Information processing device, operation method for information processing device, and operation program for information processing device

Abstract

This information processing device predicts, using a prediction model, an evaluation result of the Ames mutagenicity test for a candidate substance of a product. The information processing device comprises a processor that uses, as a prediction model, a first prediction model and a second prediction model, and performs acquiring candidate substance information related to the candidate substance, inputting the candidate substance information or information based on the candidate substance information to the first prediction model and the second prediction model, causing the first prediction model and the second prediction model to output a first prediction evaluation result and a second prediction evaluation result, and presenting prediction information corresponding to the first prediction evaluation result and the second prediction evaluation result to a user, the first prediction model outputting the first prediction evaluation result indicating whether or not there is a mutagenicity related to a base-to-substitution type mutation, in which a part of a base sequence changes, in the candidate substance when the candidate substance is added to a first type strain having sensitivity to the base-to-substitution type mutation, and the second prediction model outputting the second prediction evaluation result indicating whether or not there is a mutagenicity related to a frame-shift type mutation, in which a reading frame of a base sequence is deviated, in the candidate substance when the candidate substance is added to a second type strain having sensitivity to the frame-shift type mutation.

Description

Information processing device, operating method for information processing device, and operating program for information processing device

[0001] The technology of the present disclosure relates to an information processing device, an operating method for an information processing device, and an operating program for an information processing device.

[0002] The Ames mutagenicity test (also known as the reverse mutation test) is widely used in fields such as drug discovery. The Ames mutagenicity test is a test to evaluate whether candidate substances for pharmaceutical products and other products have mutagenicity. Mutagenicity is the property of causing irreversible mutations in genes, and is one of the factors that induce cancer in cells. In the Ames mutagenicity test, a candidate substance is added to bacteria such as Salmonella typhimurium, and the subsequent bacterial growth is used to evaluate whether the candidate substance has mutagenicity (positive) or not (negative).

[0003] Gene mutations include base pair substitution mutations, in which a portion of the base sequence is changed, and frameshift mutations, in which the reading frame of a triplet of base sequences is shifted due to the insertion or deletion of bases. For example, the Ames mutagenicity test recommends the use of three strains, TA100, TA1535, and WP2uvrA, which are sensitive to base pair substitution mutations, and two strains, TA98 and TA1537, which are sensitive to frameshift mutations. If any one of these five strains is positive, the candidate substance is rejected as mutagenic.

[0004] The Ames mutagenicity test has various problems, such as the test period for one candidate substance being several weeks to several months, being relatively expensive, and requiring a relatively large amount of candidate substance, making it impossible to test with a small amount. Furthermore, depending on the product, the number of candidate substances may reach several thousand to tens of thousands, making it unrealistic to actually perform the Ames mutagenicity test on all candidate substances. Therefore, various techniques have been developed to predict the evaluation results of the Ames mutagenicity test using machine learning models. For example, M. J. Martinez, et al. "Multitask Deep Neural Networks for Ames Mutagenicity Prediction" Journal of Chemical Information and Modeling 62(24) September 2022. (hereinafter referred to as Non-Patent Document 1) describes a technology for predicting the evaluation results of an Ames mutagenicity test using a multitask deep neural network model (hereinafter simply referred to as a deep learning model).

[0005] However, the number of past Ames mutagenicity test data that can be used as training data for the deep learning model described in Non-Patent Document 1 is not very large, at around several thousand, even for data on all five strains. Furthermore, the deep learning model described in Non-Patent Document 1 is commonly used for both base pair substitution mutations and frameshift mutations, which have completely different causes of occurrence, structural features, etc. Therefore, the deep learning model described in Non-Patent Document 1 may not have sufficient prediction accuracy.

[0006] One embodiment of the technology of the present disclosure provides an information processing device, an operating method of an information processing device, and an operating program of an information processing device that can improve the prediction accuracy of the evaluation results of an Ames mutagenicity test and the accuracy of identifying whether the type of genetic mutation is a base pair substitution mutation or a frameshift mutation.

[0007] The information processing device disclosed herein is an information processing device that predicts the evaluation results of an Ames mutagenicity test on a candidate substance for a product using prediction models, and is equipped with a processor. The processor uses two prediction models: a first prediction model that outputs a first prediction evaluation result indicating whether or not a candidate substance has mutagenicity associated with base pair substitution mutations when the candidate substance is added to a first type of bacterial strain that is sensitive to base pair substitution mutations that change a portion of the base sequence; and a second prediction model that outputs a second prediction evaluation result indicating whether or not the candidate substance has mutagenicity associated with frameshift mutations when the candidate substance is added to a second type of bacterial strain that is sensitive to frameshift mutations that shift the reading frame of the base sequence. The processor obtains candidate substance information related to the candidate substance, inputs the candidate substance information or information based on the candidate substance information into the first prediction model and the second prediction model, causes the first prediction model and the second prediction model to output the first prediction evaluation result and the second prediction evaluation result, and presents prediction information corresponding to the first prediction evaluation result and the second prediction evaluation result to a user.

[0008] Preferably, if both the first and second predictive evaluation results indicate that the candidate substance is not mutagenic, the processor presents the user with predictive information indicating that the candidate substance is not mutagenic, and if at least one of the first and second predictive evaluation results indicates that the candidate substance is mutagenic, the processor presents the user with predictive information indicating that the candidate substance is mutagenic.

[0009] The processor preferably presents the first predicted evaluation result and the second predicted evaluation result themselves to the user as predicted information.

[0010] The candidate substance information is preferably information relating to the chemical structure of the candidate substance.

[0011] The predictive model is preferably constructed by any one of machine learning techniques such as support vector machines, linear separation, gradient boosting trees, Adaboost, random forests, deep learning, and ensemble learning thereof.

[0012] It is preferable that the first prediction model is constructed by deep learning, and the second prediction model is constructed by transfer learning of the trained first prediction model.

[0013] It is preferable that the first prediction model and the second prediction model are constructed using different machine learning methods.

[0014] It is preferable that the first prediction model is constructed by deep learning, and the second prediction model is constructed by a machine learning method other than deep learning.

[0015] The first prediction model and the second prediction model are constructed using the same machine learning method, but it is preferable that the internal parameters for deriving the first prediction evaluation result and the second prediction evaluation result are different.

[0016] When the predictive model is constructed using any of the machine learning techniques of support vector machines, linear separation, gradient boosting trees, Adaboost, random forests, and ensemble learning thereof, it is preferable that the processor acquires feature information of the candidate substances and inputs the feature information into the predictive model.

[0017] The feature information preferably includes at least one of a feature related to the geometric shape of the candidate substance, a feature related to the electronic properties of the candidate substance, a feature related to the physicochemical properties of the candidate substance, and a feature related to the partial structure of the candidate substance.

[0018] It is preferable that the first learning data used to train the first prediction model and the second learning data used to train the second prediction model are at least partially different from each other.

[0019] It is preferable that the first learning data used to train the first prediction model and the second learning data used to train the second prediction model are prepared based on information on the first and second species of strains.

[0020] The method of operating an information processing device disclosed herein is a method of operating an information processing device that predicts the evaluation results of an Ames mutagenicity test on a candidate substance for a product using prediction models, and includes using, as the prediction models, a first prediction model that outputs a first predicted evaluation result indicating whether or not a candidate substance has mutagenicity associated with base pair substitution mutations when the candidate substance is added to a first type of bacterial strain that is sensitive to base pair substitution mutations, in which a portion of the base sequence is changed, and a second prediction model that outputs a second predicted evaluation result indicating whether or not the candidate substance has mutagenicity associated with frameshift mutations when the candidate substance is added to a second type of bacterial strain that is sensitive to frameshift mutations, in which the reading frame of the base sequence is shifted; obtaining candidate substance information related to the candidate substance; inputting the candidate substance information or information based on the candidate substance information into the first prediction model and the second prediction model, and outputting the first predicted evaluation result and the second predicted evaluation result from the first prediction model and the second prediction model; and presenting prediction information corresponding to the first predicted evaluation result and the second predicted evaluation result to a user.

