Learning device, learning method, prediction device, prediction method, and program

Abstract

To provide a technique capable of predicting physical properties with high accuracy for compounds suitable for a learned model.SOLUTION: A learning device 10 includes a learning unit 130 and a range identification unit 150. The learning unit 130 generates a learned model by performing machine learning using a plurality of pieces of learning data. Each piece of learning data is data in which feature values relating to a compound and an index relating to physical properties of the compound are associated with each other. The range identification unit 150 identifies an applicable range of the learned model by using the plurality of pieces of learning data.SELECTED DRAWING: Figure 12

Description

[Technical Field]

[0001] The present invention relates to a learning device, a learning method, a prediction device, a prediction method, and a program. [Background technology]

[0002] Technology is being developed to predict the physical properties of compounds using machine learning models.

[0003] Patent Document 1 describes a learning model that uses molecular structure descriptors of a complex to estimate the physical properties of the complex. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2024-010914 Summary of the Invention [Problem to be solved by the invention]

[0005] However, the technology of Patent Document 1 targets complexes and is not applicable to a wide range of compounds. Furthermore, when attempting to apply property estimation using a machine learning model to a wide range of compounds, it is not obvious that highly accurate predictions are possible for all compounds.

[0006] The present invention provides a technology that can predict the physical properties of compounds that are suitable for a trained model with high accuracy. [Means for solving the problem]

[0007] According to one embodiment of the present invention, the following learning device, learning method, prediction device, prediction method, and program are provided.

[0008] 1. A learning unit that generates a trained model by performing machine learning using a plurality of training data in which feature quantities related to compounds and indexes related to the physical properties of the compounds are associated with each other; a range specification unit that specifies an application range of the trained model using the plurality of pieces of training data; Learning device. 2. In the learning device described in 1., The range identification unit identifies the application range based on an in-range prediction error, which is a prediction error of the trained model for data within the application range, and an out-of-range prediction error, which is a prediction error of the trained model for data outside the application range. Learning device. 3. In the learning device described in 2., The range specification unit identifying the in-range prediction error and the out-of-range prediction error for a tentative application range; Identifying the application range based on the relationship between the within-range prediction error and the out-of-range prediction error identified for the tentative application range and a change in the within-range prediction error with respect to a change in the tentative application range. Learning device. 4. In the learning device according to 2. or 3., The range identification unit identifies the prediction error of the trained model using at least one of a mean absolute error, a root mean square error, and a coefficient of determination. Learning device. 5. In the learning device according to any one of 1. to 4., The range specification unit specifies the application range using at least one of a k-nearest neighbor method, a one-class support vector machine, and a jagging method. Learning device. 6. In the learning device according to any one of 1. to 5., a selection unit that selects an element to be included in the feature from a plurality of candidate elements included in the provisional feature, In each of the plurality of learning data, the feature amount including the element selected by the selection unit and an index related to a physical property of the compound are associated with each other; The selection unit a first selection process for selecting two or more candidate elements from the plurality of candidate elements based on at least one of variations in each of the plurality of compounds and similarities between the plurality of candidate elements; a second selection process for selecting one or more candidate elements from the two or more candidate elements as elements to be included in the feature quantity based on the importance of each of the two or more candidate elements selected in the first selection process; Learning device. 7. In the learning device according to 6., In the first selection process, the selection unit selects the two or more candidate elements using at least the variance of each of the candidate elements in the plurality of compounds and a predetermined variance threshold. Learning device. 8. In the learning device according to 6. or 7., In the first selection process, the selection unit selects the two or more candidate elements using at least correlation coefficients between the plurality of candidate elements for the plurality of compounds and a predetermined correlation threshold. Learning device. 9. The learning device according to any one of 6. to 8. In the second selection process, the selection unit selecting i candidate elements in descending order of importance from the two or more candidate elements as elements to be included in the feature; The value of i is determined so that the prediction error of a prediction model obtained by machine learning using learning data including features including the i candidate elements satisfies a predetermined condition. Learning device. 10. A learning method implemented by one or more computers, comprising: a learning step of generating a trained model by performing machine learning using a plurality of pieces of training data in which feature amounts related to compounds and indexes related to physical properties of the compounds are associated with each other; and a range specification step of specifying an application range of the trained model using the plurality of training data. How to learn. 11. On the computer, a learning step of generating a trained model by performing machine learning using a plurality of pieces of training data in which feature amounts related to compounds and indexes related to physical properties of the compounds are associated with each other; and executing a range specification step of specifying an application range of the trained model using the plurality of pieces of training data. program. 12. A determination unit that determines whether an index related to the applicability range based on the feature amount of the target compound falls within the applicability range of the trained model, using range information indicating the applicability range of the trained model; a prediction unit that predicts an index related to a physical property of the target compound using the feature amount of the target compound and the trained model when the determination unit determines that the target compound falls within the applicable range. Prediction device. 13. A prediction method implemented by one or more computers, comprising: a determination step of determining whether an index relating to the applicability range based on the feature amount of the target compound falls within the applicability range of the trained model, using range information indicating the applicability range of the trained model; and a prediction step of predicting an index related to a physical property of the target compound using the feature amount of the target compound and the trained model when the target compound is determined to be within the applicable range in the determination step. Forecasting methods. 14. On the computer, a determination step of determining whether an index relating to the applicability range based on the feature amount of the target compound falls within the applicability range of the trained model, using range information indicating the applicability range of the trained model; If it is determined in the determination step that the target compound falls within the applicable range, a prediction step is executed in which an index related to a physical property of the target compound is predicted using the feature amount of the target compound and the trained model. program. [Effects of the Invention]

[0009] According to the present invention, a technology is provided that can predict the physical properties of compounds suitable for a trained model with high accuracy. [Brief explanation of the drawings]

[0010] [Figure 1] 1 is a diagram illustrating an overview of a learning device according to a first embodiment. [Figure 2] FIG. 10 is a diagram illustrating a computer for realizing a learning device. [Figure 3] FIG. 1 is a diagram illustrating an overview of a learning method according to a first embodiment. [Figure 4] FIG. 2 is a diagram showing a trained model obtained by the training device and training method according to the first embodiment. [Figure 5] FIG. 10 is a diagram for explaining an outline of a process performed by a selection unit. [Figure 6] FIG. 10 is a diagram illustrating an example of the relationship between the number of elements i and the prediction error. [Figure 7] FIG. 2 is a diagram illustrating an example of the functional configuration of a learning device according to the first embodiment. [Figure 8] 10 is a flowchart illustrating an example of the flow of provisional feature generation and a first process. [Figure 9] 10 is a diagram illustrating an example of the flow of a second selection process performed by a selection unit and a process performed by a learning unit. FIG. [Figure 10] 1 is a diagram illustrating an overview of a prediction device according to a first embodiment. [Figure 11] FIG. 1 is a diagram illustrating an overview of a prediction method according to a first embodiment. [Figure 12] FIG. 10 is a diagram illustrating an overview of a learning device according to a second embodiment. [Figure 13] FIG. 10 is a diagram illustrating an outline of a learning method according to a second embodiment. [Figure 14] FIG. 10 is a diagram illustrating an example of an application range. [Figure 15] FIG. 10 is a diagram illustrating an example of the functional configuration of a learning device according to a second embodiment. [Figure 16] 10 is a flowchart illustrating the flow of processing executed by a learning device according to a second embodiment. [Figure 17] FIG. 10 is a diagram illustrating an overview of a prediction device according to a second embodiment. [Figure 18] FIG. 10 is a diagram illustrating an overview of a prediction method according to a second embodiment. [Figure 19] FIG. 10 is a diagram illustrating a flow of processing executed by a prediction device according to a second embodiment. [Figure 20] FIG. 10 is a diagram illustrating an example of the functional configuration of a learning device according to a third embodiment. [Figure 21] FIG. 10 is a diagram showing the relationship between the value of i and MAE in Example 1. [Figure 22] 10 is a graph showing the relationship between X and R2, both within and outside the scope of application. [Figure 23] FIG. 1 is a diagram showing the prediction accuracy of the trained model according to Example 1 with respect to training data. [Figure 24] FIG. 10 is a diagram showing the prediction accuracy of the trained model according to Example 1 for test data within the applicable range. [Figure 25] FIG. 10 is a diagram showing the prediction accuracy of the trained model according to Example 1 for test data outside the applicable range. [Figure 26] FIG. 10 is a diagram showing the prediction accuracy of the trained model according to Example 2 for the training data. [Figure 27] FIG. 10 is a diagram showing the prediction accuracy of the trained model according to Example 2 for test data within the applicable range. [Figure 28] FIG. 10 is a diagram showing the prediction accuracy of the trained model according to Example 2 for test data outside the applicable range. DETAILED DESCRIPTION OF THE INVENTION

[0011] Hereinafter, embodiments of the present invention will be described with reference to the drawings. In all the drawings, like components are designated by like reference numerals, and their description will be omitted where appropriate.

[0012] FIG. 1 is a diagram illustrating an overview of a learning device 10 according to a first embodiment. The learning device 10 according to this embodiment includes a selection unit 110 and a learning unit 130. The selection unit 110 selects elements to be included in the features from among multiple candidate elements included in tentative features related to a compound. The learning unit 130 generates a trained model by performing machine learning using multiple training data. The training data is data in which feature elements including the elements selected by the selection unit 110 are associated with indices related to the physical properties of the compound. The selection unit 110 executes a first selection process and a second selection process. The first selection process is a process of selecting two or more candidate elements from among multiple candidate elements based on at least one of the variability of each of the multiple candidate elements among multiple compounds and the similarity between the multiple candidate elements. The second selection process is a process of selecting one or more candidate elements from among the two or more candidate elements selected in the first selection process as elements to be included in the features based on the importance of each of the two or more candidate elements selected in the first selection process.

