Prediction device, data processing device, prediction method, data processing method, computer program, and recording medium

Abstract

To provide a technique suitable for predicting properties of a molecular compound in accordance with the molecular structure of the molecular compound.SOLUTION: Z prediction models are associated with Z sets of explanatory variables which are different from each other. A set of Z explanatory variables is a combination of p explanatory variables that differ from each other among N explanatory variables. The N explanatory variables represent the molecular structure. Each of the Z prediction models substitutes a value corresponding to the molecular structure of a target molecular compound into an explanatory variable set associated with the prediction model, thereby calculating a value of an objective variable representing a property caused by the target molecular compound. A model selection part 222 selects A prediction models from the Z prediction models according to the value of the explanatory variable set corresponding to the molecular structure of the target molecular compound. A prediction execution unit 223 predicts the properties caused by the target molecular compound using the A prediction models.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] This disclosure relates to a prediction device, a data processing device, a prediction method, a data processing method, a computer program, and a recording medium. [Background technology]

[0002] In recent years, machine learning has been used in various fields (for example, Patent Documents 1 and 2). Materials informatics (MI) is an example of a technology that utilizes machine learning. [Prior art documents] [Patent Documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2018-77547 [Patent Document 2] Japanese Patent Publication No. 2006-048429 [Non-patent literature]

[0004] [Non-Patent Document 1] R. Ramprasad et al. npj Computational Materials 3, 54 (2017). [Non-Patent Document 2] Y. Mo, S.-P. Ong and G. Ceder, Chem. Mat., 24, 15, (2012). [Overview of the Initiative] [Problems that the invention aims to solve]

[0005] This disclosure provides a technology suitable for predicting the properties of molecular compounds according to their molecular structure. [Means for solving the problem]

[0006] This disclosure is, A prediction device that predicts at least one property possessed by a target molecular compound based on its molecular structure, A set of Z pre-trained predictive models, where Z is a natural number greater than or equal to 2, each of the Z predictive models is associated with a set of Z distinct explanatory variables, each of the Z sets of explanatory variables is a combination of p distinct explanatory variables out of N explanatory variables, where N is a natural number greater than or equal to 2, the N explanatory variables represent molecular structures, and p is a natural number greater than or equal to 1 and less than N, and each of the Z predictive models calculates the value of the target variable representing the characteristic brought about by the target molecular compound by substituting the value corresponding to the molecular structure of the target molecular compound into the set of explanatory variables associated with that predictive model. A model selection unit that selects A prediction models from the Z prediction models according to the values ​​of the explanatory variable set corresponding to the molecular structure of the target molecular compound, wherein A is a natural number between 1 and Z, and is greater than or equal to 1. A prediction execution unit that uses the A prediction models to predict the properties that the target molecular compound will produce, We provide a prediction device equipped with the following features. [Effects of the Invention]

[0007] The technology described herein is suitable for predicting the properties of molecular compounds according to their molecular structure. [Brief explanation of the drawing]

[0008] [Figure 1] Figure 1 shows a block diagram of the data processing device of the embodiment. [Figure 2] Figure 2 shows a flowchart illustrating the processing of the data processing device according to the embodiment. [Figure 3] Figure 3 shows an example of molecular compound data according to the embodiment. [Figure 4] Figure 4 is a distribution diagram showing the relationship between R training 2 and R validation 2 of the prediction model according to the example. [Figure 5A] Figure 5A is a distribution diagram showing the relationship between the measured flash point and the predicted flash point when using the single model with the highest R validation score (2). [Figure 5B] Figure 5B is a distribution diagram showing the relationship between the measured flash point and the predicted flash point when an ensemble model is constructed by selecting a prediction model according to the molecular structure of each organic molecular compound. [Figure 6] Figure 6 is a distribution diagram showing the distribution of predicted flash point values ​​for the QM9 molecule in the library, as determined by the estimation apparatus according to the example. [Figure 7] Figure 7 shows the top five molecular compounds in terms of predicted flash points, when the molecular compounds proposed in the examples are arranged in descending order of their predicted flash points, along with their predicted flash points. [Modes for carrying out the invention]

[0009] (Our findings) Various attempts are being made regarding materials development.

[0010] In recent years, artificial intelligence (AI) technologies based on machine learning and other methods have been developing rapidly. Along with this, research into materials informatics (MI) for material creation has become increasingly active. MI is a technology that integrates data science and materials science. In data science, data is collected. This collected, known data is then fed into AI. This enables predictions about unknown data, and new knowledge systems can be constructed. In MI, these data science methods are applied to materials science, and new materials are developed from accumulated historical data. One specific example of MI is described in Non-Patent Document 1.

[0011] Traditional materials development, which does not rely on MI (Material Informatics), heavily depends on the tacit knowledge of researchers, such as intuition and experience. In traditional materials development without MI, data was sometimes obtained by repeating experiments hundreds or thousands of times to acquire the desired properties. This requires a lot of time, effort, and expense. In contrast, in materials development using MI, database data is fed to AI. This provides highly accurate experimental candidates. Experiments are then conducted based on the provided experimental candidates. By repeating this cycle of providing experimental candidates and conducting experiments based on those candidates, the materials development period can be shortened. To give a numerical example, the period can be shortened from 10 years to about 1 or 2 years. This reduction in period leads to a reduction in the effort and cost required for development. In fact, one company developed a certain all-solid-state battery material over 5 years using materials development without MI. In contrast, another later entrant company completed the same all-solid-state battery material in 1 year using materials development with MI. This example is a recent and memorable example of how MI can shorten the development period. For more information on the advantages of MI, please refer to Non-Patent Document 2.

[0012] Here, we consider using a prediction system containing X pre-trained prediction models to predict the properties of a molecular compound from its molecular structure. X is a natural number greater than or equal to 2.

[0013] In the example prediction, r prediction models are selected from X prediction models based on their good prediction accuracy. The properties of molecular compounds are predicted from their molecular structure using these r prediction models for any of the multiple molecular compounds. In some cases, this method can accurately predict the properties. However, the prediction accuracy in this case varies depending on the molecular compound. Depending on the molecular structure, sufficient prediction accuracy may not be achieved.

[0014] Therefore, the inventors investigated a technique suitable for predicting the properties of molecular compounds according to their molecular structure.

[0015] The specification may use the expression "target molecular compound." The expression "target" is not intended to be interpreted restrictively to any specific molecular compound.

[0016] The specification may use the expression "molecular structure." "Molecular structure" is a concept that may include the number of constituent atoms within a molecule, the types of constituent atoms, the arrangement of constituent atoms, and the bonding patterns of the constituent atoms. The specification may also use the expression "molecular structure of the molecular compound." In this expression, the molecular structure may be the overall structure of the molecular compound or a partial structure of the molecular compound. A partial structure is, for example, a functional group. The same applies to expressions such as "molecular structure of the target molecular compound."

[0017] The specification may use the expression "properties brought about by the molecular compound." This expression may refer to the properties of the molecular compound itself, or to the properties of a device or other element containing the molecular compound and other components. The same applies to expressions such as "at least one property brought about by the target molecular compound."

[0018] In the specification, the expression "explanatory variable set" is sometimes used. In this expression, the number of explanatory variables included in the explanatory variable set may be one or multiple.

[0019] The specification states: * The symbol "x" is sometimes used to represent vectors. For example, "x * " is a vector.

[0020] (Embodiment) The present disclosure will be described below through embodiments, but these embodiments are not intended to limit the scope of the claims. Furthermore, not all combinations of features described in the embodiments are necessarily essential to the solutions of the present disclosure.

[0021] Figure 1 shows a block diagram of the data processing device 10 of the embodiment. The data processing device 10 includes a learning device 100 and a prediction device 200. In the learning device 100, multiple sets of explanatory variable pairs are selected using multiple explanatory variables representing molecular structures. Each set of explanatory variable pairs includes at least one explanatory variable. Next, multiple prediction models are constructed based on multiple sample data of molecular compounds. The prediction models are associated with the sets of explanatory variable pairs. The prediction models are configured to identify the value of a target variable representing at least one property of the molecular compound by substituting values ​​corresponding to the molecular structure of the molecular compound into the set of explanatory variable pairs associated with them.

[0022] The prediction device 200 generates data representing the molecular structures of multiple candidate molecular compounds (hereinafter referred to as candidate molecular compounds). Next, one or more prediction models are selected from among the multiple prediction models. Then, at least one characteristic exhibited by the candidate molecular compound is predicted using the selected prediction model. If the predicted characteristic exhibited by the candidate molecular compound falls within the target range, that candidate molecular compound is proposed. The proposal is made to the user of the prediction device 200 or the data processing device 10. Specifically, the proposal is made by displaying information.

