that meet the sparsity threshold among the blending ratios of each raw material may be used. On the other hand, if the threshold is set to a low value such as 50%, it is expected that the number of raw materials that do not meet the sparsity threshold will increase. In such cases, in addition to the blending ratio of each raw material, the sum of the blending ratios for each classification of raw materials may also be considered as a candidate.

[0072] Furthermore, estimation may be performed considering additional information in addition to the formulation information described above. For example, in addition to the items determined by the above procedure, the manufacturing method (stirring method, stirring time, stirring speed, emulsification temperature, etc.) may be used as further input features of the model. Also, since the temporal properties are affected by storage conditions, information indicating the storage conditions may be used as further input features of the model, or different models may be generated for each storage condition. Examples of storage conditions include storage temperature, humidity, and degree of airtightness.

[0073] <7. Estimation device> Next, the estimation of temporal property information using the model generated by the procedure described above will be explained. In this embodiment, the estimation device 2 estimates the formulation information using the model stored in the database DB. The estimation device 2 comprises a reception unit 21, an estimation unit 22, and an output unit 23. The estimation device 2 is configured to communicate with the database DB via wired or wireless connection and can utilize the model stored in the database DB.

[0074] The reception unit 21 accepts input of formulation information for a formulation for which the temporal property information is to be estimated. The formulation information accepted here corresponds to the items used in training the model.

[0075] The estimation unit 22 inputs the formulation information received by the reception unit 21 into the model generated by the model generation method described above, which is stored in the database DB. Then, the estimation unit 22 obtains the temporal property information as the output of the model.

[0076] The output unit 23 outputs information based on the time-series property information estimated by the estimation unit 22. For example, it is assumed that the output unit 23 processes the estimated time-series property information and sends the result to a display device, where it is displayed on the screen.

[0077] The estimation unit 22 may use multiple models. For example, different time-dependent physical properties may be estimated using multiple models, or degradation classifications from different perspectives may be estimated using multiple models. [Examples]

[0078] The following describes the learning model generation experiments conducted by the inventors. While the present invention is based on the following experimental methods and results, it is not limited to the following embodiments.

[0079] In this example, a model was generated to estimate the viscosity of the formulation three months after manufacturing, using the time of manufacture as the reference point for the physical properties over time. In this study, viscosity estimation was performed using transformed physical properties obtained by logarithmically transforming viscosity. More specifically, learning and estimation were performed using transformed physical properties obtained by log10 transforming viscosity.

[0080] Figure 6 shows an example of the training data used in Examples 1 to 3. Thus, for each formulation, training data was prepared using the following as explanatory variables: the blending ratio of each raw material, the sum of the blending ratios for each classification of raw materials, multiple manufacturing methods, and the phase state, with the viscosity three months after manufacturing the formulation as the dependent variable. Note that "blending ratio" was defined as mass percentage concentration (wt%).

[0081] Furthermore, the sum of the blending ratios for each raw material and each classification, which have a sparsity of 80% or less, were used as input features for training the model.

[0082] Under the above conditions, a model was generated for each storage temperature condition. As a result, R^2 > 0.5 was obtained for all conditions, including room temperature, high temperature, low temperature, and temperature cycling, indicating a certain level of prediction accuracy. [Examples]

[0083] In this example, a model was generated that estimates the hardness of the formulation after 3 months, with the time of manufacture as the reference point, as the dependent variable for determining the physical properties over time. The explanatory variables are the same as in Example 1. Here, in estimating hardness, we used transformed physical properties obtained by logarithmically transforming hardness for learning and estimation. More specifically, we used transformed physical properties obtained by log2 transformation for learning and estimation.

[0084] Furthermore, the blending ratios for each raw material, which have a sparsity of 70% or less, were used as input features for training the model.

[0085] Under the above conditions, a model was generated for each storage temperature condition. As a result, R^2 > 0.5 was obtained under high temperature and circulating temperature conditions, indicating a certain level of prediction accuracy. [Examples]

[0086] In this example, a model was generated that estimates the pH of the formulation three months after manufacturing as the dependent variable for determining the physical properties over time. The explanatory variables are the same as in Example 1.

[0087] Furthermore, the blending ratios for each ingredient, which have a sparsity of 90% or less, were used as input features for training the model.

