Information processing system, information processing method, information processing program, and method for producing molecular compound

Abstract

An information processing system is provided, the information processing system comprising at least one processor. The at least one processor performs the following processing: for each of a plurality of first molecules, acquiring training data indicating a plurality of feature quantities associated with the first molecule; for each of a plurality of second molecules, acquiring test data indicating a plurality of feature quantities associated with the second molecule; for each of the plurality of feature quantities, calculating an index based on a first probability distribution and a second probability distribution, the first probability distribution being the probability distribution of the feature quantity in the training data and the second probability distribution being the probability distribution of the feature quantity in the test data; and based on the index for each of the plurality of feature quantities, selecting one or more of the plurality of feature quantities as one or more input feature quantities to be used as input parameters for a predictive model based on machine learning to predict the property values ​​of molecules.

Description

Technical Field

[0001] This disclosure relates to an information processing system, an information processing method, an information processing program, and a method for generating molecular compounds. Background Technology

[0002] Patent document 1 discloses a method for identifying the amino acid sequence of an antibody with affinity for an antigen. The method includes the steps of: querying a machine learning engine for a proposed amino acid sequence of an antibody with high affinity for the antigen; and obtaining the proposed amino acid sequence from the machine learning engine.

[0003] Non-patent document 1 discloses a technique for selecting feature values ​​in training and test data. In this technique, when the score of the classification between the training data and the test data obtained by an adversarial classifier is higher than a predetermined threshold, feature values ​​that are highly important in that classification are excluded, and the adversarial classifier is retrained. When the score is lower than the threshold, the remaining feature values ​​are used to train a prediction model.

[0004] [List of Citations]

[0005] [Patent Literature]

[0006] [Patent Document 1] International Publication No. WO 2018 / 132752

[0007] [Non-patent literature]

[0008] [Non-Patent Literature 1] Pan, Jing et al., “Adversarial validation approach to concept drift problem in user targeting automation systems at uber.” arXiv:2004.03045 (2020). Summary of the Invention

[0009] [Technical Issues]

[0010] The goal is to improve the accuracy of machine learning models used to predict molecular properties.

[0011] [Solution to the problem]

[0012] The information processing system according to aspects of this disclosure includes at least one processor. The at least one processor performs the following processes: for each of a plurality of first molecules, acquiring training data representing a plurality of feature values ​​associated with the first molecule; for each of a plurality of second molecules, acquiring test data representing a plurality of feature values ​​associated with the second molecule; for each of the plurality of feature values, calculating an index based on a first probability distribution and a second probability distribution, the first probability distribution being a probability distribution in the training data and the second probability distribution being a probability distribution in the test data; and based on the index for each of the plurality of feature values, selecting one or more of the plurality of feature values ​​as one or more input feature values ​​to be used as input parameters for a predictive model based on machine learning to predict molecule property values.

[0013] The information processing method according to an aspect of this disclosure is executed by an information processing system including at least one processor. The information processing method includes the following steps: for each of a plurality of first molecules, acquiring training data representing a plurality of feature values ​​associated with the first molecule; for each of a plurality of second molecules, acquiring test data representing a plurality of feature values ​​associated with the second molecule; for each of the plurality of feature values, calculating an index based on a first probability distribution and a second probability distribution, the first probability distribution being a probability distribution in the training data and the second probability distribution being a probability distribution in the test data; and based on the index for each of the plurality of feature values, selecting one or more of the plurality of feature values ​​as one or more input feature values ​​to be used as input parameters for a predictive model for predicting molecule property values ​​based on machine learning.

[0014] The information processing procedure according to aspects of this disclosure enables a computer to perform the following steps: for each of a plurality of first molecules, acquiring training data representing a plurality of feature values ​​associated with the first molecule; for each of a plurality of second molecules, acquiring test data representing a plurality of feature values ​​associated with the second molecule; for each of the plurality of feature values, calculating an index based on a first probability distribution and a second probability distribution, the first probability distribution being a probability distribution in the training data and the second probability distribution being a probability distribution in the test data; and based on the index for each of the plurality of feature values, selecting one or more of the plurality of feature values ​​as one or more input feature values ​​to be used as input parameters for a predictive model for predicting molecule property values ​​based on machine learning.

[0015] In these respects, based on the trend of each feature value in both the training and test data, the feature values ​​to be used as input parameters for the machine learning-based prediction model are selected as input feature values. By selecting feature values ​​as described above, the accuracy of the machine learning model (prediction model) used to predict the properties of molecules can be improved.

[0016] [Beneficial effects of the invention]

[0017] According to aspects of this disclosure, the accuracy of machine learning models used to predict molecular properties can be improved. Attached Figure Description

[0018] [Figure 1] Figure 1 shows a graph used to illustrate the selection of input feature values.

[0019] [Figure 2] Figure 2 shows an example of the functional configuration of an information processing system.

[0020] [Figure 3] Figure 3 shows an example of the hardware configuration of a computer that acts as an information processing system.

[0021] [Figure 4] Figure 4 shows a graph illustrating an example of training data.

[0022] [Figure 5] Figure 5 shows a flowchart illustrating an example of a process performed by an information processing system.

[0023] [Figure 6] Figure 6 shows a flowchart illustrating an example of the process for selecting input feature values.

[0024] [Figure 7] Figure 7 shows a diagram used to illustrate cross-validation.

[0025] [Figure 8] Figure 8 shows a flowchart illustrating an example of cross-validation.

[0026] [Figure 9] Figure 9 shows a flowchart illustrating another example of the process for selecting input feature values.

[0027] [Figure 10] Figure 10 shows a diagram illustrating the first instance of verification.

[0028] [Figure 11] Figure 11 shows a diagram illustrating a second instance of verification. Detailed Implementation

[0029] In the following description, various embodiments of this disclosure will be described in detail with reference to the accompanying drawings. In the description of the drawings, the same or equivalent elements have the same reference numerals, and redundant descriptions are omitted.

[0030] [System Overview]

[0031] The information processing system according to this disclosure is a computer system that selects one or more feature values ​​of a molecule, which are used as input parameters for a predictive model used to predict the property values ​​of the molecule based on machine learning. In this disclosure, the feature values ​​used as input parameters are referred to as "input feature values". The information processing system eliminates some feature values ​​from a plurality of candidate feature values ​​for input feature values, and then selects the remaining one or more feature values ​​as one or more input feature values. The number of feature values ​​is also referred to as the number of dimensions. The process for selecting one or more input feature values ​​from a plurality of feature values ​​can also be described as a process for reducing the number of dimensions of the feature values.

[0032] In this example, the information processing system uses one or more selected input feature values ​​to perform machine learning, thereby generating a relevant predictive model. This process corresponds to the learning phase of machine learning. In another example, the information processing system inputs one or more input feature values ​​of a molecule whose feature values ​​are unknown into the generated predictive model, thereby predicting the molecule's feature values. In this disclosure, the molecule whose feature values ​​are predicted by the information processing system is also referred to as a "molecule of interest." The information processing system can perform a screening process, in which the feature value of each of a plurality of molecules of interest is predicted, and at least one molecule of interest is selected from the plurality of molecules of interest based on the predicted feature values. The prediction of feature values ​​corresponds to the prediction phase or operational phase of machine learning. In both the learning and prediction phases, the information processing system generates and uses a predictive model with one or more selected input feature values.

[0033] Machine learning is a technique used to autonomously discover patterns or rules based on given information through iterative learning. Predictive models generated by machine learning are machine learning models built using algorithms and data structures, and are also called learned models. In practice, predictive models are built using neural networks such as convolutional neural networks (CNNs).

[0034] In this disclosure, eigenvalues ​​are numerical values ​​that represent the quantitative characteristics of a molecule. In an example, multiple eigenvalues ​​are set for each molecule. Each eigenvalue may represent a characteristic of the relationship between multiple constituent units that form the molecule, a sequence of multiple constituent units, or a characteristic of a specific constituent unit.

[0035] In this disclosure, a characteristic value is a numerical value that represents the quantitative properties of a molecule. For example, a characteristic value can represent various properties related to a molecule's binding ability, affinity, pharmacological activity, physical properties, kinetics, or safety.

[0036] The information processing system selects input feature values. The prediction model is generated using machine learning with training data on the molecule. Based on the prediction model, predicted values ​​of the molecule's properties as shown by the test data are calculated. In this disclosure, the generated prediction model is also referred to as a "property prediction model".

[0037] In this disclosure, each of the molecules shown by the training data is referred to as a "first molecule." Each of the molecules shown by the test data is referred to as a "second molecule." The first molecule and the second molecule may be morphologically identical or different. For example, the first molecule is an antibody, and the second molecule is also an antibody. As another example, the first molecule is a cyclic peptide, and the second molecule is also a cyclic peptide. Furthermore, when using the energy of the interaction between the ligand and the protein as a characteristic value to predict affinity, the first molecule may be a small molecule, where the second molecule is a cyclic peptide.

[0038] The first molecular group comprises multiple first molecules. For each of the multiple first molecules, the training data includes multiple feature values ​​associated with the first molecule. The second molecular group comprises multiple second molecules. For each of the multiple second molecules, the test data includes multiple feature values ​​associated with the second molecule. The multiple first molecules in the first molecular group may be morphologically identical or different. The multiple second molecules in the second molecular group may be morphologically identical or different.

[0039] As an example, when the molecule is an antibody, training data representing the antibody sequence is used. In the case of molecules with the sequence information described above, the first molecular set can be rewritten as a sequence set for training, and the second molecular set can be rewritten as a sequence set for prediction.

[0040] For each of a plurality of first molecules, the training data includes a plurality of feature values ​​of the first molecule and at least one characteristic value of the first molecule. For each of a plurality of second molecules, the test data includes a plurality of feature values ​​of the second molecule. If there is a relatively large number of feature values ​​whose trends change between the training data and the test data, the accuracy of the predictive model will deteriorate. In this disclosure, the information processing system selects one or more input feature values ​​such that feature values ​​whose trends change between the training data and the test data do not affect the machine learning and predictive model. In one example, the information processing system selects feature values ​​whose trend differences between the training data and the test data are relatively small. In another example, the information processing system excludes feature values ​​whose trends differ significantly between the training data and the test data.

[0041] In this example, the information processing system calculates the distance between distributions for each of several feature values ​​as a selection metric. This distance is the distance between a first probability distribution (the one in the training data) and a second probability distribution (the one in the test data). The probability distribution of a feature value shows the probability of each value assumed based on that feature value. Since the probability distribution represents the trend of the feature value, it can be said that feature values ​​with relatively large differences between the first and second probability distributions have different trends in the training and test data. The selection metric is used to select input feature values.