[0023] As can be understood from the above explanation, the data processing device 10 predicts that the characteristics represented by the dependent variable will be brought about by each factor represented by each explanatory variable. The value of the dependent variable may be an experimentally measured value or a theoretically calculated value.

[0024] The learning device 100 includes an acquisition unit 110, a prediction model construction unit 120, and a model screening unit 130. The prediction device 200 includes a storage unit 210, a candidate search unit 220, a display unit 230, and a recording unit 240. The selection of explanatory variable sets and the construction of prediction models are performed by the prediction model construction unit 120. The generation of data representing the molecular structure of candidate molecular compounds, the selection of prediction models, and the prediction of the characteristics of candidate molecular compounds are performed by the candidate search unit 220. The display of candidate molecular compounds is performed by the display unit 230.

[0025] The prediction model construction unit 120 includes an explanatory variable selection unit 121 and a learning processing unit 122. The candidate search unit 220 includes a provision unit 221, a model selection unit 222, and a prediction execution unit 223. The acquisition unit 110, explanatory variable selection unit 121, learning processing unit 122, model screening unit 130, storage unit 210, provision unit 221, model selection unit 222, prediction execution unit 223, display unit 230, and recording unit 240 will be described in detail below with reference to Figure 2. Figure 2 shows a flowchart for explaining the processing of the data processing device 10 of the embodiment. Steps S11 to S17 are performed by the learning device 100. Steps S21 to S28 are performed by the prediction device 200.

[0026] In step S11, the acquisition unit 110 acquires multiple molecular compound data from the database 20. Each molecular compound data associates the molecular structure of the molecular compound with the properties that the molecular compound exhibits. In this embodiment, the properties that the molecular compound exhibits in each molecular compound data may be properties obtained by measurement or properties obtained by theory.

[0027] Next, in step S12, the acquisition unit 110 generates multiple sample data. Specifically, the acquisition unit 110 transforms the molecular structure of each molecular compound in the molecular compound data using molecular descriptors. This yields values ​​for multiple explanatory variables representing the molecular structure of the molecular compound. In addition, the values ​​of the properties brought about by the molecular compound in each molecular compound data are used as the values ​​of the target variable. In this way, multiple sample data are obtained. In each sample data, the values ​​of the multiple explanatory variables representing the molecular structure of the molecular compound are associated with the values ​​of the target variable representing the properties brought about by the molecular compound. In this embodiment, in each sample data, the value of the target variable representing the properties brought about by the molecular compound can be a characteristic value obtained by measurement or a characteristic value obtained by theory. The multiple sample data are provided to the prediction model construction unit 120.

[0028] Molecular descriptors can be calculated using tools such as RDKit, mordred, Dragon7, cinfony, PaDEL-descriptor, and Chemopy. The input values ​​for molecular descriptors can be, for example, the SMILES notation (simplified molecular input line entry system), which is a structurally unambiguous representation method that uses ASCII alphanumeric characters to represent the chemical structure of a molecule. RDKit, mordred, Dragon7, cinfony, PaDEL-descriptor, and Chemopy can calculate 208, 1825, 5270, 307, 797, and 252 molecular descriptors, respectively. Custom-created molecular descriptors may also be used.

[0029] Figure 3 shows an example of molecular compound data according to the embodiment. Figure 3 contains T molecular compound data. Each row in Figure 3 corresponds to one molecular compound data. The T molecular compound data each corresponds to T distinct molecular compounds. Specifically, the T molecular compounds are molecular compound 1, molecular compound 2, ... and molecular compound T. Figure 3 is, Molecular compound 1, due to its molecular structure, gives a characteristic value of 3.0. Molecular compound 2, due to its molecular structure, gives a characteristic value of 0.5. • Molecular compound T, due to its molecular structure, gives a characteristic value of 2.0. This demonstrates that, in the example in Figure 3, molecular compound 1, molecular compound 2, ..., and molecular compound T are specifically organic molecular compounds. The characteristic values ​​are values ​​normalized by dividing the measured values ​​by the reference values.

[0030] The properties may be the physical properties of the target molecular compound itself, the physical properties of the material containing that molecular compound, or the device properties of a device containing that molecular compound. For example, the properties may be: ·flash point, ·viscosity, ·Solubility, ·resistance, ·boiling point, • Melting point • Dielectric constant, ·Current, ·Voltage, • Current-voltage characteristics, • Frequency of operation of electronic devices, • Operating characteristics of electronic devices, • The degree of degradation of electronic devices, or • Shape of electronic devices, As stated above, the properties may be measured experimentally, calculated theoretically, or obtained by other means. If there are multiple predicted properties, a predictive model can be constructed for each property, and multiple elements from the above list (e.g., flash point and viscosity) may be used as multiple properties.

[0031] Next, in step S13, the explanatory variable selection unit 121 selects X sets of explanatory variables that are different from each other. Specifically, in this embodiment, the molecular structure of a molecular compound is represented by N explanatory variables. The explanatory variable selection unit 121 selects X sets of explanatory variables such that each set of X sets of explanatory variables is a combination of p explanatory variables out of the N explanatory variables. N is a natural number greater than or equal to 2. X is a natural number greater than or equal to 2. p is a natural number greater than or equal to 1 or greater than or equal to less than N. Specifically, the p explanatory variables are p explanatory variables that are different from each other. "X sets of explanatory variables that are different from each other" means that there is at least one difference in explanatory variables between the p explanatory variables in one set of explanatory variables and the p explanatory variables in another set of explanatory variables. For example, if p is 4, and the p explanatory variables in one set of explanatory variables are x1, x2, x3, and x4, and the p explanatory variables in another set of explanatory variables are x1, x2, x3, and x5, then the explanatory variables in these sets can be said to be different from each other.

[0032] In one specific example, X = N C p This is how we can select X sets of explanatory variables so that we cover all combinations of selecting p explanatory variables from N explanatory variables. However, X < N C pTo achieve this, we may randomly select, for example, a combination of p explanatory variables from N explanatory variables. For example, N is between 3 and 1,000,000. In one specific example, N is between 3 and 10,000, and in another specific example, N is between 10,000 and 1,000,000. For example, p is between 2 and 10,000,000. In one specific example, p is between 2 and 20, and in another specific example, p is between 3 and 10,000,000. For example, X is between 3 and 10 quadrillion. In one specific example, X is between 3 and 10,000,000, and in another specific example, X is between 50,000 and 10 quadrillion.

[0033] Next, in step S14, the explanatory variable selection unit 121 selects one set of explanatory variables from among the X sets of explanatory variables.

[0034] Next, in step S15, the learning processing unit 122 constructs a predictive model associated with the set of explanatory variables selected in step S14. Specifically, the above-mentioned multiple sample data are used as multiple training data. In each training data, the values ​​of multiple explanatory variables representing the molecular structure of the molecular compound are associated with the value of the target variable representing the properties that the molecular compound provides. In this embodiment, in each training data, the value of the target variable representing the properties that the molecular compound provides can be a characteristic value obtained by measurement or a characteristic value obtained by theory. The learning processing unit 122 constructs a predictive model associated with the set of explanatory variables selected in step S14 by learning using the multiple training data. In this embodiment, the learning performed by the learning processing unit 122 is supervised learning.

[0035] The learning process for building a predictive model is, for example, Linear functions, Holomorphic function by ridge regression, Holomorphic function by Lasso regression, Support Vector Machines Neural networks, Maximum likelihood estimation method, or Bayesian method It is based on at least one selected from the group consisting of. Typically, support vector machines and neural networks are non-linear functions. The neural network may be based on a self-organizing map. Learning may be deep learning using a neural network. Learning may be by multivariate analysis such as discriminant analysis, multiple regression analysis, etc.

[0036] Explained using mathematical formulas, in this embodiment, the prediction model is the function f(x * ). x * = {x1, x2, ···, x p} is the input vector of the function f(x * ). The function f(x * ) may be an analytical function that takes each element of the vector x * as an input value. The analytical function may be, for example, a linear function, a regular function, etc. The function f(x * ) may be a non-linear function that takes each element of the vector x * as an input value. The non-linear function may be, for example, a support vector machine, a neural network, etc. Each element of the vector x * may be based on molecular descriptors.