[0021] The operating program of an information processing device disclosed herein is an operating program of an information processing device that predicts the evaluation results of an Ames mutagenicity test on a candidate substance of a product using prediction models, and causes a computer to execute processes including: using, as the prediction models, a first prediction model that outputs a first predicted evaluation result indicating whether or not a candidate substance has mutagenicity associated with base pair substitution mutations when the candidate substance is added to a first type of bacterial strain that is sensitive to base pair substitution mutations, in which a portion of the base sequence is changed; and a second prediction model that outputs a second predicted evaluation result indicating whether or not a candidate substance has mutagenicity associated with frameshift mutations when the candidate substance is added to a second type of bacterial strain that is sensitive to frameshift mutations, in which the reading frame of the base sequence is shifted; acquiring candidate substance information related to the candidate substance; inputting the candidate substance information or information based on the candidate substance information into the first prediction model and the second prediction model, and outputting the first predicted evaluation result and the second predicted evaluation result from the first prediction model and the second prediction model; and presenting prediction information corresponding to the first predicted evaluation result and the second predicted evaluation result to a user.

[0022] According to the technology of the present disclosure, it is possible to provide an information processing device, an operating method for an information processing device, and an operating program for an information processing device that are capable of improving the prediction accuracy of the evaluation results of an Ames mutagenicity test.

[0023] 1 is a diagram illustrating an information processing server and a user terminal. FIG. 2 is a block diagram illustrating computers constituting the information processing server and the user terminal. FIG. 3 is a block diagram illustrating a processing unit of a CPU of the information processing server. FIG. 4 is a diagram illustrating feature information. FIG. 5 is a diagram illustrating processing of a prediction unit that inputs candidate substance information into a first prediction model and causes the first prediction model to output a first prediction evaluation result. FIG. 6 is a diagram illustrating processing of a prediction unit that inputs feature information into a second prediction model and causes the second prediction model to output a second prediction evaluation result. FIG. 7 is a diagram illustrating processing in the learning phase of a first prediction model. FIG. 8 is a diagram illustrating processing in the learning phase of a second prediction model. FIG. 9 is a diagram illustrating allocation of first learning data and second learning data. FIG. 10 is a table illustrating first prediction evaluation results, second prediction evaluation results, and overall prediction evaluation results. FIG. 11 is a diagram illustrating prediction information. FIG. 12 is a block diagram illustrating a processing unit of a CPU of a user terminal. FIG. 13 is a diagram illustrating an information input screen. FIG. 14 is a diagram illustrating a prediction evaluation result display screen. FIG. 15 is a flowchart illustrating the processing procedure of an information processing server. FIG. 16 is a diagram illustrating an aspect of constructing a second prediction model by transfer learning of a trained first prediction model. FIG. 17 is a diagram illustrating another method of prediction using the first prediction model and the second prediction model. FIG. 18 is a diagram illustrating another example of allocation of first learning data and second learning data. 19 is a diagram showing a first prediction model and a second prediction model, both of which are constructed using support vector machines. It is a diagram showing boundary lines in the feature space of the first prediction model shown in FIG. 19. It is a diagram showing boundary lines in the feature space of the second prediction model shown in FIG. 19.

[0024] As shown in FIG. 1 as an example, an information processing server 10 is connected to a user terminal 11 via a network 12. The information processing server 10 is an example of an "information processing device" according to the technology of the present disclosure. The user terminal 11 is installed, for example, at a pharmaceutical company that develops pharmaceutical products or at an organization that is contracted by a pharmaceutical company to develop pharmaceuticals, i.e., a contract research organization (CRO). The user terminal 11 is operated by a user U involved in pharmaceutical development at the pharmaceutical company or contract research organization. The network 12 is, for example, a wide area network (WAN) such as the Internet or a public communication network. Note that while only one user terminal 11 is connected to the information processing server 10 in FIG. 1 , in reality, multiple user terminals 11 from multiple pharmaceutical companies or contract research organizations are connected to the information processing server 10.

[0025] The user terminal 11 transmits a prediction request 13 to the information processing server 10. The prediction request 13 is a request to have the information processing server 10 predict the evaluation results of an Ames mutagenicity test on a pharmaceutical candidate substance. The prediction request 13 includes candidate substance information 14 related to the candidate substance. If there are multiple candidate substances for which the evaluation results of the Ames mutagenicity test are to be predicted, the prediction request 13 includes multiple pieces of candidate substance information 14 corresponding to the multiple candidate substances, as shown in the figure. The candidate substance information 14 is information related to the chemical structure of the candidate substance. More specifically, the candidate substance information 14 is a character string representing the chemical structure of the candidate substance using SMILES (Simplified Molecular Input Line Entry System) notation. Although not shown in the figure, the prediction request 13 also includes a terminal ID (Identification Data) for uniquely identifying the user terminal 11 that has sent the prediction request 13 .

[0026] When the information processing server 10 receives the prediction request 13, it predicts the evaluation result of the Ames mutagenicity test for the candidate substance and derives prediction information 15. The information processing server 10 distributes the prediction information 15 to the user terminal 11 that sent the prediction request 13. When the information processing server 10 receives the prediction information 15, the user terminal 11 makes the prediction information 15 available for viewing by the user U.

[0027] 2, the computers that make up the information processing server 10 and the user terminal 11 basically have the same configuration, and include a storage 20, a memory 21, a CPU (Central Processing Unit) 22, a communication unit 23, a display 24, and an input device 25. These are interconnected via a bus line 26.

[0028] The storage 20 is a hard disk drive built into the computers that make up the information processing server 10 and the user terminal 11, or connected via a cable or network. Alternatively, the storage 20 is a disk array consisting of multiple hard disk drives. The storage 20 stores control programs such as an operating system, various application programs (hereinafter referred to as APs (Application Programs)), and various data associated with these programs. Note that a solid state drive may be used instead of a hard disk drive.

[0029] The memory 21 is a work memory for the CPU 22 to execute processing. The CPU 22 loads a program stored in the storage 20 into the memory 21 and executes processing in accordance with the program. In this way, the CPU 22 comprehensively controls each part of the computer. The CPU 22 is an example of a "processor" according to the technology of the present disclosure. The memory 21 may be built into the CPU 22.

[0030] The communication unit 23 is a network interface that controls the transmission of various information via the network 12, etc. The display 24 displays various screens. The various screens are equipped with operation functions using a GUI (Graphical User Interface). The computers that make up the information processing server 10 and the user terminal 11 accept input of operation instructions from an input device 25 via the various screens. The input device 25 is a keyboard, a mouse, a touch panel, a microphone for voice input, etc.

[0031] In the following explanation, the parts of the computer that make up the information processing server 10 (storage 20 and CPU 22) are distinguished by adding the suffix "A" to their symbols, and the parts of the computer that make up the user terminal 11 (storage 20, CPU 22, display 24, and input device 25) are distinguished by adding the suffix "B" to their symbols.