[0042] Figure 1 illustrates the selection of input feature values. In this example, the information processing system calculates the inter-distribution distance as a selection metric for each of the Q feature values. This inter-distribution distance is the distance between a first probability distribution 210 in the training data and a second probability distribution 220 in the test data. The information processing system excludes feature values ​​with relatively large differences between the first probability distribution 210 and the second probability distribution 220 (e.g., feature values ​​F1, F3, F4), and then selects the remaining feature values ​​(e.g., feature values ​​F2, F5, FQ) as input feature values. Each of the selected input feature values ​​is a feature value with relatively small differences and relatively consistent trends between the first probability distribution 210 and the second probability distribution 220. Therefore, the predictive model generated by machine learning based on the input feature values ​​can predict the characteristic values ​​of the molecule of interest.

[0043] [System Configuration]

[0044] Figure 2 illustrates a functional configuration of an information processing system 10 according to an example. In this example, the information processing system 10 accesses a database 20 for storing training data used for machine learning. The database 20 may be provided in a different computer system than the information processing system 10, or it may be a component of the information processing system 10. In this example, the information processing system 10 accesses the database 20 via a communication network such as the Internet or an intranet.

[0045] The information processing system 10 includes a feature selector 11, a learner 12, and a predictor 13. The feature selector 11 is a functional module that selects one or more input feature values ​​from a plurality of feature values. The learner 12 is a functional module that generates a prediction model 30 based on the selected one or more input feature values ​​using machine learning. The predictor 13 is a functional module that performs predictions related to molecules of interest using the generated prediction model 30.

[0046] Figure 3 illustrates an example of the hardware configuration of a computer 100 acting as an information processing system 10. For example, the computer 100 includes a processor 101, main memory 102, secondary memory 103, a communication controller 104, input devices 105, and output devices 106. The processor 101 executes the operating system and applications. The main memory 102 includes, for example, ROM and RAM. The secondary memory 103 includes a hard disk or flash memory and typically stores a larger amount of data than the main memory 102. The communication controller 104 includes, for example, a network interface card (NIC) or a wireless communication module. The input devices 105 include, for example, a keyboard, a mouse, and a touch panel. The output devices 106 include, for example, a monitor and speakers.

[0047] The functional modules of the information processing system 10 are implemented by an information processing program 110 pre-stored in the auxiliary memory 103. Specifically, the functional modules are implemented by allowing the information processing program 110 to be read from the processor 101 or the main memory 102, wherein the processor 101 executes the information processing program 110. The processor 101 causes the communication controller 104, input device 105, or output device 106 to operate according to the information processing program 110, and to read and write data from the main memory 102 or the auxiliary memory 103. The data, prediction models, or databases required for processing can be stored in the main memory 102 or the auxiliary memory 103.

[0048] The information processing program 110 may be recorded in a non-transitory computer-readable storage medium (such as a CD-ROM, DVD-ROM, or semiconductor memory) and provided. Alternatively, the information processing program 110 may be provided via a communication network as a data signal superimposed on a carrier wave.

[0049] Information processing system 10 may include one computer 100, or may include multiple computers 100. When multiple computers 100 are used, the computers 100 are connected via a communication network (such as the Internet or an intranet) to logically construct an information processing system 10.

[0050] Figure 4 illustrates an example of training data stored in database 20. The training data comprises multiple data records corresponding to multiple molecules. In this example, each data record includes a molecule ID, which uniquely identifies each molecule as a data item, multiple feature values ​​of the molecule, and property values ​​of the molecule. When the molecule is a protein, the multiple feature values ​​used as training data can be obtained, for example, by evaluating a protein embedding method (TAPE). In the example of Figure 4, database 20 stores R data records, each representing Q feature values ​​(i.e., feature values ​​in Q dimensions). The number of dimensions Q of the feature values ​​can be less than 10, or it can be on the order of tens, hundreds, thousands, or tens of thousands. Information processing system 10 excludes some feature values ​​from the Q feature values ​​and then selects one or more of the remaining feature values ​​as one or more input feature values. In example 4, physical property values ​​are represented as property values ​​of the molecule, but property values ​​are not limited to physical property values.

[0051] [System Operation]

[0052] Figure 5 shows a flowchart illustrating an example of a process performed by the information processing system 10 (as process flow S1). Process flow S1 is an example of an information processing method according to this disclosure. In step S10, feature selector 11 selects one or more input feature values ​​from a plurality of feature values ​​represented by training data for use in machine learning. In this example, feature selector 11 calculates a plurality of evaluation metrics, some of which feature values ​​are selected or excluded, wherein the evaluation metrics are calculated while changing the feature values ​​to be selected or excluded. Feature selector 11 selects a set of feature values ​​as input feature values ​​when the evaluation metrics meet predetermined conditions. In step S20, learner 12 generates a prediction model 30 by performing machine learning based on the selected one or more input feature values. In step S30, predictor 13 performs a prediction using prediction model 30. Step S10 corresponds to the preprocessing of the learning phase, step S20 corresponds to the learning phase, and step S30 corresponds to the prediction phase. The one or more input feature values ​​selected in step S10 are used in steps S20 and S30.

[0053] (Selection of input feature values)

[0054] Figure 6 shows a flowchart illustrating in detail an example of the process for selecting input feature values ​​(i.e., an example of step S10).

[0055] In step S11, the feature selector 11 retrieves training data and test data from the database 20. For each of the plurality of first molecules, the training data includes multiple feature values ​​associated with the first molecule. For each of the plurality of second molecules, the test data includes multiple feature values ​​associated with the second molecule.

[0056] In step S12, based on the first probability distribution in the training data and the second probability distribution in the test data, the feature selector 11 calculates a selection metric for each of a plurality of feature values. This selection metric is used to select the input feature value. In an example, the feature selector 11 calculates the inter-distribution distance as the selection metric for each feature value. This inter-distribution distance is the distance between the first probability distribution and the second probability distribution. The inter-distribution distance indicates the degree of difference between the first probability distribution and the second probability distribution. The feature selector 11 can use an integral probability metric (IPM) as the inter-distribution distance. As an example of IPM, the feature selector 11 can calculate the Wasserstein distance, the maximum mean difference (MMD), or the Dudley metric as the inter-distribution distance.

[0057] Feature selector 11 can calculate a first probability distribution and a second probability distribution for calculating the selection metric. The feature value is Fi, the number of data records in the training data is j, and the number of data records in the test data is k. Feature selector 11 can calculate the probability distribution of j values ​​for feature value Fi as the first probability distribution and calculate the probability distribution of k values ​​as the second probability distribution.

[0058] In step S13, the feature selector 11 initializes the number of features to be excluded, n. The number of features to be excluded is the number of features excluded from the plurality of features based on a selection criterion for each of the features. The number of features to be excluded can also be described as the numerical reduction in the dimension of the features. The feature selector 11 changes the number of features to be excluded in any subsequent process. When the number of features to be excluded increases, the initial value of the numerical reduction n can be 0, 1, or more. When the number of features to be excluded decreases, the initial value of the numerical reduction n can be {(total number of features) - 1} or a smaller number.

[0059] In step S14, the eigenvalue selector 11 excludes n eigenvalues ​​from a plurality of eigenvalues ​​based on a selection criterion for each eigenvalue. When using inter-distribution distances such as Wasserstein distance, the eigenvalue selector 11 excludes n eigenvalues ​​in descending order of inter-distribution distances. That is, the eigenvalue selector 11 excludes the n eigenvalues ​​in which the distance between the first probability distribution and the second probability distribution is relatively large.

[0060] In step S15, the feature selector 11 generates and evaluates a provisional prediction model based on one or more feature values ​​remaining after the exclusion process. In this disclosure, the generation and evaluation of the provisional prediction model are also referred to as the "evaluation process." The provisional prediction model is not the prediction model 30 generated by the learner 12 and used by the predictor 13, but rather a machine learning model (the learned model) temporarily used to select one or more input feature values. The feature selector 11 may use cross-validation or a retention method as the method used for the evaluation process.

[0061] As an example of step S15, cross-validation will be described. In cross-validation, the feature selector 11 divides the training data into multiple groups, selects one of these groups as validation data, and selects the remaining groups as narrowly defined training data. Each of the independent groups is also called a "fold". The combination of validation data and narrowly defined training data is also called a "split". The feature selector 11 generates a provisional prediction model using the narrowly defined training data and evaluates the provisional prediction model using the validation data. The feature selector 11 performs the generation and evaluation of the provisional prediction model while changing the groups (folds) used as validation data, and calculates the evaluation metric for the provisional prediction model for each split. The feature selector 11 obtains statistics of multiple evaluation metrics obtained from multiple splits as the evaluation metrics for one cross-validation.

[0062] Figure 7 illustrates a diagram used to demonstrate cross-validation. In the example shown, feature selector 11 divides the training data into five groups. In the first split, feature selector 11 selects "Fold 1" as the validation data and selects the remaining four groups as narrowly defined training data. Feature selector 11 uses the narrowly defined training data to generate a provisional predictive model through machine learning and uses the validation data to compute the evaluation metric E1 of this provisional predictive model. Feature selector 11 performs the generation and evaluation of provisional predictive models for the second through fifth splits while changing the validation data, and computes four evaluation metrics E2, E3, E4, and E5 for each split. Feature selector 11 determines the statistics of the five evaluation metrics E1 through E5 as the final evaluation metrics for one cross-validation. Examples of statistics include the mean and median, but are not limited to these.

[0063] Figure 8 shows a flowchart illustrating an example of cross-validation. In step S151, the feature selector 11 sets the cross-validation split. The feature selector 11 divides the training data obtained from the database 20 into multiple groups, selects one of these groups as validation data, and selects the remaining groups as training data in a narrow sense.

[0064] In step S152, the feature selector 11 generates a provisional prediction model by performing machine learning based on one or more feature values ​​remaining after the exclusion process. The feature selector 11 generates the provisional prediction model by performing machine learning (supervised learning) using narrowly defined training data. The feature selector 11 does not use the excluded n feature values ​​and inputs the remaining one or more feature values ​​as input parameters (e.g., input vectors) into the machine learning model. In this example, the feature selector 11 updates the parameter set in the machine learning model by performing backpropagation (backpropagation method) based on the error between the predicted value calculated by the machine learning model and the correct solution (label). The learner 12 obtains the provisional prediction model by repeating this process until a given termination condition is met. The termination condition may be the processing of all data records of the narrowly defined training data.

[0065] In step S153, the feature selector 11 calculates an evaluation metric based on the remaining one or more feature values. The evaluation metric is a value that indicates how accurately the provisional prediction model can compute the feature value. The feature selector 11 does not use the excluded n feature values ​​and inputs the remaining one or more feature values ​​as input parameters (e.g., input vectors) into the provisional prediction model for each data record of the validation data. The provisional prediction model computes the feature value based on the input parameters for each data record. Hereinafter, the feature value computed by the provisional prediction model is also referred to as the predicted feature value.