[0088] Under the above conditions, a model was generated for each storage temperature condition. As a result, under the high-temperature circulation temperature condition, R^2 > 0.5 was obtained, indicating a certain level of prediction accuracy. [Examples]

[0089] In this example, a model was generated to estimate the degradation classification of the color-related sensory properties after 3 months as information on properties over time. The sum of the blending ratios for each raw material and for each classification was used as input features for training the model. Figure 7 shows an example of the training data used in Examples 4 to 7.

[0090] Under the above conditions, the recall value, one of the accuracy metrics of the model, was 0.8 or higher, indicating a certain level of prediction accuracy. Figure 8 shows some of the experimental results in this example. The sparsity condition is shown in "Feature Condition 1" in Figure 8, and whether or not to include the sum of the blending ratios for each classification of raw materials in the items is shown in "Feature Condition 2" in Figure 8. In Feature Condition 2, a "-" indicates that the learning was performed without including the sum of the blending ratios for each classification in the items. [Examples]

[0091] In this example, a model was generated to estimate the degradation category of the properties after 3 months as information on properties over time. The explanatory variables are the same as in Example 4. Here, "properties" refers to the nature and state, and refers to the physical and / or chemical structural state of the formulation, such as uniformity, aggregation / separation state, crystallization state, transparency, etc. The sum of the blending ratios for each raw material and for each classification was used as input features for training the model.

[0092] Under the above conditions, the recall value was 1, indicating a certain level of prediction accuracy. Figure 9 shows some of the experimental results in this example. The sparsity condition is shown in "Feature Condition 1" in Figure 9, and whether or not to include the sum of the blending ratios for each raw material classification as an item is shown in "Feature Condition 2" in Figure 9. [Examples]

[0093] In this example, a model was generated to estimate the deterioration classification of sensory properties related to aroma after 3 months as information on properties over time. The explanatory variables are the same as in Example 4. Only the blending ratio of each raw material was used as input features for training the model; the sum of the blending ratios for each classification was not used.

[0094] Under the above conditions, the recall value was 0.8 or higher, indicating a certain level of prediction accuracy. Figure 10 shows some of the experimental results in this example. The sparsity condition is shown in "Feature Condition 1" in Figure 10, and whether or not to include the sum of the blending ratios for each raw material classification as an item is shown in "Feature Condition 2" in Figure 10. [Examples]

[0095] In this example, a model was generated to estimate the degradation category of sensory properties related to touch after 3 months as time-dependent property information. The explanatory variables are the same as in Example 4. The sum of the blending ratios for each raw material and for each classification was used as input features for training the model.

[0096] Under the above conditions, the recall value was 0.8 or higher, indicating a certain level of prediction accuracy. Figure 11 shows some of the experimental results in this example. The sparsity conditions are shown in "Feature Condition 1" in Figure 11, and whether or not to include the sum of the blending ratios for each raw material classification as an item is shown in "Feature Condition 2" in Figure 11. In Feature Condition 1, conditions marked with "-" indicate that training was performed without any sparsity restrictions.

[0097] As described above, it was suggested that various temporal properties can be estimated with a certain degree of accuracy from the formulation information. In this way, by inferring the physical properties over time and the degradation classification from the formulation information, it becomes possible to predict the characteristics of a formulation over time without actually conducting time-based tests during the formulation design process. [Explanation of Symbols]

[0098] 1: Model Generator 11: Acquisition part 12:Item determination section 13: Training Data Creation Section 14: Learning Department 2: Estimation device 21: Reception Department 22: Estimation part 23: Output section 10: Information Processing Device 101: Control Unit 102: Storage section 103: Communications Department NW: Network

Claims

1. An acquisition process for obtaining the blending ratio of each raw material in a formulation and time-series property information indicating the properties of the formulation related to the passage of time, for multiple formulations. A process for determining items to be used for model training from the blending information based on the blending ratios of each raw material, based on the sparsity of the data based on the blending ratios of each raw material, A training data creation step that creates training data including feature quantities that include the values ​​of the aforementioned items and the temporal property information, A model generation method comprising: a learning step of training a model that uses the aforementioned training data to train a model that outputs estimation results of the time-dependent property information of a formulation based on the formulation information of the formulation.