[0037] In the first modeling example, p = 3. The function f(x * ) is a linear function. The set of explanatory variables includes explanatory variables corresponding to the 11th, 15th, and 34th RDKit molecular descriptors. f(x * ) is a linear function in three-dimensional space. Specifically, f(x * ) includes each element of the vector x * represented in three-dimensional space with the above RDKit molecular descriptors as each axis. More specifically, f(x * ) = a 11 x 11 + a 15 x 15 + a 34 x 34 + const1 a 11 , a 15and a 34 x is the coefficient. const1 is the constant term. 11 , x 15 and x 34 These are the explanatory variables for the 11th, 15th, and 34th RDKit molecular descriptors of the molecular compound, respectively. The coordinates (x) represented by the training data. 11 , x 15 , x 34 The value of the function f(x * By substituting this into ), the output value f(x * The value of ) is obtained. By doing this for multiple training data, multiple output values ​​are obtained. These output values ​​are obtained from the function f(x). * This is a predicted value representing the characteristic value produced by the molecular compound, as predicted by ). On the other hand, the value of the target variable in each of the multiple training data is the measured value representing the characteristic value. The coefficient a is set such that the sum of the squares of the absolute error of the predicted value and the measured value over the multiple training data is minimized. 11 a 15 and a 34 This optimizes the function f(x * ) is identified.

[0038] In the second modeling example, the function f(x * ) includes crossover terms for different molecular descriptors. Specifically, f(x * )=a 11 x 11 +a 15 x 15 +a 34 x 34 +b 11,15 x 11 x 15 +b 11,34 x 11 x 34 +b 15,34 x 15 x 34 +const2 b. 11,15 , b 11,34 and b 15,34 is the coefficient. const2 is the constant term. The coordinates (x) represented by the training data. 11 , x 15 , x 34) The value of is substituted into the function f(x * ) to obtain the output value f(x * ). By doing this for a plurality of training data, a plurality of output values are obtained. These output values are predicted values representing the characteristic values brought about by the molecular compound predicted by the function f(x * ). On the other hand, the value of the target variable for each of the plurality of training data is the measured value representing the characteristic value. The coefficients a 11 , a 15 , a 34 , b 11,15 , b 11,34 and b 15,34 are optimized so that the sum of the squares of the absolute value errors between the predicted values and the measured values is minimized for the plurality of training data. Thereby, the function f(x * ) is identified.

[0039] In the third modeling example, the function f(x * ) includes the squared terms of the molecular descriptors. Specifically, f(x * ) = a 11 x 11 + a 15 x 15 + a 34 x 34 + b 11,15 x 11 x 15 + b 11,34 x 11 x 34 + b 15,34 x 15 x 34 + c 11 x 11 2 + c 15 x 15 2 + c 34 x 34 2 + const3 where c 11 , c 15 and c 34 are coefficients. const3 is a constant term. The values of the coordinates (x 11 , x 15 , x 34 ) represented by the training data are substituted into the function f(x *By substituting this into ), the output value f(x * The value of ) is obtained. By doing this for multiple training data, multiple output values ​​are obtained. These output values ​​are obtained from the function f(x). * This is a predicted value representing the characteristic value produced by the molecular compound, as predicted by ). On the other hand, the value of the target variable in each of the multiple training data is the measured value representing the characteristic value. The coefficient a is set such that the sum of the squares of the absolute error of the predicted value and the measured value over the multiple training data is minimized. 11 a 15 a 34 , b 11,15 , b 11,34 , b 15,34 , c 11 , c 15 and c 34 This optimizes the function f(x * ) is identified.

[0040] In the fourth modeling example, the function f(x * ) is based on x * It is a nonlinear function of the vector x represented by the training data. * The values ​​of each element are given by the function f(x * By substituting this into ), the output value f(x * The value of ) is obtained. By doing this for multiple training data, multiple output values ​​are obtained. These output values ​​are obtained from the function f(x). * This is a predicted value representing the characteristic value produced by the molecular compound, as predicted by ). On the other hand, the value of the target variable in each of the multiple training data is the measured value representing that characteristic value. f(x) is set so that the sum of the squares of the absolute error of the predicted value and the measured value over the multiple training data is minimized. * This optimizes the function f(x * ) is identified. Specifically, in the fourth modeling example, the function f(x * ) can be a neural network or a support vector machine. Function f(x * If the function f(x) is a neural network, optimize each weight coefficient and bias term. * If the system is a support vector machine, optimize the coefficients and constant terms of each weight.

[0041] In step S15 of this embodiment, a prediction model is constructed and its accuracy is evaluated. Hereinafter, the number of sample data is denoted as J, where J is a natural number greater than or equal to 3. As can be understood from the above description, the learning processing unit 122 constructs a prediction model associated with the selected set of explanatory variables by learning using multiple training data. In this embodiment, in constructing the prediction model, all J sample data are used as the multiple training data mentioned above for learning.

[0042] In contrast, when evaluating the accuracy of a prediction model, out of J sample data, V are used as validation data and the remaining L are used as training data. V is a natural number greater than or equal to 1. L is a natural number greater than or equal to 2. Specifically, the learning processing unit 122 divides the J sample data into M sets. M is a natural number greater than or equal to 2. One of the M sets contains V validation data. M-1 of the M sets contains L training data in total. Each of the M-1 sets contains multiple training data. The number of training data in each of the M-1 sets may or may not be exactly the same. Hereafter, the number of training data in each of the M-1 sets, which may or may not be exactly the same in each set, will be denoted as [d]. [d] is a natural number greater than or equal to 2. The sample data is divided in this way, and cross-validation is performed. During the cross-validation process, the sets used as validation data are changed sequentially, so that M evaluation models are constructed for one prediction model. In other words, M evaluation models are constructed for each set of molecular descriptors.

[0043] <Obtaining evaluation of the predictive model for validation data> The learning processing unit 122 then substitutes values ​​corresponding to the validation data into p explanatory variables for each of the M evaluation models. The resulting values ​​of the target variable are predicted values ​​representing the characteristic values ​​produced by the molecular compounds related to the validation data. On the other hand, the values ​​of the target variable in the validation data are measured values ​​representing the same characteristic values. In this way, M × V pairs of predicted-measured-measured values ​​are obtained. Based on these predicted-measured-measured-value pairs, an evaluation of the predictive model for the validation data is obtained.

[0044] <Obtaining evaluation of the predictive model on training data> Furthermore, the learning processing unit 122 substitutes values ​​corresponding to the training data into p explanatory variables for each of the M evaluation models. The resulting values ​​of the target variable are predicted values ​​representing the characteristic values ​​produced by the molecular compounds related to the training data. On the other hand, the values ​​of the target variable in the training data are measured values ​​representing the same characteristic values. In this way, M × [d] pairs of predicted-measured value pairs are obtained. Based on these predicted-measured value pairs, an evaluation of the prediction model for the training data is obtained.

[0045] Alternatively, you can perform cross-validation using the one-miss method with V=1.

[0046] The evaluation metrics for a predictive model on training data are not particularly limited. For example, the evaluation metric could be based on M × [d] pairs of predicted-actual pairs. • The sum of absolute errors (MAE) between the measured value and the predicted value. • Standard deviation of the error between the measured value and the predicted value. • R-squared for training data obtained from the distribution of actual and predicted values 訓練 2 , These are then calculated. Specifically, these evaluation metrics are indicators for evaluating prediction accuracy.

[0047] The evaluation metrics for the predictive model on validation data are not particularly limited. For example, the evaluation metrics could be based on M×V pairs of predicted-actual values. • The sum of absolute errors (MAE) between the measured value and the predicted value. • Standard deviation of the error between the measured value and the predicted value. • R-squared for validation data obtained from the distribution of measured and predicted values 検証 2 , These are then calculated. Specifically, these evaluation metrics are indicators for evaluating prediction accuracy.

[0048] As an evaluation metric for the overall prediction accuracy of a prediction model, for example, the coefficient of determination R mentioned above is used. 訓練 2 and the coefficient of determination R 検証 2 At least one selected from the group consisting of the above may be adopted. In this embodiment, the coefficient of determination R is used as an evaluation index for the overall prediction accuracy of the prediction model. 訓練 2 and the coefficient of determination R 検証 2 The coefficient of determination R is a combination of these factors. 2 The group (R 訓練 2 , R 検証 2 ) will be adopted.

[0049] As can be understood from the above explanation, in step S15 of this embodiment, the values ​​of the sample data can be applied to a multidimensional space with multiple molecular descriptors as axes. The coefficients associated with the selected explanatory variables in the predictive model based on a linear function, etc., can be adjusted so as to minimize the sum of the squares of the errors with the measured values. Through such learning, a predictive model that can predict the characteristics can be constructed. Also, the coefficient of determination R 2 The group (R 訓練 2 , R 検証 2 The prediction accuracy of the prediction model can be evaluated based on this.

[0050] Next, in step S16, the predictive model building unit 120 decides whether to continue selecting explanatory variable sets. If it decides to continue, it proceeds to step S14. In step S14, a different set of explanatory variables is selected than those already selected. If it decides not to continue, it proceeds to step S17.