[0066] The feature selector 11 calculates an evaluation metric for the provisional prediction model based on the correct solutions (labels) for each data record in the predicted feature values ​​and validation data. The correct solutions (labels) in the process are the feature values ​​in the training data used as validation data. For example, the feature selector 11 can calculate the mean squared error between the predicted feature values ​​and the correct solutions, or calculate a correlation coefficient indicating the degree of correlation between the predicted feature values ​​and the correct solutions.

[0067] As shown in step S154, when there are still unexecuted segmentations (not in step S154), the process returns to step S151. In the repeated step S151, the feature selector 11 changes the group used as validation data (folding) and sets the next segmentation. In the repeated step S152, the feature selector 11 generates a provisional prediction model by performing machine learning (supervised learning) on ​​the narrowly defined training data based on the remaining one or more feature values. In the repeated step S153, the feature selector 11 calculates the evaluation metric of the provisional prediction model based on the remaining one or more feature values.

[0068] Once all segmentations have been processed (as in step S154), the process continues to step S155. In step S155, the eigenvalue selector 11 calculates the statistic for the evaluation metric of each segment as the final evaluation metric for cross-validation. For example, the eigenvalue selector 11 calculates the average mean squared error or the average correlation coefficient as the final evaluation metric.

[0069] Similarly, referring to Figure 6, in step S16, the feature selector 11 determines whether to change the exclusion quantity n. The exclusion quantity n can be increased or decreased.

[0070] In one example, the feature selector 11 determines whether to increment the exclusion quantity n based on the number of feature values ​​excluded by the exclusion process or the number of feature values ​​not yet excluded by the exclusion process. For example, the feature selector 11 determines that n increments when the number of valid dimensions (which is the number of feature values ​​remaining after the exclusion process) is equal to or greater than a predetermined threshold, and does not increment when the number of valid dimensions is less than the threshold. Alternatively, the feature selector 11 determines that n increments when the number of valid dimensions is greater than the predetermined threshold, and does not increment when the number of valid dimensions is equal to or less than the threshold. In another example, the feature selector 11 determines that n increments when the number of feature values ​​excluded by the exclusion process is equal to or less than a predetermined threshold, and does not increment when the number of feature values ​​is greater than the threshold. Alternatively, the feature selector 11 determines that n increments when the number of feature values ​​excluded by the exclusion process is less than the predetermined threshold, and does not increment when the number of feature values ​​is equal to or greater than the threshold.

[0071] In one example, the feature selector 11 determines whether to decrement the exclusion quantity n based on the number of feature values ​​excluded by the exclusion process or the number of feature values ​​not yet excluded by the exclusion process. For example, the feature selector 11 determines that n decrements when the number of valid dimensions (which is the number of feature values ​​remaining after the exclusion process) is equal to or less than a predetermined threshold, and does not decrement when the number of valid dimensions is greater than the threshold. Alternatively, the feature selector 11 determines that n decrements when the number of valid dimensions is less than the predetermined threshold, and does not decrement when the number of valid dimensions is equal to or greater than the threshold. On the other hand, the feature selector 11 determines that n decrements when the number of feature values ​​excluded by the exclusion process is equal to or greater than the predetermined threshold, and does not decrement when the number of feature values ​​is less than the threshold. In yet another example, the feature selector 11 determines that n decrements when the number of feature values ​​excluded by the exclusion process is greater than the predetermined threshold, and does not decrement when the number of feature values ​​is equal to or less than the threshold.

[0072] When the exclusion quantity n is changed (as in step S16), the process continues to step S17. In step S17, the feature selector 11 changes the exclusion quantity n by a predetermined amount. This change is either incremental or decremental. The predetermined amount is any positive integer. For example, the feature selector 11 increases the exclusion quantity n by 1, 2, 3, 4, 5, 6, 7, 8, 9, or 10. As another example, the feature selector 11 increases the number of feature values ​​excluded by an amount equal to the initial value n. Alternatively, the feature selector 11 decreases the exclusion quantity n by 1, 2, 3, 4, 5, 6, 7, 8, 9, or 10. After step S17, the process continues to step S14. In the repeated step S14, the feature selector 11 excludes n feature values ​​from a plurality of feature values ​​based on a selection criterion for each feature value. In the repeated step S15, the feature selector 11 generates and evaluates a provisional prediction model based on the remaining one or more feature values. The feature selector 11 gradually increases or decreases the number of exclusions n to calculate the evaluation index of the provisional prediction model for each number of exclusions n.

[0073] If the exclusion quantity n remains unchanged (not in step S16), the process continues to step S18. In step S18, the feature selector 11 selects one or more input feature values ​​from a plurality of feature values. In an example, the feature selector 11 determines the final exclusion quantity nFINAL based on the final evaluation index obtained for each exclusion quantity. The feature selector 11 can determine the exclusion quantity that yields the best final evaluation index as the exclusion quantity nFINAL. For example, the feature selector 11 can determine the exclusion quantity that yields the highest or lowest final evaluation index as the exclusion quantity nFINAL. As an example, when the mean squared error statistic is the evaluation index, the feature selector 11 can determine the exclusion quantity that yields the lowest final evaluation index as the exclusion quantity nFINAL. When the correlation coefficient statistic is the evaluation index, the feature selector 11 can determine the exclusion quantity that yields the highest final evaluation index as the exclusion quantity nFINAL. In other words, the eigenvalue selector 11 can determine the number of exclusions with the smallest statistic having the mean squared error or the number of exclusions with the highest statistic having the correlation coefficient as the final exclusion number nFINAL. Alternatively, the eigenvalue selector 11 can determine one of one or more exclusion numbers that is higher than the final evaluation metric when multiple eigenvalues ​​are used as the exclusion number nFINAL. Alternatively, the eigenvalue selector 11 can determine one of one or more exclusion numbers that is less than a predetermined threshold decrease in the final evaluation metric when multiple eigenvalues ​​are used as the exclusion number nFINAL. The eigenvalue selector 11 excludes nFINAL eigenvalues ​​from multiple eigenvalues ​​based on the selection metric for each eigenvalue. When using inter-distribution distances such as Wasserstein distance, the eigenvalue selector 11 excludes nFINAL eigenvalues ​​in descending order of inter-distribution distance and selects the remaining one or more eigenvalues ​​as one or more input eigenvalues. In this case, the eigenvalue selector 11 selects one or more eigenvalues ​​whose inter-distribution distance is less than a predetermined reference determined based on the final evaluation metric as one or more input eigenvalues. The predetermined reference can be determined by the user or administrator of the information processing system 10, or it can be determined automatically by any technology such as machine learning.

[0074] As described above, the feature selector 11 excludes a portion of the multiple feature values ​​based on a selection criterion for each of the multiple feature values. The feature selector 11 then selects one or more feature values ​​remaining after the exclusion process as one or more input feature values. In one example, the feature selector 11 calculates the inter-distribution distance for each of the multiple feature values, which is the distance between a first probability distribution and a second probability distribution, and then selects one or more input feature values ​​from the multiple feature values ​​based on the inter-distribution distance for each of the multiple feature values. In another example, the feature selector 11 calculates the inter-distribution distance for each of the multiple feature values, which is the distance between a first probability distribution and a second probability distribution, and then selects the feature value with the relatively small inter-distribution distance for each of the multiple feature values ​​as the input feature value. As described above, in this example, the feature selector 11 repeats the evaluation process, in which a provisional prediction model is generated and an evaluation index for the provisional prediction model is calculated based on the remaining one or more feature values, while changing the number of feature values ​​excluded from the plurality of feature values ​​based on the selection index for each of the plurality of feature values ​​(repeating steps S14 to S17). The feature selector 11 can perform the evaluation process via cross-validation. The feature selector 11 determines the number of exclusions based on the plurality of evaluation indices obtained through the repeated evaluation process (step S18). The feature selector 11 excludes a number of nFINAL feature values ​​from the plurality of feature values ​​based on the selection index for each of the plurality of feature values ​​to select one or more input feature values ​​(step S18).

[0075] Referring to Figure 9, another example of the process for selecting input feature values ​​will be described. Figure 9 shows a flowchart illustrating in detail an example of step S10A. Step S10A can also be described as a modification of step S10. The difference between step S10A and step S10 is that the generation and evaluation of the provisional prediction model are repeated, while changing the number of feature values ​​selected, m, instead of the number of feature values ​​excluded, n. The differences will be mainly described below.

[0076] Similar to step S10, feature selector 11 performs the processes of steps S11 and S12.

[0077] In step S13A, the feature selector 11 initializes the number of feature values ​​to be selected, m. The number of selections, m, is the number of feature values ​​selected from the plurality of feature values ​​based on a selection criterion for each of the feature values. When the number of selections, m, increases, its initial value can be 1, 2, or greater. When the number of selections, m, decreases, its initial value can be the total number of feature values ​​or a smaller number.

[0078] In step S14A, the eigenvalue selector 11 selects m eigenvalues ​​from a plurality of eigenvalues ​​based on a selection criterion for each eigenvalue. When using inter-distribution distances such as Wasserstein distance, the eigenvalue selector 11 selects n eigenvalues ​​in ascending order of inter-distribution distances. That is, the eigenvalue selector 11 selects the m eigenvalues ​​from which the distance between the first probability distribution and the second probability distribution is relatively small.

[0079] In step S15A, the feature selector 11 generates and evaluates a provisional prediction model based on one or more selected feature values. That is, the feature selector 11 performs the evaluation process. "One or more selected feature values" is essentially the same as "one or more remaining feature values" in step S10. Therefore, in step S15A, the following can be performed: Figure 8 The cross-validation is shown.

[0080] In step S16A, the feature selector 11 determines whether to change the selection quantity m. The selection quantity m can be increased or decreased.

[0081] In one example, feature selector 11 determines whether the selection quantity m is incremented based on the number of selected feature values ​​or the number of unselected feature values. In another example, feature selector 11 determines whether the selection quantity m is decremented based on the number of selected feature values ​​or the number of unselected feature values.

[0082] When the selection quantity m is changed (as in step S16A), the process continues to step S17A. In step S17A, the feature value selector 11 changes the selection quantity m by a predetermined amount. This change is either incremental or decremental. The predetermined amount is any positive integer. For example, the feature value selector 11 increases the selection quantity m by 1, 2, 3, 4, 5, 6, 7, 8, 9, or 10. As another example, the feature value selector 11 increases the number of feature values ​​selected by an amount equal to the initial value m. Alternatively, the feature value selector 11 decreases the selection quantity m by 1, 2, 3, 4, 5, 6, 7, 8, 9, or 10. After step S17A, the process continues to step S14A. In the repeated step S14A, the feature value selector 11 selects m feature values ​​from a plurality of feature values ​​based on a selection criterion for each feature value. In the repeated step S15A, the feature value selector 11 generates and evaluates a provisional prediction model based on one or more selected feature values. The eigenvalue selector 11 gradually increases or decreases the number of selections m to calculate the evaluation index of the provisional prediction model for each number of selections m.