[0013] The hardware configuration of learning device 10 is described below. Each functional component of learning device 10 (selection unit 110 and learning unit 130) may be realized by hardware that realizes each functional component (e.g., a hardwired electronic circuit, etc.), or by a combination of hardware and software (e.g., a combination of an electronic circuit and a program that controls it). Below, we will further describe the case where each functional component of learning device 10 is realized by a combination of hardware and software.

[0014] FIG. 2 is a diagram illustrating a computer 1000 for implementing the learning device 10. The computer 1000 may be any computer. For example, the computer 1000 may be a system on chip (SoC), a personal computer (PC), a server machine, a tablet terminal, or a smartphone. The computer 1000 may be a dedicated computer designed to implement the learning device 10, or may be a general-purpose computer. The learning device 10 may be implemented by a single computer 1000 or by a combination of multiple computers 1000.

[0015] The computer 1000 includes a bus 1020, a processor 1040, a memory 1060, a storage device 1080, an input / output interface 1100, and a network interface 1120. The bus 1020 is a data transmission path through which the processor 1040, the memory 1060, the storage device 1080, the input / output interface 1100, and the network interface 1120 transmit and receive data to and from each other. However, the method of interconnecting the processor 1040 and other components is not limited to bus connection. The processor 1040 may be any of various processors, such as a central processing unit (CPU), a graphics processing unit (GPU), or a field-programmable gate array (FPGA). The memory 1060 is a main storage device implemented using a random access memory (RAM) or the like. The storage device 1080 is an auxiliary storage device implemented using a hard disk, a solid state drive (SSD), a memory card, a read-only memory (ROM), or the like.

[0016] The input / output interface 1100 is an interface for connecting the computer 1000 to an input / output device. For example, an input device such as a keyboard and an output device such as a display are connected to the input / output interface 1100. The input / output interface 1100 may be connected to the input device or output device by wireless connection or by wired connection.

[0017] The network interface 1120 is an interface for connecting the computer 1000 to a network. Examples of this communication network include a LAN (Local Area Network) and a WAN (Wide Area Network). The network interface 1120 may connect to the network wirelessly or by wire.

[0018] Storage device 1080 stores program modules that realize each functional component of learning device 10. Processor 1040 reads each of these program modules into memory 1060 and executes them to realize the function corresponding to each program module.

[0019] FIG. 3 is a diagram illustrating an overview of a training method according to this embodiment. The training method according to this embodiment is executed by one or more computers. The training method according to this embodiment includes a selection step S10 and a training step S20. In the selection step S10, one or more computers select elements to be included in the features from among multiple candidate elements included in the provisional feature for the compound. In the training step S20, the one or more computers perform machine learning using multiple training data to generate a trained model. The training data is data in which feature elements including the selected elements are associated with indices related to the physical properties of the compound. The selection step S10 includes a first selection process and a second selection process. The first selection process is a process of selecting two or more candidate elements from among multiple candidate elements based on at least one of the variability of each of the multiple candidate elements among multiple compounds and the similarity between the multiple candidate elements. The second selection process is a process of selecting one or more candidate elements from among the two or more candidate elements selected in the first selection process as elements to be included in the feature based on the importance of each of the two or more candidate elements selected in the first selection process.

[0020] The learning step S20 may be performed after the selection step S10, or may be performed during the selection step S10, as will be described in detail later.

[0021] The learning method according to this embodiment can be executed by the learning device 10 according to this embodiment. The learning device 10 and the learning method according to this embodiment will be described in detail below.

[0022] 4 is a diagram showing a trained model 30 obtained by the learning device 10 and the learning method according to this embodiment. The trained model 30 is, for example, a model including a neural network. In this case, the layer structure of the trained model 30 is not particularly limited.

[0023] The input data of the trained model 30 includes a feature F related to a compound C. This feature F includes multiple elements. For example, each element is a scalar, and the feature F is a vector. Note that, below, a case will be described in detail where the feature F is a feature related to the structure of the compound C, and all elements of the feature F are related to the structure of the compound C, but the feature F is not limited to the following example. Some or all of the feature F may be elements related to something other than the structure of the compound C.

[0024] The compound is not particularly limited. The compound may be an organic compound or an inorganic compound. The compound may also be any of a molecular compound, an intermetallic compound, and a coordination compound. When the compound is a molecule, the compound may be a low molecular weight (e.g., a molecular weight of 500 or less), a medium molecular weight (e.g., a molecular weight of 500 or more and 2000 or less), or a high molecular weight (e.g., a molecular weight of 2000 or more).

[0025] The output data of the trained model 30 includes at least an index p related to the physical properties of the compound C. The index related to the physical properties of the compound C (hereinafter also referred to as a "physical property index") is not particularly limited. The physical property index may be, for example, an index related to biological activity, an index related to chemical properties, or an index related to physical properties. Examples of the index related to biological activity include an index related to anticancer activity and an index related to the inhibition of neurotransmitters. The index related to biological activity may be an index indicating the presence or absence of a specific biological activity, or an index indicating the degree of a specific biological activity.

[0026] 5 is a diagram for explaining an outline of the process performed by the selection unit 110. The selection unit 110 selects a compound C a Temporary feature F for (a=1,2,...,n) ta =(f 1a ,f 2a ,···,f ma ) contains multiple candidate elements (f 1a ,f 2a ,···,f ma ) Compound C aThe feature value F a Here, a is a number that identifies the compound, n is a positive integer that indicates the number of compounds, and m is the provisional feature F ta It is a positive integer that indicates the number of elements in the feature F. a The elements that make up the provisional feature F ta It can be said that this is an extracted part of the elements that make up (f 1a ,f 2a ,···,f ma ), among which (f 1a ,f 5a ,···,f m-2a ) is selected and the feature F a However, the candidate elements to be selected are not limited to this example.

[0027] The training data used in machine learning to obtain the trained model 30 is compound C a The feature F including the element selected by the selection unit 110 a and Compound C a Physical property index p a In machine learning, the physical property index p a is the feature F a is used as the correct answer data for multiple compounds C1, C2,...,C n A set of training data for the above can be used as a training dataset. The training unit 130 can generate a trained model 30 by performing machine learning using the training dataset.

[0028] When predicting the physical properties of compounds using machine learning, various features are used. For example, various descriptors are used as features related to the structure of a compound. Examples of descriptors include molecular fingerprints (FPs), molecular graphs, structural descriptors, and electronic descriptors. In molecular fingerprints, the presence or absence of various substructures in a molecule is expressed as 0 or 1. In molecular graphs, chemical structures are used as graphs. Structural descriptors express information about the two-dimensional (2D) and three-dimensional (3D) structures of compounds. Electronic descriptors indicate the properties and reactivity of compounds.

[0029] However, the appropriate descriptor may vary depending on the type of property to be predicted, and also on the form of prediction, such as whether one wants to predict the presence or absence of a certain property, its level, or the value of a property.

[0030] Although it is possible to use a combination of multiple descriptors, the increase in the number of descriptors poses problems such as increased computational costs, poor interpretability, and reduced prediction accuracy for unknown compounds due to overfitting. Therefore, particularly when attempting to predict the physical properties of a wide range of compounds with high accuracy, it is important to appropriately select important descriptors depending on the physical properties to be predicted.

[0031] According to this embodiment, a trained model 30 capable of predicting the physical properties of a wide range of compounds with high accuracy is obtained using feature quantities consisting of elements appropriately selected by the selection unit 110. Furthermore, the learning device 10 and learning method according to this embodiment are not limited in the type of physical property to be predicted or the form of prediction, and can be applied to a variety of predictions.

[0032] As described above, the first selection process according to this embodiment is a process for selecting two or more candidate elements from a plurality of candidate elements based on at least one of the variation of each of the plurality of candidate elements among a plurality of compounds and the similarity between the plurality of candidate elements. The first selection process can be performed, for example, according to the Filter method.

[0033] The first selection process may include a process of selecting candidate elements based on similarities between multiple candidate elements (hereinafter also referred to as "similarity-based process"). Specifically, in the similarity-based process, the selection unit 110 selects two or more candidate elements using correlation coefficients between multiple candidate elements for multiple compounds and a predetermined correlation threshold. That is, the selection unit 110 selects two or more candidate elements based on a comparison result between the correlation coefficients between the multiple candidate elements and the correlation threshold.

[0034] The similarity-based treatment is described in detail below. Compound C a (a=1,2,...,n) provisional feature F ta =(f 1a ,f 2a ,···,f ma ), for example, for candidate element f 1a and candidate element f 2a The selection unit 110 selects a plurality of compounds C1, C2, . . . , C n candidate element f 1a A vector (f 11 ,f 12 ,···,f 1n ) and several compounds C1, C2,...,C n candidate element f 2a A vector (f 21 ,f 22 ,···,f 2n ) and the correlation coefficient with the candidate element f 1a and candidate element f 2a If the calculated correlation coefficient is equal to or greater than a predetermined correlation threshold, the candidate element f 1a and candidate element f 2a Either one of these is used as the feature F a The candidate element to be included in the provisional feature vector is excluded from the candidates to be included in the provisional feature vector. Here, there is no particular limitation on which of the two candidate elements to exclude. For example, the earlier element in the arrangement order of the elements in the provisional feature vector may be excluded, or the later element may be excluded. Alternatively, which element to exclude may be determined randomly. The correlation threshold is preferably 0.90 or more and 0.99 or less.