[0032] As an example, as shown in FIG. 3 , an operating program 30 is stored in the storage 20A of the information processing server 10. The operating program 30 is an AP for causing a computer to function as the information processing server 10. In other words, the operating program 30 is an example of an "operating program of an information processing device" according to the technology of the present disclosure. The storage 20 also stores a first prediction model 311 and a second prediction model 312. The first prediction model 311 and the second prediction model 312 are examples of a "prediction model" according to the technology of the present disclosure. In the following description, unless there is a particular need to distinguish between them, the first prediction model 311 and the second prediction model 312 will be collectively referred to as the prediction model 31.

[0033] When the operating program 30 is started, the CPU 22A of the computer constituting the information processing server 10 cooperates with the memory 21 and the like to function as a request receiving unit 35, a read / write (hereinafter abbreviated as RW (Read Write)) control unit 36, a feature derivation unit 37, a prediction unit 38, and a screen distribution control unit 39.

[0034] The request receiving unit 35 receives various requests from the user terminal 11, including the prediction request 13. As described above, the prediction request 13 includes the candidate substance information 14. Therefore, by receiving the prediction request 13, the request receiving unit 35 acquires the candidate substance information 14. When the prediction request 13 is received, the request receiving unit 35 outputs the candidate substance information 14 included in the prediction request 13 to the RW control unit 36. In addition, the request receiving unit 35 outputs the terminal ID of the user terminal 11 included in the prediction request 13 to the screen distribution control unit 39.

[0035] The RW control unit 36 ​​controls the storage of various data in the storage 20A and the reading of various data from the storage 20A. In particular, the RW control unit 36 ​​controls the storage of candidate substance information 14 in the storage 20A and the reading of the candidate substance information 14 from the storage 20A. The RW control unit 36 ​​outputs the read candidate substance information 14 to the feature derivation unit 37 and the prediction unit 38. The RW control unit 36 ​​also reads the first prediction model 311 and the second prediction model 312 from the storage 20A and outputs the read first prediction model 311 and second prediction model 312 to the prediction unit 38.

[0036] The feature derivation unit 37 derives feature information 42 for a candidate substance from the candidate substance information 14. More specifically, the feature derivation unit 37 derives one piece of feature information 42 from the candidate substance information 14 for one candidate substance. Therefore, the feature derivation unit 37 derives the same number of pieces of feature information 42 as the number of pieces of candidate substance information 14 included in the prediction request 13. The feature derivation unit 37 derives the feature information 42, for example, using a machine learning model that outputs feature information 42 when candidate substance information 14 is input. The feature derivation unit 37 outputs the feature information 42 to the prediction unit 38.

[0037] The prediction unit 38 causes the first prediction model 311 and the second prediction model 312 to predict the evaluation result of the Ames mutagenicity test for the candidate substance based on the candidate substance information 14 and the feature amount information 42. The prediction unit 38 generates prediction information 15 and outputs the prediction information 15 to the screen distribution control unit 39.

[0038] The screen delivery control unit 39 controls the delivery of various screens to the user terminal 11. Specifically, the screen delivery control unit 39 delivers and outputs various screens to the user terminal 11 that has sent the various requests in the form of screen data for web delivery created using a markup language such as XML (Extensible Markup Language). At this time, the screen delivery control unit 39 identifies the user terminal 11 that has sent the various requests based on the terminal ID from the request receiving unit 35. Note that instead of XML, other data description languages ​​such as JSON (Javascript (registered trademark) Object Notation) may be used.

[0039] The various screens include an information input screen 85 (see FIG. 13) for inputting candidate substance information 14, and a prediction evaluation result display screen 95 (see FIG. 14) for displaying prediction information 15. In addition to these processing units 35 to 39, the CPU 22A also includes an instruction receiving unit that receives various operation instructions from the input device 25.

[0040] 4 , the feature information 42 includes a feature 45 related to the geometric shape of the candidate substance, a feature 46 related to the electronic properties of the candidate substance, a feature 47 related to the physicochemical properties of the candidate substance, and a feature 48 related to the partial structure of the candidate substance. The feature 45 related to the geometric shape includes the number of bonds 49 of the candidate substance and the number 50 of benzene rings of the candidate substance. The feature 46 related to the electronic properties includes a surface charge density distribution 51 of the candidate substance and a HOMO (Highest Occupied Molecular Orbital)-LUMO (Lowest Unoccupied Molecular Orbital) energy gap 52 of the candidate substance.

[0041] The feature quantities 47 related to the physicochemical properties include the molecular weight 53 of the candidate substance, and the water solubility 54 indicating the hydrophilicity and hydrophobicity of the candidate substance. The feature quantities 48 related to the partial structure include the Klekota-Roth fingerprint 55 of the candidate substance derived by the partial structure extraction algorithm, and the MACCS Keys fingerprint 56 of the candidate substance. Note that the feature quantities 48 related to the partial structure may include a topological fingerprint, a Morgan fingerprint, a MinHash fingerprint, an Avalon fingerprint, an atom pair fingerprint, a topological dihedral angle fingerprint, a Pubchem fingerprint, and the like.

[0042] The various features in the feature information 42 have been selected by the developer of the operating program 30 as being useful for predicting whether a candidate substance has mutagenicity associated with frameshift mutations. The number of features included in the feature information 42 is preferably 200 or more, and more preferably 1,000 or more. Therefore, the feature information 42 can be said to be multidimensional feature data having several hundred to several thousand dimensions.

[0043] 5 as an example, the prediction unit 38 inputs the candidate substance information 14 into a first prediction model 311 and causes the first prediction model 311 to output a first predicted evaluation result 601. The first predicted evaluation result 601 indicates whether or not the candidate substance has mutagenicity related to base pair substitution mutations when added to a first species strain that is sensitive to base pair substitution mutations that cause a partial change in the base sequence.

[0044] 6 , the prediction unit 38 inputs feature information 42 derived from the candidate substance information 14 into a second prediction model 312, and causes the second prediction model 312 to output a second prediction evaluation result 602. The feature information 42 is an example of "information based on candidate substance information" according to the technology of the present disclosure. The second prediction evaluation result 602 indicates whether the candidate substance has mutagenicity related to frameshift mutations when added to a second type of bacterial strain that is sensitive to frameshift mutations, which shift the reading frame of a base sequence.

[0045] Here, as shown in FIG. 5 , the first prediction model 311 is a machine learning model corresponding to base pair substitution mutations constructed using a deep neural network, i.e., deep learning. In contrast, as shown in FIG. 6 , the second prediction model 312 is a machine learning model corresponding to frameshift mutations constructed using a support vector machine. As such, the first prediction model 311 and the second prediction model 312 are constructed using different machine learning methods. Furthermore, the first prediction model 311 is constructed using deep learning, while the second prediction model 312 is constructed using a machine learning method other than deep learning (here, a support vector machine). The first prediction model 311 and the second prediction model 312 may be constructed using any of the following machine learning methods: support vector machine, linear separation, gradient boosting tree, Adaboost, random forest, deep learning, and ensemble learning of these.

[0046] As is clear from the above explanation, the first prediction model 311 and the second prediction model 312 are not models that are experimentally constructed by a method that does not use bacterial strains but uses only nucleic acids and compounds, such as isothermal titration calorimetry or ultraviolet-visible spectroscopy. Furthermore, the first prediction model 311 and the second prediction model 312 are not models that are constructed only by simulation, such as docking simulation or quantum chemical calculation.

[0047] As an example, as shown in FIG. 7 , the first prediction model 311 is trained using first training data 621. The first training data 621 is a set of training candidate substance information 14L and first supervised answer data 601CA. The training candidate substance information 14L is candidate substance information 14 for candidate substances for which an Ames mutagenicity test was actually conducted in the past. The first supervised answer data 601CA is the result of an evaluation of whether or not the candidate substance provided by the training candidate substance information 14L is mutagenic in an Ames mutagenicity test that was actually conducted in the past.