[0083] When the number of selections *m* remains unchanged (No in step S16A), the process continues to step S18. In step S18, the feature value selector 11 selects one or more input feature values ​​from a plurality of feature values. In this example, the feature value selector 11 determines the final number of selections *mFINAL* based on the final evaluation metric obtained for each number of selections. The feature value selector 11 selects *mFINAL* feature values ​​from the plurality of feature values ​​based on the selection metric for each feature value. When using inter-distribution distances such as Wasserstein distance, the feature value selector 11 selects *mFINAL* feature values ​​as input feature values ​​in ascending order of inter-distribution distances. That is, the feature value selector 11 selects one or more feature values ​​whose inter-distribution distance is less than a predetermined reference determined based on the final evaluation metric as one or more input feature values. As mentioned above, the predetermined reference can be determined manually or automatically.

[0084] As described above, the feature selector 11 performs a repeated evaluation process in which a provisional prediction model is generated and an evaluation index for the provisional prediction model is calculated based on one or more selected feature values, while varying the number of feature values ​​selected from the plurality of feature values ​​based on the selection index for each of the plurality of feature values ​​(repeating steps S14A to S17A). The feature selector 11 can perform the evaluation process via cross-validation. The feature selector 11 selects one or more input feature values ​​based on the plurality of evaluation indices obtained through the repeated evaluation process.

[0085] (Learning Phase)

[0086] The process for generating a predictive model 30 based on one or more selected input feature values ​​will be described, specifically step S20. Learner 12 generates the predictive model 30 by performing machine learning (supervised learning) using one or more input feature values ​​from the training data obtained by feature selector 11. In machine learning, feature selector 11 does not use the excluded nFINAL feature values ​​and inputs one or more feature values ​​from the training data as input parameters (e.g., input vectors) to the machine learning model. In this example, feature selector 11 updates the parameter set in the machine learning model by performing backpropagation (backpropagation method) based on the error between the predicted value calculated by the machine learning model and the correct solution (label). Learner 12 obtains the predictive model 30 by repeating this process until a given termination condition is met. The termination condition may be processing all data records (first quantifier) ​​of the training data. It is important to note that the generated predictive model 30 is assumed to be the optimal computational model and is not necessarily the "actually optimal computational model".

[0087] (Prediction Phase)

[0088] The prediction performed using prediction model 30, i.e., step S30, will be described. Predictor 13 acquires one or more input feature values ​​for each of one or more molecules of interest. Each molecule of interest can be a molecule specified by a user of information processing system 10 through an input operation or a selection operation. Predictor 13 can read input feature values ​​from a predetermined database or receive input feature values ​​from another computer (such as a user terminal) for each molecule of interest. Predictor 13 inputs one or more input feature values ​​of the molecule of interest into prediction model 30. When one or more other feature values ​​are acquired in addition to one or more input feature values ​​of the molecule of interest, predictor 13 does not use the one or more other feature values ​​and inputs one or more input feature values ​​as input parameters (e.g., input vectors) into prediction model 30. Predictor model 30 calculates feature values ​​of the molecule of interest based on the input feature values, and predictor 13 acquires feature values ​​from prediction model 30. Therefore, predictor 13 uses prediction model 30 to acquire feature values ​​for each of one or more molecules of interest.

[0089] Predictor 13 can output the feature value of each of one or more molecules of interest as a prediction result. When obtaining feature values ​​for each of a plurality of molecules of interest, predictor 13 can select at least one of the plurality of molecules of interest based on the feature values. For example, predictor 13 selects molecules of interest whose feature values ​​satisfy a predetermined reference. Such selection can also be described as screening. Predictor 13 can output information about the selected at least one molecule of interest as a prediction result. The information may include at least one of the following: the name, structure, sequence information, and feature values ​​of the molecule of interest.

[0090] In step S30, predictor 13 can treat at least one of the plurality of second molecules as the molecule of interest. Predictor 13 can acquire one or more feature values ​​for each molecule of interest, and output a characteristic value for each molecule of interest obtained by inputting the acquired one or more input feature values ​​into the generated prediction model. When acquiring a characteristic value for each of the plurality of second molecules, predictor 13 can select at least one of the plurality of second molecules as a candidate molecule based on the characteristic value, and output information about the selected at least one candidate molecule.

[0091] The predictor 13 can store the prediction results in a storage device (such as auxiliary memory 103), display the prediction results on an output device 106, or send the prediction results to another computer (such as a user terminal).

[0092] (Methods for producing molecular compounds)

[0093] Based on the information about at least one molecule of interest or candidate molecule output during the prediction phase, a molecular compound having the molecular sequence of at least one molecule of interest or candidate molecule can be generated.

[0094] When the molecular compound is an antibody, the antibody can be produced by, for example, the recombinant method or configuration described in U.S. Patent No. 4,816,567. An example of a production method is a method for preparing an antibody, comprising culturing a host cell containing a nucleic acid encoding the antibody under conditions suitable for expressing an antibody that is a candidate molecular compound described herein, or collecting the antibody from the host cell (or host cell culture medium). The isolated nucleic acid encoding the antibody may encode an amino acid sequence of the VL containing the antibody and / or an amino acid sequence of the VH containing the antibody (e.g., the light chain and / or heavy chain of the antibody). The host cell containing such nucleic acid contains (1) a vector containing a nucleic acid encoding an amino acid sequence of the VL containing the antibody and an amino acid sequence of the VH containing the antibody, or (2) a first vector containing a nucleic acid encoding an amino acid sequence of the VL containing the antibody and a second vector containing a nucleic acid encoding an amino acid sequence of the VH containing the antibody (e.g., the host cell is transformed). In this example, the host cell is a eukaryotic cell (e.g., Chinese hamster ovary (CHO) cell) or a lymphocyte (e.g., Y0, NSO, Sp2 / O cell). Suitable host cells for cloning or expressing vectors encoding antibodies include prokaryotic or eukaryotic cells. For example, bacteria can be used to produce antibodies, especially when glycosylation and Fc effector function are not required. Antibody fragments and peptides can be expressed in bacteria, as per references such as US 5648237, US 5789199, and US 5840523. Furthermore, antibody fragments can be expressed in *E. coli*, as per reference “Charlton, Methods in Molecular Biology, Vol. 248 (BKC Lo, ed., Humana Press, Totowa, NJ, 2003), pp. 245-254”. After expression, the antibody can be separated from the bacterial cell paste into a soluble fraction or further purified.

[0095] When the molecular compound is a peptide or cyclic peptide, it can be prepared by liquid-phase synthesis, solid-phase synthesis involving Fmoc synthesis, Boc synthesis, etc., or combinations thereof. Solid-phase synthesis is a method in which a compound is bound to a solid and the compound reacts chemically with reagents on a solid resin to synthesize the compound of interest. Solid-phase synthesis of peptides is a method in which the desired amino acid or peptide is bound to a solid resin and conjugated with an amino acid or peptide bound to the solid resin, thereby sequentially linking the desired amino acid or peptide to extend the peptide chain, thus synthesizing the peptide. The desired peptide is obtained by isolating the peptide bound to the solid resin from the solid resin.

[0096] [molecular]

[0097] Molecules can be small molecules (low molecular weight compounds), medium molecules (medium molecular weight compounds), or large molecules (macromolecule compounds). In this disclosure, small molecules or low molecular weight compounds are compounds having a molecular weight of less than 500 g / mol. In this disclosure, medium molecular weight compounds are compounds having a molecular weight of 500 g / mol or greater but less than 30,000 g / mol. In this disclosure, macromolecules are compounds having a molecular weight of 30,000 g / mol or greater.

[0098] Molecules can be biological or non-biological molecules. They can be nucleic acids, peptides, cyclic peptides, proteins, or antigen-binding molecules such as antibodies, or molecules that bind to a target molecule (target-binding molecules). Molecules can be drug candidate molecules. When a molecule is a peptide, cyclic peptide, protein, or antibody, its building blocks are amino acids. When a molecule is a nucleic acid, its building blocks are nucleosides or nucleotides.

[0099] In this disclosure, the desired characteristics are those required for novel targets suitable for drug candidates and can be arbitrarily set. Examples of characteristics include, but are not limited to, the ability to bind to a predetermined in vivo target, pharmacological activity, physical properties, kinetics, and safety. Examples of physical properties include thermal stability, chemical stability, solubility, viscosity, photostability, long-term storage stability, nonspecific adsorption, lipophilicity, and membrane permeability. As an example, when the designed molecule is mRNA (messenger RNA), the characteristic is the ability to translate proteins. The molecule is a molecule that binds to a target molecule, such as an antigen-binding molecule, and the characteristic can be the ability to bind to the target molecule.

[0100] - Antigen-binding molecules

[0101] In this disclosure, the term "antigen-binding molecule containing an antigen-binding domain" is used in the broadest sense. Specifically, antigen-binding molecules include various molecular types, as long as they contain an antigen-binding domain. An antigen-binding molecule can be a molecule consisting solely of an antigen-binding domain, or a molecule containing an antigen-binding domain and another domain. If the antigen-binding molecule is a molecule formed by an antigen-binding domain and an Fc region bound together, examples of such molecules include complete antibodies and antibody fragments. Antibodies can include monoclonal antibodies (including agonist and antagonist antibodies), human antibodies, humanized antibodies, and chimeric antibodies. The antigen-binding molecules in this disclosure can include scaffold molecules, which are formed by using only a portion of the three-dimensional structure serving as the scaffold (such as existing stable α / β barrel protein structures) as a library for constructing the antigen-binding domain.

[0102] - Combined with competency assessment

[0103] There are no limitations on the techniques for assessing the ability of a target molecule-binding molecule to bind to a target molecule. The assessment of binding ability can be performed by quantitatively evaluating the binding of the target molecule-binding molecule to the target molecule. The target molecule is, for example, a target protein. The target molecule-binding molecule is, for example, an antigen-binding molecule, and the target molecule is, for example, an antigen. For example, when the target molecule is an antigen, the assessment can be made by measuring the binding activity of the antigen-binding molecule to the antigen. "Binding activity" is the total strength of the non-covalent interactions between one or more binding sites of a molecule (e.g., an antibody) and its binding partner (e.g., an antigen). Here, "binding activity" is not strictly limited to a 1:1 interaction between members of a binding pair (e.g., antibody and antigen). For example, when the members of a binding pair reflect a monovalent 1:1 interaction, binding activity means intrinsic binding affinity (sometimes simply referred to as "affinity"). If the members of a binding pair are capable of both monovalent and multivalent binding, then the binding activity is the sum of their binding forces. The binding activity of molecule X to its partner Y is typically expressed as a dissociation constant (KD) or "the amount of analyte bound per unit amount of ligand". For example, an octet value is one indicator of binding capacity and is measured as the amount of analyte bound per unit amount of ligand. Binding activity can be measured using any of the conventional methods known in the art, including those described herein. Those skilled in the art can appropriately determine conditions other than the concentration of compounds specific to target tissues.