[0035] This calculation and determination of correlation coefficients is performed for all pairs that can be formed among multiple candidate elements. However, for candidate elements that have been excluded in the previous determination, it is not necessary to further calculate correlation coefficients between them and other candidate elements.

[0036] This process makes it possible to eliminate one of the candidate elements that are highly similar to each other, i.e., to reduce the number of candidate elements so as to avoid overlapping information.

[0037] The first selection process may also include a process of selecting candidate elements based on the variability of each of the candidate elements across multiple compounds (hereinafter also referred to as "variation-based process"). Specifically, in the variability-based process, the selection unit 110 selects two or more candidate elements using the variance of each of the candidate elements across multiple compounds and a predetermined variance threshold. That is, the selection unit 110 selects two or more candidate elements based on the comparison result between the variance of each of the candidate elements and the variance threshold.

[0038] The variability-based treatment is described in detail below. Compound C a (a=1,2,...,n) provisional feature F ta =(f 1a ,f 2a ,···,f ma ), for example, for candidate element f 1a The selection unit 110 selects a plurality of compounds C1, C2, . . . , C n candidate element f 11 ,f 12 ,···,f 1n The variance of the candidate element f 1a If the calculated variance is equal to or less than a predetermined variance threshold, the candidate element f 1a The feature F a The variance threshold is preferably between 0.00 and 0.10.

[0039] This calculation and determination of variance is performed for each of the multiple candidate elements. By performing this process, candidate elements that make little contribution to the difference between compounds can be eliminated.

[0040] The selection unit 110 may perform either the variation-based processing or the similarity-based processing, without performing the other. However, it is preferable that the selection unit 110 perform both the variation-based processing and the similarity-based processing. The variation-based processing and the similarity-based processing may be performed in this order, or in the reverse order. When both the variation-based processing and the similarity-based processing are performed, the subsequent processing may be performed on the multiple candidate elements remaining in the previous processing, i.e., the multiple candidate elements that were not excluded. It is more preferable that the selection unit 110 performs the variation-based processing and then the similarity-based processing. By performing the variation-based processing before the similarity-based processing, it is possible to reduce processing related to elements that should not be included in the feature, such as processing based on the similarity between candidate elements that have little contribution to the difference between compounds.

[0041] As described above, the second selection process according to this embodiment is a process of selecting one or more of the two or more candidate elements selected in the first selection process as elements to be included in the feature, based on the importance of each of the two or more candidate elements.

[0042] In the second selection process, the selection unit 110 selects i candidate elements from two or more candidate elements in descending order of importance as elements to be included in the feature. Here, the selection unit 110 determines the value of i so that the prediction error of a prediction model obtained by machine learning using training data including features containing the i candidate elements satisfies a predetermined condition. The predetermined condition is, for example, that the magnitude of change in the prediction error is equal to or less than a predetermined threshold.

[0043] The second selection process will be described in detail below. In the second selection process, the selection unit 110 uses a machine learning regression method to rearrange the candidate elements in order of importance in accordance with the Embedded method, which performs feature selection during model learning. Examples of the regression method used here include random forest regression, LightGBM, and LASSO regression. However, the regression method used is not limited to these examples. The regression method used here is also used in the trained model 30. Furthermore, the method for rearranging the candidate elements in order of importance is not particularly limited, and various existing methods can be used.

[0044] Next, the selection unit 110 identifies elements to be included in the feature in accordance with the Wrapper method, which evaluates combinations of feature using machine learning. Specifically, a provisional model is generated each time a candidate element is added to the feature one by one in descending order of importance, starting with the candidate element with the highest importance. Then, the prediction error of the generated provisional model is calculated. Then, when the magnitude of the change in the prediction error with respect to the increase in candidate elements becomes equal to or less than a threshold, the selection unit 110 selects the candidate element that was included in the feature as the final feature F a Here, indicators of the prediction error include, for example, the mean absolute error (MAE), the root mean square error (RMSE), and the coefficient of determination (R 2 ) can be used. The selection unit 110 may use any one of them alone or may use two or more of them in combination. Note that the larger the MAE and RMSE, the larger the prediction error. On the other hand, R 2 The smaller the σ, the larger the prediction error. It is preferable to use a test data set different from the training data set to calculate the prediction error.

[0045] FIG. 6 is a diagram illustrating the relationship between the number i of candidate elements included in a feature and the prediction error. As shown in the example of FIG. 6, it is expected that the prediction error will decrease as the number i of elements increases from its initial value. Furthermore, once the number i of elements reaches a certain level, the change in the prediction error relative to the change in the number i of elements decreases, and the prediction error converges. In other words, even if the number i of elements is further increased, the effect of improving prediction accuracy is not significant. Therefore, it is effective to determine the value of i on the condition that the magnitude of the change in the prediction error is equal to or less than a predetermined threshold.

[0046] By performing the second selection process, candidate elements are selected in order of importance, and an appropriate number of elements are selected as the feature F a can be included in

[0047] All of the data on the multiple compounds targeted in the first selection process may be used for training the provisional model in the second selection process, or some of the data may not be used for training the provisional model. Of the data on the multiple compounds targeted in the first selection process, data on some of the compounds may be used for training the provisional model, and data on the remaining compounds may be used for testing to calculate the prediction error of the provisional model.

[0048] The selection unit 110 according to this embodiment executes the first selection process and the second selection process in this order. That is, similar candidate elements are removed, and then candidate elements with high importance are identified. In this way, it is possible to identify elements that should be included in the feature to obtain high prediction accuracy. In other words, by executing the first selection process before the second selection process, it is possible to avoid, for example, having multiple similar candidate elements in succession when the candidate elements are sorted in order of importance.

[0049] FIG. 7 is a diagram illustrating an example of the functional configuration of a learning device 10 according to this embodiment. In the example of FIG. 7, the learning device 10 further includes a generation unit 100. The generation unit 100 generates provisional features. In the example of FIG. 7, the learning device 10 further includes a training data storage unit 111 and a model storage unit 131. The training data storage unit 111 stores at least one of training data and provisional training data. The model storage unit 131 stores a trained model 30.

[0050] When the learning device 10 includes the generation unit 100, the storage device 1080 of the computer 1000 that implements the learning device 10 further stores a program module that implements the functions of the generation unit 100 according to this embodiment.

[0051] The learning data storage unit 111 may be provided in the learning device 10 as shown in FIG. 7, or may be provided external to the learning device 10. When the learning data storage unit 111 is provided in the learning device 10, the learning data storage unit 111 is realized by a storage device 1080 of a computer 1000 that realizes the learning device 10 according to this embodiment. The model storage unit 131 may be provided in the learning device 10 as shown in FIG. 7, or may be provided external to the learning device 10. When the model storage unit 131 is provided in the learning device 10, the model storage unit 131 is realized by a storage device 1080 of a computer 1000 that realizes the learning device 10 according to this embodiment.

[0052] FIG. 8 is a flowchart illustrating the generation of provisional features and the flow of the first process.

[0053] In S101, the generation unit 100 generates provisional features for each of a plurality of compounds. Specifically, for example, the generation unit 100 can acquire information about the compounds from a database about compounds and use it to generate provisional features. An existing open database can be used as the database about compounds. The information about the compounds includes, for example, SMILES, which represents the structure of the compound as a character string using ASCII codes, and physical property information about the compound.

[0054] The generation unit 100 uses SMILES to calculate descriptors that indicate the structure of the compound and includes them as elements in the provisional feature. The generation unit 100 can calculate the descriptors using existing methods or programs. The generation unit 100 may calculate multiple types of descriptors and connect them to form the provisional feature. The multiple types of descriptors preferably include one or more of FP, 2D structural descriptors, and 3D structural descriptors, preferably two or more of FP, 2D structural descriptors, and 3D structural descriptors, and more preferably include all of FP, 2D structural descriptors, and 3D structural descriptors.

[0055] Furthermore, the generating unit 100 identifies a physical property index of the compound using the physical property information of the compound. The generating unit 100 may use the physical property information of the compound acquired from the database as the physical property index as is, or may obtain the physical property index by performing existing processing, calculation, or the like on the physical property information of the compound.

[0056] The generation unit 100 associates provisional feature amounts with physical property indices for each compound. A set of provisional feature amounts and physical property indices that are associated with each other is called "provisional training data." The generation unit 100 generates provisional training data for each of a plurality of compounds.

[0057] The selection unit 110 acquires multiple pieces of temporary training data and executes the first selection process. The selection unit 110 can acquire the temporary training data from the generation unit 100. Alternatively, the selection unit 110 may read and acquire the temporary training data generated by the generation unit 100 and temporarily stored in the training data storage unit 111. However, the method by which the selection unit 110 acquires the temporary training data is not limited to this example. The learning device 10 may not include the generation unit 100. In that case, instead of acquiring the temporary training data generated by the generation unit 100, the selection unit 110 may acquire temporary training data that has been prepared in advance by some method and stored in the training data storage unit 111 accessible by the selection unit 110. As another example, the learning device 10 may acquire the temporary training data from another functional component of the learning device 10 or from a device other than the learning device 10.

[0058] In the example of Fig. 8, the processes from S102 to S105 are the variability-based processes of the first selection process. In the example of Fig. 8, the processes from S106 to S108 are the similarity-based processes of the first selection process. However, as described above, the first selection process may include at least one of the similarity-based processes and the variability-based processes.