[0048] The candidate substance information for learning 14L is input to the first prediction model 311. As a result, a first prediction evaluation result for learning 601L is output from the first prediction model 311. The first prediction evaluation result for learning 601L is compared with the first supervised answer data 601CA, and a loss calculation is performed for the first prediction model 311 using a loss function based on the comparison result. Then, internal parameters such as the filter coefficients of the first prediction model 311 are updated according to the result of the loss calculation, and the first prediction model 311 is updated according to the update setting.

[0049] The above-described series of processes, including input of the candidate substance information 14L for training to the first prediction model 311, output of the first prediction evaluation result for training 601L from the first prediction model 311, loss calculation, update setting, and update of the first prediction model 311, are repeatedly performed while the first training data 621 is changed. The repetition of the above-described series of processes is terminated when the prediction accuracy of the first prediction evaluation result for training 601L for the first supervised answer data 601CA reaches a preset level. The first prediction model 311 whose prediction accuracy has thus reached the preset level is stored in storage 20A. Note that training may be terminated after a predetermined number of repetitions of the above-described series of processes, regardless of the prediction accuracy. Furthermore, training of the first prediction model 311 may continue even after storage in storage 20A.

[0050] As an example, as shown in FIG. 8 , the second prediction model 312 is generated based on second training data 622. The second training data 622 is a set of training feature information 42L and second supervised answer data 602CA. The training feature information 42L is feature information 42 of candidate substances for which an Ames mutagenicity test was actually conducted in the past. The second supervised answer data 602CA is the result of an evaluation of whether or not the candidate substance that provided the training feature information 42L is mutagenic in an Ames mutagenicity test that was actually conducted in the past.

[0051] There are multiple second training data 622, including data in which the second correct answer data 602CA indicates that the candidate substance is mutagenic (positive) and data in which the second correct answer data 602CA indicates that the candidate substance is not mutagenic (negative). In a graph plotting these multiple second training data 622 in a feature space 65 represented by multiple features constituting the training feature information 42L, a boundary line 66 is determined along which the second training data 622 for which the second correct answer data 602CA indicates that the candidate substance is mutagenic and the second training data 622 for which the second correct answer data 602CA indicates that the candidate substance is not mutagenic are classified. In this process, an optimization problem is solved to minimize misclassification while maximizing the margin, i.e., the distance from the support vector to the boundary line 66, of the second training data 622 closest to the boundary line 66. By determining the boundary line 66 in this manner, a second prediction model 312 is generated. The generated second prediction model 312 is stored in storage 20A. Note that learning of the second prediction model 312 may continue even after the feature space 65 is stored in the storage 20A. Also, in Fig. 8, for convenience of explanation, the dimension of the feature space 65 is shown as two-dimensional with a Z1 axis and a Z2 axis, but the actual dimension of the feature space 65 is several hundred to several thousand, as described above. Similarly, for convenience of explanation, the dimension of the feature space 65 is shown as two-dimensional in Figs. 20 and 21 below.

[0052] As an example, as shown in FIG. 9 , information 70 of an Ames mutagenicity test actually conducted in the past, which is the basis for the first learning data 621 and the second learning data 622, includes candidate substance information 14, evaluation results indicating whether the candidate substance is mutagenic, and bacterial strain information. As described above, the bacterial strains include first-type strains that are sensitive to base pair substitution mutations and second-type strains that are sensitive to frameshift mutations. In this example, the first-type strains are three types: TA100, TA1535, and WP2uvrA, and the second-type strains are two types: TA98 and TA1537. WP2uvrA / pKM101 may be used instead of WP2uvrA. Similarly, TA98NR may be used instead of TA98. The first-type strain may also include TA102. Additionally, the second strain may include TA97 or TA97a, or TA1538.

[0053] As shown in Table 71, among the Ames mutagenicity tests actually conducted in the past, those for which the first type strain is registered are assigned to the first learning data 621, and those for which the second type strain is registered are assigned to the second learning data 622. In other words, the first learning data 621 and the second learning data 622 are prepared based on information on the first and second type strains. Therefore, the first learning data 621 and the second learning data 622 are different.

[0054] The candidate substance information 14 for the Ames mutagenicity test assigned to the first learning data 621 becomes training candidate substance information 14L for the first learning data 621, and the evaluation result becomes first supervised data 601CA. Furthermore, feature information 42 derived from the candidate substance information 14 for the Ames mutagenicity test assigned to the second learning data 622 becomes training feature information 42L for the second learning data 622, and the evaluation result becomes second supervised data 602CA. Information 70 on Ames mutagenicity tests actually conducted in the past may be public information that is widely available to the public, or may be information independently accumulated by a pharmaceutical company or a contract research organization. Furthermore, the information may be composed of both public information and information independently accumulated by a pharmaceutical company or a contract research organization.

[0055] 10, if both the first predicted evaluation result 601 and the second predicted evaluation result 602 indicate that the candidate substance is not mutagenic, the prediction unit 38 outputs an overall predicted evaluation result 76 indicating that the candidate substance is not mutagenic. On the other hand, if at least one of the first predicted evaluation result 601 and the second predicted evaluation result 602 indicates that the candidate substance is mutagenic, as shown in the columns other than the top column of Table 75, the prediction unit 38 outputs an overall predicted evaluation result 76 indicating that the candidate substance is mutagenic.

[0056] 11, the prediction unit 38 outputs a first predicted evaluation result 601, a second predicted evaluation result 602, and an overall predicted evaluation result 76 as prediction information 15. FIG. 11 illustrates a case where both the first predicted evaluation result 601 and the second predicted evaluation result 602 indicate that the candidate substance is not mutagenic.

[0057] As an example, as shown in FIG. 12 , a prediction AP 80 is stored in the storage 20B of the user terminal 11. The prediction AP 80 is installed in the user terminal 11 by the user U. The prediction AP 80 is an AP for predicting the evaluation results of an Ames mutagenicity test. When the prediction AP 80 is launched, the CPU 22B of the user terminal 11 functions as a browser control unit 82 in cooperation with the memory 21 and the like. The browser control unit 82 controls the operation of a web browser dedicated to the prediction AP 80.

[0058] The browser control unit 82 reproduces various screens based on various screen data from the information processing server 10 and displays the reproduced various screens on the display 24B. The browser control unit 82 also accepts various operation instructions input by the user U from the input device 25B via the various screens. The browser control unit 82 transmits various requests, including the prediction request 13, to the information processing server 10 in response to the operation instructions.

[0059] When the prediction AP 80 is launched, an information input screen 85, as shown in FIG. 13 for example, is displayed on the display 24B under the control of the browser control unit 82. The information input screen 85 has input boxes 86 for the candidate substance information 14 of a plurality of candidate substances. In the input boxes 86, the chemical structural formulas of the candidate substances can be written using a description tool that appears when a description tool display button 87 is selected, or a file of the chemical structural formula of the candidate substance can be dropped. The input boxes 86 can be added by selecting the add buttons 88A and 88B at the bottom. The add button 88A is a button for adding one input box 86 at a time, and the add button 88B is a button for adding ten input boxes 86 at a time.

[0060] The user U inputs the chemical structural formula of a desired candidate substance in the input box 86, and then selects the prediction button 89. When the prediction button 89 is selected, the browser control unit 82 generates a prediction request 13 including candidate substance information 14 corresponding to the chemical structural formula input in the input box 86, and transmits the generated prediction request 13 to the information processing server 10.