[0104] In one mode, for antibody binding activity, a ligand capture method is used, such as BIACORE (R) T200 or BIACORE (R) 4000 (GE Healthcare, Uppsala, Sweden), which uses surface plasmon resonance analysis as the measurement principle.

[0105] In this mode, the BIACORE® evaluation software is used to analyze the measurement results. The calculation of kinetic parameters is performed by simultaneously fitting the sensor maps of association and dissociation using a 1:1 combined model. Through this process, the association rate (kon or ka), dissociation rate (koff or kd), and equilibrium dissociation constant (KD) can be calculated.

[0106] As a value for antigen-binding activity, KD (dissociation rate constant) can be used when the antigen is a soluble molecule, and apparent kd (apparent dissociation rate constant) can be used when the antigen is a membrane-associated molecule. Both kd (dissociation rate constant) and apparent KD (apparent dissociation rate constant) can be measured using methods known to those skilled in the art. For example, Biacore (GE Healthcare), flow cytometry, etc., can be used.

[0107] Another characterization model involves, for example, the selection of antigen-binding molecules using display libraries. One example of this model is panning using phage display. For instance, in affinity assessment, phages presenting antigen-binding molecules that interact with the target antigen can be concentrated through a process in which a phage library presenting multiple different antigen-binding molecules is prepared, the target antigen is contacted with the prepared phages, and then unbound phages are washed out. Sequences with affinity for the target antigen can be identified by analyzing the nucleic acid sequences encoding the antigen-binding molecules contained in the concentrated phages. Another example of this model is panning using mammalian cell display. In pharmacological activity assessments using display systems, cells containing genes for antigen-binding molecules with the desired pharmacological activity can be isolated using a process such as flow cytometry, in which a library containing multiple different antigen-binding molecules is expressed in the targeted mammalian cells, and the reporter gene activity is altered based on its effects on the cells. In physical property assessments using a display system, cells possessing genes for antigen-binding molecules that can be stably expressed at high levels can be isolated using a process such as flow cytometry. In this process, a library containing multiple different antigen-binding molecules is expressed in targeted mammalian cells, and the expression level is examined by staining with antibodies specific to that antigen-binding molecule. Characterization of antigen-binding molecules through panning is not limited to techniques using bacteriophages or mammalian cells; various techniques can be used as long as they allow for the presentation of antigen-binding molecules. Examples include techniques that allow ribosome presentation, techniques that allow mRNA presentation, techniques that allow viral presentation (other than bacteriophages), and techniques that allow bacterial presentation (such as *E. coli*).

[0108] Another characterization modality is, for example, obtaining antibody gene sequences from immune cells derived from an individual or obtaining antibody protein sequences from serum. In affinity assessments involving the extraction of antibody gene sequences from immune cells, sequences with affinity for the target antigen can be identified by processes involving: administering the target antigen protein to an individual to induce immune sensitization, and extracting the antibody gene, which binds to the target antigen, from immune cells possessing the gene.

[0109] For antigens that cause immune sensitization, techniques that utilize not only proteins but also genes encoding proteins or cells expressing proteins can be used.

[0110] Examples of individuals who may be subjects include, but are not limited to, humans, mice, rats, hamsters, rabbits, monkeys, chickens, camels, llamas, and alpacas.

[0111] Examples of techniques for analyzing nucleic acid sequence or appearance frequencies include, but are not limited to, techniques for cloning recombinant organisms with nucleic acid sequences having different antigen-binding molecules and analyzing them using capillary electrophoresis via the Sanger method and techniques for analysis using next-generation sequencers.

[0112] In nucleic acid sequence analysis, the strength of a property can be determined based on its frequency of occurrence. For example, in nucleic acid sequence analysis after concentration, the property of an antigen-binding molecule encoded by a sequence with a high frequency of occurrence can be assessed as high. On the other hand, the property of an antigen-binding molecule encoded by a sequence with a low frequency of occurrence after concentration can be assessed as lower than the property of an antigen-binding molecule encoded by a sequence with a high frequency of occurrence.

[0113] Techniques for obtaining information about antigen-binding molecules derived from display libraries or individuals are applicable to various types of characterization in addition to those mentioned above.

[0114] - Pharmacological activity assessment

[0115] Techniques for assessing molecular pharmacological activity are not limited to this. Pharmacological activity can be assessed, for example, by measuring the neutralizing, agonist, or cytotoxic activity exhibited by a molecule. Examples of assessments of cytotoxic activity as a type of pharmacological activity assessment include the assessment of antibody-dependent cell-mediated cytotoxicity (ADCC), complement-dependent cytotoxicity (CDC), T cell-dependent cytotoxicity (TDCC), and antibody-dependent phagocytic cell (ADCP) activity. CDC activity is cytotoxic activity due to the complement system. ADCC activity is activity in which an immune cell binds to the Fc region of an antigen-binding molecule containing an antigen-binding domain, which binds to a membrane-associated molecule expressed on the cell membrane of a target cell via an Fcγ receptor expressed on the immune cell, and the immune cell damages the target cell. TDCC activity is an activity in which T cells damage target cells by bringing them into close proximity using a bispecific antibody containing an antigen-binding domain that binds to a membrane-associated molecule expressed on the target cell membrane and an antigen-binding domain that targets any of the constituent subunits of the T cell receptor (TCR) complex on the T cell, particularly the CD3ε chain. Whether the antigen-binding molecule of interest possesses ADCC, CDC, TDCC, or ADCP activity can be determined by known methods.

[0116] Neutralizing activity is the activity of inhibiting a ligand that has biological activity against cells (e.g., viruses and toxins). In other words, a substance with neutralizing activity binds to a ligand or a receptor to which the ligand binds to inhibit the binding between the ligand and the receptor. Receptors whose binding to the ligand has been inhibited by neutralizing activity will be prohibited from exhibiting the biological activity mediated by that receptor. Neutralizing activity is not limited to the inhibition of ligand-receptor binding, and the activity of inhibiting the function of biologically active proteins is also understood as neutralizing activity. Examples of protein function include enzyme activity.

[0117] - Physical property assessment

[0118] As examples, the assessment of stability (such as thermal stability, chemical stability, light stability, stability to mechanical stimuli, and long-term storage stability) can be performed by measuring the decomposition, chemical modification, and association of molecules before and after treatments intended for stability assessment (such as heat treatment, exposure to low pH environments, exposure to light, mechanical stirring, and long-term storage). Examples of measurement methods used to assess stability include techniques involving chromatography, such as ion-exchange chromatography or size exclusion chromatography, mass spectrometry, and electrophoresis. Another measurement method may also be used.

[0119] Other examples of assessing physical properties include evaluating protein solubility via polyethylene glycol precipitation, evaluating viscosity via small-angle X-ray scattering, and assessing nonspecific binding based on evaluation of binding to the extracellular matrix (ECM).

[0120] Other examples of physical property assessments include the assessment of protein expression levels, the assessment of binding to resins or ligands used for purification, and the assessment of surface charge.

[0121] - Dynamics assessment

[0122] Molecular dynamics assessment can be performed by administering the molecule to an animal (such as a mouse, rat, monkey, or dog) and measuring the amount of the molecule in the blood over time after administration. Alternatively, kinetic assessment can be performed via pharmacokinetic (PK) evaluation. As a technique for directly assessing PK, the kinetic behavior of a molecule can be predicted from its amino acid sequence by using software to calculate the molecule's surface charge, isoelectric point, etc.

[0123] - Security Assessment

[0124] Examples of assessing molecular safety include immunogenicity prediction tools (such as ISPRI's network-based immunogenicity screening (EpiVax, Inc.)), HLA binding of fragment peptides of antigen-binding molecules, MAP (MHC-related peptide proteomics), and the detection of T cell epitopes and the assessment of immunogenicity related to T cell growth. Safety assessments can be performed as long as they can be measured using techniques such as those that assess immune responses and platelet aggregation by binding to rheumatoid factor (RF), PBMCs, or whole blood.

[0125] In this example, a drug discovery system including information processing system 10 provides a method for generating novel targets with predetermined properties, such as specific physiological activities (e.g., binding to specific proteins). Examples of drugs include potential activators, such as low-molecular-weight drugs, medium-molecular-weight drugs, biological agents, cellular agents, nucleic acid drugs, biopharmaceuticals, and other activators. Targets include molecular structures having desired or defined biological activities. Biological activity may, for example, preferentially bind to a specific protein over other proteins. Molecules that are drug candidates include biomolecules and compounds, including a variety of molecules such as nucleic acids, peptides, cyclic peptides, proteins, antibodies, target-binding molecules, high-molecular-weight compounds, medium-molecular-weight compounds, and low-molecular-weight compounds.

[0126] Drug discovery systems may include devices for selecting molecules that interact with a drug's target and devices for generating lead molecules. Drug discovery systems may be, for example, information processing systems that include the disclosures in WO 2020 / 246617.

[0127] Drug discovery systems include molecular design devices. These devices search for candidate molecules with desired properties and output information about the identified candidates. The drug discovery system uses this information from the molecular design devices to select new targets suitable for the drug candidates.

[0128] [Verification Example]

[0129] (First verification example)

[0130] The following will describe a verification example of the information processing system according to this disclosure.

[0131] In a first validation example, a dataset representing measurements of the binding affinity of each of a plurality of antibody sequences is used to validate the usefulness of the information processing system according to this disclosure. The dataset includes sequences of bispecific antibodies that recognize MarvelD3 and CD3 as antigens, along with characteristic values ​​of their binding affinity. The subject of the first validation example is improving the binding affinity of antibodies to MarvelD3, which is presumed to be an antigen. Binary eight-bit values ​​are used as characteristic values ​​of the binding affinity.

[0132] MarvelD3 is a tight junction protein with a four-transmembrane structure. In the first validation instance, MarvelD3 was designated as a target candidate for an anticancer drug. The development of bispecific antibodies bridging cancer antigens and T-cell antigens holds promise for cancer therapy. In the first validation instance, a learned predictive model was used when identifying anti-MarvelD3 candidates with better properties from lead antibodies.

[0133] In the first validation instance, the octet values ​​of the antibody sequences were measured multiple times, and the antibody sequences and their octet values ​​obtained through measurement were used as input to the prediction model. Batch measurements were performed to obtain the binary octet values ​​of typically 100 or fewer antibody sequences in a single measurement.