[0051] In step S16 of this embodiment, if neither the first nor the second condition is met, it is determined to continue. If at least one selected from the group consisting of the first and second conditions is met, it is determined not to continue. The first condition is that the number of times step S15 has been executed has reached a predetermined number of times. The predetermined number of times may be X or less than X. The second condition is that the number of prediction models constructed in step S15 whose evaluation is equal to or greater than the threshold evaluation has reached a threshold number. In this embodiment, the threshold number is a multiple number.

[0052] By step S17, Y prediction models have been constructed. Y is a natural number between 2 and X (inclusive).

[0053] In step S17, the model screening unit 130 screens the prediction models. Specifically, the model screening unit 130 selects a prediction model that satisfies the screening conditions from Y prediction models. In this embodiment, the screening conditions include an overfitting condition, which states that the prediction model is not overfitted, and an underfitting condition, which states that the prediction model is not underfitted. Specifically, R 訓練 2 R 検証 2 If the value is greater than or equal to the overfitting threshold, the predictive model is judged to be overfitted. 検証 2 R 訓練 2 If the value is greater than or equal to the untrained threshold, the predictive model is determined to be untrained.

[0054] The overfitting threshold is, for example, between 0.05 and 0.40, and in one numerical example, 0.15. The underfitting threshold is, for example, between 0.05 and 0.40, and in one numerical example, 0.15. The overfitting threshold may be greater than the underfitting threshold, the same as the underfitting threshold, or less than the underfitting threshold. If you want to be stricter in determining overfitting, you can set a small value as the overfitting threshold. If you want to be lenient in determining overfitting, you can set a large value as the overfitting threshold. If you want to be stricter in determining underfitting, you can set a small value as the underfitting threshold. If you want to be lenient in determining underfitting, you can set a large value as the underfitting threshold.

[0055] The screening conditions may include additional conditions in addition to the overfitting and underfitting conditions. Specifically, in this embodiment, the screening conditions include the overfitting, underfitting, and correlation conditions. The correlation conditions are R 検証 2 The condition is that the value is greater than the correlation threshold. The correlation threshold is, for example, between 0.1 and 0.9, and in one numerical example, it is 0.4.

[0056] The screening in step S17 has selected Z predictive models, where Z is a natural number between 2 and Y (inclusive).

[0057] Next, in step S21, the storage unit 210 stores Z prediction models. These Z prediction models form a population of prediction models.

[0058] Next, in step S22, the providing unit 221 prepares data representing the molecular structures of C candidate molecular compounds. C is a natural number of 1 or more. In this embodiment, C is a natural number of 2 or more. Typically, the candidate molecular compounds are molecular compounds with unknown properties.

[0059] In this embodiment, the supply unit 221 may include a molecule generator. The molecule generator generates data representing the molecular structure of a candidate molecular compound. Specifically, the molecule generator randomly arranges a predetermined number of atoms from a predetermined number of predetermined atomic species, randomly generates chemical bonds between these atoms, and then, if the resulting molecule is chemically valid, collects that molecule as a candidate molecular compound, repeating this process.

[0060] The process of collecting chemically valid molecules as candidate molecular compounds is carried out according to a predetermined algorithm to satisfy predetermined rules that define acceptable molecular structures. These predetermined rules include, for example, those relating to the number of bonds between atoms. For example, there is one bond between hydrogen and other atoms, three bonds between nitrogen and other atoms, and four bonds between carbon and other atoms.

[0061] The supply unit 221 may collect candidate molecular compounds from known databases. These databases may include Pubchem, CAS, ChemSpider, etc. Typically, the candidate molecular compounds collected from the databases are those not used in building the prediction model in the learning processing unit 122.

[0062] Next, in step S23, the supply unit 221 selects one candidate molecular compound from among the C candidate molecular compounds.

[0063] Next, in step S24, the model selection unit 222 selects A prediction models from the population of prediction models according to the molecular structure of the candidate molecular compound selected in step S23. A is a natural number between 1 and Z (inclusive). In this embodiment, A is a natural number between 2 and Z (inclusive). In this embodiment, A is a predetermined number.

[0064] In this embodiment, A prediction models are selected according to a selection criterion. The selection criterion is to prioritize the selection of prediction models that satisfy the interpolation condition and have a high evaluation. By including the interpolation condition in the selection criterion, prediction models that are well-suited to the molecular structure of the molecular compound can be selected. This can contribute to good prediction of characteristics. Prioritizing the selection of prediction models with high evaluations can also contribute to good prediction of characteristics. In this embodiment, the interpolation condition is that candidate coordinates are interpolated to a set of reference coordinates in a p-dimensional explanatory variable space. Candidate coordinates are coordinates defined by the values ​​of explanatory variable sets corresponding to the molecular structure of the candidate molecular compound. The set of reference coordinates is a plot of coordinates defined by the values ​​of explanatory variable sets corresponding to the training data used when training the prediction model, for multiple training data sets. The above evaluation is typically an evaluation of prediction accuracy. Specifically in this embodiment, the above evaluation is the coefficient of determination R 検証 2 That is the case.

[0065] Instead of the interpolation condition, other conditions that show the relationship between the candidate coordinates and the set of reference coordinates may be used. Other conditions include, for example, distance conditions. Specifically, within the p-dimensional explanatory variable space, the reference coordinate with the shortest distance from the candidate coordinates is defined as the specific coordinate. The distance between the candidate coordinates and the specific coordinate is defined as the specific distance. The distance condition is that the specific distance is less than or equal to a predetermined value. The above evaluation is expressed by the coefficient of determination R 検証 2 It does not have to be that way. The above evaluation may be the sum of absolute errors (MAE) of the errors between the measured and predicted values ​​for the validation data, or it may be the standard deviation of the errors between the measured and predicted values ​​for the validation data.

[0066] If the number of predictive models that satisfy the selection criteria in step S24 is less than A, appropriate actions may be taken. In the first example, in this case, the process returns to step S17. Then, at least one action selected from the group consisting of increasing the overfitting threshold, increasing the underfitting threshold, and decreasing the correlation threshold is performed, and then step S17 and subsequent steps are executed again. In the second example, in this case, the steps from step S24 onward are executed using the fewer than A predictive models selected in step S24 according to the selection criteria. If the second example is adopted, in the explanation of steps S24 onward, "A" is read as the above number less than A.

[0067] Next, in step S25, the prediction execution unit 223 uses the A prediction models selected in step S24 to perform predictions regarding the candidate molecular compounds selected in step S23. Specifically, the prediction execution unit 223 uses the A prediction models to predict the properties that the candidate molecular compounds will possess. These properties may be properties of the candidate molecular compounds themselves. These properties may also be properties of a device that includes the candidate molecular compounds and other elements.

[0068] In this embodiment, A is 2 or more. The prediction execution unit 223 predicts the properties that a candidate molecular compound will possess using an ensemble model that includes A prediction models. Specifically, it calculates the predicted value of the property using each of the A prediction models. The average of the obtained A predicted values ​​is treated as the predicted value of the property by the ensemble model. In this way, prediction using the ensemble model is performed. The average can be, for example, the arithmetic mean, harmonic mean, geometric mean, trimmed mean, etc.

[0069] Next, in step S26, the prediction execution unit 223 determines whether the number of executions in step S25 has reached C. If the number of executions has reached C, the process proceeds to step S27. If the number of executions has not reached C, the process proceeds to step S23.

[0070] At step S27, for B candidate molecular compounds, the properties predicted to be brought about by their co-molecular compounds are within the target range. B is an integer between 0 and C, inclusive. In this embodiment, B is a natural number between 1 and C, inclusive.

[0071] In step S27, the prediction execution unit 223 selects candidate molecular compounds whose predicted properties fall within the target range. In this embodiment, B candidate molecular compounds are selected.

[0072] Next, in step S28, the prediction device 200 proposes the candidate molecular compounds selected in step S27 to the user of the prediction device 200 or the data processing device 10. In this proposal, the prediction device 200 notifies the user of the candidate molecular compounds along with the characteristics that the candidate molecular compounds are predicted to bring about. In this embodiment, the candidate molecular compounds selected in step S27 and the characteristics that the candidate molecular compounds are predicted to bring about are displayed on the display unit 230.

[0073] Specifically, the target value exists within the target range. In step S27, B candidate molecular compounds are sorted in a table format or the like. The prediction execution unit 223 re-sorts the B candidate molecular compounds in order of the characteristic value predicted to be produced by each candidate molecular compound being closest to the target value. In step S28, the re-sorted B candidate molecular compounds are displayed on the display unit 230 along with the characteristics predicted to be produced by each candidate molecular compound.