[0061] Furthermore, when the evaluation result of the Ames mutagenicity test is predicted in the information processing server 10, a predicted evaluation result display screen 95 shown in Fig. 14 as an example is displayed on the display 24B under the control of the browser control unit 82. The predicted evaluation result display screen 95 displays a list of predicted information 15 for each candidate substance. In this way, the predicted information 15 is presented to the user U in the form of screen data distribution.

[0062] A chemical structure formula display button 96 is provided at the top of the predicted evaluation result display screen 95. When the chemical structure formula display button 96 is selected, a screen listing the chemical structure formulas of the candidate substances is displayed. In addition, a save button 97 and an OK button 98 are provided at the bottom of the predicted evaluation result display screen 95. When the save button 97 is selected, the candidate substance information 14 and the prediction information 15 are stored in association with each other in the storage 20B of the user terminal 11. When the OK button 98 is selected, the display of the predicted evaluation result display screen 95 is cleared.

[0063] Next, the operation of the above configuration will be described with reference to the flowchart shown in Fig. 15 as an example. When the operating program 30 is started in the information processing server 10, the CPU 22A of the information processing server 10 functions as a request receiving unit 35, a RW control unit 36, a feature derivation unit 37, a prediction unit 38, and a screen delivery control unit 39, as shown in Fig. 3. When the prediction AP 80 is started in the user terminal 11, the CPU 22B of the user terminal 11 functions as a browser control unit 82, as shown in Fig. 12.

[0064] 13 is displayed on the display 24B of the user terminal 11 under the control of the browser control unit 82. When the user U inputs the chemical structural formula of a desired candidate substance in an input box 86 on the information input screen 85 and selects the prediction button 89, a prediction request 13 is transmitted from the browser control unit 82 to the information processing server 10. As shown in FIG. 1, the prediction request 13 includes candidate substance information 14, which is a character string representing the chemical structure of the candidate substance in SMILES notation, the terminal ID of the user terminal 11, and the like.

[0065] In the information processing server 10, the request receiving unit 35 receives the prediction request 13 (YES in step ST100). The candidate substance information 14 included in the prediction request 13 is output from the request receiving unit 35 to the RW control unit 36 ​​and stored in the storage 20A under the control of the RW control unit 36 ​​(step ST110). In addition, the terminal ID of the user terminal 11 included in the prediction request 13 is output from the request receiving unit 35 to the screen distribution control unit 39.

[0066] The candidate substance information 14 is read from the storage 20A by the RW control unit 36 ​​(step ST120). The candidate substance information 14 is output from the RW control unit 36 ​​to the feature derivation unit 37 and the prediction unit 38.

[0067] The feature derivation unit 37 derives feature information 42 from the candidate substance information 14 (step ST130). As shown in Fig. 4, the feature information 42 includes a feature 45 related to the geometric shape of the candidate substance, a feature 46 related to the electronic properties of the candidate substance, a feature 47 related to the physicochemical properties of the candidate substance, and a feature 48 related to the partial structure of the candidate substance. The feature information 42 is output from the feature derivation unit 37 to the prediction unit 38.

[0068] In the prediction unit 38, as shown in FIG. 5 , the candidate substance information 14 is input to the first prediction model 311. As a result, a first prediction evaluation result 601 is output from the first prediction model 311 (step ST140_1). Concurrently, in the prediction unit 38, as shown in FIG. 6 , the feature amount information 42 is input to the second prediction model 312. As a result, a second prediction evaluation result 602 is output from the second prediction model 312 (step ST140_2). Then, as shown in FIG. 11 , prediction information 15 is generated, which includes the first prediction evaluation result 601, the second prediction evaluation result 602, and the overall prediction evaluation result 76. The prediction information 15 is output from the prediction unit 38 to the screen distribution control unit 39.

[0069] The screen distribution control unit 39 generates screen data for the predicted evaluation result display screen 95 shown in Fig. 14 based on the prediction information 15. Under the control of the screen distribution control unit 39, the screen data for the predicted evaluation result display screen 95 is distributed to the user terminal 11 that sent the prediction request 13 (step ST150).

[0070] In the user terminal 11, under the control of the browser control unit 82, the screen data of the predicted evaluation result display screen 95 is reproduced, and the reproduced predicted evaluation result display screen 95 is displayed on the display 24B. In this way, the predicted information 15 is presented to the user U.

[0071] As described above, the information processing server 10 uses the first prediction model 311 and the second prediction model 312 as the prediction model 31 for predicting the evaluation results of an Ames mutagenicity test for a candidate substance. The first prediction model 311 outputs a first predicted evaluation result 601 indicating whether or not the candidate substance has mutagenicity related to base pair substitution mutations when added to a first type of bacterial strain that is sensitive to base pair substitution mutations that cause a partial change in the base sequence. The second prediction model 312 outputs a second predicted evaluation result 602 indicating whether or not the candidate substance has mutagenicity related to frameshift mutations when added to a second type of bacterial strain that is sensitive to frameshift mutations that cause a shift in the reading frame of the base sequence.

[0072] The CPU 22A of the information processing server 10 functions as a request receiving unit 35, a prediction unit 38, and a screen distribution control unit 39. The request receiving unit 35 receives a prediction request 13 from the user terminal 11 and acquires candidate substance information 14 included in the prediction request 13. The prediction unit 38 inputs the candidate substance information 14 to a first prediction model 311 and inputs feature amount information 42 derived from the candidate substance information 14 to a second prediction model 312. The CPU 22A then causes the first prediction model 311 and the second prediction model 312 to output a first prediction evaluation result 601 and a second prediction evaluation result 602, respectively. The screen distribution control unit 39 presents the prediction information 15 to the user U by distributing screen data of a prediction evaluation result display screen 95 including prediction information 15 corresponding to the first prediction evaluation result 601 and the second prediction evaluation result 602 to the user terminal 11. This makes it possible to improve the accuracy of predicting the evaluation results of the Ames mutagenicity test compared to using a prediction model that is common to both base pair substitution mutations and frameshift mutations, which have completely different causes, structural features, etc. As a result, it also makes it possible to improve the accuracy of distinguishing whether the type of gene mutation is a base pair substitution mutation or a frameshift mutation.

[0073] 10 , 11 , and 14 , if both the first predicted evaluation result 601 and the second predicted evaluation result 602 indicate that the candidate substance is not mutagenic, the screen distribution control unit 39 presents the fact that the candidate substance is not mutagenic to the user U as prediction information 15. On the other hand, if at least one of the first predicted evaluation result 601 and the second predicted evaluation result 602 indicates that the candidate substance is mutagenic, the screen distribution control unit 39 presents the fact that the candidate substance is mutagenic to the user U as prediction information 15. Therefore, the user U can correctly understand the predicted evaluation results of the Ames mutagenicity test, i.e., whether or not the candidate substance is mutagenic.

[0074] Furthermore, the screen distribution control unit 39 presents the first predicted evaluation result 601 and the second predicted evaluation result 602 themselves to the user U as prediction information 15. This allows the user U to understand whether or not the candidate substance has mutagenicity associated with base pair substitution mutations, and whether or not the candidate substance has mutagenicity associated with frameshift mutations. For example, the user U can compare the chemical structure of a candidate substance that has mutagenicity associated with base pair substitution mutations with the chemical structure of a candidate substance that does not have mutagenicity associated with base pair substitution mutations, and use this information to help with subsequent candidate substance design, such as by using it as a guideline for designing candidate substances that do not have mutagenicity associated with base pair substitution mutations.

[0075] 1, the candidate substance information 14 is information about the chemical structure of the candidate substance. The information about the chemical structure most clearly represents the properties of the candidate substance. Therefore, the first prediction evaluation result 601 that reflects the properties of the candidate substance can be output from the first prediction model 311.