[0059] In S102, the selection unit 110 identifies one candidate element from among multiple candidate elements included in the provisional feature. Next, in S103, the selection unit 110 calculates the variance of the identified candidate element across multiple compounds. Then, it determines whether the calculated variance is equal to or less than a variance threshold. If the variance is equal to or less than the variance threshold (Yes in S103), the selection unit 110 excludes the candidate element from candidates to be included in the feature (i.e., deletes it from the provisional feature) (S104). On the other hand, if the variance is not equal to or less than the variance threshold (No in S103), the selection unit 110 does not exclude the candidate element. In other words, the selection unit 110 leaves the candidate element as a candidate to be included in the feature and performs the process of S105.

[0060] In S105, it is determined whether the variances of all of the candidate elements in the provisional feature have been determined. If the determination of all candidate elements has been completed (Yes in S105), the selection unit 110 then performs the process of S106. On the other hand, if the determination of the variance of at least any candidate element has not been completed (No in S105), the process returns to S102, and the selection unit 110 identifies the next candidate element to be determined.

[0061] After S105 and before S106, the selection unit 110 may cause the learning data storage unit 111 to store provisional learning data including the provisional feature amounts at that time.

[0062] In S106, the selection unit 110 identifies a pair of candidate elements by selecting two from among the candidate elements remaining in the provisional feature. Next, in S107, the selection unit 110 calculates a correlation coefficient for the pair. Then, the selection unit 110 determines whether the calculated correlation coefficient is equal to or greater than a correlation threshold. If the correlation coefficient is equal to or greater than the correlation threshold (Yes in S107), the selection unit 110 excludes one of the pair from the candidates to be included in the feature (i.e., deletes it from the provisional feature) and leaves the other as a candidate to be included in the feature in S108. On the other hand, if the correlation coefficient is not equal to or greater than the correlation threshold (No in S107), the selection unit 110 does not exclude either candidate element. That is, the selection unit 110 leaves both of the pair as candidates to be included in the feature and performs the process of S109.

[0063] In S109, it is determined whether the correlation coefficients of all pairs of candidate elements in the provisional feature have been determined. If the correlation coefficients of all pairs have been determined (Yes in S109), the selection unit 110 ends the first selection process. On the other hand, if the correlation coefficients of at least one pair have not been determined (No in S109), the process returns to S106, and the selection unit 110 identifies the next pair to be determined.

[0064] The selection unit 110 may store provisional training data including provisional features at the time when the first selection process is completed in the training data storage unit 111.

[0065] 9 is a diagram illustrating the flow of the second selection process performed by the selection unit 110 and the process performed by the learning unit 130. After the first selection process ends, the selection unit 110 according to this embodiment starts the second selection process for the multiple candidate elements remaining in the tentative feature.

[0066] In S201, the selection unit 110 sorts two or more candidate elements included in the provisional feature in order of importance, as described above. Next, in S202, the selection unit 110 initializes a variable i. The initial value of the variable i is determined in advance and may be 1 or a value greater than 1. However, the initial value is set to a value that is sufficiently smaller than the expected number of elements in the feature.

[0067] In S203, the selection unit 110 extracts candidate elements 1 to i in descending order of importance from among two or more candidate elements included in the provisional feature, and generates a feature including the extracted candidate elements for each of the multiple compounds. Then, the selection unit 110 generates data for each compound that associates the generated feature with the physical property index of the compound. In S204, the learning unit 130 generates a provisional model by performing machine learning using the data generated in S203 as training data. The configuration of this provisional model is the same as the configuration of the trained model 30 described above. Furthermore, the configurations of the input data and output data of the provisional model are the same as the configurations of the input data and output data of the trained model 30 described above.

[0068] In S204, in the machine learning, the learning unit 130 updates multiple parameters of the provisional model until a predetermined termination condition is satisfied. The learning unit 130 can perform machine learning using an existing method. The learning unit 130 stores the generated provisional model in the model storage unit 131.

[0069] In S205, the selection unit 110 determines whether the prediction error of the provisional model obtained in S204 satisfies a condition. As described above, this condition can be that the magnitude of change in the prediction error is equal to or less than a predetermined threshold. If the prediction error of the provisional model does not satisfy the condition (No in S205), the selection unit 110 increments the value of i (S206). That is, the selection unit 110 increases the value of i by 1. Then, the selection unit 110 executes the process of S203 again.

[0070] On the other hand, if the prediction error of the provisional model satisfies the condition (Yes in S205), the second selection process ends.

[0071] The latest provisional model stored in the model storage unit 131 at the time when the second selection process is completed can be used as the trained model 30 in the prediction method described later. Finally, the i candidate elements extracted by the selection unit 110 in S203 are finally selected by the selection unit 110 as the feature F a corresponds to the elements you select to include in

[0072] Furthermore, the final tentative model generation step (S204) in the second selection process corresponds to the learning step S20 according to this embodiment. In other words, it can also be understood that the learning step S20 is performed within the selection step S10.

[0073] The selection unit 110 selects the training data at the time when the second selection process is completed, that is, the training data used to generate the trained model 30, and the feature quantity F a The learning data storage unit 111 may store learning data including the above.

[0074] The prediction device 20 according to this embodiment will be described below. The prediction device 20 according to this embodiment can predict the physical property index of a compound using the trained model 30 generated by the learning device 10 according to this embodiment.

[0075] 10 is a diagram showing an overview of a prediction device 20 according to this embodiment. The prediction device 20 according to this embodiment includes a prediction unit 230. The prediction unit 230 predicts a physical property index of a target compound using a trained model 30 generated by the learning device 10 according to this embodiment and feature quantities related to the target compound.

[0076] The hardware configuration of a computer that realizes the prediction device 20 according to this embodiment is shown in, for example, Fig. 2, similar to the learning device 10. However, a program module that realizes the function of the prediction unit 230 according to this embodiment is stored in the storage device 1080 of the computer 1000 that realizes the prediction device 20 according to this embodiment.

[0077] 10 , the prediction unit 230 can read out the trained model 30 from the model storage unit 131 and use it for prediction. In the example of FIG. 10 , the model storage unit 131 is provided outside the prediction device 20, but the model storage unit 131 may be provided in the prediction device 20. When the model storage unit 131 is provided in the prediction device 20, the model storage unit 131 is realized using a storage device 1080 of the computer 1000 that realizes the prediction device 20.

[0078] 11 is a diagram showing an overview of the prediction method according to this embodiment. The prediction method according to this embodiment is executed by one or more computers. The prediction method according to this embodiment includes a prediction step S30. In the prediction step S30, one or more computers predict the physical property index of the target compound using the trained model 30 generated by the learning device 10 according to this embodiment and feature quantities related to the target compound.

[0079] The target compound may be an existing compound or a new compound, and the trained model 30 according to this embodiment can be applied to a variety of compounds.

[0080] The prediction unit 230 acquires the feature quantities of the target compound. The prediction unit 230 may acquire the feature quantities of the target compound from a device other than the prediction device 20, or from another functional component of the prediction device 20. Alternatively, the prediction unit 230 may read and acquire the feature quantities of the target compound from a storage unit in which the feature quantities are previously stored. This storage unit may be provided in the prediction device 20, or may be provided externally to the prediction device 20. When this storage unit is provided in the prediction device 20, this storage unit is realized using a storage device 1080 of the computer 1000 that realizes the prediction device 20.

[0081] The feature quantities of the target compound may be generated by the prediction unit 230. In this case, the prediction unit 230, for example, calculates a plurality of descriptors indicating the structure of the target compound, similar to the generation unit 100 described above, extracts a plurality of necessary elements from among them, i.e., a plurality of elements corresponding to the plurality of elements selected by the selection unit 110, and generates a feature quantity to be input to the trained model 30. Alternatively, the prediction unit 230 may calculate only a plurality of descriptors corresponding to the plurality of elements selected by the selection unit 110, for example, to generate a feature quantity to be input to the trained model 30.

[0082] The prediction unit 230 inputs the feature amount of the target compound into the trained model 30, and causes the trained model 30 to output the physical property index of the target compound. In this way, the predicted physical property index of the target compound can be obtained.

[0083] The prediction unit 230 outputs the physical property indices obtained for the target compound. For example, the prediction unit 230 may display the physical property indices on a display connected to the prediction device 20, or may output the physical property indices to another device. Alternatively, the prediction unit 230 may store the physical property indices in a storage unit accessible from the prediction unit 230.

[0084] Next, the operation and effect of this embodiment will be described. According to this embodiment, the selection unit 110 executes a first selection process and a second selection process to select an element to be included in the feature from among multiple candidate elements included in the provisional feature related to the compound. In this way, the physical properties of a wide range of compounds can be predicted with high accuracy.

[0085] (Second embodiment) FIG. 12 is a diagram illustrating an overview of a learning device 10 according to a second embodiment. The learning device 10 according to this embodiment includes a learning unit 130 and a range specification unit 150. The learning unit 130 generates a trained model by performing machine learning using a plurality of pieces of training data. The training data is data in which feature quantities related to compounds and indicators related to the physical properties of the compounds are associated with each other. The range specification unit 150 specifies the application range of the trained model using a plurality of pieces of training data.

[0086] The hardware configuration of the computer that realizes the learning device 10 according to this embodiment is similar to that of the learning device 10 according to the first embodiment, and is shown, for example, in Figure 2. However, the storage device 1080 of the computer 1000 that realizes the learning device 10 according to this embodiment stores program modules that realize the functions of the learning unit 130 and range identification unit 150 according to this embodiment.