[0074] In one specific example, in step S25, the prediction execution unit 223 predicts additional information about a candidate molecular compound if the predicted properties of the candidate molecular compound are within the target range. The additional information includes, for example, the similarity between the candidate molecular compound and molecular compounds corresponding to the training data used to train A prediction models. The additional information also includes, for example, evaluation metrics for A prediction models and evaluations based on those evaluation metrics. Other examples of information included in the additional information are the molecular weight of the candidate molecular compound, the ease of synthesis of the candidate molecular compound, and the registration number of the candidate molecular compound in the molecular database. The prediction of additional information is typically performed by calculation. Then, in step S28, the prediction execution unit 223 notifies the user of the additional information or displays it on the display unit 230.

[0075] The display unit 230 is, for example, a display screen for a smartphone, personal computer, television, etc. The display screen is, for example, a liquid crystal display, an organic electroluminescent display, etc.

[0076] As can be understood from the above description, in the data processing device 10 according to one specific example of this embodiment, multiple prediction models can be selected by screening. For each candidate molecular compound, one or more prediction models that satisfy conditions dependent on the molecular structure of the candidate molecular compound can be selected from among the selected multiple prediction models. From the predicted values ​​by the selected prediction models, one or more candidate molecular compounds whose predicted characteristics fall within the target range can be selected. With this configuration, a prediction model corresponding to the molecular structure of each candidate molecular compound can be selected. Therefore, predictions according to the molecular structure, and ideally, predictions individually optimized for each molecular structure, are made. This makes it possible to predict the characteristics of candidate molecular compounds with various molecular structures with high overall reliability. The selected candidate molecular compounds, along with the characteristics they produce, can be notified to the user.

[0077] This section explains the advantages of constructing an ensemble model that includes multiple predictive models based on linear functions. Predictive models based on linear functions are categorized as weak predictors. Weak predictors can be constructed even with a limited amount of training data. Ensembling models can improve prediction accuracy. Therefore, an ensemble model containing multiple predictive models based on linear functions can predict the properties of molecular compounds with good accuracy based on a limited amount of training data. In one numerical example, accurate predictions are possible with approximately 10 to 1000 training data points. In one example, the number of training data points is between 10 and 1000, and in another specific example, it is between 20 and 200.

[0078] Furthermore, it is also possible to utilize factor information representing factors that are expected to contribute to the characteristics. Specifically, in this embodiment, as described above, the model screening unit 130 screens to select Z predictive models that satisfy the screening conditions from Y predictive models. The screening conditions are an overfitting condition, where the predictive model is not overfitted, an underfitting condition, where the predictive model is not underfitted, and R 検証 2 The correlation condition is that the value is greater than the correlation threshold. The overfitting condition, underfitting condition, and correlation condition are all indicators of prediction accuracy. Therefore, Z prediction models may have higher prediction accuracy than prediction models that are excluded by screening. In this embodiment, among the N explanatory variables, the explanatory variable that appears most frequently in the set of Z explanatory variables corresponding to the Z prediction models is called the specific explanatory variable. Factor information based on the specific explanatory variable has a high probability of contributing to the characteristics. Therefore, information based on the specific explanatory variable is used as factor information representing factors that are predicted to contribute to the characteristics. Such use can contribute to the creation of new molecular compounds or devices containing molecular compounds.

[0079] For example, factor information is recorded in the recording unit 240 and / or displayed in the display unit 230. The recording unit 240 allows the factor information to be used later in various ways. The display unit 230 allows the user of the prediction device to check the factor information. For example, factor information can be provided from the model screening unit 130 to the recording unit 240. Alternatively, factor information can be provided from the recording unit 240 to the display unit 230.

[0080] In predictive models based on linear functions, the correspondence between explanatory variables related to molecular structure and dependent variables related to the properties produced by the molecular compound is easily understood. This is advantageous in that it increases the probability that the factor information actually represents the factors that contribute to the properties.

[0081] The technology relating to this disclosure will be further explained below with reference to examples. In the examples, a prediction device was configured in accordance with the above embodiment. Specifically, the configured prediction device can select multiple prediction models based on linear functions according to the molecular structure of each molecular compound, and construct an ensemble model including the selected prediction models. Then, a search for molecular compounds with high flash points was performed using this prediction device. When using flammable liquids as materials for electronic devices, it is necessary to enhance safety from the standpoint of the Fire Service Act and international distribution. Under the Fire Service Act, as a general rule, liquids with a flash point of less than 250°C are treated as hazardous materials. For this reason, it is desirable to use liquids with the highest possible flash point. Many flammable liquids, depending on their flash point, may fall under Class 4 hazardous materials (flammable liquids) or designated flammable materials (flammable liquids) under the Fire Service Act. Many organic solvents widely used in electronic devices, etc., have a flash point of 200°C or less. For example, the flash points of ethylene carbonate (EC), dimethyl carbonate (DMC), fluoroethylene carbonate (FEC), ethyl methyl carbonate (EMC), and propylene carbonate (PC), which are organic solvents widely used in electronic devices, are 143°C, 17°C, 102°C, 23.9°C, and 132°C, respectively. All of these chain carbonates belong to either the second or third class of petroleum. It is believed that if a safe solvent with a high flash point that does not belong to either the second or third class of petroleum can be found, management costs can be reduced. [Examples]

[0082] <Learning to construct a predictive device for predicting flash points> Data for 112 molecular compounds was prepared. These molecular compounds are organic molecules whose molecular structure and flash point are known. The 112 data were collected from publicly known databases and literature. For each of the 112 data, the molecular structure of the organic molecule and its flash point are correlated.

[0083] For each explanatory variable set, we used combinations of three explanatory variables from the 200 molecular descriptors in RDKit. The number of explanatory variable sets used was: 200 There are 1,313,400 pairs of explanatory variables in C3. These 1,313,400 pairs of explanatory variables are all distinct from one another.

[0084] The molecular structures of 112 organic molecular compounds were converted into values ​​for multiple explanatory variables representing the molecular structure of each organic molecular compound using RDKit molecular descriptors. This resulted in 112 sample data points for each molecular compound. In each sample data point, the values ​​of the multiple explanatory variables representing the molecular structure of the organic molecular compound are correlated with the value of the objective variable representing the flash point of the organic molecular compound.

[0085] Of the 112 sample data points, 87 were used as training data. The remaining 25 data points were used as test data.

[0086] Through training using 87 data points, 1,313,400 predictive models were constructed, each corresponding to 1,313,400 pairs of explanatory variables. These predictive models predict flash points using linear functions. In each predictive model, a set of explanatory variables representing the molecular structure of an organic molecular compound is associated with a target variable representing the flash point of the organic molecular compound.

[0087] For the 1,313,400 predictive models, follow the explanations in the above sections on "<Obtaining evaluation of predictive models on validation data>" and "<Obtaining evaluation of predictive models on training data>" to set up the pairs (R 訓練 2 , R 検証 2 ) was calculated. Specifically, the pair (R) was calculated by cross-validation using the one-miss method. 訓練 2 , R 検証 2 The coefficient of determination R for the training data of the prediction model according to the example was calculated. Figure 4 shows the coefficient of determination R for the training data of the prediction model according to the example. 訓練 2 And the coefficient of determination R for the validation data of the same prediction model. 検証 2 This is a distribution diagram showing the relationship between and . In other words, Figure 4 shows the pair (R 訓練 2 , R 検証 2 This is a distribution map of ).

[0088] In Figure 4, the region enclosed by the dotted line DL distributed above the figure contains 14,921 pairs of predictive models out of 1,313,400 predictive models that are neither overfitted nor underfitted and have high correlation (R 訓練 2 , R 検証 2 Points representing ) are plotted. In the example, "neither overfitting nor underfitting" is R 訓練 2 From R 検証 2 This means that the absolute value of the difference after subtracting is less than 0.106. "High correlation" means that R 検証 2 This means that it is greater than 0.4.

[0089] In the example, a screening was performed to narrow down 1,313,400 predictive models to 14,921 predictive models within the dotted line DL. The prediction device in the example selects 10 predictive models from the 14,921 predictive models according to selection criteria and constructs an ensemble model containing the 10 predictive models for each organic molecular compound. The selection criteria in the example is that among the predictive models that satisfy the interpolation condition, R 検証 2 This method prioritizes selecting 10 items with high values. By including interpolation conditions in the selection criteria, a predictive model that fits the molecular structure of the organic molecular compound can be selected, and an ensemble model that fits that molecular structure can be constructed. 検証 2 By prioritizing the selection of models with high performance, it becomes possible to predict the flash point with good accuracy. The predicted value output by the ensemble model is the arithmetic mean of the predicted values ​​output by the 10 selected prediction models.

[0090] <Evaluation of prediction device> 14,921 predictive models were evaluated using 25 test datasets. Regarding the test datasets, • The coefficient of determination R obtained from the distribution of measured and predicted values テスト 2 It is 0.51, The sum of the absolute errors (MAE) between the measured and predicted values ​​is 30.77. The standard deviation of the error between the measured value and the predicted value was 36.48.