[0134] The antibody used in the assay was obtained through the following procedure. First, a plasmid encoding a pre-designed heavy or light chain was provided, and the recombinant antibody was transiently expressed using Expi293F cells. The antibody was removed from the culture supernatant using protein A and dissolved in a buffer solution. The dissolved buffer solution was mixed under reducing conditions to prepare the MarvelD3 / CD3 bispecific antibody. In the preparation, charge repulsion was applied between identical heavy chains to achieve selective heavy chain heterodimerization. The concentration of the antibody in the buffer solution was determined by absorbance at 280 nm. Subsequently, the buffer solution containing the bispecific antibody was subjected to ion-exchange chromatography, confirming the provision of the expected MarvelD3 / CD3 antibody.

[0135] The Octet HTX system is used for octet measurements. Extracellular vesicles bearing the CD81 protein and human Marvel D3 on their surface are captured on a sensor chip using an anti-CD81 antibody. After a 600-second baseline step in D-PBS(-) solution containing 0.1% BSA, association and dissociation responses are measured, respectively, for 900 seconds and 1,500 seconds in the same buffer containing 20 nM antibody. Antibody binding capacity is expressed as a wavelength change between the baseline step and the end of the association phase. Measurements are performed at 30°C and a vibration rate of 1,000 Hz during the baseline step, association, and dissociation phases.

[0136] The specific method is as follows. First, 1,144 antibody sequences were extracted as a training sequence set. Next, training data representing the training sequence set was used to prepare a property prediction model for predicting binding affinity. A virtual sequence set containing 196,608 virtual sequences was input into the property prediction model as the prediction sequence set. This virtual sequence set was prepared from candidate combinations for amino acid changes and did not have measurement values. The prediction sequence set for prediction was screened based on the prediction values ​​of the property prediction model for each antibody sequence contained in the prediction sequence set, and it was verified whether antibody sequences with binding affinity could be selected.

[0137] In the preparation of the characteristic prediction model, the protein language model TAPE was used to convert each antibody sequence into a 768-dimensional vector for each of the VH and VL regions, and the 1,536-dimensional vector data obtained through vector combination was defined as the feature value of each antibody sequence.

[0138] To verify the usefulness of the invention, input feature values ​​for use in generating a predictive model were selected from 1536 dimensional feature values ​​using two different methods. When constructing a feature prediction model using the selected feature values, the binding affinity levels of the antibody sequences selected by the model were compared. In the first method, LightGBM (Optical Gradient Boosting Machine) was used. In the first method, a classification model was prepared to distinguish between the sequence sets used for training and the sequence sets used for prediction, and the importance of each dimension among the 1536 dimensional feature values ​​was calculated based on the feature importance obtained from the model. In the first method, feature values ​​were selected such that the number of features with low importance (i.e., features that estimate small differences between the sequence sets used for training and those used for prediction) was increased by a factor of 10 x 10. In the second method, the interdistribution distance was defined by the Wasserstein distance of each feature value in each dimension of the sequence sets used for training and those used for prediction. In the second method, feature values ​​were selected such that the number of features with small Wasserstein distances (i.e., features with relatively consistent distributions between the sequence sets used for training and those used for prediction) was increased sequentially. Cross-validation using mean squared error as the metric is applied to each feature set progressively selected by each of the first and second methods to determine the feature set with the highest prediction performance for the sequence groups used for training. The performance of property predictions for binding forces using LightGBM is then evaluated using these feature sets.

[0139] Figure 10 illustrates the distribution of measurements of antibody sequences selected as the top 16 antibody sequences when selected from a group of sequences used for prediction in descending order of predicted binding strength for each of the following: where no feature selection is performed and all feature values ​​are used as input feature values ​​(baseline); where feature values ​​are selected using feature importance obtained from a class classification model (classifier); and where feature values ​​are selected using Wasserstein distance. Figure 10 shows that feature values ​​selected using Wasserstein distance can screen for antibody sequences with higher binding strength measurements compared to using all feature values ​​and using feature values ​​selected using feature importance. In other words, Figure 10 demonstrates the usefulness of the information processing system (feature selection technique) according to this disclosure.

[0140] (Second verification example)

[0141] Next, a second verification example will be described. In the second verification example, the usefulness of the information processing system according to this disclosure of AqSoIDB is utilized. AqSoIDB provides a dataset of solubility for 9,982 low-molecular-weight compounds, which is downloaded using Therapeutics Data Common, a platform for providing the ability to download public datasets. Using 1,000 molecules randomly extracted from all compounds as the first molecular group, a property prediction model for predicting solubility is prepared. Next, using the remaining 8,982 molecules as the second molecular group, the second molecular group is selected based on the prediction values ​​of the property prediction model. Subsequently, it is verified whether molecules with high solubility measurements can be selected. For verification, the number of random types is set to 10 when randomly extracting molecules, and 10 verifications are performed. In the verification, feature values ​​are selected according to step S10A. In step S11, training data and test data including 2,048-dimensional feature values ​​extracted using MolCLR (Molecular Contrastive Learning of Representations via Graph Neural Networks) are obtained. In steps S12 and S14A, the Wasserstein distance and the feature importance of the model used for classification between the first and second molecular groups are used, respectively, and the two types of selection metrics are compared. In step S13A, the initial value of the number of feature selections, m, is set to 18. In step S15A, cross-validation is used as the method for evaluating the performance of the provisional prediction model, mean squared error is used as the performance evaluation metric, and LightGBM (Optical Gradient Boosting Machine) is used as the supervised learning technique. In step S16A, if the number of feature selections, m, is less than 2,048, the process continues to step S17A, and if the number of selections, m, is 2,048, the process continues to step S18. In step S17A, the number of feature selections, m, is incremented by 10. In step S18, the set of feature values ​​with the smallest mean squared error of the provisional prediction model is selected as the input feature values.

[0142] Figure 11 shows the distribution of the average values ​​of the measurements of the top 100 compounds selected when choosing from a descending order of predicted solubility values ​​for each of the following: where no eigenvalue selection is performed and all eigenvalues ​​are used as input eigenvalues ​​(baseline); where eigenvalues ​​are selected using eigenvalue importance as the selection criterion (classifier); and where eigenvalues ​​are selected using Wasserstein distance as the selection criterion (Wasserstein). Figure 11 shows that using eigenvalues ​​selected using Wasserstein distance can screen for compounds with higher solubility measurements compared to using all eigenvalues ​​and using eigenvalues ​​selected using eigenvalue importance. In other words, Figure 11 demonstrates the usefulness of the information processing system (eigenvalue selection technique) according to this disclosure.

[0143] [Variation]

[0144] The technology according to this disclosure has been described in detail above based on various embodiments thereof. However, this disclosure is not limited to the examples described above. Various modifications can be made to the technology according to this disclosure without departing from the spirit of this disclosure.

[0145] In the above example, the information processing system 10 includes a learner 12 and a predictor 13 in addition to the feature selector 11; however, the information processing system may not have at least one of the learner and predictor. The prediction model is portable between computer systems. Therefore, the prediction model generated by the information processing system can be used in another computer system. Alternatively, another computer system can use one or more input feature values ​​selected by the information processing system to generate a prediction model through machine learning, where the information processing system uses the prediction model to perform the prediction phase.

[0146] The information processing system can be built as a server in a client-server system or implemented in a standalone computer. Alternatively, the information processing system can be implemented in a user terminal that can access a database storing training and test data via a communication network.

[0147] The process of a method executed by at least one processor is not limited to the examples described above. For example, a step or part of the process described above may be omitted, and these steps may be executed in a different order. Any two or more of the steps described above may be combined, and a part of the steps may be modified or deleted. Alternatively, other steps may be performed in addition to those described above.

[0148] When comparing the magnitudes of two values ​​in this disclosure, either of the two criteria (i.e., "more than" and "more than") can be used, and either of the two criteria (i.e., "or less than" and "less than") can be used.

[0149] In this disclosure, the phrase "at least one processor executes a first process, executes a second process, ... and executes an nth process," or equivalent wording, indicates a concept in which the execution of the n processes from the first to the nth process by the processor changes in between. That is, such wording includes both the concept of all n processes being executed by the same processor and the concept of changes occurring among the n processes due to any strategy.

[0150] In this disclosure, the term "to" indicates a range that includes the values ​​at both ends. For example, "A to B" means a range of A or greater and B or less.

[0151] In this disclosure, the term “about” when used in conjunction with a numerical value means the range between +10% and -10% of that value.

[0152] In this specification, the term “and / or” is used to indicate the items described before and after “and / or”, or any combination thereof. For example, “A, B and / or C” includes the subjects “A”, “B” and “C”, as well as the combinations “A and B”, “A and C”, “B and C” and “A and B and C”.

[0153] [Replenish]

[0154] As can be understood from the various examples above, this disclosure includes the following aspects.

[0155] (Supplement 1)

[0156] An information processing system includes at least one processor, wherein the at least one processor performs the following processing:

[0157] For each of the multiple first molecules, obtain training data representing multiple feature values ​​associated with that first molecule;

[0158] For each of the multiple second molecules, acquire test data representing multiple feature values ​​associated with that second molecule;

[0159] For each of the plurality of feature values, an index is calculated based on a first probability distribution and a second probability distribution, wherein the first probability distribution is the probability distribution in the training data and the second probability distribution is the probability distribution in the test data; and

[0160] Based on the index of each of the plurality of feature values, one or more of the plurality of feature values ​​are selected as one or more input feature values ​​to be used as input parameters for a predictive model for predicting the property values ​​of molecules based on machine learning.

[0161] (Supplement 2)

[0162] According to the information processing system described in Supplement 1, the at least one processor performs the following processing:

[0163] For each of the plurality of eigenvalues, the inter-distribution distance is calculated as the index, wherein the inter-distribution distance is the distance between the first probability distribution and the second probability distribution; and

[0164] One or more input feature values ​​are selected from the plurality of feature values ​​based on the inter-distribution distance for each of the plurality of feature values.

[0165] (Supplement 3)

[0166] According to the information processing system described in Supplement 2, the at least one processor selects one or more feature values ​​in which the distance between the distributions is less than a predetermined reference as the one or more input feature values.

[0167] (Supplement 4)

[0168] According to the information processing system described in Supplement 2 or 3, the at least one processor calculates the Wasserstein distance as the distance between the distributions.

[0169] (Supplement 5)

[0170] According to any one of Supplements 2 to 4, the information processing system wherein the at least one processor performs the following processing:

[0171] Based on the inter-distribution distance for each of the plurality of feature values, a portion of the plurality of feature values ​​are excluded, and the remaining one or more feature values ​​are selected as the one or more input feature values.