2. The aforementioned temporal property information includes physical property values ​​after a predetermined period has elapsed from the reference point, The model generation method according to claim 1, wherein in the learning step, the model is trained with the formulation information as explanatory variables and the physical properties as the target variable.

3. In the above-mentioned training data creation step, the training data is created using the transformed physical property values ​​obtained by log-converting the physical property values, The model generation method according to claim 2, wherein in the learning step, the model is trained using learning data in which the transformed physical property value is the target variable.

4. The aforementioned temporal property information includes a deterioration classification indicating whether or not the properties have deteriorated after a predetermined period has elapsed from the reference point. The model generation method according to claim 1, wherein the learning step involves learning a model that classifies the deterioration categories of a formulation based on the formulation information.

5. The model generation method according to claim 4, wherein the aforementioned properties include functional properties.

6. The model generation method according to claim 4, wherein in the learning step, the learning data of the degradation category indicating that degradation exists is treated as an outlier, and the model is trained using an anomaly detection algorithm.

7. The model generation method according to claim 1, wherein in the item determination step, the sum of the blending ratios of raw materials belonging to the same classification is calculated as the blending information for each formulation, and the sparsity of the blending ratio for each raw material and the sum of the blending ratios for each classification are each used as candidates for items to be used to train the model, and the items to be used to train the model are determined by calculating the sparsity of each.

8. The model generation method according to claim 7, wherein in the item determination step, raw materials with the same blending purpose are classified as the same category, and the blending information is calculated showing the sum of the blending ratios of raw materials belonging to the same category.

9. A model generation device comprising an acquisition unit, an item determination unit, a learning data creation unit, and a learning unit, The acquisition unit acquires, for each of several formulations, the blending ratio of each raw material in the formulation and time-dependent property information indicating the properties of the formulation related to the passage of time. The item determination unit determines, based on the sparsity of the data based on the blending ratios for each raw material, which items to be used for training the model from the blending information based on the blending ratios. The aforementioned training data creation unit creates training data that includes feature quantities including the values ​​of the items and the temporal property information, The learning unit is a model generation device that uses the learning data to train a model that outputs an estimated result of the temporal property information of a formulation based on the formulation information of the formulation.

10. A model generation program that causes a computer to function as an acquisition unit, an item determination unit, a learning data creation unit, and a learning unit, The acquisition unit acquires, for each of several formulations, the blending ratio of each raw material in the formulation and time-dependent property information indicating the properties of the formulation related to the passage of time. The item determination unit determines, based on the sparsity of the data based on the blending ratios for each raw material, which items to be used for training the model from the blending information based on the blending ratios. The aforementioned training data creation unit creates training data that includes feature quantities including the values ​​of the items and the temporal property information, The learning unit is a model generation program that uses the learning data to train a model that outputs an estimated result of the time-dependent property information of a formulation based on the formulation information of the formulation.

11. An estimation method that performs estimation based on the relationship between the mixing ratio of the raw materials of a preparation and the properties of the preparation over time, An estimation method for estimating the time-dependent properties of a pharmaceutical preparation from the formulation information, by inputting formulation information relating to the mixing ratio of raw materials of the preparation into a model generated by the model generation method described in claim 1, and obtaining estimation results relating to time-dependent property information indicating the properties of the preparation related to the passage of time as the output of the model.

12. An estimation device comprising a reception unit and an estimation unit, The aforementioned reception unit receives input of formulation information regarding the mixing ratio of raw materials for the formulation. The estimation device performs estimation of the time-dependent properties of a pharmaceutical preparation from the formulation information by inputting formulation information relating to the mixing ratio of raw materials of the preparation into a model generated by the model generation method described in claim 1, and obtaining estimation results relating to time-dependent properties indicating the properties of the preparation in relation to the passage of time as the output of the model.

13. An estimation program that causes a computer to function as a reception unit and an estimation unit, The aforementioned reception unit receives input of formulation information regarding the mixing ratio of raw materials for the formulation. The estimation unit is an estimation program that performs estimation of the time-dependent properties of a pharmaceutical preparation from the formulation information by inputting formulation information relating to the mixing ratio of raw materials of the preparation into a model generated by the model generation method described in claim 1, and obtaining estimation results relating to time-dependent property information indicating the properties of the preparation related to the passage of time as the output of the model.

Images