[0091] We confirmed an improvement in prediction accuracy by selecting multiple prediction models according to the molecular structure of each organic molecular compound and constructing an ensemble model. Specifically, out of 14,921 prediction models, R 検証 2 The prediction accuracy of the single model with the highest accuracy was evaluated using 25 test data sets. The results are shown in Figure 5A. Figure 5A shows the measured flash point and R 検証 2This is a distribution diagram showing the relationship between the predicted flash point when using the single model with the highest value and the actual flash point. In addition, the prediction accuracy when constructing an ensemble model by selecting multiple prediction models according to the molecular structure of each organic molecular compound was evaluated using 25 test data. The results are shown in Figure 5B. Figure 5B is a distribution diagram showing the relationship between the measured flash point and the predicted flash point when constructing an ensemble model by selecting multiple prediction models according to the molecular structure of each organic molecular compound. In Figures 5A and 5B, the horizontal axis represents the measured flash point, and the vertical axis represents the predicted flash point.

[0092] As can be seen from Figures 5A and 5B, R 検証 2 The error between the observed and predicted values ​​is smaller when constructing an ensemble model by selecting multiple prediction models according to the molecular structure of each organic molecular compound, compared to using the single model with the highest efficiency. Table 1 shows the coefficient of determination R obtained from the distribution of observed and predicted values ​​for the test data. テスト 2 The sum of absolute errors (MAE) between the measured and predicted values ​​for the test data, and the standard deviation of the errors between the measured and predicted values ​​for the test data, are described for both cases. From Table 1, R 検証 2 It can be seen that the value of the index representing prediction accuracy is better when constructing an ensemble model by selecting multiple prediction models according to the molecular structure of each organic molecular compound, compared to when using the single model with the highest value. From Table 1, when predicting the flash point of various candidate organic molecular compounds, R 検証 2 It will be understood that, rather than using the single model with the highest accuracy, it is easier to ensure overall prediction reliability by selecting multiple prediction models according to the molecular structure of each organic molecular compound and constructing an ensemble model. [Table 1]

[0093] Using the estimation device constructed in the example, the flash points of 133,885 molecular compounds registered in Library QM9 were predicted. Library QM9 contains molecular structures of small molecule compounds widely used in the MI field. The prediction results are shown in Figure 6. Figure 6 is a distribution map showing the distribution of predicted flash points of molecular compounds in Library QM9 by the estimation device according to the example. In Figure 6, the horizontal axis represents the molecular weight of the molecular compound. The vertical axis represents the predicted flash point. "Training data" corresponds to the training data. "New data" corresponds to the predicted data from the prediction device for molecular compounds in Library QM9.

[0094] From 133,885 molecular compounds, those with a predicted flash point of 190°C or higher were extracted. 190°C is the highest flash point among the 87 data points, i.e., the training data, mentioned above. As a result of this extraction, 45 molecular compounds were found from the 133,885 molecular compounds. The highest predicted flash point of these 45 molecular compounds was approximately 219°C. Figure 7 shows the five molecular compounds with the highest predicted flash points among the 45 molecular compounds. It is possible to configure a prediction device so that these molecular compounds and their predicted flash points are as proposed. In short, Figure 7 shows the top five molecular compounds in terms of predicted flash points, when the molecular compounds that could be proposed in the example are arranged in descending order of their predicted flash points, along with their predicted flash points.

[0095] Furthermore, for the 14,921 prediction models mentioned above, a histogram of the frequency of each molecular descriptor is created, and the top 5 are shown in Table 2. Specifically, Table 2 shows the prediction accuracy, specifically R 検証 2 Table 2 lists the top 5 molecular descriptors in the histogram of the highly predictive models. The results in Table 2 show that the factor related to the sum of the area values ​​of polar regions on the molecular surface (TPSA) contributes significantly to the flash point. [Table 2]

[0096] Many of the high-flashpoint molecular compounds shown in Figure 7 contain fluorine. This is thought to be because molecular descriptors, called TPSAs, which represent the area of ​​both polar surfaces of a molecule, are a major factor contributing to a high flashpoint (see Table 2).

[0097] As can be understood from the above explanation, the estimation apparatus according to the example makes it possible to list many candidate molecular compounds that can produce a flash point within the target range from rare historical accumulated data.

[0098] The estimation device according to the embodiment is configured to predict the flash point. The estimation device can also be configured to predict other properties instead of the flash point. Such a prediction model can be reconstructed by training again using the training data. Other properties include, for example, viscosity and solubility. Furthermore, the estimation device can be configured to predict the properties of a device that includes other elements along with the compound molecule.

[0099] Furthermore, as can be understood from the above explanation, the estimation apparatus in the example allows us to determine which factors of the molecular structure of the compound molecule contribute to its properties. In the example, as can be understood from Figures 6 and 7 and Table 2, it is considered that the factor related to the sum of the area values ​​of the polarized parts of the molecular surface (TPSA) contributes to the properties. In the example, the explanatory variable related to this factor corresponds to the specific explanatory variable.

[0100] Although the present disclosure has been described above using embodiments and examples, the technical scope of the present disclosure is not limited to the scope described herein. It will be apparent to those skilled in the art that various modifications or improvements can be made to the embodiments and examples. Such modified or improved forms may also be included in the technical scope of the present disclosure. For example, instead of generating sample data in the acquisition unit 110, sample data may be provided to the learning device 100 from outside the learning device 100. Prediction and display of factors expected to contribute to the characteristics are optional. A configuration may also be adopted in which an information signal is transmitted from the prediction device 200 to an external device of the data processing device 10, and the external device displays, based on the information signal, candidates whose expected characteristics fall within the target range, or displays factors that are expected to contribute to the characteristics.

[0101] It should be noted that the execution order of operations, procedures, steps, and stages in the apparatus, systems, programs, and methods shown in the claims, specifications, and drawings is not explicitly stated as "before" or "prior to," and that these can be implemented in any order unless the output of a previous process is used in a later process. Even if the operation flow in the claims, specifications, and drawings is described using phrases such as "first," "next," etc., for convenience, this does not mean that it is mandatory to perform the operations in that order.

[0102] (Note) This disclosure provides for the following technologies:

[0103] (Technology 1) A prediction device that predicts at least one property possessed by a target molecular compound based on its molecular structure, A set of Z pre-trained predictive models, where Z is a natural number greater than or equal to 2, each of the Z predictive models is associated with a set of Z distinct explanatory variables, each of the Z sets of explanatory variables is a combination of p distinct explanatory variables out of N explanatory variables, where N is a natural number greater than or equal to 2, the N explanatory variables represent molecular structures, and p is a natural number greater than or equal to 1 and less than N, and each of the Z predictive models calculates the value of the target variable representing the characteristic brought about by the target molecular compound by substituting the value corresponding to the molecular structure of the target molecular compound into the set of explanatory variables associated with that predictive model. A model selection unit that selects A prediction models from the Z prediction models according to the values ​​of the explanatory variable set corresponding to the molecular structure of the target molecular compound, wherein A is a natural number between 1 and Z, and is greater than or equal to 1. A prediction execution unit that uses the A prediction models to predict the properties that the target molecular compound will produce, A prediction device equipped with the following features.

[0104] Technology 1 is suitable for predicting the properties of molecular compounds based on their molecular structure.

[0105] At least one characteristic may be one characteristic or multiple characteristics. p may be a natural number between 2 and N. A may be a natural number between 2 and Z. A may be a predetermined number. The target molecular compound may contain organic molecular compounds or consist only of organic molecular compounds. The target molecular compound may contain inorganic molecular compounds or consist only of inorganic molecular compounds. The target molecular compound may be a metal-organic framework (MOF) or consist only of metal-organic frameworks. The target molecular compound may contain organometallic complexes or consist only of organometallic complexes. The target molecular compound may contain polymer compounds or consist only of polymer compounds. The target molecular compound may contain at least two selected from the group consisting of organic molecular compounds, inorganic molecular compounds, metal-organic frameworks, organometallic complexes, and polymer compounds.

[0106] (Technology 2) Each of the aforementioned Z prediction models is: Linear functions, Holomorphic function by ridge regression, Holomorphic function by Lasso regression, Support Vector Machines Neural networks, Maximum likelihood estimation method, and, Bayesian method It is based on at least one selected from the group consisting of, A prediction device as described in Technology 1.

[0107] The prediction model in Technology 2 is an example of a prediction model.

[0108] (Technology 3) The prediction execution unit, when A is a natural number of 2 or more, uses an ensemble model containing the A prediction models to predict the properties that the target molecular compound will produce. A prediction device according to Technology 1 or 2.