[0172] (Supplement 6)

[0173] According to the information processing system described in Supplement 5, the at least one processor performs the following processing:

[0174] The evaluation process involves altering the number of feature values ​​excluded from the plurality of feature values ​​based on the inter-distribution distance for each of the feature values, while simultaneously generating a provisional prediction model using the training data through the machine learning process and calculating an evaluation metric for the provisional prediction model based on the remaining one or more feature values.

[0175] Based on multiple evaluation indicators, the number of feature values ​​to be excluded from these feature values ​​is determined as the exclusion quantity; and

[0176] The number of feature values ​​to be excluded is excluded based on the inter-distribution distance for each of the plurality of feature values, and one or more input feature values ​​are selected.

[0177] (Supplement 7)

[0178] According to the information processing system described in Supplement 6, the plurality of feature values ​​are N feature values, where N is a natural number greater than 2, and

[0179] The at least one processor performs the following processing:

[0180] The number of eigenvalues ​​to be excluded from these N eigenvalues ​​is set to n, where n is a natural number;

[0181] Based on the inter-distribution distance for each of the N eigenvalues, n eigenvalues ​​are excluded from the plurality of eigenvalues;

[0182] An evaluation process is performed, in which a first provisional prediction model is generated using the training data through the machine learning based on (Nn) feature values, and an evaluation metric for the first provisional prediction model is calculated.

[0183] Further, based on the inter-distribution distance for each of the (Nn) eigenvalues, n eigenvalues ​​are excluded from the (Nn) eigenvalues; and

[0184] An evaluation process is performed, in which a second provisional prediction model is generated using the training data through the machine learning based on (N-2n) feature values, and an evaluation metric for the second provisional prediction model is calculated.

[0185] (Supplement 8)

[0186] According to the information processing system of Supplement 6 or 7, the at least one processor determines the number of times the highest evaluation index is obtained as the exclusion number.

[0187] (Supplement 9)

[0188] According to the information processing system of Supplement 6 or 7, the at least one processor determines the number of times the lowest evaluation index is obtained as the exclusion number.

[0189] (Supplement 10)

[0190] According to the information processing system described in Supplement 5,

[0191] The at least one processor performs the following processing:

[0192] The evaluation process involves altering the number of feature values ​​selected from the plurality of feature values ​​based on the inter-distribution distance for each of the feature values, while simultaneously generating a provisional prediction model using the training data through the machine learning process based on one or more selected feature values, and calculating an evaluation metric for the provisional prediction model.

[0193] Based on multiple evaluation indicators, the number of feature values ​​selected from these multiple feature values ​​is determined as the selection quantity; and

[0194] Based on the inter-distribution distance for each of the plurality of feature values, the selected number of feature values ​​are chosen as the one or more input feature values.

[0195] (Supplement 11)

[0196] According to the information processing system described in Supplement 10,

[0197] These multiple feature values ​​are N feature values, where N is a natural number greater than 2.

[0198] The at least one processor performs the following processing:

[0199] The number of feature values ​​to be selected from these N feature values ​​is set to m, where m is a natural number;

[0200] Based on the inter-distribution distance for each of the N feature values, select m feature values ​​from the plurality of feature values;

[0201] An evaluation process is performed, in which a first provisional prediction model is generated using the training data through the machine learning based on m feature values, and an evaluation metric for the first provisional prediction model is calculated.

[0202] Further, based on the inter-distribution distance for each of the m feature values, add the m feature values ​​to the m feature values; and

[0203] An evaluation process is performed, in which a second provisional prediction model is generated using the training data through the machine learning based on 2m feature values, and an evaluation metric for the second provisional prediction model is calculated.

[0204] (Supplement 12)

[0205] According to the information processing system of Supplement 11, the at least one processor determines the number of times the highest evaluation index is obtained as the selection number.

[0206] (Supplement 13)

[0207] According to the information processing system of Supplement 11, the at least one processor determines the number of times the minimum evaluation index is obtained as the selection number.

[0208] (Supplement 14)

[0209] According to any one of Supplements 6 to 13, the information processing system wherein the at least one processor performs the evaluation processing through cross-validation.

[0210] (Supplement 15)

[0211] According to any one of Supplements 1 to 14, the information processing system wherein the at least one processor performs the machine learning to generate the predictive model by using the one or more input feature values ​​of the training data.

[0212] (Supplement 16)

[0213] According to the information processing system described in Supplement 15, the at least one processor performs the following processing:

[0214] Obtain one or more input feature values ​​for a molecule of interest, where the properties of the molecule are unknown; and

[0215] Output the characteristic value of the molecule of interest, which is obtained by inputting one or more of the acquired input feature values ​​into the generated prediction model.

[0216] (Supplement 17)

[0217] According to the information processing system described in Supplement 15, the at least one processor performs the following processing:

[0218] Obtain one or more input feature values ​​for each of a plurality of molecules of interest, wherein the feature values ​​of each of the plurality of molecules of interest are unknown;

[0219] One or more input feature values ​​obtained for each of the plurality of molecules of interest are input into the generated prediction model, and the feature value of the molecule of interest is obtained from the prediction model;

[0220] Based on the characteristic value of each of the plurality of molecules of interest, at least one molecule of interest is selected from the plurality of molecules of interest; and

[0221] Output information about at least one selected molecule of interest.

[0222] (Supplement 18)

[0223] According to any one of Supplements 1 to 17, each of the first molecule and the second molecule is selected from nucleic acids, peptides, cyclic peptides, proteins, antibodies, and low molecular weight compounds.

[0224] (Supplement 19)

[0225] According to the information processing system described in Supplement 15, the at least one processor performs the following processing:

[0226] For each of at least one of the plurality of second molecules, obtain the one or more input feature values; and

[0227] For each of the at least one second molecule, output the characteristic value of that second molecule obtained by inputting one or more of the acquired input feature values ​​into the generated prediction model.

[0228] (Supplement 20)

[0229] According to the information processing system described in Supplement 15, the at least one processor performs the following processing:

[0230] For each of the plurality of second molecules, obtain one or more input feature values;

[0231] The acquired one or more input feature values ​​are input into a prediction model generated for each of the plurality of second molecules, and the feature value of the molecule of interest is obtained from the prediction model.

[0232] Based on the characteristic value of each of the plurality of second molecules, at least one second molecule is selected as a candidate molecule from the plurality of second molecules; and

[0233] Output information about at least one selected candidate molecule.

[0234] (Supplement 21)

[0235] According to any one of Supplements 1 to 20, the information processing system wherein the characteristic value is selected from at least one of affinity, pharmacological activity, physical properties, kinetics, and safety.

[0236] (Supplement 22)

[0237] According to any one of Supplements 1 to 20, the information processing system

[0238] The first molecule and the second molecule are antigen-binding molecules, and

[0239] This characteristic value represents the ability of the antigen-binding molecule to bind to the antigen.

[0240] (Supplement 23)

[0241] An information processing method, performed by an information processing system including at least one processor, includes the following steps:

[0242] For each of the multiple first molecules, obtain training data representing multiple feature values ​​associated with that first molecule;

[0243] For each of the plurality of second molecules, test data representing the plurality of feature values ​​associated with that second molecule is obtained;

[0244] For each of the plurality of feature values, an index is calculated based on a first probability distribution and a second probability distribution, wherein the first probability distribution is the probability distribution in the training data and the second probability distribution is the probability distribution in the test data; and

[0245] Based on the index of each of the plurality of feature values, one or more of the plurality of feature values ​​are selected as one or more input feature values ​​to be used as input parameters for a predictive model for predicting molecular property values ​​based on machine learning.

[0246] (Supplement 24)

[0247] An information processing program that enables a computer to perform the following steps:

[0248] For each of the multiple first molecules, obtain training data representing multiple feature values ​​associated with that first molecule;

[0249] For each of the plurality of second molecules, test data representing the plurality of feature values ​​associated with that second molecule is obtained;

[0250] For each of the plurality of feature values, an index is calculated based on a first probability distribution and a second probability distribution, wherein the first probability distribution is the probability distribution in the training data and the second probability distribution is the probability distribution in the test data; and

[0251] Based on the index of each of the plurality of feature values, one or more of the plurality of feature values ​​are selected as one or more input feature values ​​to be used as input parameters for a predictive model for predicting molecular property values ​​based on machine learning.

[0252] (Supplement 25)

[0253] A method for generating a molecular compound, the method comprising a generation step of generating a molecular compound having a molecular sequence of the at least one molecule of interest based on information output from an information processing system according to Supplement 17.

[0254] (Supplement 26)

[0255] A method for generating a molecular compound, the method comprising a generation step of generating a molecular compound having a molecular sequence of at least one candidate molecule based on information output from an information processing system according to Supplement 20.

[0256] (Supplement 27)

[0257] A non-transitory, computer-readable recording medium configured to store an information processing program that enables a computer to perform the following steps:

[0258] For each of the multiple first molecules, obtain training data representing multiple feature values ​​associated with that first molecule;

[0259] For each of the plurality of second molecules, test data representing the plurality of feature values ​​associated with that second molecule is obtained;

[0260] For each of the plurality of feature values, an index is calculated based on a first probability distribution and a second probability distribution, wherein the first probability distribution is the probability distribution in the training data and the second probability distribution is the probability distribution in the test data; and

[0261] Based on the index of each of the plurality of feature values, one or more of the plurality of feature values ​​are selected as one or more input feature values ​​to be used as input parameters for a predictive model for predicting molecular property values ​​based on machine learning.

[0262] As described in Supplements 1, 23, 24, and 27, the feature values ​​to be used as input features for the machine learning-based prediction model are selected based on the trend of each feature value in both the training and test data. By selecting feature values ​​as described above, the accuracy of the machine learning model (prediction model) used to predict molecular properties can be improved. Improving the accuracy of the machine learning model allows for better prediction of molecular properties.

[0263] According to Supplement 2, feature values ​​are selected based on the inter-distribution distance between the first probability distribution and the second probability distribution. By introducing the inter-distribution distance, which represents the quantified difference in trend between the two probability distributions, input feature values ​​can be selected based on objective criteria. Therefore, further improvements in the accuracy of the machine learning model can be expected.

[0264] According to Supplement 3, feature values ​​with relatively small inter-distribution distances are selected as input feature values. As mentioned above, the selected input feature values ​​show relatively consistent trends between the training and test data. Therefore, machine learning based on these input feature values ​​can improve the accuracy of machine learning models (predictive models).

[0265] According to Supplement 4, the inter-distribution distance can be appropriately calculated by introducing the Wasserstein distance as the inter-distribution distance. Therefore, the input feature values ​​can be selected more appropriately.

[0266] According to Supplement 5, input features can be selected by focusing attention on features that are evaluated as unsuitable as input features.