[0076] 5 and 6, the prediction model 31 is constructed by one of the machine learning methods: support vector machine, linear separation, gradient boosting tree, Adaboost, random forest, deep learning, and ensemble learning of these. Since these are all very common machine learning methods, it is possible to easily construct a prediction model 31 with relatively high prediction accuracy.

[0077] The first prediction model 311 and the second prediction model 312 are constructed using different machine learning methods. Therefore, it is possible to construct the first prediction model 311 suitable for predicting whether or not a candidate substance has mutagenicity associated with base pair substitution mutations, and the second prediction model 312 suitable for predicting whether or not a candidate substance has mutagenicity associated with frameshift mutations.

[0078] Furthermore, the first prediction model 311 is constructed by deep learning, and the second prediction model 312 is constructed by a machine learning method other than deep learning.

[0079] Base pair substitution mutations are the primary cause of mutagenicity, and there is more reference data available for them than frameshift mutations, which tend to be a secondary cause of mutagenicity. In other words, the first training data 621 is larger than the second training data 622. Therefore, the first prediction model 311 is suitable for deep learning, which requires a large amount of training data to improve prediction accuracy. In contrast, the second prediction model 312 is suitable for machine learning methods other than deep learning, which can improve prediction accuracy by using selected features even with relatively little training data. In this way, by constructing the first prediction model 311 and the second prediction model 312 using their respective preferred machine learning methods, it is possible to further improve the prediction accuracy of the evaluation results of the Ames mutagenicity test. It is also possible to construct multiple prediction models 31 using different machine learning methods and adopt the prediction model 31 with the highest prediction accuracy among them.

[0080] The CPU 22A of the information processing server 10 further functions as a feature derivation unit 37. The feature derivation unit 37 acquires the feature information 42 by deriving the feature information 42 from the candidate substance information 14. The prediction unit 38 inputs the feature information 42 into a second prediction model 312 constructed by a support vector machine. This allows the second prediction model 312 to output a second prediction evaluation result 602 without delay.

[0081] 4, the feature amount information 42 includes a feature amount 45 related to the geometric shape of the candidate substance, a feature amount 46 related to the electronic properties of the candidate substance, a feature amount 47 related to the physicochemical properties of the candidate substance, and a feature amount 48 related to the partial structure of the candidate substance. This can improve the prediction accuracy of the second predicted evaluation result 602. Note that the feature amount information 42 only needs to include at least one of the feature amounts 45 to 48.

[0082] 9 , the first learning data 621 used to train the first prediction model 311 and the second learning data 622 used to train the second prediction model 312 are at least partially different from each other. This allows learning appropriate for each of the first prediction model 311 and the second prediction model 312.

[0083] 9 , the first training data 621 and the second training data 622 are prepared based on information on the first and second species of bacterial strains, so that the first training data 621 and the second training data 622 suitable for the first prediction model 311 and the second prediction model 312, respectively, can be prepared.

[0084] (Variation 1) As an example, as shown in FIG. 16 , a second prediction model 312 corresponding to a frameshift mutation may be constructed by transfer learning a first prediction model 311 corresponding to a base pair substitution mutation constructed using a deep neural network. In transfer learning, the first prediction model 311, which has been trained using first training data 621 as shown in FIG. 7 , is subjected to the same process as shown in FIG. 7 using second training data 622. The first prediction model 311 is trained using the first training data 621, which is larger in quantity than the second training data 622, and this is used as the second prediction model 312 by transfer learning. Therefore, compared to when the first prediction model 311 and the second prediction model 312 constructed using a deep neural network are separately prepared, it is possible to suppress a decrease in the prediction accuracy of the second prediction model 312 due to the small amount of second training data 622. In other words, this configuration also solves the conventional problem of the inability to improve the prediction accuracy of the prediction model 31 due to the small amount of data from past Ames mutagenicity tests available as training data. In this case, the processing can be simplified because there is no need to derive the feature amount information 42. In this case, naturally, the first prediction model 311 and the second prediction model 312 have different internal parameters, such as filter coefficients.

[0085] (Variation 2) In the above example, the prediction unit 38 uses the first prediction model 311 to predict whether or not a candidate substance has mutagenicity associated with base pair substitution mutations, and the second prediction model 312 to predict whether or not a candidate substance has mutagenicity associated with frameshift mutations. However, this is not limiting. As an example, prediction may be performed using the procedure shown in FIG. 17 . That is, first, the first prediction model 311 is used to predict whether or not a candidate substance has mutagenicity associated with base pair substitution mutations. Then, only if the first prediction evaluation result 601 indicates that the candidate substance does not have mutagenicity associated with base pair substitution mutations, the second prediction model 312 is used to predict whether or not the candidate substance has mutagenicity associated with frameshift mutations.

[0086] In this case, as in the above example, if both the first prediction evaluation result 601 and the second prediction evaluation result 602 indicate that the candidate substance is not mutagenic, the prediction unit 38 outputs an overall prediction evaluation result 76 indicating that the candidate substance is not mutagenic. On the other hand, if at least one of the first prediction evaluation result 601 and the second prediction evaluation result 602 indicates that the candidate substance is mutagenic, the prediction unit 38 outputs an overall prediction evaluation result 76 indicating that the candidate substance is mutagenic. If the first prediction evaluation result 601 indicates that the candidate substance is mutagenic, prediction using the second prediction model 312 is not performed, thereby reducing processing time. However, as in the above example, it is preferable to always perform prediction using the first prediction model 311 and prediction using the second prediction model 312, as this increases the reliability of the overall prediction evaluation result 76.

[0087] (Variation 3) The method of sorting the first learning data 621 and the second learning data 622 is not limited to the example shown in Fig. 9. As an example, as shown in table 100 in Fig. 18, among Ames mutagenicity tests actually conducted in the past, those registered as having mutagenicity in the evaluation results are sorted based on the strain information as in the case of Fig. 9, but those registered as not having mutagenicity in the evaluation results may be sorted into both the first learning data 621 and the second learning data 622. In other words, it is sufficient that the first learning data 621 and the second learning data 622 are at least partially different.

[0088] 5 and 6 illustrate an example in which the first prediction model 311 and the second prediction model 312 are constructed using different machine learning techniques (i.e., the first prediction model 311 is constructed using a deep neural network, and the second prediction model 312 is constructed using a support vector machine). However, this is not limiting. As an example, as shown in Fig. 19, the first prediction model 311 and the second prediction model 312 may be constructed using the same machine learning technique. Fig. 19 illustrates an example in which the first prediction model 311 and the second prediction model 312 are both constructed using a support vector machine.

[0089] However, in the case of FIG. 19 , even if the machine learning method is the same (support vector machine), as shown in FIGS. 20 and 21 , the first boundary line 661 in the feature space 65 of the first prediction model 311 is naturally different from the second boundary line 662 in the feature space 65 of the second prediction model 312. The first boundary line 661 and the second boundary line 662 are examples of “internal parameters” according to the technology of the present disclosure. Even if the first prediction model 311 and the second prediction model 312 are constructed using the same machine learning method, the first prediction model 311, which is suitable for predicting whether a candidate substance has mutagenicity associated with base pair substitution mutations, and the second prediction model 312, which is suitable for predicting whether a candidate substance has mutagenicity associated with frameshift mutations, can be used. In this case, the first training data 621 is composed of a set of training feature information 42L and first supervised answer data 601CA, as shown in FIG. 20 .