[0109] As in Technique 3, using an ensemble model makes it less likely that performance will vary between prediction models, as this variation will not be reflected in the prediction of the characteristics.

[0110] (Technology 4) A providing unit that provides data representing the molecular structures of C molecular compounds to the model selection unit, wherein C is a natural number of 1 or more, and each of the molecular structures of the C molecular compounds is given as the molecular structure of the target molecular compound, further comprising the providing unit, The aforementioned supply unit is, (i) Includes a molecular generator that generates data representing the molecular structures of the C molecular compounds in such a way that it satisfies predetermined rules that define acceptable molecular structures, (ii) The system is configured to receive data representing the molecular structures of the C molecular compounds from a known database. A prediction device as described in any one of the three technical specifications.

[0111] According to technique 4, for each of the C molecular compounds, the properties that the molecular compound will possess can be predicted. C may be a natural number greater than or equal to 2. The molecular generator in (i) generates data representing the molecular structures of the C molecular compounds, for example, according to a predetermined algorithm.

[0112] (Technology 5) Each of the A prediction models mentioned above was trained using multiple training datasets. Each of the aforementioned training data sets includes a training input value set which is a combination of the values ​​of the N explanatory variables, Each of the A prediction models satisfies the interpolation condition that the target coordinates are interpolated to the reference coordinate set in the p-dimensional explanatory variable space representing the set of explanatory variables associated with the prediction model. The aforementioned target coordinates are defined by the values ​​of the explanatory variable set corresponding to the molecular structure of the target molecular compound. The aforementioned reference coordinate set is a plot of coordinates defined by the values ​​of the explanatory variable set corresponding to the training data used during the training of the prediction model, for each of the multiple training data sets. A prediction device as described in any one of the technologies 1 to 4.

[0113] Since the A predictive models in Technique 5 satisfy the interpolation conditions, there is a high probability that they are well-suited to the molecular structure of the target molecular compound.

[0114] (Technology 6) When the Z prediction models include the A prediction models that satisfy the interpolation condition, along with at least one prediction model that satisfies the interpolation condition, the evaluation of the A prediction models is higher than the evaluation of the at least one prediction model. A prediction device as described in Technology 5.

[0115] According to the A prediction models of Technology 6, good predictions worthy of high praise can be achieved.

[0116] (Technology 7) The system further includes a candidate display unit that displays the target molecular compound when the characteristics predicted by the prediction execution unit as being produced by the target molecular compound fall within the target range. A prediction device according to any one of the technologies described in items 1 to 6.

[0117] According to Technology 7, the user of the prediction device can identify molecular compounds that may produce characteristics within the target range by looking at the candidate display section.

[0118] (Technology 8) The system further comprises at least one selected from the group consisting of a factor recording unit and a factor display unit, The factor recording unit records factor information that represents factors that are predicted to contribute to the characteristic, based on specific explanatory variables included in the N explanatory variables. The factor display unit displays the factor information. A prediction device as described in any one of the technologies described in items 1 to 7.

[0119] According to Technology 8, factor information can contribute to the creation of new molecular compounds or devices containing molecular compounds. Factor information may be the specific explanatory variables themselves, or it may be information that represents the molecular structure represented by the specific explanatory variables in a form other than the specific explanatory variables. Examples of forms other than the specific explanatory variables include the form of letters, the form of chemical structural formulas, etc.

[0120] One display unit may serve as both the candidate display unit and the factor display unit. One display unit is, for example, the display unit 230 described above. The candidate display unit and the factor display unit may be different display units from each other.

[0121] (Technology 9) A prediction device described in any one of the technologies 1 to 8, A learning device that provides multiple training data sets, wherein in each of the multiple training data sets, the values ​​of the N explanatory variables representing the molecular structure of the training molecular compound are associated with the value of the target variable representing the characteristic brought about by the training molecular compound. The learning device is An explanatory variable selection unit that selects X sets of explanatory variables that are different from each other, wherein each of the X sets of explanatory variables is a combination of p different explanatory variables from the N explanatory variables, and each of the X sets of explanatory variables includes the Z set of explanatory variables, A learning processing unit constructs the Z predictive models, each associated with the Z set of explanatory variables, by learning using the aforementioned multiple training data. A data processing device equipped with the following features.

[0122] According to Technique 9, Z predictive models can be constructed.

[0123] (Technology 10) X is N C p That is, A data processing device as described in Technical 9.

[0124] According to technique 10, even if N is small, it is easy to make X sufficiently large. Therefore, it is likely that one of the X predictive models will be suitable for the molecular structure of the target molecular compound.

[0125] (Technology 11) N is a natural number greater than or equal to 4. p is 3. A data processing device as described in Technical 10.

[0126] (Technology 12) N is a natural number greater than or equal to 3. p is 2. A data processing device as described in Technical 10.

[0127] In techniques 11 and 12, p is sufficiently small. This is advantageous from the standpoint of increasing computation speed.

[0128] (Technology 13) The learning device further includes a model screening unit that selects Z prediction models from Y prediction models by screening based on at least one indicator of prediction accuracy, wherein Y is a natural number between Z and X (inclusive). The prediction device further comprises at least one selected from the group consisting of a factor recording unit and a factor display unit, The factor recording unit records factor information based on a specific explanatory variable that has the most cumulative appearances among the Z sets of explanatory variables out of the N explanatory variables. The factor display unit displays the factor information. A data processing device as described in any one of the technical specifications 9 to 12.

[0129] In Technology 13, the Z predictive models selected through screening may have high predictive accuracy. Among the Z explanatory variable sets corresponding to these Z predictive models, the specific explanatory variable that appears most frequently cumulatively is highly likely to contribute to the characteristics. Furthermore, the factor information based on this specific explanatory variable is also highly likely to contribute to the characteristics. This factor information can contribute to the creation of new molecular compounds or devices containing molecular compounds. Specifically, in Technology 13, the screening by the model screening unit is a screening process that selects predictive models in which at least one indicator related to predictive accuracy is relatively good. Specifically, the "factor" refers to a factor inherent in the molecular structure. Y may be greater than Z.

[0130] In technique 13, each of the Z predictive models may be based on a linear function. A linear function makes the correspondence between explanatory variables relating to molecular structure and dependent variables relating to the properties produced by the molecular compound clear. This is advantageous in terms of increasing the probability that the factor information actually represents factors contributing to the properties. Each of the Z predictive models may be based solely on a linear function.

[0131] (Technology 14) A computer-based prediction method for predicting at least one property of a target molecular compound based on its molecular structure, The prediction method is performed in the presence of Z pre-trained prediction models, where Z is a natural number greater than or equal to 2, and each of the Z prediction models is associated with a set of Z distinct explanatory variables, each of which is a combination of p distinct explanatory variables out of N explanatory variables, where N is a natural number greater than or equal to 2, and the N explanatory variables represent molecular structures, and p is a natural number greater than or equal to 1 and less than N, and each of the Z prediction models calculates the value of the target variable representing the characteristic brought about by the target molecular compound by substituting the value corresponding to the molecular structure of the target molecular compound into the set of explanatory variables associated with that prediction model. The aforementioned prediction method is, Select A predictive models from the Z predictive models according to the values ​​of the explanatory variable set corresponding to the molecular structure of the target molecular compound, where A is a natural number between 1 and Z (inclusive). Using the A prediction models mentioned above, predict the properties that the target molecular compound will produce, A prediction method comprising the following features.

[0132] According to Technique 14, the same effect as Technique 1 can be obtained.

[0133] (Technology 15) A data processing method performed by a computer, Implement the prediction method described in Technical 14, Implementing the learning method, Equipped with, The aforementioned learning method is Multiple training data sets are provided, and in each of these training data sets, the values ​​of the N explanatory variables representing the molecular structure of the training molecular compound are associated with the value of the objective variable representing the characteristic that the training molecular compound brings about. Selecting X sets of explanatory variables that are distinct from each other, where each of the X sets of explanatory variables is a combination of p distinct explanatory variables from the N explanatory variables, and each of the X sets of explanatory variables includes the Z set of explanatory variables. By learning using the aforementioned multiple training data, Z predictive models are constructed, each associated with the Z set of explanatory variables. A data processing method that includes this.

[0134] According to Technique 15, the same effect as Technique 9 can be obtained.

[0135] (Technology 16) A computer program that, when executed by a computer, includes instructions to cause the computer to execute the prediction method described in Technical 14.

[0136] According to the computer program of technology 16, the prediction method of technology 14 can be implemented.

[0137] (Technology 17) A computer program that, when executed by a computer, includes instructions to cause the computer to execute the data processing method described in Technical 15.

[0138] According to the computer program of Technology 17, the data processing method of Technology 15 can be executed.