[0087] FIG. 13 is a diagram showing an overview of the learning method according to this embodiment. The learning method according to this embodiment is executed by one or more computers. The learning method according to this embodiment includes a learning step S40 and a range specification step S50. In the learning step S40, the one or more computers perform machine learning using a plurality of pieces of training data to generate a trained model. The training data is data in which feature quantities related to compounds are associated with indicators related to the physical properties of the compounds. In the range specification step S50, the one or more computers specify the application range of the trained model using the plurality of pieces of training data.

[0088] The learning method according to this embodiment can be executed by the learning device 10 according to this embodiment. The learning device 10 and the learning method according to this embodiment will be described in detail below. The trained model 30, compounds, and physical property indices according to this embodiment are as described in the first embodiment. However, the training data used to generate the trained model 30 by the learning device 10 according to this embodiment is not limited to data that has undergone element selection as described in the first embodiment. In other words, the learning device 10 according to this embodiment does not need to be equipped with the selection unit 110.

[0089] To predict the physical properties of compounds, machine learning is usually performed using information on known compounds as training data. However, it must be said that the number of compounds used in machine learning is overwhelmingly small compared to the number of potential compounds, including unknown compounds. Therefore, it is not obvious whether applying a trained model to a certain compound will result in accurate prediction of the physical properties.

[0090] According to this embodiment, the learning device 10 identifies the application range of the generated trained model 30. Therefore, based on the application range, it is possible to determine whether the physical properties of a target compound can be accurately predicted using the trained model 30. In other words, it is possible to determine whether the target compound is suitable for the trained model 30, and to predict the physical properties of suitable compounds with high accuracy.

[0091] The function of the range specifying unit 150 according to this embodiment will be described in detail below.

[0092] The range specification unit 150 according to this embodiment specifies the application range based on the in-range prediction error and the out-of-range prediction error. The in-range prediction error is the prediction error of the trained model 30 for data within the application range. The out-of-range prediction error is the prediction error of the trained model 30 for data outside the application range. The range specification unit 150 calculates the prediction error of the trained model 30 using the mean absolute error (MAE), the root mean square error (RMSE), and the coefficient of determination (R 2) can be used to identify the prediction error. It is preferable to use a test data set that is different from the training data set to calculate the prediction error.

[0093] More specifically, the range determination unit 150 determines the in-range prediction error and the out-of-range prediction error for the tentative application range, and determines the application range based on the relationship between the in-range prediction error and the out-of-range prediction error determined for the tentative application range and the change in the in-range prediction error in response to the change in the tentative application range.

[0094] A specific description will be given below. As described above, the training data is data in which the feature quantities (input data) of a compound are associated with physical property indicators (correct data). The range identification unit 150 identifies the application range using data equivalent to the input data of the trained model 30, which is included in the training dataset used in the machine learning to generate the trained model 30. The portion of the training data other than the correct data, i.e., the data included in the training data equivalent to the input data, will be referred to as "training input data" hereinafter. The range identification unit 150 identifies the application range using the training input data. The application range can be defined as a range based on the training input data. The training input data is, for example, the feature quantities of a compound. It is preferable that the range identification unit 150 identifies the application range using all elements of the feature quantities included in the training dataset used in the machine learning to generate the trained model 30.

[0095] The range specification unit 150 sets a provisional application range, determines whether the provisional application range satisfies predetermined conditions, and determines the provisional application range as the application range for the trained model 30 when the provisional application range is changed and the predetermined conditions are satisfied.

[0096] In the following example, X is used as a parameter for determining the width of the provisional application range. The value of X is not particularly limited as long as it is an index representing a proportion of the entire training input data. For example, it may be a percentage of the entire training input data or a proportion of the entire training input data. The range identification unit 150 sets a threshold for the distance corresponding to X% of all training input data so that X% of the training input data falls within the provisional application range. In setting this threshold, the range identification unit 150 can use at least one of the k-nearest neighbor algorithm, the one-class support vector machine algorithm, and the jagging algorithm. For example, the k-nearest neighbor algorithm selects multiple training input data sets with the shortest distances from multiple training input data sets included in the training dataset and calculates their average distance. Then, data whose distance from at least one of the multiple training input data sets within the provisional application range is within this threshold is determined to be data within the application range. On the other hand, data whose distance from any of the multiple training input data sets within the provisional application range is greater than this threshold is determined to be data outside the application range. For example, the distance can be Euclidean distance. Range specifying unit 150 changes the value of X in increments of a predetermined magnitude from an initial value (for example, zero). If X is a percentage, this magnitude can be determined, for example, within the range of 1 to 100.

[0097] The following describes the conditions used to determine whether to confirm a provisional application range as the application range for the trained model 30. As described above, the range identification unit 150 identifies the application range based on the relationship between the in-range prediction error and the out-of-range prediction error identified for the provisional application range, and on the change in the in-range prediction error in response to a change in the provisional application range. That is, if the relationship between the in-range prediction error and the out-of-range prediction error satisfies a first condition and the change in the in-range prediction error in response to a change in the provisional application range satisfies a second condition, the provisional application range is confirmed as the application range.

[0098] The first condition may be, for example, that the in-range prediction error is smaller than the out-of-range prediction error, or that the ratio of the in-range prediction error to the out-of-range prediction error is less than or equal to a first predetermined criterion.

[0099] The second condition is, for example, that the ratio of the magnitude of the change in the within-range prediction error to the magnitude of the change in the tentative application range is equal to or less than a second predetermined criterion. k Provisional scope of application for AD k For the in-range prediction error E k Also, the previous parameter X k-1 Provisional scope of application for AD k-1 For the in-range prediction error E k-1 In this case, (E k -E k-1 ) / (X k -X k-1 ) is equal to or less than the second criterion, the parameter X k It can be said that the second condition is satisfied.

[0100] Note that the range specifying unit 150 may similarly specify the application range based on the change in the out-of-range prediction error in response to the change in the provisional application range, instead of the change in the in-range prediction error in response to the change in the provisional application range.

[0101] After identifying the application range, range identification unit 150 generates range information indicating the application range. The configuration of the range information is not particularly limited, but the range information may be, for example, a combination of a plurality of learning input data within the application range and a distance (i.e., the coordinates of the black circles and the distance D in FIG. 14, which will be described later). Alternatively, the range information may be information indicating the boundaries between the inside and outside of the application range (i.e., information indicating the dashed lines in FIG. 14, which will be described later).

[0102] FIG. 14 is a diagram showing an example of the application range. For ease of understanding, in the example of FIG. 14, the feature quantity consists of only element A and element B. In FIG. 14, black circles indicate learning input data within the application range, white circles indicate learning input data outside the application range, and dashed lines indicate the boundary between the inside and outside of the application range. Stars indicate data that is the subject of a determination of whether it is within the application range or outside the application range. The distance D indicated by the arrow in the figure is the distance determined by the range identification unit 150 based on X. Data P1 is located within the range of distance D from the point indicated by the black circle, and therefore can be said to be within the application range. On the other hand, data P2 is not located within the range of distance D from the point indicated by the black circle, and therefore can be said to be outside the application range.

[0103] Fig. 15 is a diagram illustrating an example of the functional configuration of a learning device 10 according to this embodiment. In the example of Fig. 15, the learning device 10 further includes a learning data storage unit 111 and a model storage unit 131. The learning data storage unit 111 stores learning data. The model storage unit 131 stores a generated trained model 30.

[0104] The learning data storage unit 111 may be provided in the learning device 10 as shown in FIG. 15, or may be provided external to the learning device 10. When the learning data storage unit 111 is provided in the learning device 10, the learning data storage unit 111 is realized by a storage device 1080 of a computer 1000 that realizes the learning device 10 according to this embodiment. The model storage unit 131 may be provided in the learning device 10 as shown in FIG. 15, or may be provided external to the learning device 10. When the model storage unit 131 is provided in the learning device 10, the model storage unit 131 is realized by a storage device 1080 of a computer 1000 that realizes the learning device 10 according to this embodiment.

[0105] FIG. 16 is a flowchart illustrating the flow of processing executed by the learning device 10 according to this embodiment.

[0106] In S301, the learning unit 130 according to this embodiment performs machine learning using a training dataset including a plurality of pieces of training data to generate a trained model 30. When a training dataset is stored in advance in the training data storage unit 111, the learning unit 130 can read and acquire the training dataset from the training data storage unit 111. Alternatively, the learning unit 130 may acquire the training dataset from another device or from another functional configuration unit provided in the learning device 10.

[0107] In machine learning, the learning unit 130 updates multiple parameters of the model until a predetermined termination condition is met, thereby generating a trained model 30. The learning unit 130 can perform machine learning using an existing method. The learning unit 130 stores the generated trained model 30 in the model storage unit 131.

[0108] In S302, the range identification unit 150 initializes a parameter X used to set a provisional application range ("provisional range" in the drawing). The initial value of the parameter X is determined in advance and may be 0 or a value greater than 0. However, the initial value is set to a value that is sufficiently smaller than the value assumed for X in the final application range.

[0109] In S303, the range specifying unit 150 sets a provisional application range using the parameter X. That is, as described above, the range specifying unit 150 sets a threshold value related to distance so that X% of all the learning input data included in the learning dataset falls within the provisional application range.

[0110] In S304, the range specifying unit 150 calculates the in-range prediction error and the out-of-range prediction error for the provisional application range as described above.