[0090] Both the first prediction model 311 and the second prediction model 312 may be constructed by deep learning such as a deep neural network. In this case, the internal parameters are, for example, coefficients set in an intermediate layer (hidden layer) or coefficients of a filter used in a convolution operation.

[0091] The candidate substance information 14 is not limited to a character string representing the chemical structure of the candidate substance using the SMILES notation shown as an example. It may also be a MOL (Molecular Design Limited) file or an SDF (Structure-Data File) representing the chemical structure of the candidate substance. In any case, a description method that can uniquely determine a three-dimensional structure such as an isomer is preferable, and a description method that can represent three-dimensional coordinate information of a molecule is even more preferable.

[0092] The feature information 42 may be derived by a device other than the information processing server 10 and input to the information processing server 10. Alternatively, the user U may input the feature information 42 via the input device 25B of the user terminal 11.

[0093] The prediction information 15 may be only the overall prediction evaluation result 76. Conversely, the prediction information 15 may be only the set of the first prediction evaluation result 601 and the second prediction evaluation result 602.

[0094] The information processing server 10 may be installed in a pharmaceutical company or a pharmaceutical contract research organization, or may be installed in a data center independent from the pharmaceutical company or the pharmaceutical contract research organization.

[0095] Instead of delivering screen data of the predicted evaluation result display screen 95 including the prediction information 15 to the user terminal 11, the prediction information 15 itself may be delivered to the user terminal 11. In this case, the user terminal 11 generates the predicted evaluation result display screen 95 based on the prediction information 15 under the control of the browser control unit 82.

[0096] The method of presenting the prediction information 15 to the user U is not limited to the example of delivering screen data. The prediction information 15 may be presented to the user U by printing it on a paper medium, or by attaching it to an e-mail and sending it to the user terminal 11.

[0097] The products are not limited to pharmaceuticals, but can also be cosmetics, etc.

[0098] The hardware configuration of the computer constituting the information processing server 10 according to the technology of the present disclosure can be modified in various ways. For example, the information processing server 10 can be configured with multiple computers separated as hardware in order to improve processing power and reliability. For example, the functions of the request reception unit 35 and the RW control unit 36 ​​and the functions of the feature derivation unit 37, the prediction unit 38, and the screen delivery control unit 39 can be distributed and performed by two computers. In this case, the information processing server 10 is configured with two computers. Some or all of the functions of the information processing server 10 may be performed by the user terminal 11.

[0099] In this way, the hardware configuration of the computer of the information processing server 10 can be changed as appropriate depending on the required performance such as processing power, safety, reliability, etc. Furthermore, not only the hardware but also APs such as the operating program 30 can be duplicated or stored in a distributed manner in multiple storage devices in order to ensure safety and reliability.

[0100] In the above embodiment, for example, the hardware structure of the processing unit that executes various processes, such as the request receiving unit 35, the RW control unit 36, the feature derivation unit 37, the prediction unit 38, the screen delivery control unit 39, and the browser control unit 82, can be any of the various processors listed below. As described above, the various processors include the CPUs 22A and 22B, which are general-purpose processors that execute software (the operating program 30 and the prediction AP 80) and function as various processing units, as well as programmable logic devices (PLDs) that are processors whose circuit configuration can be changed after manufacture, such as a field programmable gate array (FPGA), and dedicated electrical circuits that are processors having a circuit configuration designed specifically for executing specific processing, such as an application specific integrated circuit (ASIC).

[0101] A single processing unit may be configured with one of these various processors, or may be configured with a combination of two or more processors of the same or different types (e.g., a combination of multiple FPGAs and / or a combination of a CPU and an FPGA).Furthermore, multiple processing units may be configured with a single processor.

[0102] Examples of configuring multiple processing units with a single processor include, first, a form in which one processor is configured with a combination of one or more CPUs and software, as typified by computers such as client and server, and this processor functions as multiple processing units. Second, a form in which a processor is used to realize the functions of the entire system including multiple processing units with a single IC (Integrated Circuit) chip, as typified by systems on chips (SoCs). In this way, various processing units are configured using one or more of the above-mentioned various processors as a hardware structure.

[0103] Furthermore, more specifically, the hardware structure of these various processors can be an electric circuit (circuitry) that combines circuit elements such as semiconductor elements.

[0104] From the above description, the technology described in the following supplementary paragraphs can be understood.

[0105] [Supplementary Item 1] An information processing device that predicts evaluation results of an Ames mutagenicity test on a product candidate substance using prediction models, comprising a processor, wherein the processor uses as the prediction models: a first prediction model that, when the candidate substance is added to a first type of bacterial strain that is sensitive to base pair substitution mutations that change a part of a base sequence, outputs a first predicted evaluation result indicating whether or not the candidate substance has mutagenicity associated with the base pair substitution mutation; and a second prediction model that, when the candidate substance is added to a second type of bacterial strain that is sensitive to frameshift mutations that shift the reading frame of the base sequence, outputs a second predicted evaluation result indicating whether or not the candidate substance has mutagenicity associated with the frameshift mutation; obtains candidate substance information related to the candidate substance; inputs the candidate substance information or information based on the candidate substance information into the first prediction model and the second prediction model; causes the first prediction model and the second prediction model to output the first predicted evaluation result and the second predicted evaluation result; and presents prediction information according to the first predicted evaluation result and the second predicted evaluation result to a user. [Supplementary Item 2] The information processing device according to Supplementary Item 1, wherein the processor, when both the first predictive evaluation result and the second predictive evaluation result indicate that the candidate substance is not mutagenic, presents to the user as the prediction information that the candidate substance is not mutagenic, and when at least one of the first predictive evaluation result and the second predictive evaluation result indicates that the candidate substance is mutagenic, presents to the user as the prediction information that the candidate substance is mutagenic. [Supplementary Item 3] The information processing device according to Supplementary Item 1 or Supplementary Item 2, wherein the processor presents the first predictive evaluation result and the second predictive evaluation result themselves to the user as the prediction information. [Supplementary Item 4] The information processing device according to any one of Supplementary Items 1 to 3, wherein the candidate substance information is information about the chemical structure of the candidate substance.[Supplementary Item 5] The information processing device of any one of Supplementary Items 1 to 4, wherein the prediction model is constructed by one of machine learning methods including support vector machines, linear separation, gradient boosting trees, Adaboost, random forests, deep learning, and ensemble learning thereof. [Supplementary Item 6] The information processing device of Supplementary Item 5, wherein the first prediction model is constructed by deep learning, and the second prediction model is constructed by transfer learning of the trained first prediction model. [Supplementary Item 7] The information processing device of Supplementary Item 5, wherein the first prediction model and the second prediction model are constructed by different machine learning methods. [Supplementary Item 8] The information processing device of Supplementary Item 7, wherein the first prediction model is constructed by deep learning, and the second prediction model is constructed by a machine learning method other than deep learning. [Supplementary Item 9] The information processing device of Supplementary Item 5 or Supplementary Item 6, wherein the first prediction model and the second prediction model are constructed by the same machine learning method, but wherein internal parameters for deriving the first prediction evaluation result and the second prediction evaluation result are different. [Supplementary Item 10] The information processing device according to any one of Supplementary Items 5, 7, 8, and 9, wherein, when the prediction model is constructed by any one of machine learning methods including support vector machines, linear separation, gradient boosting trees, Adaboost, random forests, and ensemble learning thereof, the processor acquires feature information of the candidate substances and inputs the feature information to the prediction model. [Supplementary Item 11] The information processing device according to Supplementary Item 10, wherein the feature information includes at least one of feature amounts related to the geometric shape of the candidate substance, feature amounts related to the electronic properties of the candidate substance, feature amounts related to the physicochemical properties of the candidate substance, and feature amounts related to the partial structure of the candidate substance. [Supplementary Item 12] The information processing device according to any one of Supplementary Items 1 to 11, wherein first training data used to train the first prediction model and second training data used to train the second prediction model are at least partially different.[Supplementary Item 13] An information processing device according to any one of Supplementary Items 1 to 12, wherein first learning data used to train the first prediction model and second learning data used to train the second prediction model are prepared based on information on the first species of strain and the second species of strain.