[0139] (Technology 18) A computer-readable, non-transient recording medium on which the computer program described in Technical 16 is recorded.

[0140] According to the recording medium of Technology 18, the computer program of Technology 16 can be stored in a computer that does not currently have the program recorded, and the prediction method of Technology 14 can be executed by running the computer.

[0141] (Technology 19) A computer-readable, non-transient recording medium on which the computer program described in Technical 17 is recorded.

[0142] According to the recording medium of Technology 19, the computer program of Technology 17 can be stored in a computer that does not currently have the program recorded, and the data processing method of Technology 15 can be executed by running the program.

[0143] A computer may include a processor and memory. Processors include, for example, a CPU (Central Processing Unit), an ASIC (Application Specific Integrated Circuit), and an FPGA (Field Programmable Gate Array).

[0144] Recording media include, for example, semiconductor recording media, magnetic recording media, magneto-optical recording media, and optical recording media. Examples of semiconductor recording media include HDDs (Hard Disk Drives), SD (Secure Digital) cards, and USB (Universal Serial Bus) memory. Examples of magnetic recording media include flexible disks. Examples of magneto-optical recording media include MO (Magneto Optical Disks). Examples of optical recording media include CDs (Compact Discs), DVDs (Digital Versatile Discs), and SSDs (Solid State Drives). [Industrial applicability]

[0145] The prediction device, data processing device, prediction method, data processing method, computer program, and recording medium of this disclosure are useful for the discovery of novel materials such as high flash point solvents. [Explanation of Symbols]

[0146] 10 Data Processing Devices 20 Databases 100 Learning Devices 110 Acquisition Department 120 Predictive Model Construction Department 121 Explanatory Variable Selection Section 122 Learning Processing Unit 130 Model Screening Department 200 Prediction Devices 210 Savings Department 220 Candidate Search Department 221 Provision Department 222 Model Selection Section 223 Prediction Execution Unit 230 Display section 240 Records Section

Claims

1. A prediction device that predicts at least one property of a target molecular compound from the molecular structure of the target molecular compound, Z pre-trained predictive models, where Z is a natural number greater than or equal to 2, each of the Z predictive models is associated with a set of Z distinct explanatory variables, each of the Z sets of explanatory variables is a combination of p distinct explanatory variables out of N explanatory variables, where N is a natural number greater than or equal to 2, the N explanatory variables represent molecular structures, and p is a natural number greater than or equal to 1 and less than N, and each of the Z predictive models calculates the value of the target variable representing the characteristic brought about by the target molecular compound by substituting the value corresponding to the molecular structure of the target molecular compound into the set of explanatory variables associated with that predictive model, A model selection unit that selects A prediction models from the Z prediction models according to the values ​​of the explanatory variable set corresponding to the molecular structure of the target molecular compound, wherein A is a natural number between 1 and Z. A prediction execution unit that uses the A prediction models to predict the properties that the target molecular compound will produce, A prediction device equipped with the following features.

2. Each of the Z prediction models is: Linear functions, Holomorphic function by ridge regression, Holomorphic function by Lasso regression, Support Vector Machines Neural networks, Maximum likelihood estimation method, and, Bayesian method It is based on at least one selected from the group consisting of, The prediction device according to claim 1.

3. The prediction execution unit, when A is a natural number of 2 or more, uses an ensemble model containing the A prediction models to predict the properties that the target molecular compound will produce. The prediction device according to claim 1.

4. A providing unit that provides data representing the molecular structures of C molecular compounds to the model selection unit, wherein C is a natural number of 1 or more, and each of the molecular structures of the C molecular compounds is given as the molecular structure of the target molecular compound, further comprising the providing unit, The aforementioned supply unit is, (i) A molecular generator that generates data representing the molecular structures of the C molecular compounds in such a way that it satisfies predetermined rules that define acceptable molecular structures, or (ii) configured to receive data representing the molecular structures of the C molecular compounds from a known database, The prediction device according to claim 1.

5. Each of the A prediction models mentioned above was trained using multiple training datasets. Each of the aforementioned training data sets includes a training input value set which is a combination of the values ​​of the N explanatory variables, Each of the A prediction models satisfies the interpolation condition that the target coordinates are interpolated relative to the reference coordinate set in the p-dimensional explanatory variable space representing the set of explanatory variables associated with the prediction model. The aforementioned target coordinates are defined by the values ​​of the explanatory variable set corresponding to the molecular structure of the target molecular compound. The aforementioned reference coordinate set is a plot of coordinates defined by the values ​​of the explanatory variable set corresponding to the training data used during the training of the prediction model, for each of the multiple training data sets. The prediction device according to claim 1.

6. When the Z prediction models include the A prediction models that satisfy the interpolation condition, along with at least one prediction model that satisfies the interpolation condition, the evaluation of the A prediction models is higher than the evaluation of the at least one prediction model. The prediction device according to claim 5.

7. The system further includes a candidate display unit that displays the target molecular compound when the characteristics predicted by the prediction execution unit as being produced by the target molecular compound fall within the target range. The prediction device according to claim 1.

8. The system further comprises at least one selected from the group consisting of a factor recording unit and a factor display unit, The factor recording unit records factor information that represents factors that are predicted to contribute to the characteristic, based on specific explanatory variables included in the N explanatory variables. The factor display unit displays the factor information. The prediction device according to claim 1.

9. A prediction device according to any one of claims 1 to 8, A learning device that provides multiple training data sets, wherein in each of the multiple training data sets, the values ​​of the N explanatory variables representing the molecular structure of the training molecular compound are associated with the value of the target variable representing the characteristic brought about by the training molecular compound. The learning device is An explanatory variable selection unit that selects X sets of explanatory variables that are different from each other, wherein each of the X sets of explanatory variables is a combination of p different explanatory variables from the N explanatory variables, and each of the X sets of explanatory variables includes the Z set of explanatory variables, A learning processing unit constructs Z predictive models, each associated with a set of Z explanatory variables, by learning using the aforementioned multiple training data sets. A data processing device equipped with the following features.

10. X is, N C p That is, The data processing device according to claim 9.

11. N is a natural number greater than or equal to 4. p is 3. The data processing device according to claim 10.

12. N is a natural number greater than or equal to 3, p is 2. The data processing device according to claim 10.

13. The learning device further includes a model screening unit that selects Z prediction models from Y prediction models by screening based on at least one indicator of prediction accuracy, wherein Y is a natural number between Z and X (inclusive). The prediction device further comprises at least one selected from the group consisting of a factor recording unit and a factor display unit, The factor recording unit records factor information based on a specific explanatory variable that has the most cumulative appearances among the Z sets of explanatory variables out of the N explanatory variables. The factor display unit displays the factor information. The data processing device according to claim 9.

14. A computer-based prediction method for predicting at least one property of a target molecular compound based on its molecular structure, The prediction method is performed in the presence of Z pre-trained prediction models, where Z is a natural number greater than or equal to 2, and each of the Z prediction models is associated with a set of Z distinct explanatory variables, each of which is a combination of p distinct explanatory variables out of N explanatory variables, where N is a natural number greater than or equal to 2, and the N explanatory variables represent molecular structures, and p is a natural number greater than or equal to 1 and less than N, and each of the Z prediction models calculates the value of the target variable representing the characteristic brought about by the target molecular compound by substituting the value corresponding to the molecular structure of the target molecular compound into the set of explanatory variables associated with that prediction model. The aforementioned prediction method is Select A predictive models from the Z predictive models according to the values ​​of the explanatory variable set corresponding to the molecular structure of the target molecular compound, where A is a natural number between 1 and Z (inclusive). Using the A prediction models mentioned above, predict the properties that the target molecular compound will produce, A prediction method comprising the following features.

15. A data processing method performed by a computer, Performing the prediction method described in claim 14, Implementing the learning method, Equipped with, The aforementioned learning method is Multiple training data sets are provided, and in each of these training data sets, the values ​​of the N explanatory variables representing the molecular structure of the training molecular compound are associated with the value of the objective variable representing the characteristic that the training molecular compound brings about. Selecting X sets of explanatory variables that are different from each other, where each of the X sets of explanatory variables is a combination of p different explanatory variables from the N explanatory variables, and each of the X sets of explanatory variables includes the Z set of explanatory variables. By learning using the aforementioned multiple training data, Z predictive models are constructed, each associated with the Z set of explanatory variables. A data processing method that includes this.

16. A computer program that, when executed by a computer, includes instructions to cause the computer to execute the prediction method described in claim 14.

17. A computer program that, when executed by a computer, includes instructions to cause the computer to execute the data processing method described in claim 15.

18. A computer-readable, non-transient recording medium on which the computer program described in claim 16 is recorded.

19. A computer-readable, non-transient recording medium on which the computer program described in claim 17 is recorded.

Images