[0111] In S305, the range determination unit 150 determines whether the relationship between the in-range prediction error and the out-of-range prediction error satisfies the first condition described above. If the relationship between the in-range prediction error and the out-of-range prediction error satisfies the first condition described above (Yes in S305), the range determination unit 150 then performs the process of S306.

[0112] If the relationship between the in-range prediction error and the out-of-range prediction error does not satisfy the first condition (No in S305), the range determination unit 150 changes the value of the parameter X by a predetermined amount (S308). Then, the range determination unit 150 performs the process of S303 again.

[0113] In S306, the range determination unit 150 determines whether the change in the intra-range prediction error in response to the change in the tentative application range satisfies the second condition described above. If the change in the intra-range prediction error in response to the change in the tentative application range satisfies the second condition (Yes in S306), the range determination unit 150 then performs the process of S307.

[0114] If the change in the in-range prediction error relative to the change in the provisional application range does not satisfy the second condition (No in S306), the range identification unit 150 then performs the process of S308.

[0115] In S307, the range identification unit 150 identifies the provisional application range as the final application range and generates range information indicating the application range. Then, the range identification unit 150 associates the range information with the trained model 30 and further stores it in the model storage unit 131. However, the range identification unit 150 may output the trained model 30 and the range information to another device.

[0116] The prediction device 20 according to this embodiment will be described below. The prediction device 20 can predict the physical property index of a compound using a trained model 30 generated by the learning device 10 according to this embodiment.

[0117] 17 is a diagram showing an overview of the prediction device 20 according to this embodiment. The prediction device 20 according to this embodiment includes a determination unit 210 and a prediction unit 230. The determination unit 210 uses range information indicating the application range for the trained model 30 to determine whether an index related to the application range, based on the feature amount of the target compound, falls within the application range for the trained model 30. When the determination unit 210 determines that the index related to the application range of the target compound falls within the application range, the prediction unit 230 predicts a physical property index of the target compound using the feature amount of the target compound and the trained model 30.

[0118] The hardware configuration of a computer that realizes the prediction device 20 according to this embodiment is shown in Fig. 2, for example, similar to the learning device 10. However, the storage device 1080 of the computer 1000 that realizes the prediction device 20 according to this embodiment stores program modules that realize the functions of the determination unit 210 and prediction unit 230 according to this embodiment.

[0119] In the example of FIG. 17 , the determination unit 210 reads range information from the model storage unit 131 and can use it for determination. In the example of FIG. 17 , the prediction unit 230 reads the trained model 30 from the model storage unit 131 and can use it for prediction. In the example of FIG. 17 , the model storage unit 131 is provided outside the prediction device 20, but the model storage unit 131 may be provided in the prediction device 20. When the model storage unit 131 is provided in the prediction device 20, the model storage unit 131 is realized using a storage device 1080 of the computer 1000 that realizes the prediction device 20.

[0120] FIG. 18 is a diagram showing an overview of the prediction method according to this embodiment. The prediction method according to this embodiment is executed by one or more computers. The prediction method according to this embodiment includes a determination step S60 and a prediction step S70. In the determination step S60, the one or more computers use range information indicating the applicability range for the trained model 30 to determine whether an index related to the applicability range based on the feature amount of the target compound falls within the applicability range for the trained model 30. In the prediction step S70, if it is determined in the determination step S60 that the index related to the applicability range of the target compound falls within the applicability range, the one or more computers predict a physical property index of the target compound using the feature amount of the target compound and the trained model 30.

[0121] The target compound may be an existing compound or a new compound, and the trained model 30 according to this embodiment can be applied to a variety of compounds.

[0122] FIG. 19 is a diagram illustrating the flow of processing executed by the prediction device 20 according to this embodiment.

[0123] In S401, the determination unit 210 acquires the feature quantities of the target compound. The determination unit 210 may acquire the feature quantities of the target compound from a device other than the prediction device 20, or from another functional configuration unit of the prediction device 20. Alternatively, the determination unit 210 may read and acquire the feature quantities of the target compound from a storage unit in which the feature quantities are previously stored. This storage unit may be provided in the prediction device 20, or may be provided externally to the prediction device 20. When this storage unit is provided in the prediction device 20, this storage unit is realized using the storage device 1080 of the computer 1000 that realizes the prediction device 20.

[0124] The feature amount of the target compound may be generated by the determination unit 210. The determination unit 210 may, for example, calculate a descriptor indicating the structure of the target compound and include it in the feature amount. The determination unit 210 may include multiple descriptors in the feature amount. The feature amount of the target compound has the same configuration as the learning input data included in the learning data used in the machine learning to obtain the trained model 30.

[0125] In S402, the determination unit 210 determines whether an index related to the application range of the target compound falls within the application range indicated in the application information of the trained model 30. The index related to the application range is, for example, a distance from the training input data. When the range information is a combination of multiple training input data within the application range and a distance, the determination unit 210 identifies the shortest distance between the feature of the target compound and the multiple training input data included in the range information. Then, the determination unit 210 compares the identified shortest distance with the distance included as a threshold in the range information. When the shortest distance is equal to or less than the threshold distance, the determination unit 210 determines that the index related to the application range of the target compound falls within the application range. When the shortest distance is longer than the threshold distance, the determination unit 210 determines that the index related to the application range of the target compound does not fall within the application range.

[0126] If the index relating to the applicability range of the target compound falls within the applicability range (Yes in S402), then the prediction unit 230 inputs the feature amount of the target compound into the trained model 30, and causes the trained model 30 to output the physical property index of the target compound (S403). Furthermore, the prediction unit 230 outputs the physical property index predicted by the trained model 30 as output information. In this way, the physical property index predicted for the target compound can be obtained.

[0127] On the other hand, if the index relating to the applicability range of the target compound does not fall within the applicability range (No in S402), the prediction unit 230 does not predict the physical property index, and the processing ends. However, if the index relating to the applicability range of the target compound does not fall within the applicability range, the prediction unit 230 may output information indicating that a prediction is not possible as output information. Furthermore, even if the index relating to the applicability range of the target compound does not fall within the applicability range, the prediction unit 230 may input the feature amount of the target compound to the trained model 30 and cause the trained model 30 to output the physical property index of the target compound. Then, the predicted physical property index may be output as output information together with information indicating that the prediction accuracy is not high.

[0128] When the index relating to the range of application of the target compound does not fall within the range of application, it is preferable to obtain the physical property index of the target compound, for example, by actually conducting an experiment. The obtained physical property index can be used to further perform machine learning on the trained model 30, thereby improving the prediction accuracy of the trained model 30.

[0129] The method of output by the prediction unit 230 is not particularly limited, but the prediction unit 230 may, for example, display the output information on a display connected to the prediction device 20, or may output the output information to a device other than the prediction device 20. Alternatively, the prediction unit 230 may store the information in a storage unit accessible from the prediction unit 230.

[0130] Next, the operation and effect of this embodiment will be described. According to this embodiment, the range specification unit 150 specifies the application range of the trained model 30 using multiple pieces of training data. In this way, it is possible to determine whether a target compound is suitable for the trained model 30, and to predict the physical properties of suitable compounds with high accuracy.

[0131] (Third embodiment) 20 is a diagram illustrating the functional configuration of a learning device 10 according to the third embodiment. It includes a selection unit 110, a learning unit 130, and a range specification unit 150 according to the present embodiment. That is, the learning device 10 according to the present embodiment can be said to be the learning device 10 according to the first embodiment further including a range specification unit 150. It can also be said that the learning device 10 according to the present embodiment further includes a selection unit 110 in the learning device 10 according to the second embodiment.

[0132] In the learning method according to this embodiment, as described in the first embodiment, the selection unit 110 selects the provisional feature F ta Among the multiple candidate elements included in a Then, as described in the first embodiment, the learning unit 130 generates the trained model 30 using a plurality of pieces of training data. In each of the plurality of pieces of training data used by the learning unit 130, a feature amount F including the element selected by the selection unit 110 is selected. a and an index relating to the physical properties of the compound.

[0133] Furthermore, as described in the second embodiment, the range identification unit 150 identifies the application range of the trained model 30 using multiple pieces of training data used to update the model parameters when generating the trained model 30.

[0134] In this embodiment, all of the data on the multiple compounds targeted in the first selection process may be used for training to generate the trained model 30, or some of the data may not be used for training to generate the trained model 30. Of the data on the multiple compounds targeted in the first selection process, data on some of the compounds may be used for training to generate the trained model 30, and data on the remaining compounds may be used for testing to calculate the in-range prediction error and out-of-range prediction error of the trained model 30.

[0135] In the learning device 10 according to this embodiment, the selection unit 110 may perform one of the first selection process and the second selection process, but not the other. For example, the candidate elements selected in the first selection process may be selected based on the feature F a Alternatively, if the first selection process is not performed and the provisional feature F generated by the generating unit 100 is ta A second selection process may be performed on the

[0136] The prediction device 20 and prediction method according to this embodiment are the same as the prediction device 20 and prediction method according to the second embodiment.

[0137] Next, the operation and effect of this embodiment will be described. In this embodiment, at least the same operation and effect as in the second embodiment can be obtained. [Example]

[0138] The present embodiment will be described in detail below with reference to examples, but the present embodiment is not limited to the descriptions of these examples.

[0139] Example 1 In Example 1, a trained model for predicting the anticancer activity of a compound against human cancer cells was generated using the training method according to the third embodiment, as follows.