[0267] According to Supplement 6, the generation and evaluation of the provisional prediction model are repeated, while varying the number of excluded features, and dynamically determining this number based on multiple evaluation metrics obtained through the repetition. The determined number of excluded features is used to select input features. In the example, if an excessive number of features with relatively large inter-distribution distances are excluded, even features that actually contribute to predicting the feature values ​​may be excluded. By dynamically determining this number of exclusions as described above, one or more input features can be appropriately selected to improve the accuracy of the machine learning model.

[0268] According to Supplement 7, the percentage increase in the number of exclusions is constant in the repetition of this evaluation treatment, and therefore the evaluation treatment can be efficiently repeated using convenient techniques.

[0269] According to Supplement 8, the number of features that receive the highest evaluation metric value is determined as the final exclusion number. According to Supplement 9, the number of features that receive the lowest evaluation metric value is determined as the final exclusion number. Therefore, one or more input features can be selected, with the expectation that these one or more input features will give the machine learning model the highest accuracy.

[0270] According to Supplement 10, the generation and evaluation of the provisional prediction model are repeated, while varying the number of selected feature values ​​and dynamically determining the number of selections based on multiple evaluation metrics obtained through the repetition. The determined number of selected feature values ​​is then used as input. In the example, if too many feature values ​​with relatively large inter-distribution distances are selected, even a number of feature values ​​that actually contribute to predicting the feature values ​​can be chosen. By dynamically determining the number of selections as described above, one or more input feature values ​​can be appropriately selected to improve the accuracy of the machine learning model.

[0271] According to Supplement 11, in the repetition of this evaluation process, the percentage increase in the number of selections is constant, and therefore the evaluation process can be efficiently repeated using convenient techniques.

[0272] According to Supplement 12, the number of feature values ​​selected to achieve the highest evaluation metric is determined as the final selection number. According to Supplement 13, the number of feature values ​​selected to achieve the lowest evaluation metric is determined as the final selection number. Therefore, one or more input feature values ​​can be selected, with the expectation that these one or more input feature values ​​will give the machine learning model the highest accuracy.

[0273] According to Supplement 14, cross-validation is introduced into the evaluation process used to obtain the evaluation metrics, and thus, even when the amount of training data is limited, each provisional prediction model can be accurately evaluated by making efficient use of the limited data.

[0274] According to Supplement 15, machine learning is performed to generate a predictive model using one or more input feature values ​​selected based on the trend of each of the feature values ​​in both the training and test data. Therefore, a predictive model with good accuracy can be obtained.

[0275] According to Supplement 16, the properties of molecules of interest can be predicted with good accuracy by using a predictive model generated by machine learning using one or more selected input feature values.

[0276] According to Supplement 17, the properties of each molecule of interest can be predicted with good accuracy using a predictive model generated by machine learning using one or more selected input feature values. Therefore, at least one molecule of interest can be appropriately selected from a plurality of molecules of interest.

[0277] According to Supplement 18, the accuracy of machine learning models (predictive models) used to predict the properties of nucleic acids, peptides, cyclic peptides, proteins, antibodies, and low molecular weight compounds can be improved.

[0278] According to Supplement 19, the properties of the second molecule can be predicted with good accuracy by using a predictive model generated by machine learning using one or more selected input feature values.

[0279] According to Supplement 20, the properties of each second molecule can be predicted with good accuracy using a predictive model generated by machine learning using one or more selected input feature values. Therefore, at least one candidate molecule can be appropriately selected from multiple candidate molecules.

[0280] According to Supplement 21, the accuracy of machine learning models (predictive models) used to predict the affinity, pharmacological activity, physical properties, kinetics, or safety of molecules can be improved.

[0281] According to Supplement 22, the accuracy of machine learning models (predictive models) used to predict the ability of antigen-binding molecules to bind to antigens.

[0282] According to Supplement 25, molecular compounds that can be expected to have desired properties can be generated based on information about at least one molecule of interest appropriately selected using a predictive model.

[0283] According to Supplement 26, molecular compounds that can be expected to have desired properties can be generated based on information about at least one candidate molecule appropriately selected using a predictive model.

[0284] [List of Reference Symbols]

[0285] 10……Information processing system, 11……Feature value selector, 12……Learner, 13……Predictor, 20……Database, 30……Predictive model, 110……Information processing program, 210……First probability distribution, 220……Second probability distribution.

Claims

1. An information processing system comprising at least one processor, The at least one processor performs the following processing: For each of the plurality of first molecules, training data representing multiple feature values ​​associated with the first molecule is obtained; For each of the plurality of second molecules, test data representing multiple feature values ​​associated with the second molecule are obtained; For each of the plurality of feature values, an index is calculated based on a first probability distribution and a second probability distribution, wherein the first probability distribution is the probability distribution in the training data and the second probability distribution is the probability distribution in the test data; as well as Based on the index of each of the plurality of feature values, one or more of the plurality of feature values ​​are selected as one or more input feature values ​​to be used as input parameters for a predictive model for predicting molecular property values ​​based on machine learning.

2. The information processing system according to claim 1, The at least one processor performs the following processing: For each of the plurality of feature values, the inter-distribution distance is calculated as the index, wherein the inter-distribution distance is the distance between the first probability distribution and the second probability distribution; and One or more input feature values ​​are selected from the plurality of feature values ​​based on the inter-distribution distance for each of the plurality of feature values.

3. The information processing system according to claim 2, wherein the at least one processor selects one or more feature values ​​in which the distance between the distributions is less than a predetermined reference as the one or more input feature values.

4. The information processing system according to claim 2 or 3, wherein the at least one processor calculates the Wasserstein distance as the inter-distribution distance.

5. The information processing system according to any one of claims 2 to 4, The at least one processor performs the following processing: A subset of the plurality of feature values ​​is excluded based on the inter-distribution distance for each of the plurality of feature values, and the remaining one or more feature values ​​are selected as the one or more input feature values.

6. The information processing system according to claim 5, The at least one processor performs the following processing: The evaluation process involves altering the number of feature values ​​excluded from the plurality of feature values ​​based on the inter-distribution distance for each of the plurality of feature values, while simultaneously generating a provisional prediction model using the training data through the machine learning process and calculating an evaluation metric for the provisional prediction model based on the remaining one or more feature values. Based on multiple evaluation indicators, the number of feature values ​​to be excluded from the multiple feature values ​​is determined as the exclusion quantity; as well as The number of feature values ​​to be excluded is excluded from the plurality of feature values ​​based on the inter-distribution distance for each of the plurality of feature values, and one or more input feature values ​​are selected.

7. The information processing system according to claim 6, The plurality of feature values ​​are N feature values, where N is a natural number greater than 2, and The at least one processor performs the following processing: The number of feature values ​​to be excluded from the N feature values ​​is set to n, where n is a natural number; Based on the inter-distribution distance for each of the N feature values, n feature values ​​are excluded from the plurality of feature values; An evaluation process is performed, in which a first provisional prediction model is generated using the training data through the machine learning based on (Nn) feature values, and an evaluation metric for the first provisional prediction model is calculated. Further, based on the inter-distribution distance for each of the (Nn) eigenvalues, n eigenvalues ​​are excluded from the (Nn) eigenvalues; as well as An evaluation process is performed, in which a second provisional prediction model is generated using the training data through the machine learning based on (N-2n) feature values, and an evaluation metric for the second provisional prediction model is calculated.

8. The information processing system according to claim 6 or 7, wherein the at least one processor determines the number of times the highest evaluation index is obtained as the exclusion quantity, or determines the number of times the lowest evaluation index is obtained as the exclusion quantity.

9. The information processing system according to any one of claims 2 to 4, The at least one processor performs the following processing: The evaluation process involves altering the number of feature values ​​selected from the plurality of feature values ​​based on the inter-distribution distance for each of the plurality of feature values, while simultaneously generating a provisional prediction model using the training data through the machine learning process based on one or more selected feature values, and calculating an evaluation metric for the provisional prediction model. Based on multiple evaluation indicators, the number of feature values ​​selected from the multiple feature values ​​is determined as the selection quantity; as well as Based on the inter-distribution distance for each of the plurality of feature values, the selected number of feature values ​​are chosen as the one or more input feature values.

10. The information processing system according to any one of claims 6 to 9, wherein the at least one processor performs the evaluation processing via cross-validation.

11. The information processing system according to any one of claims 1 to 10, wherein the at least one processor generates the predictive model by performing the machine learning using one or more input feature values ​​of the training data.

12. The information processing system according to claim 11, The at least one processor performs the following processing: Obtain one or more input feature values ​​for a molecule of interest, wherein the characteristic values ​​of the molecule of interest are unknown; and The characteristic value of the molecule of interest is output, which is obtained by inputting one or more acquired input feature values ​​into the generated prediction model.

13. The information processing system according to claim 11, The at least one processor performs the following processing: Obtain one or more input feature values ​​for each of a plurality of molecules of interest, wherein the feature values ​​of each of the plurality of molecules of interest are unknown; One or more input feature values ​​obtained for each of the plurality of molecules of interest are input into the generated prediction model, and the characteristic values ​​of the molecules of interest are obtained from the prediction model; At least one molecule of interest is selected from the plurality of molecules of interest based on the characteristic value of each of the plurality of molecules of interest; as well as Output information about at least one selected molecule of interest.

14. The information processing system according to any one of claims 1 to 13, wherein each of the first molecule and the second molecule is selected from nucleic acids, peptides, cyclic peptides, proteins, antibodies and low molecular weight compounds.

15. An information processing method, executed by an information processing system including at least one processor, the information processing method comprising the following steps: For each of the plurality of first molecules, training data representing multiple feature values ​​associated with the first molecule is obtained; For each of the plurality of second molecules, test data representing multiple feature values ​​associated with the second molecule are obtained; For each of the plurality of feature values, an index is calculated based on a first probability distribution and a second probability distribution, wherein the first probability distribution is the probability distribution in the training data and the second probability distribution is the probability distribution in the test data; as well as Based on the index of each of the plurality of feature values, one or more of the plurality of feature values ​​are selected as one or more input feature values ​​to be used as input parameters for a predictive model for predicting molecular property values ​​based on machine learning.

16. An information processing program that enables a computer to perform the following steps: For each of the plurality of first molecules, training data representing multiple feature values ​​associated with the first molecule is obtained; For each of the plurality of second molecules, test data representing multiple feature values ​​associated with the second molecule are obtained; For each of the plurality of feature values, an index is calculated based on a first probability distribution and a second probability distribution, wherein the first probability distribution is the probability distribution in the training data and the second probability distribution is the probability distribution in the test data; as well as Based on the index of each of the plurality of feature values, one or more of the plurality of feature values ​​are selected as one or more input feature values ​​to be used as input parameters for a predictive model for predicting molecular property values ​​based on machine learning.

17. A method for manufacturing a molecular compound, the method comprising: The step of generating a molecular compound having the molecular sequence of the at least one molecule of interest based on the information output from the information processing system of claim 13.

Images