[0106] The technology of the present disclosure can be appropriately combined with the various embodiments and / or various modified examples described above. Furthermore, it is not limited to the above embodiments, and various configurations can be adopted without departing from the spirit of the present disclosure. Furthermore, the technology of the present disclosure extends not only to programs, but also to storage media that non-temporarily store programs, and computer program products that include programs.

[0107] The above-described description and illustrations are a detailed explanation of the parts related to the technology of the present disclosure and are merely an example of the technology of the present disclosure. For example, the above description of the configuration, functions, actions, and effects is an explanation of an example of the configuration, functions, actions, and effects of the parts related to the technology of the present disclosure. Therefore, it goes without saying that unnecessary parts may be deleted, new elements may be added, or replacements may be made to the above-described description and illustrations within the scope of the gist of the technology of the present disclosure. Furthermore, to avoid confusion and facilitate understanding of the parts related to the technology of the present disclosure, the above-described description and illustrations omit explanations of common technical knowledge that do not require particular explanation to enable the implementation of the technology of the present disclosure.

[0108] In this specification, "A and / or B" is synonymous with "at least one of A and B." In other words, "A and / or B" means that it may be only A, only B, or a combination of A and B. Furthermore, in this specification, the same concept as "A and / or B" is also applied when three or more things are expressed by connecting them with "and / or."

[0109] All publications, patent applications, and technical standards mentioned in this specification are herein incorporated by reference to the same extent as if each individual publication, patent application, or technical standard was specifically and individually indicated to be incorporated by reference.

Claims

1. An information processing device that predicts evaluation results of an Ames mutagenicity test for a candidate substance of a product using a prediction model, comprising a processor, wherein the processor uses as the prediction models a first prediction model that outputs a first predicted evaluation result indicating whether or not the candidate substance has mutagenicity associated with a base pair substitution mutation when the candidate substance is added to a first type of bacterial strain that is sensitive to a base pair substitution mutation in which a part of a base sequence is changed, and a second prediction model that outputs a second predicted evaluation result indicating whether or not the candidate substance has mutagenicity associated with a frameshift mutation when the candidate substance is added to a second type of bacterial strain that is sensitive to a frameshift mutation in which the reading frame of the base sequence is shifted, obtains candidate substance information regarding the candidate substance, inputs the candidate substance information or information based on the candidate substance information into the first prediction model and the second prediction model, causes the first prediction model and the second prediction model to output the first predicted evaluation result and the second predicted evaluation result, and presents prediction information corresponding to the first predicted evaluation result and the second predicted evaluation result to a user.

2. The information processing device of claim 1, wherein the processor, when both the first prediction evaluation result and the second prediction evaluation result indicate that the candidate substance is not mutagenic, presents to the user a message indicating that the candidate substance is not mutagenic as the prediction information, and when at least one of the first prediction evaluation result and the second prediction evaluation result indicates that the candidate substance is mutagenic, presents to the user a message indicating that the candidate substance is mutagenic as the prediction information.

3. The information processing device according to claim 1, wherein the processor presents the first prediction evaluation result and the second prediction evaluation result themselves to the user as the prediction information.

4. The information processing device according to claim 1, wherein the candidate substance information is information regarding the chemical structure of the candidate substance.

5. The information processing device according to claim 1, wherein the predictive model is constructed by any one of machine learning techniques including support vector machines, linear separation, gradient boosting trees, AdaBoost, random forests, deep learning, and ensemble learning thereof.

6. The information processing device according to claim 5, wherein the first prediction model is constructed by deep learning, and the second prediction model is constructed by transfer learning of the first prediction model that has already been learned.

7. The information processing device according to claim 5, wherein the first prediction model and the second prediction model are constructed using different machine learning techniques.

8. The information processing device according to claim 7, wherein the first prediction model is constructed by deep learning, and the second prediction model is constructed by a machine learning method other than deep learning.

9. An information processing device as described in claim 5, wherein the first prediction model and the second prediction model are constructed using the same machine learning method, but the internal parameters for deriving the first prediction evaluation result and the second prediction evaluation result are different.

10. The information processing device of claim 5, wherein when the predictive model is constructed by any one of machine learning techniques including a support vector machine, a linear separation, a gradient boosting tree, an AdaBoost, a random forest, and an ensemble learning thereof, the processor acquires feature information of the candidate substance and inputs the feature information into the predictive model.

11. The information processing device according to claim 10, wherein the feature information includes at least one of features relating to the geometric shape of the candidate substance, features relating to the electronic properties of the candidate substance, features relating to the physicochemical properties of the candidate substance, and features relating to the partial structure of the candidate substance.

12. An information processing device according to claim 1, wherein the first learning data used to train the first prediction model and the second learning data used to train the second prediction model are at least partially different from each other.

13. An information processing device as described in claim 1, wherein the first learning data used to train the first prediction model and the second learning data used to train the second prediction model are prepared based on information on the first species of bacterial strain and the second species of bacterial strain.

14. A method for operating an information processing device that predicts evaluation results of an Ames mutagenicity test for a candidate substance of a product using prediction models, comprising: using as the prediction models a first prediction model that outputs a first predicted evaluation result indicating whether or not the candidate substance has mutagenicity associated with a base pair substitution mutation, in which a part of a base sequence is changed, when the candidate substance is added to a first type of bacterial strain that is sensitive to the base pair substitution mutation, and a second prediction model that outputs a second predicted evaluation result indicating whether or not the candidate substance has mutagenicity associated with a frameshift mutation, in which the reading frame of the base sequence is shifted, when the candidate substance is added to a second type of bacterial strain that is sensitive to a frameshift mutation, in which the reading frame of the base sequence is shifted; obtaining candidate substance information regarding the candidate substance; inputting the candidate substance information or information based on the candidate substance information into the first prediction model and the second prediction model, and outputting the first predicted evaluation result and the second predicted evaluation result from the first prediction model and the second prediction model; and presenting prediction information corresponding to the first predicted evaluation result and the second predicted evaluation result to a user.

15. An operating program for an information processing device that predicts evaluation results of an Ames mutagenicity test for a candidate substance of a product using a prediction model, comprising: using as the prediction models a first prediction model that outputs a first predicted evaluation result indicating whether or not the candidate substance has mutagenicity associated with a base pair substitution mutation, in which a part of a base sequence is changed, when the candidate substance is added to a first type of bacterial strain that is sensitive to the base pair substitution mutation, in which the reading frame of the base sequence is changed; and a second prediction model that outputs a second predicted evaluation result indicating whether or not the candidate substance has mutagenicity associated with the frameshift mutation, in which the reading frame of the base sequence is shifted, when the candidate substance is added to a second type of bacterial strain that is sensitive to the frameshift mutation, in which the reading frame of the base sequence is shifted; obtaining candidate substance information regarding the candidate substance; inputting the candidate substance information or information based on the candidate substance information into the first prediction model and the second prediction model, and outputting the first predicted evaluation result and the second predicted evaluation result from the first prediction model and the second prediction model; and presenting prediction information corresponding to the first predicted evaluation result and the second predicted evaluation result to a user.

Images