[0140] In this example, we utilized drug sensitivity data from an anticancer drug screening panel for 60 human cancer cell lines established by the National Cancer Institute (NCI). The physical property index to be predicted was the 50% growth inhibitory concentration (GI50). Training data for approximately 25,000 compounds with an average molecular weight of approximately 370 was used to generate trained models for each cell line.

[0141] An example of generating a trained model for the non-small cell lung cancer cell line NCI-H23 will be described in detail below. First, provisional features consisting of 7922 candidate elements were generated for each compound. A first selection process was performed on these provisional features, with a variance threshold of 0.02 and a correlation threshold of 0.99. As a result, 1961 candidate elements were selected. Further, a second selection process was performed, and finally, 16 candidate elements were selected, resulting in a trained model. Figure 21 shows the relationship between the value of i (horizontal axis) and MAE (vertical axis) in this example.

[0142] Next, we identified the scope of application of this trained model, using 80% of the data for the approximately 25,000 compounds mentioned above as training data and the remaining 20% ​​as test data.

[0143] Figure 22 shows the relationship between X and the in-scope and out-of-scope R 2 In this graph, the horizontal axis represents X (the proportion of the training data that falls within the application range (AD)), and the vertical axis represents R 2 This graph shows the R 2 and R for out-of-scope test data 2 It is shown that:

[0144] In this example, the distance corresponding to X=40 was set as the threshold for determining the applicable range. As can be seen from FIG. 22, within the range of X from 0 to 40, the R 2 increases, and the R for the out-of-scope data 2 This is because as X increases, more correctly predicted data falls within the scope of application. On the other hand, when X exceeds 40, the R for the data within the scope of application decreases as X increases. 2 This is because the amount of data with large errors increases within the applicable range.

[0145] As described above, we generated a trained model and identified its scope of application.

[0146] The trained model obtained and the evaluation results of the applicability range are shown in Figs. 23 to 25. In the graphs of Figs. 23 to 25, the horizontal axis indicates the -log(GI50) value obtained by experiment, and the vertical axis indicates the -log(GI50) value predicted by the trained model. Fig. 23 shows the prediction accuracy for the training data, Fig. 24 shows the prediction accuracy for the test data that falls within the applicability range, and Fig. 25 shows the prediction accuracy for the test data that falls outside the applicability range. In each of Figs. 23 to 25, MAE, RMSE, and R are used as statistical values ​​indicating the prediction error. 2 The values ​​of are also shown.

[0147] As can be seen from the results shown in Figures 23 to 25, GI50 was predicted with very high accuracy for the training data and test data that fell within the applicable range. On the other hand, for test data that fell outside the applicable range, R 2 Therefore, it was confirmed that the trained model according to this example can predict the GI50 value of an unknown compound with high accuracy, and also can obtain information in advance as to whether or not accurate prediction is possible.

[0148] The results of similar evaluations of eight cell lines from different sites (non-small cell lung NCI-H23, colonic SW-620, breast MCF7, ovarian IGROV1, blood RPMI-8226, kidney SN12C, skin SK-MEL-28, and nervous system SF-539) are summarized in Table 1. 2 was 0.8 or higher for data within the applicable range and around 0.4 for data outside the applicable range, confirming that predictions can be made with high accuracy for data within the applicable range regardless of the type of cancer cell.

[0149] [Table 1]

[0150] <Example 2> In Example 2, as described below, a trained model for predicting the inhibitory properties of a compound against acetylcholinesterase, a type of neurotransmitter in the body, was generated using the training method according to the third embodiment.

[0151] In this example, we used ChEMBL, an open data database for biologically active small molecules such as pharmaceuticals and pharmaceutical candidate compounds. The physical property index to be predicted was the 50% inhibitory concentration (IC50). A trained model was generated using training data for approximately 4,500 compounds with an average molecular weight of approximately 400.

[0152] The generation of the trained model according to this embodiment will be described in detail below. First, provisional features consisting of 7922 candidate elements were generated for each compound. A first selection process was performed on these provisional features, with the variance threshold set to 0.02 and the correlation threshold set to 0.99. As a result, 247 candidate elements were selected. Furthermore, a second selection process was performed, and as a result, 23 candidate elements were finally selected, resulting in a trained model.

[0153] Next, the application range of this trained model was identified. In this example, the distance corresponding to X=40 was set as the threshold for determining the application range using the same procedure as in Example 1.

[0154] As described above, we generated a trained model and identified its scope of application.

[0155] The trained model obtained and the evaluation results of the range of application are shown in Figs. 26 to 28. In the graphs of Figs. 26 to 28, the horizontal axis indicates the log(IC50) value obtained by the experiment, and the vertical axis indicates the log(IC50) value predicted by the trained model. Fig. 26 shows the prediction accuracy for the training data, Fig. 27 shows the prediction accuracy for the test data that falls within the range of application, and Fig. 28 shows the prediction accuracy for the test data that falls outside the range of application. In each of Figs. 26 to 28, MAE, RMSE, and R are used as statistical values ​​indicating the prediction error. 2 The values ​​of are also shown.

[0156] As can be seen from the results shown in Figures 26 to 28, IC50 was predicted with very high accuracy for the training data and test data that fell within the applicable range. On the other hand, for test data that fell outside the applicable range, R 2 Therefore, it was confirmed that the trained model according to this example can predict the inhibition of neurotransmitters with high accuracy and can obtain information in advance as to whether or not accurate prediction is possible.

[0157] Although the embodiments of the present invention have been described above with reference to the drawings, these are merely examples of the present invention, and various other configurations may be adopted. For example, in the sequence diagrams and flowcharts used in the above description, multiple steps (processes) are described in order, but the order of execution of the steps performed in each embodiment is not limited to the order described. In each embodiment, the order of the steps shown in the drawings may be changed to the extent that the content is not affected. Furthermore, the above-described embodiments may be combined to the extent that the content is not contradictory.

[0158] Below, examples of reference forms are given. 1-1. A learning unit that generates a trained model by performing machine learning using a plurality of pieces of training data in which feature quantities related to compounds and indexes related to the physical properties of the compounds are associated with each other; a range specification unit that specifies an application range of the trained model using the plurality of pieces of training data; Learning device. 1-2. In the learning device described in 1-1, The range identification unit identifies the application range based on an in-range prediction error, which is a prediction error of the trained model for data within the application range, and an out-of-range prediction error, which is a prediction error of the trained model for data outside the application range. Learning device. 1-3. In the learning device described in 1-2, The range specification unit identifying the in-range prediction error and the out-of-range prediction error for a tentative application range; Identifying the application range based on the relationship between the within-range prediction error and the out-of-range prediction error identified for the tentative application range and a change in the within-range prediction error with respect to a change in the tentative application range. Learning device. 1-4. In the learning device according to 1-2. or 1-3., The range identification unit identifies the prediction error of the trained model using at least one of a mean absolute error, a root mean square error, and a coefficient of determination. Learning device. 1-5. The learning device according to any one of 1-1 to 1-4, The range specification unit specifies the application range using at least one of a k-nearest neighbor method, a one-class support vector machine, and a jagging method. Learning device. 1-6. The learning device according to any one of 1-1 to 1-5, a selection unit that selects an element to be included in the feature from a plurality of candidate elements included in the provisional feature, In each of the plurality of learning data, the feature amount including the element selected by the selection unit and an index related to a physical property of the compound are associated with each other; The selection unit a first selection process for selecting two or more candidate elements from the plurality of candidate elements based on at least one of variations in each of the plurality of compounds and similarities between the plurality of candidate elements; a second selection process for selecting one or more candidate elements from the two or more candidate elements as elements to be included in the feature quantity based on the importance of each of the two or more candidate elements selected in the first selection process; Learning device. 1-7. The learning device according to 1-6. In the first selection process, the selection unit selects the two or more candidate elements using at least the variance of each of the candidate elements in the plurality of compounds and a predetermined variance threshold. Learning device. 1-8. The learning device according to 1-6. or 1-7. In the first selection process, the selection unit selects the two or more candidate elements using at least correlation coefficients between the plurality of candidate elements for the plurality of compounds and a predetermined correlation threshold. Learning device. 1-9. The learning device according to any one of 1-6 to 1-8, In the second selection process, the selection unit selecting i candidate elements in descending order of importance from the two or more candidate elements as elements to be included in the feature; The value of i is determined so that the prediction error of a prediction model obtained by machine learning using learning data including features including the i candidate elements satisfies a predetermined condition. Learning device. 1-10. A learning method implemented by one or more computers, comprising: a learning step of generating a trained model by performing machine learning using a plurality of pieces of training data in which feature amounts related to compounds and indexes related to physical properties of the compounds are associated with each other; and a range specification step of specifying an application range of the trained model using the plurality of training data. How to learn. 1-11. To the computer, a learning step of generating a trained model by performing machine learning using a plurality of pieces of training data in which feature amounts related to compounds and indexes related to physical properties of the compounds are associated with each other; and executing a range specification step of specifying an application range of the trained model using the plurality of pieces of training data. program. 1-12. A determination unit that determines whether an index related to the applicability range based on the feature amount of a target compound falls within the applicability range of the trained model, using range information indicating the applicability range of the trained model; a prediction unit that predicts an index related to a physical property of the target compound using the feature amount of the target compound and the trained model when the determination unit determines that the target compound falls within the applicable range. Prediction device. 1-13. A prediction method implemented by one or more computers, comprising: a determination step of determining whether an index relating to the applicability range based on the feature amount of the target compound falls within the applicability range of the trained model, using range information indicating the applicability range of the trained model; and a prediction step of predicting an index related to a physical property of the target compound using the feature amount of the target compound and the trained model when the target compound is determined to be within the applicable range in the determination step. Forecasting methods. 1-14. To the computer, a determination step of determining whether an index relating to the applicability range based on the feature amount of the target compound falls within the applicability range of the trained model, using range information indicating the applicability range of the trained model; If it is determined in the determination step that the target compound falls within the applicable range, a prediction step is executed in which an index related to a physical property of the target compound is predicted using the feature amount of the target compound and the trained model. program. 2-1. A selection unit that selects an element to be included in the feature from among a plurality of candidate elements included in the provisional feature related to the compound; a learning unit that generates a trained model by performing machine learning using a plurality of pieces of training data in which the feature amounts including the selected elements and indexes related to the physical properties of the compound are associated with each other, The selection unit a first selection process for selecting two or more candidate elements from the plurality of candidate elements based on at least one of variations in each of the plurality of compounds and similarities between the plurality of candidate elements; a second selection process for selecting one or more candidate elements from the two or more candidate elements as elements to be included in the feature quantity based on the importance of each of the two or more candidate elements selected in the first selection process; Learning device. 2-2. In the learning device described in 2-1, In the first selection process, the selection unit selects the two or more candidate elements using at least the variance of each of the candidate elements in the plurality of compounds and a predetermined variance threshold. Learning device. 2-3. In the learning device according to 2-1. or 2-2., In the first selection process, the selection unit selects the two or more candidate elements using at least correlation coefficients between the plurality of candidate elements for the plurality of compounds and a predetermined correlation threshold. Learning device. 2-4. The learning device according to any one of 2-1 to 2-3, In the second selection process, the selection unit selecting i candidate elements in descending order of importance from the two or more candidate elements as elements to be included in the feature; The value of i is determined so that the prediction error of a prediction model obtained by machine learning using learning data including features including the i candidate elements satisfies a predetermined condition. Learning device. 2-5. The learning device according to any one of 2-1 to 2-4, The apparatus further includes a range specification unit that specifies an application range of the trained model using the plurality of pieces of training data. Learning device. 2-6. In the learning device described in 2-5, The range identification unit identifies the application range based on an in-range prediction error, which is a prediction error of the trained model for data within the application range, and an out-of-range prediction error, which is a prediction error of the trained model for data outside the application range. Learning device. 2-7. In the learning device according to 2-6, The range specification unit identifying the in-range prediction error and the out-of-range prediction error for a tentative application range; Identifying the application range based on the relationship between the within-range prediction error and the out-of-range prediction error identified for the tentative application range and a change in the within-range prediction error with respect to a change in the tentative application range. Learning device. 2-8. The learning device according to 2-6. or 2-7. The range identification unit identifies the prediction error of the trained model using at least one of a mean absolute error, a root mean square error, and a coefficient of determination. Learning device. 2-9. The learning device according to any one of 2-5 to 2-8, The range specification unit specifies the application range using at least one of a k-nearest neighbor method, a one-class support vector machine, and a jagging method. Learning device. 2-10. A learning method implemented by one or more computers, comprising: a selection step of selecting an element to be included in the feature from a plurality of candidate elements included in the provisional feature related to the compound; a learning step of performing machine learning using a plurality of pieces of learning data in which the feature amounts including the selected elements and indexes related to the physical properties of the compound are associated with each other, thereby generating a trained model; The selecting step includes: a first selection process for selecting two or more candidate elements from the plurality of candidate elements based on at least one of variations in each of the plurality of compounds and similarities between the plurality of candidate elements; and a second selection process for selecting one or more candidate elements from the two or more candidate elements as elements to be included in the feature quantity based on the importance of each of the two or more candidate elements selected in the first selection process. How to learn. 2-11. To the computer, a selection step of selecting an element to be included in the feature from a plurality of candidate elements included in the provisional feature related to the compound; a learning step of generating a trained model by performing machine learning using a plurality of pieces of training data in which the feature amounts including the selected elements and indexes related to the physical properties of the compound are associated with each other; The selecting step includes: a first selection process for selecting two or more candidate elements from the plurality of candidate elements based on at least one of variations in each of the plurality of compounds and similarities between the plurality of candidate elements; and a second selection process for selecting one or more candidate elements from the two or more candidate elements as elements to be included in the feature quantity based on the importance of each of the two or more candidate elements selected in the first selection process. program. 2-12. A prediction unit that predicts an index related to the physical properties of a target compound using a trained model generated by the learning device according to any one of 2-1 to 2-9 and a feature quantity related to the target compound. Prediction device. 2-13. A prediction method implemented by one or more computers, comprising: A prediction step of predicting an index related to the physical properties of a target compound using a trained model generated by the learning device according to any one of 2-1 to 2-9 and a feature amount related to the target compound. Forecasting methods. 2-14. To the computer, A prediction step is executed in which an index related to the physical properties of a target compound is predicted using a trained model generated by the learning device according to any one of 2-1 to 2-9 and a feature amount related to the target compound. program. [Explanation of symbols]

[0159] 10 Learning Device 20 Prediction Device 30 pre-trained models 100 Generator 110 Selection Section 111 Learning data storage unit 130 Learning Department 131 Model Memory Unit 150 Range specification section 210 Judgment section 230 Prediction Department 1000 calculator 1020 Bus 1040 processor 1060 memory 1080 storage device 1100 Input / Output Interface 1120 Network Interface

Claims

1. a learning unit that generates a trained model by performing machine learning using a plurality of pieces of training data in which feature amounts related to compounds and indexes related to the physical properties of the compounds are associated with each other; a range specification unit that specifies an application range of the trained model using the plurality of pieces of training data; Learning device.

2. 2. The learning device according to claim 1, The range identification unit identifies the application range based on an in-range prediction error, which is a prediction error of the trained model for data within the application range, and an out-of-range prediction error, which is a prediction error of the trained model for data outside the application range. Learning device.

3. 3. The learning device according to claim 2, The range specification unit identifying the in-range prediction error and the out-of-range prediction error for a tentative application range; Identifying the application range based on the relationship between the within-range prediction error and the out-of-range prediction error identified for the tentative application range and a change in the within-range prediction error with respect to a change in the tentative application range. Learning device.

4. 4. The learning device according to claim 2 or 3, The range identification unit identifies the prediction error of the trained model using at least one of a mean absolute error, a root mean square error, and a coefficient of determination. Learning device.

5. The learning device according to any one of claims 1 to 4, The range specification unit specifies the application range using at least one of a k-nearest neighbor method, a one-class support vector machine, and a jagging method. Learning device.

6. The learning device according to any one of claims 1 to 5, a selection unit that selects an element to be included in the feature from a plurality of candidate elements included in the provisional feature, In each of the plurality of learning data, the feature amount including the element selected by the selection unit and an index related to a physical property of the compound are associated with each other; The selection unit a first selection process for selecting two or more candidate elements from the plurality of candidate elements based on at least one of variations in each of the plurality of compounds and similarities between the plurality of candidate elements; a second selection process for selecting one or more candidate elements from the two or more candidate elements as elements to be included in the feature quantity based on the importance of each of the two or more candidate elements selected in the first selection process; Learning device.

7. 7. The learning device according to claim 6, In the first selection process, the selection unit selects the two or more candidate elements using at least a variance of each of the candidate elements in the plurality of compounds and a predetermined variance threshold. Learning device.

8. 8. The learning device according to claim 6 or 7, In the first selection process, the selection unit selects the two or more candidate elements using at least correlation coefficients between the plurality of candidate elements for the plurality of compounds and a predetermined correlation threshold. Learning device.

9. The learning device according to any one of claims 6 to 8, In the second selection process, the selection unit selecting i candidate elements from the two or more candidate elements in descending order of importance as elements to be included in the feature quantity; The value of i is determined so that the prediction error of a prediction model obtained by machine learning using learning data including features including the i candidate elements satisfies a predetermined condition. Learning device.

10. 1. A method of learning implemented by one or more computers, comprising: a learning step of generating a trained model by performing machine learning using a plurality of pieces of training data in which feature amounts related to compounds and indexes related to physical properties of the compounds are associated with each other; and a range specification step of specifying an application range of the trained model using the plurality of training data. How to learn.

11. On the computer, a learning step of generating a trained model by performing machine learning using a plurality of pieces of training data in which feature amounts related to compounds and indexes related to physical properties of the compounds are associated with each other; and executing a range specification step of specifying an application range of the trained model using the plurality of pieces of training data. program.

12. a determination unit that determines whether an index related to the applicability range based on the feature amount of the target compound falls within the applicability range of the trained model, using range information indicating the applicability range of the trained model; a prediction unit that predicts an index related to a physical property of the target compound using the feature amount of the target compound and the trained model when the determination unit determines that the target compound falls within the applicable range. Prediction device.

13. 1. A prediction method implemented by one or more computers, comprising: a determination step of determining whether an index relating to the applicability range based on the feature amount of the target compound falls within the applicability range of the trained model, using range information indicating the applicability range of the trained model; and a prediction step of predicting an index related to a physical property of the target compound using the feature amount of the target compound and the trained model when the target compound is determined to be within the applicable range in the determination step. Forecasting methods.

14. On the computer, a determination step of determining whether an index relating to the applicability range based on the feature amount of the target compound falls within the applicability range of the trained model, using range information indicating the applicability range of the trained model; If it is determined in the determination step that the target compound falls within the applicable range, a prediction step is executed in which an index related to a physical property of the target compound is predicted using the feature amount of the target compound and the trained model. program.

Images